Next Article in Journal
Numerical Simulation of Subcooled Flow Boiling in a Threaded Tube and Investigation of Heat Transfer and Bubble Behavior
Previous Article in Journal
Simple Discrete Control of a Single-Phase Voltage Source Inverter in a UPS System for Low Switching Frequency
 
 
Font Type:
Arial Georgia Verdana
Font Size:
Aa Aa Aa
Line Spacing:
Column Width:
Background:
Review

A Review for Green Energy Machine Learning and AI Services

1
Department of Applied Data Science, San Jose State University, San Jose, CA 95192, USA
2
BRI, San Francisco, CA 94104, USA
3
Department of Computer Engineering, San Jose State University, San Jose, CA 95192, USA
*
Author to whom correspondence should be addressed.
Energies 2023, 16(15), 5718; https://doi.org/10.3390/en16155718
Submission received: 29 June 2023 / Revised: 21 July 2023 / Accepted: 24 July 2023 / Published: 31 July 2023
(This article belongs to the Section K: State-of-the-Art Energy Related Technologies)

Abstract

:
There is a growing demand for Green AI (Artificial Intelligence) technologies in the market and society, as it emerges as a promising technology. Green AI technologies are used to create sustainable solutions and reduce the environmental impact of AI. This paper focuses on describing the services of Green AI and the challenges associated with it at the community level. This article also highlights the accuracy levels of machine learning algorithms for various time periods. The process of choosing the appropriate input parameters for weather, locations, and complexity is outlined in this paper to examine the ML algorithms. For correcting the algorithm performance parameters, metrics like RMSE (root mean square error), MSE (mean square error), MAE (mean absolute error), and MPE (mean percentage error) are considered. Considering the performance and results of this review, the LSTM (long short-term memory) performed well in most cases. This paper concludes that highly advanced techniques have dramatically improved forecasting accuracy. Finally, some guidelines are added for further studies, needs, and challenges. However, there is still a need for more solutions to the challenges, mainly in the area of electricity storage.

1. Introduction

An emerging solution to the problem of power-hungry algorithms is Green AI or making AI development more sustainable. By reducing the energy consumption of data centers and other AI infrastructure, Green AI can help reduce the hidden costs of technology development. Green AI is leading the way to creating a more sustainable and responsible form of technology development, allowing us to achieve more accurate results with less energy consumption. Green AI can also be more cost-effective and efficient, in addition to being more environmentally friendly. For instance, a data center powered by renewable energy can save money on energy costs. As a concept in AI, Green AI refers to research that achieves novel results without increasing computational costs. This paper focuses on Green AI services. The services include green energy power generation, load forecasting, load profiling, demand and response, electricity price, storage assessment, outage, and power consumption. All these services motivate using renewable energy sources in various service sectors. There has been various research conducted under green energy services. The paper [1,2] focuses on solar power generation and its uses. Paper [3] studies power generation via wind energy. There are various parameters considered for load forecasting, like in ECM (energy cloud management) for distribution of power and support operation [4]; load prediction is made at various levels and different time durations like short, mid, and long terms [5,6], considering parameters like temperature, humidity, and climatic changes [5,6,7,8]. While in [9] we study the load forecasting for solar power and wind energy. The state-of-the-art technique is reviewed and used for renewable energy load forecasting [10]. One of the major roles in load forecasting is played via load profiling assessment [11]. It shows the variation in electrical load over time.
The energy is generated and priced based on demand and response from the market [12,13]. The important parameter of demand and response is studied in this paper based on real pricing schemes and load balancing considering user involvement. A large set of fundamental drives like system loads and technical indicators are used for spot price prediction [14,15] based on a long review of data-driven models for around a decade [16]. The prices are raised or dropped based on their power consumption and usage analysis. There are various machine-learning models and techniques discussed in this paper. The analysis [17] discussed fundamental data-driven methods that are used to analyze the structure of energy consumption [18], a derived hybrid machine learning model, and [19] state-of-the-art models. All these services focus on predicting and analyzing various parameters of Green AI services.
As Green AI becomes a widely accepted measure of research quality in addition to accuracy, researchers can develop models that are environmentally friendly and inclusive of all communities. To utilize technology to its full potential, we need to apply various machine learning techniques along with deep learning algorithms and forecasting techniques to guarantee the accuracy of the models. Generally, there are three types: Data-driven, Classical, and Artificially intelligent. A classical classifier, such as a SARIMA or ARIMA (autoregressive integrated moving average), can also be built with artificial intelligence models, such as CNNs, LSTMs, or FNNs. In addition to ensuring efficient electricity usage, these forecasting models are designed to predict electricity costs.
Our paper differs from others regarding Green AI in several ways. Our first focus is community-level energy management, which has received less attention than building on grid-level energy management. The comparisons between the various machine learning models regarding green AI services are also included. The data presented in this paper are designed to be flexible and adaptable to different types of communities. This contrasts with other models that may be better suited to certain types of communities or energy systems. In addition, we evaluate our model using real-world data, which enhances its practical relevance and applicability. This review paper aims to present a novel Green AI model that can be applied to the management of energy at the community level, which is adaptable, practical, and effective. Based on our paper, we can help communities develop more resilient and sustainable energy systems.
The next parts of this paper are divided into the following sections. Section 2 explains the green energy AI taxonomy, Section 3 focuses on green energy power generation: solar and wind, Section 4 reviews green energy storage, Section 5 focuses on green energy load, Section 6 is for power consumption, Section 7 focuses on electricity price forecasting, Section 8 directs to the future research, and Section 9 is the conclusion of this paper.

2. Green Energy AI Taxonomy

In recent years, there has been an increased demand for environmentally friendly and sustainable solutions, and communities are looking for ways to reduce their carbon footprints. Green AI services can enable this transition by optimizing energy usage, reducing waste, and promoting renewable resource usage. The use of artificial intelligence services like Green Energy Power Generation, Green Energy Load Forecasting, Load Profiling Assessment, Demand Response Analysis, Electricity Price Forecasting, Green Energy Storage Assessment, Electricity Outage Analysis, and Power Consumption/Usage Prediction can enable communities to make informed decisions about energy usage and resource allocation through data-driven insights and tools. All the mentioned services are targeted to help communities achieve the ultimate goal of using energy efficiently, which would reduce the impact on the environment would lead to the sustainable development of the community.
The taxonomy of the green energy cloud AI service is presented in Figure 1, which is given below in detail. The table shows the commonly used machine learning models that will be explained in detail in a later part of this paper. This model uses various input parameters and has different complexity levels, and usage varies based on the time frames that are considered for the researchers.
Table 1 lists the services and ML models commonly used. The models can be used to predict customer behavior, detect anomalies, and perform data analysis. The services help to deploy models quickly and scale the machine learning applications easily. They also help to manage the models and monitor their performance.
Green energy power generation—The electricity generated by renewable energy sources is called green energy, and the process of converting green energy into electricity is called green energy power generation. The methods like pumped storage are used before the electricity is transmitted or delivered to the end user. Green energy power generation involves wind, solar, geothermal, and tidal. There is research that explains the features, application, and future scope that would enhance the forecasting methods, which would lead to higher accuracy. Akhter et al. [1] discuss methods for forecasting PV (photovoltaic) power output using metaheuristics and machine learning methods, focusing on solar power generation, considering factors like (1) temperature, (2) sunrays, (3) humidity, and (4) atmospheric pressure. This uses the PV output of the power forecast based on time horizons. This study uses MCP (measure-correlate-predict) models for forecasting power generation and alternative methods such as neural networks or hybrid models. In a study by Utpal et al. [2]. Considering wind speed and meteorological variables, Soraida et al. [3] provided an overview of wind electricity generation. The study focuses on medium- and long-term trends in wind speed and power, which can be considered in summative planning. At the same time, short-term forecasts are used mainly for operational purposes.
They examined machine learning models for solar energy generation based on accuracy, reliability, the model’s computational cost, and the model’s complexity. The importance of the correlation among input data, output data, and preprocessing of model discussion is provided for various forecasting time frames. Along with PV-assembled smart buildings, well-organized managed systems, EV (electric vehicle) charging, and smart grids are covered in the paper.
Green energy storage assessment—This is the storage of electricity on a larger scale using various collection methods of a power grid. This assessment generally considers when the power is in abundant form and is cheap, mainly from renewable sources, or while demand is less, and we know that demand will rise and prices will be higher soon. The evaluation of storage analysis includes analysis and prediction mainly of the SoC (state of charge), lifespan, and SoH (state of health)
Green energy load forecasting—This is the part that studies the consumption of electricity by the electric circuits and appliances in the community [4]. Load management is required, and ECM (Energy Conservation Measures) is used for the efficient distribution, supporting operations, and planning of various processes for managing electrical load. The pricing of retails and dynamic powers is based on the consumption and prediction of electricity. The accuracy of consumption is significant at lower ends compared to a higher level [5]. In the term load forecasting, the “load” refers to the demand, which is measured in kilowatts (kW), or in energy which is measured in kWh (kilowatt-hours). The hourly data uses the same power and energy magnitude, so no distinction is made between demand and energy. Load forecasting involves the prediction of magnitudes and geographic localities accurately for a planning horizon divided into different periods. Generally, the hourly total system load is in the basic quantity of interest. However, load forecasting also considers values of peaks and load hourly, day-wise, weekly, and month-wise [6]. Green energy load forecasting includes the motor, industry equipment, fridge, air conditioner, lighting, water heater, and elevator. Intensive research is conducted on load forecasting, considering a variety of parameters for a more accurate evaluation, leading to greater use of renewable energy, and promoting Green AI.
According to the paper [7], various electricity load forecasting models were presented by Yildiz et al. 2017 that focus on regression models. Considering all the seasons and various climatic changes and conditions, day-ahead hourly loads are predicted using the data obtained from buildings and campuses. The review concluded that using the regression model is not just limited to forecasting the total energy load of buildings, but also includes HVAC (heating, venting, and air conditioning) loads and retrofit savings. The forecasting methods for wind power, solar, and load are reviewed by H. Wang et al. There has been a study of wind-speed/irradiance corrections adopted by NWP (numerical weather prediction), as well as wind and solar power forecasting methods, and load forecasting methods used by the NWP for demand forecasting, as well as wind and solar power forecasting methods for demand forecasting. The paper describes the NWP as one of the most critical and important factors that would affect the accuracy of forecasts by wind and solar, mainly for short-time forecasting. The paper also focuses on the challenges and future research direction of solar, load, and wind [9]. The review [8] describes the use of machine learning in forecasting loads in district heating and cooling systems. A study was conducted on the prediction of heating loads and demands as well as on the design, maintenance, and scheduling of heating systems. The review considers parameters such as weather/climate changes, socioeconomic features, historic energy consumptions, and thermal load for predicting both the short and long-time load. The paper also discusses using ML-based algorithms to link DHC to smart electricity grids. Sheraz et al. conducted a comprehensive review of the renewable energy and load forecasting literature using deep learning models and state-of-the-art techniques at residential and commercial sectors [10]. This survey provides an overview and categorization of deep learning models used in smart energy management systems. Models of forecasting are evaluated according to the types of energy (wind and solar), building types (commercial and noncommercial), and temporal granularities (five minutes, ten minutes, fifteen minutes, thirty minutes).
Load profiling assessment—The load profile shows the graph of the variation between electrical load v/s time. The profiling can change or vary based on the type of customer. The customer type includes residential customers, commercial plots, and industrial areas. Temperature and holiday seasons also affect the load profile assessment graphs. During electricity generation, the producers use the information to plan the need for electricity and make it available at a given time [11]. The load profile includes consumer learning, training, classifying, predicting, consumer load pattern, consumer energy consumption behavior, and peak time.
Demand response analysis—It is defined as “Changes in electric usage by end-use customers from their normal consumption patterns in response to changes in the price of electricity over time, or to incentive payments designed to induce lower electricity use at times of high wholesale market prices or when system reliability is jeopardized” [12]. One of the most critical tasks in ECM is DRM (demand response management). The DRM helps balance the gap between the demand and supply of the energies. Using the DRM system, we can shave off the valley or peak in demand for electric power in real-time, which benefits the companies and consumers in electricity bills and real-time pricing schemes, balancing the loads and increasing user involvement [20]. According to the paper, Jebaraj et al. discussed energy modeling issues dealing with different forms of energy like solar, wind, biomass, and bioenergy.
Additionally, they reviewed and presented various models, including energy planning, demand-supply, and forecasting models. Along with it, they also reviewed various optimized models, renewable energy models, and models that help in reducing emissions [20]. The paper focuses on problems like the demand and supply of energies and weather impact on the models. It focuses on the optimized cost, multi-level optimization: self-sufficiency, conservation, and sustainability, and models that use AI and show symbolic reasoning, flexibility, and explanation capabilities. Suganthi et al. reviewed the energy forecasting models demand management, and econometric models for planning future needs, identifying conservation measures, framing policy decisions, optimizing energy utilization, and reducing emissions. The review states that researchers can utilize sophisticated modeling techniques, including gray prediction, genetic algorithms, fuzzy logic, SVR (support vector regression), ACO (ant colony optimization), and PSO (particle swarm optimization) for macroeconomic planning to predict energy demand accurately [14].
Electricity price forecasting—The price of electricity is based on the seasonality at a daily level, weekly level, and to a certain extent, an annual level. The majority of studies conducted on the EPF, along with various short-term time horizons, mainly focus on day-to-day markets. Mid-term is preferred for managing risk, balance sheet calculations, and derivative pricing. In most cases, the evaluation is conducted on the distribution of prices based on future periods rather than actual point forecasts. This type of modeling technique is mainly used in long-standing traditions in finance, and the inflow of finance solutions is measured. Long-term forecasting is mainly based on investments, and profitability analysis is conducted. Based on long-term forecasting, planning and determining future sites or new sources of fuel power plants are considered. Similarly, spot pricing approaches include statistical methods like econometric and technical analysis, multiagent factors, and CI (computational intelligence) models. The spot price is dependent on fundamental drivers that include demands and consumption figures, weather parameters, cost of fuels, and reserve margin; in other words, surplus generation is available generation minus/over predicted demand, an important power grid component or market is scheduled for maintenance or forced to go down [15,21].
There is dynamic forecasting, along with the time parameter used for forecasting. Dynamic forecasting prices use techniques like time of use pricing, pricing in real-time, critical peak time pricing, and inclined block rate. A fine-resolved prediction of electricity market prices is especially important in electricity sectors with a high percentage of non-dispatchable renewable energy. Demir et al. reviewed the electricity price based on technical indicators that can help improve the accuracy of forecasts for day-ahead electricity markets. This paper shows that traders’ behavioral biases can be captured by TIs, resulting in statistically significant reductions in forecast error with ML models [16]. Lu et al. used a decade-long review that considered input parameters like natural gas and crude oil, carbon prices, and electricity prices as input parameters for data-driven models. The paper covers the aspects of accuracy, time horizon, and input variables for energy price prediction [17].
Power Consumption/Usage Analysis—The article by Yixuan et al. discussed prevailing data-driven methodologies used to evaluate building community energy consumption across granularities and archetypes, including machine learning methods for predictions and clustering for classification. The paper addresses several building-related applications using data-driven methodologies that assist the problems of loads, energy patterns, energy consumption based on regions, building stock benchmarking, and retrofit strategies on global levels [18]. The paper explains applications of the state-of-art technique in machine learning models for energy systems explained by Mosavi et al. They are used for energy systems, applications, and taxonomy, contributing to efficient energy use and handling the energy grid’s sustainability. A hybrid machine learning model for energy systems is presented in the paper to help identify ML modeling techniques and energy types and to understand how accurate the model performs, the robustness of the model, its precision, and its generalization abilities. Extensive reviews of ML models in diverse application areas have demonstrated the popularity and effectiveness of ML models in almost any energy domain [19]. The purpose of the study is to review recent studies related to designing energy models and estimating energy consumption in various areas. Using novel hybrid and ensemble prediction technologies, the research reports a significant increase in accuracy and performance using the state-of-art energy consumption model [22].
Along with all the above Green AI services, the work on the concept of VPP (virtual power plant) and duck curve plays an important role in green energy generation, storage, and load forecasting, it contributes to the aspects of energy saving, economic aspect, power transactions, communications between technologies, and resource allocations. Additionally, it enhances the power plant’s or consumer’s readiness and balances the performance energy range and storage [13]. There is vast research on this topic. The concept is explained using applications and challenges using the case studies of Europe, the USA, and Australia. The challenges in these regions using VPP were mainly resource allocation, communication between the technologies, power transactions, operations, and control systems [23]. The calculation and power flow of VPP of low and medium voltage in distributed networks are presented in the article. It uses the concept of DER (distributed energy resources) to see its impact on when capacity is at its maximum with storage units. The paper describes data acquired for a single VPP instance, and it is not easy to generalize these findings to other VPP locations [24]. The researchers in the paper have attempted to address the scheduling techniques, technical and economic aspects, and optimization of VPP. The paper concludes by suggesting that deep reinforcement learning handles feature extraction and scalability, as they are generated by combining reinforcement learning and deep learning [25]. The paper focuses on the economic aspect of VPP. It analyzes the possibility of combining the DERs and ESS in VPP and considers them as a single plant, this will influence the factors like price and production [26]. The study of VPP in Ireland, Belgium, and the Netherlands was conducted at the community level, where community energy meets intelligent grids. The three cVPPs had to abide by the existing energy system, making it challenging for them to play the chosen roles in the energy system, work on a local scale, and maintain their own needs and values, as explained in the paper [27]. This paper presents a comprehensive, integrated techno-economic modeling approach for an urban VPP that fully exploits the DER’s aggregated flexibility and results in attractive business cases [28]. A VPP model was developed to maximize operating profit to meet demand. It was applied to real data from an irrigation system comprising several Aragon (Spain) water pumping stations. It includes a wind farm, six hydroelectric power plants injecting the generated electricity directly into the distribution network, and on-site photovoltaic plants prioritizing self-consumption [29]. The paper proposes a multi-time scale stochastic optimization scheduling strategy for a new energy virtual power plant based on a robust stochastic optimization theory. The results show that the models can improve operating profit and new energy consumption capacity [30]. Based on the SARIMA-KF(Kalman-filter) hybrid algorithm for real-time optimal operation of VPPs, an adaptive and predictive energy management strategy for energy storage with renewable energy is proposed [31]. The proposed VPP mitigates issues related to DER penetration by providing grid frequency and voltage support, load forecasting, and power flow control. The input parameters considered for the DER issue are historical load and weather data, weather forecast inputs, and models of VPP players [32]. In a way, there are a lot of benefits to users, suppliers, and stakeholders using VPP. It supports the Green AI services by supporting the weather and load forecast management, real-time processing of historical data, giving alerts during market fluctuations during the day, and providing sufficient resources to smart grids at the time of peak hours. Currently, Tesla introduced Powerwall in CA, USA in the year 2015 and plans to launch its own virtual power plant in California. Swell Energy is also on the domestic scene with VPP designs ready to power homes throughout California, Hawaii, and New York City.

3. Green Energy Power Generation

3.1. Solar Energy Generation and Prediction

Sunlight is an easy-to-get and sustainable energy resource. Individuals and companies can use the sun to generate electricity by installing solar panels on the roof or bare ground in most places on Earth. Physical and statistical models are not enough to forecast solar energy generation for the wide variations of solar energy. With sufficient data, ML and DL algorithms perform better in temporal pattern learning and solar energy generation forecasting. During days when solar energy production is high and grid demand is low, the Duck Curve shows the electricity demand from the grid. A graph showing these curves reveals a distinctly duck-like shape [33]. For about a half-century now, the California Independent System Operator (CAISO) has been monitoring the Duck Curve and its future expectations, and their biggest finding has been the widening gap between morning and evening prices [34]. In a Boiteux-Turvey-style model, the long-run equilibrium value of storage capacity minimizes the expected system cost conditional on generation capacities solving the duck curve problems [35]. A duck chart was used to examine how much photovoltaic generation and wind power might need to be cut off if additional grid flexibility operations are not considered. Energy storage systems allow greater penetration of solar and wind resources, achieving a 50% renewable portfolio standard [36].
According to the scales of prediction horizon, solar energy power forecasting can be categorized into:
Those three are used to support the decision-making of the electricity market and power system operation. In EMS (energy management system), short-term forecasting plays an important part [50].
  • For parameters used to forecast:
  • Solar parameters: global solar radiation, solar irradiance, and UV Index;
  • Weather data: temperature, humidity, wind speed, wind direction, etc.;
  • Power parameters: AC current, AC voltage, DC current, DC voltages, frequency, and phases;
  • Solar position: solar zenith angle and solar azimuth angle.
Solar parameters and weather parameters are the main factors of solar power generation [37,38,39]. Studies used solar and weather information as inputs to build their forecasting models. Solar parameters are directly linked with PV power generation [38]. Temperature can affect the performance of PV cells [51]. High temperature and high relative humidity negatively affect PV cell performance [52]. However, solar irradiance data or weather is not available in some cases. Li et al. proposed a forecasting model that uses solar position data instead of solar irradiance data [42]. The algorithm will learn the relationship between solar position and solar power generation. With the development of AI-driven IoT technology, a novel solar power generation framework has been introduced [40]. They use power data, like current and voltage, to predict power generation by solar energy.
Solar energy generation forecasting is a multivariate time-series forecasting. ML and DL algorithms can explore the relationships between solar power generation and other factors and learn the temporal and spatial patterns. The various model and their performance is explained in Table 2 along with the results.
  • Machine learning algorithms: Mahmud et al. implemented LR (linear regression), PR (polynomial regression), DTR (decision tree regression), SVR, RFR (random forest regression), LSTM, and MPR (multilayer perceptron regression) to do a solar power generation forecast, and compared the models [37]. The results show RFR has the best performance [37]. Babbar et al. proposed 10-day-ahead solar power generation forecasting using hybrid Adaboost, indicating an %MAPE 8.88 [38]. Sharma et al. built an SVM (support vector machine) model to predict solar generation and found the SVM performs better than the PPF (past-predicted future) model [38].
  • Deep learning algorithms: Li et al. showed that ANN (artificial neural networks) produce satisfactory accuracy when predicting short-term solar energy production [42]. LSTM can learn the temporal pattern from historical data. Liu et al. proposed simplified LSTM models for short-term solar energy generation prediction [38]. Zhou et al. studied a novel CNN-ALSTM model which combines CNN, LSTM, and attention mechanisms [40]. The study by Tian et al. on short-term PV generation of the power transformer model has a better performance than the DNN (deep neural network) model and GRU (gated recurrent unit) model [46].

3.2. Wind Energy Generation and Prediction

The variability of wind power poses a major obstacle to its combination into the grid. The discussion is made on the data sources and types in which we can forecast the wind power forecasting model. Data from wind power plants and weather data are used in the analyses. In the case of high-permeability intermittent power supplies, accurate wind power forecasts can substantially reduce its impact on grid operations. Five components make up the general framework for wind speed forecasting: (1) preprocessing, (2) feature-based analysis, (3) model structure optimization using deep learning, (4) model structure optimization using reinforcement learning, and (5) model performance evaluation. In the studies, it has been shown that the change in climatic conditions affects wind power and its forecasting. As the weather changes, the wind power prediction considers the parameters accordingly.
The table compares various models based on the performance metrics used for wind power generation. To determine the most appropriate fit model for predicting wind power production, M. Eugen et al. compared the models. The Ridge Polynomial Regression Model was found to have the highest precision compared to the other four supervised learning models [53]. Transfer learning is suggested and used in the paper’s prediction system of the deep neural-based method. Suggested DNN-MRT techniques overcome the deficiencies of the formerly proposed methods. The study suggested by the technique is evaluated for datasets from five wind farms across different regions [54].
Deep learning models are used in a study by X. Deng et al. to predict short-term wind generation. As part of the analysis, various model structures were examined and compared throughout different seasons to enhance understanding and application of architectural design. The paper investigates the literature on DL, enforcement learning, and transfer learning in wind speed and wind energy forecasting [55]. The paper used the pre-trained TCN (temporal convolutional network) model with historical weather data and wind turbine power generation outflows from Scada to Turkey. LSTM and GRU models have been demonstrated to be superior to TCN models in terms of predicting long-term sequence data and extracting features. To solve the long-distance dependencies of predicting wind forecasts, substantial quantities of time-spatial data like 1-year wind energy were incorporated into the TCN model [56]. The assessment of 6 months ahead of WPF (wind power forecasting) is presented in the paper using tree-based algorithms. A. Ahmadi et al. used this tree-based algorithm to solve the overfitting and underfitting issue.
Three models were developed to study the impact of data entered on forecast accuracy. The results of experiments revealed that forecasted models suffer considerable accuracy degradation when longer time intervals are considered, and height extrapolation is used [57]. An Ecemis et al. study examined how ML algorithms could be used to predict wind energy production over the long term. In the study presented, machine learning algorithms were successfully used to predict the feasibility of establishing wind farms at unknown geographical locations by comparing the model of the underlying location. The study’s findings are useful for the development of a new wind power plant in an unknown place [58]. An effective data-driven model for forecasting wind energy is developed in the study. The paper examines the yield of enhanced ML models for forecasting wind generation time series. A study showed that machine learning models perform better when dynamic information is incorporated. The results revealed that data lagged by one month, and the contribution of input variables could help forecast wind power generation prediction more accurately [59]. The application of the sparse ML technique is used to predict the next upcoming hour of wind energy generation. A novel nonparametric density estimate approach was used for the probability band of prediction.
A sparse model can then be used to calculate a range of confidence in your predictions, and a nonparametric conditional density estimation method can be used to demonstrate positive accuracy when estimating very short-term wind power [60]. The comprehensive and comparative analysis of various algorithms and techniques like ANN, SVR, RT, and RF are presented and studied, along with the pros and cons in the paper. The reason for the use of all of the ML algorithms was to identify patterns in multidimensional data and to test the robustness and tolerance to the outliers and errors along with the above measures; it is also focused on the lower carbon emissions and economy and handling noisy data [61]. Using wind speed data, this study applies three deep learning algorithms to predict short-term wind energy generation. The approach relies on EEMD (empirical mode decomposition) and LSTM networks. An innovative deep-interval prediction model representing a Long LSTM proof is created to assess and contrast the existing methodologies with the proposed model. Loss functions are tuned for root average square backpropagation, and comparison tests are carried out [62].
The paper aims to support local electricity distributors by focusing on day-ahead wind power prediction so they can plan the operation of other production plants and use energy markets to meet demand. To learn the relation between wind production, NWP, and a day time information combination of regression models with machine learning algorithms, large datasets are used [63].
Wind energy parameters have been used to investigate the relationship between initial relative energy and wind energy characteristics. Wind speed [53,54,55,56,57,58,59,61,62,63], wind direction [53,54,55,56,57,59,62,63], air density, and weather conditions like temperature and humidity [55,56,57,61,62,63] are among the parameters involved in the operation of wind farms. The table above uses various parameters to examine the performance of the models.
  • Wind Parameters: speed, direction, historical power generation data;
  • Weather Parameters: temperature, humidity.
Among the parameters that M. Eugen considered to estimate wind power production was historical data from wind production and weather forecasts [53]. The current weather forecast is heavily reliant on earlier forecasts and power metrics. Thus, previous forecasts of the last 24 h and related features (selected through mutual information) are fed into the model to explore power measurement dependence [54]. The next instant must be calculated on a local atmospheric situation to forecast wind speed and wind power around a wind farm. These conditions include pressure, along with temperature, roughness, and obstacles [55]. In the prediction experiment, wind speed and wind direction were used along with temperature and humidity as key model parameters [56]. By aggregating every 10-min sampled wind power value, every hour’s total wind power has been calculated by converting actual wind speed values into hourly mean and standard deviation values [57]. During the study, wind power forecasting was performed using daily wind speeds, daily standard deviations of hourly winds, and generated daily wind power data as inputs to machine learning algorithms [58]. For improved wind prediction accuracy, multiple input variables such as wind speed and direction are used to improve the performance and are further examined by machine learning models. A significant improvement in wind power prediction accuracy was revealed by using input variables [59]. Jiaqing et al. described a sparse machine learning nonparametric density estimation approach to estimate very short-term wind power. The model is based on forecasts, real-time observations collected from the past hours, and information from nearby power generators [60]. In order to generate wind power, Adrian et al. considered air density, pressure, and temperature [61]. In the method, wind speed, air density, surface heat flux, surface temperature, and relative humidity were the meteorology variables that correlated with wind power production. The method provided a prominent level of accuracy compared to the existing model [62]. In addition to NWP (numerical weather predictions), the model uses historical wind turbine production data. A power-grid operative at Valais, SEIC-Télédis Group collects production data, such as wind parameters, temperature, and humidity through smart meters. The primary variables in this equation that directly impact production are wind speed, air temperature, humidity, and density [63].
The models estimate the amount of wind energy produced accurately, solving a crucial problem for the renewable energy sector. By analyzing real-world data from wind turbines, a methodology is proposed to compare various machine learning algorithms and select the best accurate final model for predicting wind power generation. Various machine learning models are presented in the table that can be used to predict wind power generation:
  • Machine learning algorithms: ANN, ARX, ARMAX, AdaBoost, LSSVM, ARIMA, Bagging, FNN, DT, GBoost, GPR, GRU, KNN, LASSO, MSS, RF, RNN, RT, RVM, SVR, TCN, XGBoost;
  • Deep learning algorithms: MLP, LSTM, TRA;
  • Hybrid algorithm: HybridNN.
There are multiple models used to predict the wind power generation for long-term, short-term, and mid-term forecasting like AdaBoost [57], ANN [61], ARIMA [54], Bagging [57], DT [57,60], FNN [55], GBoost [57,63], GPR [59], GRU [56,62], HybridNN [55], kNN [58], LASSO [58,60], LR [60], LSTM [56,60,62], MSS [55], RF [57,58,61], Ridge [60], RNN [56,62], RPR [53], RT [61], RVM [60], SVR [54,58,60,61] TCN [56], TRA [55], and XGBoost [57,58]. RPR forecasts short-term wind production (MPR) by using four models: LR, RR (ridge regression), PRR (polynomial ridge regression), and multilayer perceptron regression. Based on a comparison of the four models, the PRR has the most accurate data, with a 16.41% error rate [53]. Baseline regression models like ARIM and SVR can check the baseline performance of the farms. It is a way of comparing the performance of the DNN-MRT (deep belief network is used as a meta-regressor) with baseline regressors [54]. These models include the Traditional approach without model selection (TRA), the FNN (fuzzy model neural network), the MSS (model structure selection), and the hybridNN (hybrid neural network) [55]. To forecast long-term wind power, Wen-Hui et al. aimed to achieve MAPE of not more than 10% for predictions occurring 24 to 72 h in the future. The study compared the productivity of four DLN (deep learning network)-based power forecasting models: the TCN, LSTM, RNN (recurrent neural network), and GRU. Based on data input volume, error reduction stability, and forecast accuracy, the TCN surpasses the three remaining models for wind energy prediction [56]. Two distinct sites are compared, Ghadamgah and Khaf, using models including DT, Bagging, RF, AdaBoost, GBoost (gradient boosting), and XGBoost version 1.7.(extreme boosting). Training these six models was based on the average as well as the standard deviation of wind speed measurements taken at those locations, as well as height measurements [57]. A long-term model of wind power was developed with the help of 5 ML algorithms. Combinations of algorithms were used, including RF, LASSO (least absolute shrinkage selector operators), kNN (k closest neighbors), and SVR. The results showed that the XGBoost, SVR, and RF algorithms deliver accurate long-term predictions of daily wind power. Several algorithms can be used to perform this task, but RF is the best algorithm [58]. An evaluation of the effectiveness of machine learning models to predict time series data for univariate wind power by A Alkesaiberi et al. proves that the most effective wind power can be predicted using machine learning techniques, such as the GPR (gaussian process regression) model. A total of seven machine learning approaches, including kernel-based methods (such as SVR and GPR models) and ensemble learning techniques (including Boosting, Bagging, Random Forest, and XGBoost), were evaluated [59]. Comparing the predictive accuracy with that of other competitive approaches was performed in the paper. There are two types of regression methods: classical linear regression and Ridge regression. Predicted values are based on a meteorological/physical approach. Other approaches to predicting wind power include SVR, RVM (relevance vector machine), DT method, time-series model, and LSTM method [60]. Various prediction methods can be used to forecast wind energy production, ranging from physical, statistical, and machine learning to hybrid (i.e., a combination of the first three). For forecasting wind energy production SVR, RT, RF, and ANN are chosen [61]. In their study, V. Chandran et al. used deep learning algorithms to forecast short-term wind power using LSTM, gated reference units, and RNN. Compared to the wind farms’ outputs, these models are applied six times to extract complicated data and nonlinear aspects. However, the outcomes of these practical tests demonstrate that the LSTM, GRU, and RNN can be effectively deployed in those locations before establishing wind farms if the location is appropriate [62]. Short-term wind power estimation can be improved using a machine learning method like a GBoost. It may be possible to anticipate wind forecast errors with the GMB model. The GBoost model was used to improve wind power forecasting using historical power.

4. Green Energy Storage Analysis and Prediction

Renewable and reusable batteries are widely used as mini-grids to lower carbon emissions and mitigate the greenhouse effect. Reusable batteries, like lead-acid batteries and lithium-ion batteries, are used to store energy for late use. Battery performance will decrease over time because of battery degradation. When the degradation reaches a specific capacity (70–80%), it can cause battery operation problems and even failure. Thus, users need battery performance information for decision-making. SoC, SoH, and lifespan are used to indicate the battery’s performance and describe battery degradation. SoC represents the level of charge of a battery. It is the ratio of the remaining charge divided by the maximum charge. SoH is the nominal battery capacity ratio to the current maximum capacity. Lifespan represents the life cycle of the battery. RUL (remaining useful life) is an indicator of the battery lifetime. SoC and SoH are important for battery users and necessary for the BMS (battery management system) [64]. SoC [65,66,67,68,69,70], SoH [70,71,72,73,74,75], and lifespan [76,77,78,79] can be predicted using machine learning or deep learning algorithms, which are shown in detail in Table 3.
  • SoC estimation—Chemali et al. [65] explored the features of batteries and built an LSTM model to estimate SoC under different ambient temperatures. Ni and Yang [66] considered physical mechanisms and proposed a physics constrained NN (neural network) that performs better than a normal NN. The result of [67] shows that the SoC estimation NN algorithm using voltage increment performs better than a traditional NN. Huang et al. built a CNN-GRU (convolution-gated recurrent unit) network to predict SoC, which is easy to implement and can directly learn from all the parameters. [66]. Yang et al. suggested a GRU encoder-decoder network with a dual-stage attention mechanism and found that the model with an attention mechanism is effective and robust [69];
  • SoH estimation—To calculate the SoH, Khan et al. suggested a MCCPs (multiple channel CPs)-based BMS [71]. A flexible SoH prediction scheme for lithium-ion batteries using a predefined voltage range with LSTM and TL (transfer learning) to constitute a cell mean model with partial training data is discussed in the paper [72]. A novel walk-forward algorithm was proposed, using gradient boosting regression for SoH estimation, which provided better results with little prior information about the batteries [75].
Lifespan—To estimate the lifespan of a battery, an artificial RNN architecture with LSTM is proposed, considering each cell voltage and load voltage along with temperature and charge-discharge cycle [77]. With advances in deep learning, predicting RUL can play a crucial role in good battery health systems for management. Ren et al. proposed a deep learning approach integrating autoencoder with DNN. The battery health features were extracted using an autoencoder and fed into DNN to predict RUL. Early life span prediction of the lithium-ion battery will ensure battery care and safety, reliability, and accelerate the battery development lifecycle [78]. Fei et al. proposed a machine-learning framework that analyzes various parameters, selects optimal low-dimensional feature subsets, and feeds the features into six representative machine-learning models to anticipate the battery life [79].
The traditional methods used to estimate SoC are open circuit voltage, coulomb counting, and Kalman filter [64]. Some common methods to estimate the SoH are the Internal resistance or impedance method, measuring energy on a full charge, Kalman filters, and electrochemical methods [71]. There is no direct way to measure battery lifespans. Most of the research about SoC, SoH, and lifespan is based on lab data [80,81], and only a small amount of data is collected in real-world situations [82,83]. Voltage, current, cell temperature, and time are used to predict SoC, SoH [82], and combined aging factors calculated from raw data and health metrics extracted from data to predict end-of-life failure. Table 4 summarizes the purpose, battery type, models, and its performance for storage analysis and prediction.

5. Green Energy Load Forecasting

A utility can minimize risk by forecasting the future consumption of commodities it transmits or delivers. Several techniques are used, including price elasticity analysis, weather analysis, and demand response/load analysis. Regional customer load data and time series profiles of those loads must be used to forecast customer loads. Seasonality must be adjusted for accurate forecasting. As part of distribution circuit load measurements, distribution load forecasts need to be reconciled with distribution network configuration.
RES energy demand and load forecasting is a relatively young research topic, but it has already drawn a lot of interest. Several surveys and reviews have examined ML techniques for energy and load forecasting from different viewpoints and objectives, while research projects have offered ML-based systems. Technological development, business conditions, and demand levels influence load forecasting. For these reasons, organizations use the economic, technological, and demand types of forecasting in planning. The focus of load forecasting can be drilled down to a single building, to an entire region, or even more. For the forecasting of the load generated by power stations [84], generator units [85], smart meters in the campus [86], regional grids [87], and demands from the electricity market [88], the input parameters vary based on locations. Some studies have been conducted based on parameters like linear and nonlinear [89] and time-series input [90] to analyze the impact of parameters in predicting the load.
  • Load Parameters: historic load patterns (hourly, previous day, weekly, quarterly, user-defined period);
  • Weather Parameters: temperature, humidity, wind speed;
  • Electricity Parameters: electricity consumption, electricity generation, electricity price, currency.
The following Table 5 presents the research on load forecasting, considering the factors mentioned below:
According to Kamil, the research was conducted to predict the electricity load generated at power stations for the week and day ahead of the Turkish electricity market. Furthermore, the same model can be applied to mid-term forecasting [84]. The research considered both linear and nonlinear components of electricity load. The research was conducted to predict electricity load in Australian markets on an hourly, daily, and monthly basis. The proposed model prevented the loss of information, and for each effective component, the original load data is used to decompose the load [89]. It was suggested by Tong et al. that the power load forecasting module and the deep feature extraction module be combined for predicting electricity loads for cities in California, Los Angeles, New York, and Florida. The feature extraction modules assist in extracting a load of 24 h on an hourly basis [85]. S. Rai et al. suggest short-term and mid-term load forecasting for weekly and the following six months’ data based on smart-metered data from a real-world distribution system at the NIT Patna campus [86]. Hamid et al., in their research, used a supervised learning algorithm to predict hourly load for New York Independent System Operators. To predict load and maximize profit, they used smart grids and feature selection and extraction techniques [88]. The aim was to propose an algorithm to improve the characteristic of gradually describing the data information, and improve FOA (fruit fly optimization algorithm) to increase the scope and range of the component, reducing the chance of failure [87]. The paper presents an experiment analyzing energy load forecasting a time series of two observed events (electricity consumption and outside temperature) measured at the University of Murcia’s selected department building [90].
For the growth of load forecasting, the following factors play important roles. Forecasts are affected by factors including historical load patterns [84,87,88], weather, temperature [85,88,89,90], electricity consumption [90], wind speed [88], humidity [89], calendar seasonal information, economic events, and geographic information. The table presents all the important parameters for load forecasting.
A large amount of data is gathered from generation, along with previous load estimation plans of electricity via natural gas, lignite, river, imported coal, wind, solar, fuel oil, geothermal energy, etc. The collected data is aggregated and used to test the models and economic factors, like how price and currency were used in [84] to forecast the load. Historic load data and future weather patterns like temperature, humidity, and wind speed are the main parameters that affect the results of the proposed hybrid model [89]. To predict load in deep feature extraction modules, the feature extraction modules assist in extracting the load of 24 h on an hourly basis, the last 3 days’ hourly load data, and weather parameters like temperature, humidity, and type of day (weekday/weekend) [85]. The load forecasting model used meteorological variables, and past loads were correlated; temperature and humidity proved the most impactful factors in electricity load prediction [86]. The model was tested using hourly load and price data from NY-ISO for prediction [88]. The historical load data was used by Nui et al. for the proposed load forecasting algorithm [87]. The electricity consumption and temperature were used to predict the load by the university campus [90].
Expert systems, along with fuzzy, genetic algorithms, ANN, and genetic algorithms, are used to forecast loads. Various machine learning models like LSTM [84], ANN [84,89,90], SVR [86,90], ARIMA [89], MLR [86,88], SVM [88], MLP [88], LSSVM [87], RF, XGB, FNT [90] are used in calculating the load based on various time horizons. Along with a single model, to enhance the effectiveness of forecasting, some hybrid models are proposed to overcome the disadvantage of different models. It was noticed that the hybrid model [87,89] gives better performance as compared to single machine learning models.
  • Machine learning algorithms: ANN, SVR, ARIMA, SVM, LSSVM, RF, XGB, FNT;
  • Deep learning algorithms: LSTM, MLP;
  • Hybrid algorithms: (IEMD + ARIMA + WNN + FOA), IFOA-w-LSSVM.
Machine learning models like R2, MAE, and RMSE using various algorithms, such as LSTM, GRN, and CNN, were used to forecast load. It was found that LSTM gave the best result [84]. A robust hybrid model was proposed using IEMD (improved empirical mode decomposition), ARIMA, and WNN optimized by FOA, that used linear and non-linear components to optimize the accuracy of the load [89]. The paper uses a hybrid model combining the deep feature extraction module with the electricity load prediction module, such as SVR and ANNs models, demonstrating significant performance improvements [85]. ANN, MLR, and SVR were used to forecast the load generated by smart meters. The SVR results were the best for the test system when compared to ANN and MLR [86]. An enhanced version of SVM, MLP, and LR was proposed based on data preprocessing to remove redundancy, missing values, and poor-quality data [88]. Nui et al. proposed a power load forecasting IFOA-w-LSSVM (improved fruit fly optimization algorithm applied to the wavelet least square support vector machine) to improve the monthly load prediction for the regional grid and Zhejiang province in IFOA-w-LSSVM; the traditional Gaussian kernel was replaced with a wavelet kernel function [87]. Energy load forecasting using machine learning algorithms like ANN, SVR, RF, and eXtreme Gradient Boosting was used by T. Vantuch et al. They applied data indicating a time series of two observed events (electricity consumption and outside temperature) [90].

6. Power Consumption/Usage Prediction

Electricity demand, measured in units of W (or kW), is the rate at which electrical energy is used to achieve the desired output rating, whereas energy consumption, measured in units of Wh (or kWh), reflects the amount of electrical energy spent over a certain time period. Forecasting is used to predict future energy consumption to achieve an equilibrium between the demand and supply of energy.
Electricity is vital to the economic activities of the industry. Electricity forecasting is, therefore, very important. As a result, energy resources can be better planned and managed. There have been several methods proposed for predicting power consumption. Actual energy consumption is the amount of energy they consume from the existing electric grid in order to meet their transportation, residential, industrial, commercial, and other miscellaneous needs. Monitoring power consumption at various levels and time horizons includes consumption via smart buildings already utilizing Green AI technology [4], renewable and non-renewable resources [91], as well as consumption via selected cities [6], buildings [7], or any region. The prediction refers to the demand for electricity and power consumption. The paper [92,93] discusses short-term energy consumption, paper [91] talks about the mid-term, and [94] highlights long-term energy consumption. Table 6 shows the usage analysis with the help of machine learning models and its performance with respect to time horizons.
Several complex factors can affect energy consumption. Energy prediction methods for SGs are becoming more and more accurate as more renewable energy sources are used, contributing to the efficiency of SG planning. A study led by Z. Wu et al. in California calculated the energy consumption from energy-saving houses, i.e., smart buildings. The random forest gave about 50% better output than the other algorithms [92]. At Jeju, Korea, P. W. Khan et al. developed a novel machine learning-based hybrid model to enable Jeju Energy Corporation (JEC) to meet its all-power needs with renewable sources instead of imported electricity from the Korean Peninsula. This model analyzed joint power consumption trends in renewable and non-renewable energy for monthly forecasting considering data parameters [91]. A. Salam et al. performed a study on power consumption in Tetouan City, Morocco to calculate electricity consumption for every 10 min and hourly intervals. The RF was found to be the most accurate out of the four linear algorithms [93]. Guangdong, China has a high annual electricity consumption rate for residential buildings. A field study conducted by Q. Li et al. examined building performance parameters. The random forest was found to be the most accurate out of the linear algorithms used [94].
Statistical methods (econometric, technical analysis) were used forecast prices by combining past prices and/or previous and present values of exogenous factors, such as consumption, production, or weather parameters. The Table 6 above presents the parameters that play a major role in forecasting.
  • Weather Parameters: temperature, humidity, wind speed;
  • Electricity Parameters: historical electricity consumption (hourly, the previous day, weekly, quarterly, user-defined period), electricity generation, power load.
The study collected data on temperature, humidity, and electricity consumed by electric lights [92]. To analyze joint trends of power consumption, data parameters like power load and energy generated via fossil fuels, solar, and wind are considered [91]. After calculating using a smart grid, the hour and temperature were the two most prominent variables. To conduct the research, the data was collected from three different electricity suppliers in the city [93]. The researchers used performance criteria like heat transfer coefficients, inert thermal indexes of walls, and heat transfer coefficients of roofs [94].
In the literature, various methods for forecasting energy consumption were suggested, including the EDM (energy demand model), the ARIMA, and ANN, SVR for predicting.
  • Machine learning algorithms: ANN, SVR, SVM, LSSVM, RF, DT;
  • Deep learning algorithms: MLP;
  • Hybrid algorithms: (MLP + SVR + CatBoost).
For Energyblem, customers energy usage is predicted to achieve an equilibrium between the demand and supply of previous energy consumption. The implications of machine learning techniques to various problems, such as time series analyses and regression analyses, have demonstrated promising results. In the literature, a variety of techniques for predicting energy use have been put forth, such as SVR [91,92,93], ANN [92,93], RF [92,93], DT [93], SVM [94], and some hybrid algorithms, which are proposed models combining (MLP + SVR + CatBoost) [91]. The researcher have applied to various machine learning classification models like SVR, ANN, and RF for forecasting. After using the different ML models, the results from the random forest algorithm were found to overlap most with the data validation techniques that are used for smart buildings [92]. The hybrid model was proposed using MLP, SVR, and CatBoost algorithms. Compared with the rest of the models, the proposed MAPE gave much better results in finding out the consumption of renewable and non-renewable resources [91]. The data collected—followed by the ETL process and linear algorithms like RF, DT, SVR, and MLP—were applied to the data. In the study performed at Tetouan City, Morocco, it was observed that after the random forest, MLP could give the most accurate results [93]. ANNs and SVMs were used in the study to predict electricity usage per unit area of buildings in the community, and RMSE and MRE were used to calculate the errors. According to the study, SVM is the best model to calculatethe annual data [94].

7. Electricity Price Forecasting

  • Weather Parameters: temperature, humidity;
  • Electricity Parameters: historic electricity consumption (hourly, the previous day, weekly, quarterly, user-defined period), electricity generation, electricity price, electricity demand;
  • Other Parameters: price of natural gas, coke, biofuels, and geothermal generations.
The following Table 7 represents the price forecasting of electricity:
The study used data collected from power plants, power consumption, and weather conditions like temperature and humidity to calculate the price of electricity [95]. There are several factors to consider in this analysis, including regional aggregation of power generation and demand, recent prices, large hourly weather forecast records, and other chronological information to get the explanatory variables that impact the prices [96]. The data was collected from California system operators and the Berkeley Institute to predict the price. It also utilized historical data up until hour 24 of the previous day and day-ahead load predictions for each hour [97]. Utilizing historical information from the PJM interconnection system, predicting performance in low and peak pricing zones was evaluated for forecasting [98]. Historical correlation coefficients for electricity prices electricity that is generated from renewable and non-renewable sources. It includes weather conditions, price history of natural gas, electricity generated by Coke, electricity generation, and renewables (solar, wind, geothermal, and biofuels) [99]. Essential factors such as previous hourly load, hourly natural gas, and hourly weather conditions are considered for price prediction [100].
Energy costs alter as a result of variations in supply and demand. Prices often decrease when the energy supply increases, whereas they typically increase when there is a shortage. Prices go up as demand for energy rises, and down as demand declines. The models listed below aid in price prediction.
  • Machine learning algorithms: ANN, ARX, ARMAX, LSSVM, ARIMA;
  • Deep learning algorithms: MLP;
  • Hybrid algorithms: (LSSVM + ARMAX).
Banadaki used MLP with different layers to calculate and predict the results using Tensorflow, a distributed and highly scalable machine learning platform. An MAE with six hidden layers was observed to provide the most accurate result [94]. ANN-based algorithms like EMPF (explanatory model for price estimations) and novel REMPF (reference explanatory model for price estimations) were used to estimate the market price. Based on the explanatory variables, the MAPE of the REMPF model for the market was the lowest [95]. To predict the price in California for hours 1 to 24 of that day, ARX and ARMAX were used using time series data [96]. Yan et al. proposed a hybrid model combining LSSVM and ARMAX. During the study, in comparison to plain LSSVM and ARMAX, proposed models MAE and MSRE improved regression computation performance by 1.85% [97]. The research proposed a new approach for analyzing energy-related commodities using data analytics using S-ARIMA, Logistic Regression, SWM, KNN, and Random Forest, where Seasonal ARIMA gave the best results when measuring MAE, RMSE, and MSE [98]. In the case of Australian markets, the researcher used MNN to get accuracy, ranking algorithms of feed-forward neural networks with ten hidden layers in hybrid parallel topology; the Levenberg–Marquardt algorithm showed the best accuracy compared to MSE and MAPE. The research explained that the cascade–parallel-in cascade topology outperformed all the other topologies [100].

8. Future Research Direction, Challenge, and Needs

Environment challenges include climate change, energy efficiency, and renewable energy integration. These are the future research directions, challenges, and needs in Green AI:
Issue #1: Lack of a real-time green energy trading model. Supply-demand in the community must be addressed. Assuming each building in each class is an independent model, the model should be hierarchical. The issue pertains to the lack of a real-time model to predict green energy availability and adjust accordingly. The absence of a real-time trading model for renewable electricity refers to the lack of an efficient and effective system. Most renewable energy is currently sold through long-term contracts that can be rigid and cannot be adjusted according to supply and demand in real-time. Renewable electricity sources and green energy demand cannot be accurately analyzed in real-time without a real-time trading model. Consequently, resource use is inefficient, and renewable energy is less likely to be adopted.
Issue #2: Lack of a green energy emergency model; a model for energy management and decision-making that addresses emergencies such as earthquakes or floods. If this community has a power outage, how can they quickly replace it with power elsewhere? The problem is that the green energy supply chain lacks an emergency model for rapid response. Power outages can disrupt a stable and reliable energy supply.
Issue #3: Lack of a green energy community supply and demand model. In the absence of a sustainable community model that incorporates green energy, it can prove to be an issue for those who wish to implement sustainable practices at the community level and reduce the amount of energy they consume. Developing sustainable solutions without a green energy AI supply and demand community model is difficult. Energy planning needs to be community-based, energy efficiency and conservation need to be promoted, renewable energy increased, and energy storage and distribution systems improved.
Issue #4: The regulation policy standard of green energy data. To facilitate the development of sustainable and energy-efficient AI technologies, policies and regulations are necessary to encourage access, sharing, and the standardization of green energy data. AI technologies need to be allowed to be built on a reliable data source and to create more efficient AI systems. It would also encourage the development of green energy AI, which could help reduce usage and dependency on fossil fuels.
The corresponding needs for the issues are given as follows.
Need #1: Distributed Model for developing Green Energy data storage. Our smart communities, energy trading, and predictions need a distributed hierarchical energy model. A distributed energy model that can provide sustainable and reliable energy to communities must consider comprehensive factors, not single ones. Renewable energy sources, like solar and wind, can be generated and distributed using a decentralized infrastructure.
Need #2: Real-time energy supply and demand decision-making model; a model that supports the decision-making process. Monitoring and managing energy supply and demand in real-time optimizes efficiency. Machine learning algorithms are needed to predict and adjust energy supply constructed from the data collected from smart meters as well as sensors. Providing support to energy service companies to make the right decisions. A decision-making model is required to support energy services in making informed decisions. Data analysis and machine learning algorithms can predict energy demand and supply.
Need #3: Cyber Green energy System Infrastructure. To support big data, the existing utility infrastructure should be converted to cloud-based infrastructure. We can manage trading data by collecting, monitoring, driving, and analyzing data. Energy consumption and production data need to be collected, managed, and analyzed via cloud infrastructure. Based on cloud-connecting infrastructure, we can connect diversifying cloud-based infrastructure. Developing cloud-based infrastructure is the most important need to increase the cyber security and effectiveness of Green AI. This will reduce the cost of green system management.
Need #4: New model to address green energy emergencies. An emergency model is required in the event of a natural disaster or other event disrupting the energy supply chain. To ensure the continued flow of energy to the community, contingency plans need to be developed and implemented.
Need #5: Smart community green energy system. A rising demand for energy-efficient and sustainable solutions for communities has led to developing community green energy intelligent systems. These built systems offer real-time data on usage, production, and distribution. They also focus on energy saving and greenhouse issues related to emissions and reduction. Communities may also be able to work together towards achieving shared energy goals and promoting sustainable development with the help of community green energy smart systems.
Need #6: Green AI data accessibility and policy remaking. Establishing partnerships with data providers, advocating for data sharing policies, encouraging data standardization, developing alternative data types, and utilizing synthetic data are essential to addressing this challenge. It is possible to develop more effective and innovative solutions for green energy by using these solutions, which can provide greater access to proprietary data and APIs.
Need #7: Security and privacy of green energy data. For Green AI to be successful, data security must protect sensitive information, ensure system reliability and availability, and promote confidence in AI for sustainable energy solutions. A secure data platform is also important for data processing, storing, and analyzing. In addition to detecting and responding to malicious activities, data security must prevent data breaches. Data privacy in sustainable energy solutions is the most important factor in protecting personal and sensitive data, preventing unauthorized access and use, and maintaining trust.
The corresponding challenges for the issues are given as follows.
Challenge #1: Access to green energy system data. As part of the Green AI Challenge, participants of the competition have established data sharing agreements and APIs, established public-private partnerships, developed synthetic data, collected crowdsourced data, and implemented data anonymization techniques to solve the issue of access to proprietary data. Access to proprietary data can be provided while protecting the interests of data owners and maximizing societal benefits through these solutions.
Challenge #2: Ensuring transparent and inclusive decision-making processes. Decision-making processes can be improved by involving various stakeholders and affected communities. A more informed decision can be made with this approach, which ensures that their concerns are considered. To prevent conflicts of interest, transparency in decision-making is necessary to ensure honesty and openness. Green energy AI initiatives can be ethical, effective, and sustainable if transparent and inclusive decision-making processes are considered.
Challenge #3: Green energy model trade-offs between performance and sustainability. It is common for Green AI models to sacrifice performance for sustainability. Despite its potential for acceptance, this can limit AI’s effectiveness in addressing environmental challenges. To maximize the impact of Green AI, it is essential to develop models that can balance performance and sustainability. The development of these models should be focused on finding solutions that are both efficient and effective. Furthermore, it is important to consider Green AI’s social and ethical implications.
Challenge #4: Green energy trading and market volatility. A volatile market results in sudden price changes and uncertainty for traders and buyers of green energy. Several factors can influence the price of environmental commodities, including regulatory changes, supply and demand, and technological advancements. Managing market volatility through risk management strategies such as hedging and insurance can mitigate price fluctuations for buyers and sellers.
The corresponding Future Research Directions:
Research #1: Optimizing algorithms. AI can be applied to analyzing complex issues related to the environment and energy, including weather patterns, renewable energy resources, and consumer behavior.
Research #2: Explore Green AI applications. From waste management to carbon capture and biodiversity conservation, there are numerous areas in which Green AI can be applied. These areas include exploring new applications of Green AI.
Research #3: Increasing the interpretability and explaining the ability of Green AI models. To successfully implement Green AI models, policymakers and stakeholders need to ensure that the models are transparent and comprehensible, so they can understand how predictions are made and make informed decisions in response to them.
Research #4: Collaborations between interdisciplinary experts. To advance research in Green AI and develop effective solutions to environmental problems, it is essential to extend cooperation between experts in artificial intelligence, environmental science, and other relevant fields.
Research #5: Using AGCEM as shown in Figure 2, programs like demand response, energy trading, and peer-to-peer energy sharing can be developed. In this way, AGCEM can help reduce energy costs and CO2 emissions by identifying the best time to consume energy and generate renewable energy.
Research #6: GCEDM (green community energy decision-making model): A green energy decision-making model such as GCEDM as shown in Figure 3 facilitates green energy decision-making. Combining data analytics, machine learning, and stakeholder engagement allows decision-makers to evaluate different energy scenarios and make informed decisions based on their goals. In addition to considering factors such as cost-effectiveness, energy security, and environmental impact, the model can help communities develop energy strategies tailored to their unique needs and circumstances.
The aim of this paper is to focus on community problems. Cities, communities, and smart grids can benefit from microgrids to solve the problem of community supply. A smart grid or microgrid can address a community’s energy supply and demand issue. In addition to facilitating more efficient and reliable energy distribution, these technologies can also incorporate renewable energy sources. Along with population growth and various weather patterns, energy consumption patterns also affect the supply and demand in a community. One should consider all the factors mentioned above, while addressing any kind of problem or developing a solution related to community development or specific community needs.

9. Conclusions

There is insufficient use of renewable energy resources in generating electricity and its relative factors like storage, usage, and pricing. Considering the quality of living and dependency on electric appliances, the availability of electricity is most important. At the same time, performing a literature review, this paper aimed to identify the most relevant research to the topic and contribute to the field of Green AI. We identified publications that show significant growth in the Green AI future.
Most researchers conclude that developing solutions that will support the sustainable development of Green AI will reduce the harm to the artificial systems and increase the efficiency and effectiveness of artificial systems. To achieve this sustainability for the development of the community, Green AI services play a major role. By monitoring usage patterns, prices, supply and demand, and considering the weather parameters and locations, we can use the ML and DL models to forecast electricity usage for various periods. Various topics fall under the umbrella of Green AI services, including monitoring, hyperparameter tuning, model benchmarking, and deployment.
Several Green AI services are emerging technologies, including power generation for green energy, load forecasting for green energy, demand response analysis for demand response, price forecasting for electricity, storage assessment for green energy, and power consumption and usage prediction. Furthermore, new data analysis, predictive analytics, and machine learning algorithms are also being developed to aid in developing these services.
It has been observed that most researchers can achieve an accuracy of 78% at minimum. Still, there are a few more parts where further research needs to be made, where the flexibility of the models, uncertain peaks, model reusability, and selecting the input parameters come into consideration. There is much room for researchers to experiment with effective, feasible, and measurable green strategies.
A major part of all the other factors of green services is power consumption, as it helps in economic planning and finding out how much electricity needs to be generated, and how many resources are required. Based on the above factors, the policies are generated, power generation planning is designed, and models are defined and considered. Duck curve and VPP provide huge support in managing all the green AI services. The researchers observed results that additional training of DNN would make the model one of the best suited for short and mid-term forecasting, along with LSTM. The few results highlighted empirical mode decomposition, and ARIMA gave better accuracy and efficiency by extracting the linear component of electricity load; FOA’s WNN is excellent at fitting the nonlinear component of original electricity loads, which is especially used in mid-term forecasting, and normalization helps remove the irrelevant features. It was also observed that reducing the dataset’s dimensions improves the computation time and cost of the system. The LSTM and MLP models clearly represent load and price forecasts. Conversely, some researchers observed that GRU and CNN contribute less towards the energy demand policies, as there is a lack of correlation between data while using these models. Despite all the factors that are considered for model selection, the research shows that the applied models should be less complex, as this adds to the repetitive retraining on the expanding database to control the complex systems.
Considering the best use of the strength of the models, electricity generation companies can increase the accuracy of forecasting, which will directly impact maximizing profitability and help in formulating long-term stability. But increasing the electricity generation brings up the challenge of electricity storage, as it is expensive. Therefore, generating electricity that precisely matches the market’s demand is necessary, and maintaining a backup for any uncertain situations like blackouts or electricity accidents is necessary. It is found that industries and domestic sectors rely heavily on electricity, and power cut-offs would lead to heavy losses; therefore, it is a challenge to supply the electricity in these areas with a sufficient amount of power, while considering the lower storage charges.
More challenges are needed to consider, while emphasizing Green AI. This paper concludes that more research is required at the community level, and various parameters need to be considered. There is a lack of solution papers offering tools or software packages that the community can use directly. In our view, Green AI is a hot area of research where growth and improvement are inexorable. Therefore, this paper aims to encourage researchers to find collective solutions to the challenges of Green AI, mainly at the community level; and it suggests they build software that industries and communities can use to track the usage, which will help them to find out where they can save power consumption and use the electricity efficiently. This paper has a limitation that highlights the problems related to Green AI services; as future work, we can focus on the available solutions for Green AI. While acknowledging the field’s maturity, we conclude that it is approaching a critical point to focus on and understand the usage of Green Energy and use it in our daily lives.

Funding

This research received no external funding.

Data Availability Statement

Data available in a publicly accessible repository that does not issue DOIs but can be accessed from [81,83].

Conflicts of Interest

The authors declare no conflict of interest.

References

  1. Akhter, M.N.; Mekhilef, S.; Mokhlis, H.; Mohamed Shah, N. Review on Forecasting of Photovoltaic Power Generation Based on Machine Learning and Metaheuristic Techniques. IET Renew. Power Gener. 2019, 13, 1009–1023. [Google Scholar] [CrossRef] [Green Version]
  2. Das, U.K.; Tey, K.S.; Seyedmahmoudian, M.; Mekhilef, S.; Idris, M.Y.I.; Van Deventer, W.; Horan, B.; Stojcevski, A. Forecasting of Photovoltaic Power Generation and Model Optimization: A Review. Renew. Sustain. Energy Rev. 2018, 81, 912–928. [Google Scholar] [CrossRef]
  3. Vargas, S.A.; Esteves, G.R.T.; Maçaira, P.M.; Bastos, B.Q.; Oliveira, F.L.C.; Souza, R.C. Wind power generation: A review and a research agenda. J. Clean. Prod. 2019, 218, 850–870. [Google Scholar] [CrossRef]
  4. Karady, G.G.; Holbert, K.E. Electrical Energy Conversion and Transport: An Interactive Computer-Based Approach; John Wiley & Sons: Hoboken, NJ, USA, 2013. [Google Scholar]
  5. Kumari, A.; Gupta, R.; Tanwar, S.; Kumar, N. Blockchain and AI Amalgamation for Energy Cloud Management: Challenges, Solutions, and Future Directions. J. Parallel Distrib. Comput. 2020, 143, 148–166. [Google Scholar] [CrossRef]
  6. Weron, R. Modeling and Forecasting Electricity Loads and Prices: A Statistical Approach; John Wiley & Sons: Hoboken, NJ, USA, 2007. [Google Scholar]
  7. Yildiz, B.; Bilbao, J.I.; Sproul, A.B. A Review and Analysis of Regression and Machine Learning Models on Commercial Building Electricity Load Forecasting. Renew. Sustain. Energy Rev. 2017, 73, 1104–1122. [Google Scholar] [CrossRef]
  8. Ntakolia, C.; Anagnostis, A.; Moustakidis, S.; Karcanias, N. Machine Learning Applied on the District Heating and Cooling Sector: A Review. Energy Syst. 2022, 13, 1–30. [Google Scholar] [CrossRef]
  9. Wang, H.; Zhang, N.; Du, E.; Yan, J.; Han, S.; Liu, Y. A Comprehensive Review for Wind, Solar, and Electrical Load Forecasting Methods. Glob. Energy Interconnect. 2022, 5, 9–30. [Google Scholar] [CrossRef]
  10. Aslam, S.; Herodotou, H.; Mohsin, S.M.; Javaid, N.; Ashraf, N.; Aslam, S. A Survey on Deep Learning Methods for Power Load and Renewable Energy Forecasting in Smart Microgrids. Renew. Sustain. Energy Rev. 2021, 144, 110992. [Google Scholar] [CrossRef]
  11. Load Profile. Available online: https://en.wikipedia.org/wiki/Load_profile (accessed on 30 December 2022).
  12. Benefits of Demand Response in Electricity Markets and Recommendations for Achieving Them; U.S. Department of Energy: Washington, DC, USA, 2006. Available online: https://www.energy.gov/sites/prod/files/oeprod/DocumentsandMedia/DOE_Benefits_of_Demand_Response_in_Electricity_Markets_and_Recommendations_for_Achieving_Them_Report_to_Congress.pdf (accessed on 18 July 2023).
  13. VPP Explained: What Is a Virtual Power Plant? (n.d.). Next-Kraftwerke.Com. Available online: https://www.next-kraftwerke.com/vpp/virtual-power-plant (accessed on 18 July 2023).
  14. Suganthi, L.; Samuel, A.A. Energy Models for Demand Forecasting—A Review. Renew. Sustain. Energy Rev. 2012, 16, 1223–1240. [Google Scholar] [CrossRef]
  15. Weron, R. Electricity price forecasting: A review of the state-of-the-art with a look into the future. Int. J. Forecast. 2014, 30, 1030–1081. [Google Scholar] [CrossRef] [Green Version]
  16. Demir, S.; Mincev, K.; Kok, K.; Paterakis, N.G. Introducing Technical Indicators to Electricity Price Forecasting: A Feature Engineering Study for Linear, Ensemble, and Deep Machine Learning Models. Appl. Sci. 2019, 10, 255. [Google Scholar] [CrossRef] [Green Version]
  17. Lu, H.; Ma, X.; Ma, M.; Zhu, S. Energy Price Prediction Using Data-Driven Models: A Decade Review. Comput. Sci. Rev. 2021, 39, 100356. [Google Scholar] [CrossRef]
  18. Wei, Y.; Zhang, X.; Shi, Y.; Xia, L.; Pan, S.; Wu, J.; Han, M.; Zhao, X. A Review of Data-Driven Approaches for Prediction and Classification of Building Energy Consumption. Renew. Sustain. Energy Rev. 2018, 82, 1027–1047. [Google Scholar] [CrossRef]
  19. Mosavi, A.; Salimi, M.; Faizollahzadeh Ardabili, S.; Rabczuk, T.; Shamshirband, S.; Varkonyi-Koczy, A. State of the Art of Machine Learning Models in Energy Systems, a Systematic Review. Energies 2019, 12, 1301. [Google Scholar] [CrossRef] [Green Version]
  20. Jebaraj, S.; Iniyan, S. A Review of Energy Models. Renew. Sustain. Energy Rev. 2006, 10, 281–311. [Google Scholar] [CrossRef]
  21. Ventosa, M.; Baıllo, A.; Ramos, A.; Rivier, M. Electricity market modeling trends. Energy Policy 2005, 33, 897–913. [Google Scholar] [CrossRef]
  22. Mosavi, A.; Bahmani, A. Energy Consumption Prediction Using Machine Learning; A Review. Preprints 2019. [Google Scholar] [CrossRef] [Green Version]
  23. Wang, X.; Liu, Z.; Zhang, H.; Zhao, Y.; Shi, J.; Ding, H. A Review on Virtual Power Plant Concept, Application and Challenges. In Proceedings of the 2019 IEEE Innovative Smart Grid Technologies—Asia (ISGT Asia), Chengdu, China, 21–24 May 2019; pp. 4328–4333. [Google Scholar]
  24. Sikorski, T.; Jasiński, M.; Ropuszyńska-Surma, E.; Węglarz, M.; Kaczorowska, D.; Kostyla, P.; Leonowicz, Z.; Lis, R.; Rezmer, J.; Rojewski, W.; et al. A Case Study on Distributed Energy Resources and Energy-Storage Systems in a Virtual Power Plant Concept: Technical Aspects. Energies 2020, 13, 3086. [Google Scholar] [CrossRef]
  25. Rouzbahani, H.M.; Karimipour, H.; Lei, L. A Review on Virtual Power Plant for Energy Management. Sustain. Energy Technol. Assess. 2021, 47, 101370. [Google Scholar] [CrossRef]
  26. Sikorski, T.; Jasiński, M.; Ropuszyńska-Surma, E.; Węglarz, M.; Kaczorowska, D.; Kostyła, P.; Leonowicz, Z.; Lis, R.; Rezmer, J.; Rojewski, W.; et al. A Case Study on Distributed Energy Resources and Energy-Storage Systems in a Virtual Power Plant Concept: Economic Aspects. Energies 2019, 12, 4447. [Google Scholar] [CrossRef] [Green Version]
  27. van Summeren, L.F.M.; Wieczorek, A.J.; Bombaerts, G.J.T.; Verbong, G.P.J. Community Energy Meets Smart Grids: Reviewing Goals, Structure, and Roles in Virtual Power Plants in Ireland, Belgium and the Netherlands. Energy Res. Soc. Sci. 2020, 63, 101415. [Google Scholar] [CrossRef]
  28. Wang, H.; Riaz, S.; Mancarella, P. Integrated Techno-Economic Modeling, Flexibility Analysis, and Business Case Assessment of an Urban Virtual Power Plant with Multi-Market Co-Optimization. Appl. Energy 2020, 259, 114142. [Google Scholar] [CrossRef]
  29. Naval, N.; Sánchez, R.; Yusta, J.M. A Virtual Power Plant Optimal Dispatch Model with Large and Small-Scale Distributed Renewable Generation. Renew. Energy 2020, 151, 57–69. [Google Scholar] [CrossRef]
  30. Dahai, Z.; Yunyun, Y.; Xiaojun, W.; Jinghan, H. Multi-time scale of new energy scheduling optimization for virtual power plant considering uncertainty of wind power and photovoltaic power. Acta Energiae Solaris Sin. 2022, 43, 529. [Google Scholar] [CrossRef]
  31. Mohy-ud-din, G.; Muttaqi, K.M.; Sutanto, D. Adaptive and Predictive Energy Management Strategy for Real-Time Optimal Power Dispatch from VPPs Integrated with Renewable Energy and Energy Storage. IEEE Trans. Ind. Appl. 2021, 57, 1958–1972. [Google Scholar] [CrossRef]
  32. Essakiappan, S.; Shoubaki, E.; Koerner, M.; Rees, J.-F.; Enslin, J. Dispatchable Virtual Power Plants with Forecasting and Decentralized Control, for High Levels of Distributed Energy Resources Grid Penetration. In Proceedings of the 2017 IEEE 8th International Symposium on Power Electronics for Distributed Generation Systems (PEDG), Florianópolis, Brazil, 17–20 April 2017; pp. 1–8. [Google Scholar]
  33. Synergy. Everything You Need to Know about the Duck Curve. Synergy. Available online: https://www.synergy.net.au/Blog/2021/10/Everything-you-need-to-know-about-the-Duck-Curve (accessed on 18 July 2023).
  34. Renewables and Emissions Reports. Caiso.com. Available online: https://www.caiso.com/market/Pages/ReportsBulletins/RenewablesReporting.aspx (accessed on 18 July 2023).
  35. Schmalensee, R. Competitive Energy Storage and the Duck Curve. Energy J. 2022, 43. [Google Scholar] [CrossRef]
  36. Torabi, R.; Gomes, A.; Morgado-Dias, F. The Duck Curve Characteristic and Storage Requirements for Greening the Island of Porto Santo. In Proceedings of the 2018 Energy and Sustainability for Small Developing Economies (ES2DE), Funchal, Portugal, 9–12 July 2018; pp. 1–7. [Google Scholar]
  37. Mahmud, K.; Azam, S.; Karim, A.; Zobaed, S.; Shanmugam, B.; Mathur, D. Machine Learning Based PV Power Generation Forecasting in Alice Springs. IEEE Access 2021, 9, 46117–46128. [Google Scholar] [CrossRef]
  38. Liu, C.-H.; Gu, J.-C.; Yang, M.-T. A Simplified LSTM Neural Networks for One Day-Ahead Solar Power Forecasting. IEEE Access 2021, 9, 17174–17195. [Google Scholar] [CrossRef]
  39. Jebli, I.; Belouadha, F.-Z.; Kabbaj, M.I.; Tilioua, A. Prediction of Solar Energy Guided by Pearson Correlation Using Machine Learning. Energy 2021, 224, 120109. [Google Scholar] [CrossRef]
  40. Zhou, H.; Liu, Q.; Yan, K.; Du, Y. Deep Learning Enhanced Solar Energy Forecasting with AI-Driven IoT. Wirel. Commun. Mob. Comput. 2021, 2021, 1–11. [Google Scholar] [CrossRef]
  41. Anuradha, K.; Erlapally, D.; Karuna, G.; Srilakshmi, V.; Adilakshmi, K. Analysis of Solar Power Generation Forecasting Using Machine Learning Techniques. E3S Web Conf. 2021, 309, 01163. [Google Scholar] [CrossRef]
  42. Li, Z.; Rahman, S.M.; Vega, R.; Dong, B. A Hierarchical Approach Using Machine Learning Methods in Solar Photovoltaic Energy Production Forecasting. Energies 2016, 9, 55. [Google Scholar] [CrossRef] [Green Version]
  43. Yousif, J.H.; Al-Balushi, H.A.; Kazem, H.A.; Chaichan, M.T. Analysis and Forecasting of Weather Conditions in Oman for Renewable Energy Applications. Case Stud. Therm. Eng. 2019, 13, 100355. [Google Scholar] [CrossRef]
  44. Dairi, A.; Harrou, F.; Sun, Y.; Khadraoui, S. Short-Term Forecasting of Photovoltaic Solar Power Production Using Variational Auto-Encoder Driven Deep Learning Approach. Appl. Sci. 2020, 10, 8400. [Google Scholar] [CrossRef]
  45. Qing, X.; Niu, Y. Hourly Day-Ahead Solar Irradiance Prediction Using Weather Forecasts by LSTM. Energy 2018, 148, 461–468. [Google Scholar] [CrossRef]
  46. Tian, F.; Fan, X.; Wang, R.; Qin, H.; Fan, Y. A Power Forecasting Method for Ultra-Short-Term Photovoltaic Power Generation Using Transformer Model. Math. Probl. Eng. 2022, 2022, 1–15. [Google Scholar] [CrossRef]
  47. Sharma, N.; Sharma, P.; Irwin, D.; Shenoy, P. Predicting Solar Generation from Weather Forecasts Using Machine Learning. In Proceedings of the 2011 IEEE International Conference on Smart Grid Communications (SmartGridComm), Brussels, Belgium, 17–20 October 2011. [Google Scholar] [CrossRef]
  48. Babbar, S.M.; Lau, C.Y.; Thang, K.F. Long Term Solar Power Generation Prediction Using Adaboost as a Hybrid of Linear and Non-Linear Machine Learning Model. Int. J. Adv. Comput. Sci. Appl. 2021, 12. [Google Scholar] [CrossRef]
  49. Jung, Y.; Jung, J.; Kim, B.; Han, S. Long Short-Term Memory Recurrent Neural Network for Modeling Temporal Patterns in Long-Term Power Forecasting for Solar PV Facilities: Case Study of South Korea. J. Clean. Prod. 2020, 250, 119476. [Google Scholar] [CrossRef]
  50. Gao, Y.; Li, J.; Hong, M. Machine Learning Based Optimization Model for Energy Management of Energy Storage System for Large Industrial Park. Processes 2021, 9, 825. [Google Scholar] [CrossRef]
  51. Sun, C.; Zou, Y.; Qin, C.; Zhang, B.; Wu, X. Temperature Effect of Photovoltaic Cells: A Review. Adv. Compos. Hybrid Mater. 2022, 5, 2675–2699. [Google Scholar] [CrossRef]
  52. Ramli, M.A.M.; Prasetyono, E.; Wicaksana, R.W.; Windarko, N.A.; Sedraoui, K.; Al-Turki, Y.A. On the Investigation of Photovoltaic Output Power Reduction Due to Dust Accumulation and Weather Conditions. Renew. Energy 2016, 99, 836–844. [Google Scholar] [CrossRef]
  53. Tiboaca, M.E.; Costinas, S.; Radan, P. Design of Short-Term Wind Production Forecasting Model Using Machine Learning Algorithms. In Proceedings of the 2021 12th International Symposium on Advanced Topics in Electrical Engineering (ATEE), Bucharest, Romania, 25–27 March 2021. [Google Scholar]
  54. Qureshi, A.S.; Khan, A.; Zameer, A.; Usman, A. Wind Power Prediction Using Deep Neural Network Based Meta Regression and Transfer Learning. Appl. Soft Comput. 2017, 58, 742–755. [Google Scholar] [CrossRef]
  55. Deng, X.; Shao, H.; Hu, C.; Jiang, D.; Jiang, Y. Wind Power Forecasting Methods Based on Deep Learning: A Survey. Comput. Model. Eng. Sci. 2020, 122, 273–301. [Google Scholar] [CrossRef]
  56. Lin, W.-H.; Wang, P.; Chao, K.-M.; Lin, H.-C.; Yang, Z.-Y.; Lai, Y.-H. Wind Power Forecasting with Deep Learning Networks: Time-Series Forecasting. Appl. Sci. 2021, 11, 10335. [Google Scholar] [CrossRef]
  57. Ahmadi, A.; Nabipour, M.; Mohammadi-Ivatloo, B.; Amani, A.M.; Rho, S.; Piran, M.J. Long-Term Wind Power Forecasting Using Tree-Based Learning Algorithms. IEEE Access 2020, 8, 151511–151522. [Google Scholar] [CrossRef]
  58. Demolli, H.; Dokuz, A.S.; Ecemis, A.; Gokcek, M. Wind Power Forecasting Based on Daily Wind Speed Data Using Machine Learning Algorithms. Energy Convers. Manag. 2019, 198, 111823. [Google Scholar] [CrossRef]
  59. Alkesaiberi, A.; Harrou, F.; Sun, Y. Efficient Wind Power Prediction Using Machine Learning Methods: A Comparative Study. Energies 2022, 15, 2327. [Google Scholar] [CrossRef]
  60. Lv, J.; Zheng, X.; Pawlak, M.; Mo, W.; Miśkowicz, M. Very Short-Term Probabilistic Wind Power Prediction Using Sparse Machine Learning and Nonparametric Density Estimation Algorithms. Renew. Energy 2021, 177, 181–192. [Google Scholar] [CrossRef]
  61. Buturache, A.-N.; Stancu, S. Wind Energy Prediction Using Machine Learning. Low Carbon Econ. 2021, 12, 1–21. [Google Scholar] [CrossRef]
  62. Chandran, V.; Patil, C.K.; Merline Manoharan, A.; Ghosh, A.; Sumithra, M.G.; Karthick, A.; Rahim, R.; Arun, K. Wind Power Forecasting Based on Time Series Model Using Deep Machine Learning Algorithms. Mater. Today 2021, 47, 115–126. [Google Scholar] [CrossRef]
  63. Barque, M.; Martin, S.; Vianin, J.E.N.; Genoud, D.; Wannier, D. Improving Wind Power Prediction with Retraining Machine Learning Algorithms. In Proceedings of the 2018 International Workshop on Big Data and Information Security (IWBIS), Jakarta, Indonesia, 12–13 May 2018. [Google Scholar]
  64. Vidal, C.; Malysz, P.; Kollmeyer, P.; Emadi, A. Machine Learning Applied to Electrified Vehicle Battery State of Charge and State of Health Estimation: State-of-the-Art. IEEE Access 2020, 8, 52796–52814. [Google Scholar] [CrossRef]
  65. Chemali, E.; Kollmeyer, P.J.; Preindl, M.; Ahmed, R.; Emadi, A. Long Short-Term Memory Networks for Accurate State-of-Charge Estimation of Li-Ion Batteries. IEEE Trans. Ind. Electron. 2018, 65, 6730–6739. [Google Scholar] [CrossRef]
  66. Ni, Z.; Yang, Y. A Combined Data-Model Method for State-of-Charge Estimation of Lithium-Ion Batteries. IEEE Trans. Instrum. Meas. 2022, 71, 1–11. [Google Scholar] [CrossRef]
  67. Zhao, B.; Hu, J.; Xu, S.; Wang, J.; Zhu, Y.; Zhang, L.; Gao, C. Estimation of the SOC of Energy-Storage Lithium Batteries Based on the Voltage Increment. IEEE Access 2020, 8, 198706–198713. [Google Scholar] [CrossRef]
  68. Huang, Z.; Yang, F.; Xu, F.; Song, X.; Tsui, K.-L. Convolutional Gated Recurrent Unit–Recurrent Neural Network for State-of-Charge Estimation of Lithium-Ion Batteries. IEEE Access 2019, 7, 93139–93149. [Google Scholar] [CrossRef]
  69. Yang, K.; Tang, Y.; Zhang, S.; Zhang, Z. A Deep Learning Approach to State of Charge Estimation of Lithium-Ion Batteries Based on Dual-Stage Attention Mechanism. Energy 2022, 244, 123233. [Google Scholar] [CrossRef]
  70. Varshney, A.; Singh, A.; Pradeep, A.A.; Joseph, A.; Gopakumar, P. Monitoring State of Health and State of Charge of Lithium-Ion Batteries Using Machine Learning Techniques. In Proceedings of the 2021 IEEE 5th International Conference on Condition Assessment Techniques in Electrical Systems (CATCON), Kozhikode, India, 3–5 December 2021. [Google Scholar]
  71. Khan, N.; Ullah, F.U.M.; Ullah, A.; Lee, M.Y.; Baik, S.W. Batteries State of Health Estimation via Efficient Neural Networks with Multiple Channel Charging Profiles. IEEE Access 2021, 9, 7797–7813. [Google Scholar] [CrossRef]
  72. Shu, X.; Shen, J.; Li, G.; Zhang, Y.; Chen, Z.; Liu, Y. A Flexible State-of-Health Prediction Scheme for Lithium-Ion Battery Packs with Long Short-Term Memory Network and Transfer Learning. IEEE Trans. Transp. Electrif. 2021, 7, 2238–2248. [Google Scholar] [CrossRef]
  73. Bamati, S.; Chaoui, H. Lithium-Ion Batteries Long Horizon Health Prognostic Using Machine Learning. IEEE Trans. Energy Convers. 2022, 37, 1176–1186. [Google Scholar] [CrossRef]
  74. Deng, Z.; Hu, X.; Lin, X.; Xu, L.; Che, Y.; Hu, L. General Discharge Voltage Information Enabled Health Evaluation for Lithium-Ion Batteries. IEEE ASME Trans. Mechatron. 2021, 26, 1295–1306. [Google Scholar] [CrossRef]
  75. Huotari, M.; Arora, S.; Malhi, A.; Främling, K. Comparing Seven Methods for State-of-Health Time Series Prediction for the Lithium-Ion Battery Packs of Forklifts. Appl. Soft Comput. 2021, 111, 107670. [Google Scholar] [CrossRef]
  76. Hosen, M.S.; Youssef, R.; Kalogiannis, T.; Van Mierlo, J.; Berecibar, M. Battery Cycle Life Study through Relaxation and Forecasting the Lifetime via Machine Learning. J. Energy Storage 2021, 40, 102726. [Google Scholar] [CrossRef]
  77. Thomas, J.K.; Crasta, H.R.; Kausthubha, K.; Gowda, C.; Rao, A. Battery Monitoring System Using Machine Learning. J. Energy Storage 2021, 40, 102741. [Google Scholar] [CrossRef]
  78. Ren, L.; Zhao, L.; Hong, S.; Zhao, S.; Wang, H.; Zhang, L. Remaining Useful Life Prediction for Lithium-Ion Battery: A Deep Learning Approach. IEEE Access 2018, 6, 50587–50598. [Google Scholar] [CrossRef]
  79. Fei, Z.; Yang, F.; Tsui, K.-L.; Li, L.; Zhang, Z. Early Prediction of Battery Lifetime via a Machine Learning Based Framework. Energy 2021, 225, 120205. [Google Scholar] [CrossRef]
  80. Severson, K.A.; Attia, P.M.; Jin, N.; Perkins, N.; Jiang, B.; Yang, Z.; Chen, M.H.; Aykol, M.; Herring, P.K.; Fraggedakis, D.; et al. Data-Driven Prediction of Battery Cycle Life before Capacity Degradation. Nat. Energy 2019, 4, 383–391. [Google Scholar] [CrossRef] [Green Version]
  81. Birkl, C. Oxford Battery Degradation Dataset 1. 2017. Available online: https://www.semanticscholar.org/paper/Oxford-Battery-Degradation-Dataset-1-Birkl/58168615ec09229255e5119f26d8526582a287f2 (accessed on 30 December 2022).
  82. Aitio, A.; Howey, D.A. Predicting Battery End of Life from Solar Off-Grid System Field Data Using Machine Learning. Joule 2021, 5, 3204–3220. [Google Scholar] [CrossRef]
  83. Pozzato, G.; Allam, A.; Onori, S. Lithium-Ion Battery Aging Dataset Based on Electric Vehicle Real-Driving Profiles. Data Brief 2022, 41, 107995. [Google Scholar] [CrossRef]
  84. Ünlü, K.D. A Data-Driven Model to Forecast Multi-Step Ahead Time Series of Turkish Daily Electricity Load. Electronics 2022, 11, 1524. [Google Scholar] [CrossRef]
  85. Tong, C.; Li, J.; Lang, C.; Kong, F.; Niu, J.; Rodrigues, J.J.P.C. An Efficient Deep Model for Day-Ahead Electricity Load Forecasting with Stacked Denoising Auto-Encoders. J. Parallel Distrib. Comput. 2018, 117, 267–273. [Google Scholar] [CrossRef]
  86. Rai, S.; De, M. Analysis of Classical and Machine Learning Based Short-Term and Mid-Term Load Forecasting for Smart Grid. Int. J. Sustain. Energy 2021, 40, 821–839. [Google Scholar] [CrossRef]
  87. Dongxiao, N.; Tiannan, M.; Bingyi, L. Power Load Forecasting by Wavelet Least Squares Support Vector Machine with Improved Fruit Fly Optimization Algorithm. J. Comb. Optim. 2017, 33, 1122–1143. [Google Scholar] [CrossRef]
  88. Bano, H.; Tahir, A.; Ali, I.; Khan, R.J.U.H.; Haseeb, A.; Javaid, N. Electricity Load and Price Forecasting Using Enhanced Machine Learning Techniques. In Innovative Mobile and Internet Services in Ubiquitous Computing; Springer International Publishing: Cham, Switzerland, 2020; pp. 255–267. [Google Scholar] [CrossRef]
  89. Zhang, J.; Wei, Y.-M.; Li, D.; Tan, Z.; Zhou, J. Short Term Electricity Load Forecasting Using a Hybrid Model. Energy 2018, 158, 774–781. [Google Scholar] [CrossRef]
  90. Vantuch, T.; Vidal, A.G.; Ramallo-Gonzalez, A.P.; Skarmeta, A.F.; Misak, S. Machine Learning Based Electric Load Forecasting for Short and Long-Term Period. In Proceedings of the 2018 IEEE 4th World Forum on Internet of Things (WF-IoT), Singapore, 5–8 February 2018. [Google Scholar]
  91. Khan, P.W.; Byun, Y.-C.; Lee, S.-J.; Kang, D.-H.; Kang, J.-Y.; Park, H.-S. Machine Learning-Based Approach to Predict Energy Consumption of Renewable and Nonrenewable Power Sources. Energies 2020, 13, 4870. [Google Scholar] [CrossRef]
  92. Wu, Z.; Chu, W. Sampling Strategy Analysis of Machine Learning Models for Energy Consumption Prediction. In Proceedings of the 2021 IEEE 9th International Conference on Smart Energy Grid Engineering (SEGE), Oshawa, ON, Canada, 11–13 August 2021. [Google Scholar]
  93. Salam, A.; Hibaoui, A.E. Comparison of Machine Learning Algorithms for the Power Consumption Prediction: Case Study of Tetouan City. In Proceedings of the 2018 6th International Renewable and Sustainable Energy Conference (IRSEC), Rabat, Morocco, 5–8 December 2018. [Google Scholar]
  94. Li, Q.; Ren, P.; Meng, Q. Prediction Model of Annual Energy Consumption of Residential Buildings. In Proceedings of the 2010 International Conference on Advances in Energy Engineering, Beijing, China, 19–20 June 2010. [Google Scholar]
  95. Dehghan-Banadaki, A.; Taufik, T.; Feliachi, A. Big Data Analytics in a Day-Ahead Electricity Price Forecasting Using TensorFlow in Restructured Power Systems. In Proceedings of the 2018 International Conference on Computational Science and Computational Intelligence (CSCI), Las Vegas, NV, USA, 13–15 December 2018. [Google Scholar]
  96. Monteiro, C.; Fernandez-Jimenez, L.; Ramirez-Rosado, I. Explanatory Information Analysis for Day-Ahead Price Forecasting in the Iberian Electricity Market. Energies 2015, 8, 10464–10486. [Google Scholar] [CrossRef] [Green Version]
  97. Rafał, W.; Adam, M. Forecasting spot electricity prices with time series models. In Proceedings of the European Electricity Market EEM-05 Conference, Łódź, Poland, 10–12 May 2005. [Google Scholar]
  98. Yan, X.; Chowdhury, N.A. Mid-Term Electricity Market Clearing Price Forecasting: A Hybrid LSSVM and ARMAX Approach. Int. J. Electr. Power Energy Syst. 2013, 53, 20–26. [Google Scholar] [CrossRef]
  99. Yousefi, A.; Sianaki, O.A.; Sharafi, D. Long-Term Electricity Price Forecast Using Machine Learning Techniques. In Proceedings of the 2019 IEEE Innovative Smart Grid Technologies—Asia (ISGT Asia), Chengdu, China, 21–24 May 2019. [Google Scholar]
  100. Mosbah, H.; El-hawary, M. Hourly Electricity Price Forecasting for the next Month Using Multilayer Neural Network. Can. J. Electr. Comput. Eng. 2016, 39, 283–291. [Google Scholar] [CrossRef]
Figure 1. Taxonomy of green energy cloud AI service.
Figure 1. Taxonomy of green energy cloud AI service.
Energies 16 05718 g001
Figure 2. AGCEM (aggregated green community energy model).
Figure 2. AGCEM (aggregated green community energy model).
Energies 16 05718 g002
Figure 3. DGCEM (distributed green community energy model) and GCEDM (green community energy decision-making model).
Figure 3. DGCEM (distributed green community energy model) and GCEDM (green community energy decision-making model).
Energies 16 05718 g003
Table 1. Green AI Services.
Table 1. Green AI Services.
Green AI ServicesServices DescriptionsCommonly Used Machine Learning Models
Green Energy Power GenerationThe most common green energy generation is (1) Wind power, (2) Solar Energy, (3) Tidal power using turbines, and (4) Geothermal Energy. Wind energy generated by a gas flow can generate electricity in windy areas or where the terrain differs. Tidal energy is used to power generators. However, the water movement is triggered by changing tides, so this type of energy generation is categorized as hydropower. The world’s largest natural resource is geothermal energy. Unlike other green energy generation methods, geothermal energy is less affected by environmental and climate factors. Using sunlight to generate electricity, solar energy can be used to power developing regions where sunshine is available for the bulk of the day. Aside from this, the power generation period is sufficient to meet human needs for electricity. The duck curve shows that in locations with substantial amounts of solar and wind power capacity, power production peaks in the mid-evening hours when solar power is no longer available.LR, RPR, DTR, SVR, RFR, LSTM, MLP, LR, RF, ANN, SVM, CNN, GRU, RBM, RNN, SAE, VAE, BPNN, DNN, ARX
Green Energy Load ForecastingThere are three types of load forecasting: short, mid, and long-term. Regression models, time series forecasting, artificial neural networks, statistics, and fuzzy logic are mostly used for short-term forecasting. Along with the above algorithms, the end-use and econometric algorithm approaches are for long and mid-term forecasting. A similar-day approach, like weekends and holidays, is applied for mid and long-term forecasting.LSTM, ANN, SVR, MLR, MLP, SVM, LR, IFOA-w-LSSVM, RF, FNT
Load Profiling AssessmentLoad curves or load profiles in power systems show the variation in electrical demand or load over time. This information helps generation companies determine how much power they will need at any given moment. Energy load profiles show how much electricity one’s customers use daily and/or seasonally. It will show the variation in a customer’s electrical load over time.LSTM, ANN, SVR, MLR, MLP, SVM, LR, IFOA-w-LSSVM, RF, FNT
Demand Response AnalysisDue to supply constraints, the DR (demand response) process can reduce energy load during peak demand periods. Demand response involves adjusting or reducing electricity demand to maintain system balance. A few electric system planners and operators wisely calculate the demand and provide the supply; this helps in maintaining balance. This leads to the price reduction of a wholesale market, lowering the retail rates. This can benefit the customers through time-based, critical peak, and variable peak pricing. The prices will be real-time prices that will engage the customers in rebates for peak usage and demand response efforts.
Electricity Price Forecasting (EPF)The EPF is a division of electricity that helps forecast wholesale electricity prices, both forward and spot. Electricity prices are forecasted using statistics (econometrics, technical analysis) techniques based on previous prices and exogenic factors. Typically, prices are forecast using consumption and production figures, as well as weather variables.MLP, ANN, ARX, ARMAX, S-ARIMA, MNN
Green Energy Storage AssessmentThe assessment avoids the imbalances between energy demand and production. This process captures energy at a time and uses it for the future.
This process captures energy each time and uses it for the future.
It is generally known as a battery or accumulator when it is used to store energy.
LSTM, RNN, Physics Constrained NN, MEA-BP Algo., GRU, CNN, ED, LSTM, RF, MLP, LR, PLS Reg., SVM, RVM, GPR, Gradient Boosting
Electricity Outage AnalysisIt is simply a power failure that leads to electrical power supply interruption.
Power Consumption/Usage PredictionForecasting energy consumption is a time series regression problem. The method involves predicting a customer’s energy consumption based on their history.
Various machine learning algorithms have demonstrated promising results along with time series and regression.
SVR, ANN, RF, DT, SVM
Table 2. Solar Energy Generation and Prediction.
Table 2. Solar Energy Generation and Prediction.
IDPrediction HorizonModelsModel PerformanceParameters
Solar Solar PositionWeather Power
[37]Short-termLRMAE: 0.0282 R2: 0.9657YNYN
PRMAE: 0.0159 R2: 0.9880
DTRMAE: 0.0157 R2: 0.8944
SVRMAE: 0.0137 R2: 0.9799
RFRMAE: 0.0098 R2: 0.9919
LSTMMAE: 0.1492 R2: −1.4027
MLPMAE: 0.044 R2: 0.5618
[38]Short-termLSTMRMSE: 0.512YNYY
[39]Short-termLRMAE: 0.0013 RMSE: 0.002YNYN
RFMAE: 2.64 × 10−1 RMSE: ~0.0009
SVRMAE: ~0.04 RMSE: ~0.07
ANNMAE: 0.03 RMSE: 0.08
[40]Short-termCNN-ALSTMRMSE: 1.30 MAE: 0.70NNNY
[41]Short-termSVMRMSE: 131.44 MAE: 77.16NNYN
RFRMSE: 58.57 MAE: 48.39
LRRMSE: 27.32 MAE: 12.45
[42]Short-termANNRMSE: 15 MINS- 42.15, 1 HR- 63.62, 24 HR- 182.64NYYN
SVRRMSE: 15 MINS- 43.52, 1 HR- 66.87, 24 HR- 185.44
[43]Short-termSVMMSE: 0.02637NNYN
[44]Short-termBiLSTMRMSE: 18.309YYYY
CNNRMSE: 24.105
ConvLSTM2DRMSE: 17.31
GRURMSE: 17.454
LSTMRMSE: 17.47
RBMRMSE: 17.564
RNNRMSE: 18.101
SAERMSE: 17.724
VAERMSE: 17.31
[45]Short-termLRRMSE: 0.196NNYN
BPNNRMSE: 0.150
LSTMRMSE: 0.123
[46]Short-termTransformerMAE: 0.092YNYN
GRUMAE: 0.132
DNNMAE: 0.139
[47]Short-termSVMRMSE: 163NNYN
[48]Medium-termHybrid AdaboostMAPE: 8.88YNYN
RNNMAPE: 10.88
SVMMAPE: 11.78
ARXMAPE: 13
[49]Long-termLSTMMAPE: 0.805%YYYN
Table 3. Wind Energy Generation and Prediction.
Table 3. Wind Energy Generation and Prediction.
IDPrediction HorizonModelsPerformanceParameters
Wind SpeedWind DirectionHistorical Wind Power GenerationTemperature and Humidity
[53]Short-termRPRRMSE: 9.42YYNN
[54]Short-termSVRMAE: ~0.232 RMSE: ~0.284YYYN
ARIMAMAE: ~0.447 RMSE: ~0.541
[55]Short-termTRARMSE: 1.2459YYNY
FNNRMSE: 1.1579
MSSRMSE: 1.1050
HybridNNRMSE: 1.0543
[56]Long-termTCNMAPE: 5.13YYNY
LSTMMAPE: 9.12
RNNMAPE: 173.87
GRUMAPE: 6.25
[57]Long-termDTRMSE: 29.45 R2: ~0.9996YYNY
BaggingRMSE: 25.15 R2: ~0.9996
RFRMSE: 25.27 R2: ~0.9996
AdaBoostRMSE: 25.5 R2: ~0.9996
GBoostRMSE: 25.7 R2: ~0.9996
XGBoostRMSE: 22.4 R2: ~0.9996
[58]Long-termLASSOR2: ~0.8619 RMSE: ~164.61YNYN
kNNR2: ~0.9852 RMSE: ~53.82
XGBoostR2: ~0.9939 RMSE: ~34.40
SVRR2: ~0.992 RMSE: ~38.52
RFR2: 0.995 RMSE: 30.224
[59]Very short-termGPRR2: 0.95YYNN
[60]Short-termLASSORMSE: ~0.08678NNYN
LRRMSE: ~0.11671
RidgeRMSE: ~0.10173
SVRRMSE: ~0.10652
RVMRMSE: ~0.11116
DTRMSE: ~0.09457
LSTMRMSE: ~0.09580
[61]Short-termSVRRMSE: 1286.74YNNY
RTRMSE: 2207.74
RFRMSE: 1077.82
ANNRMSE: 870.92
[62]Short-termLSTMMSE 0.1358YYNY
GRUMSE 0.130
RNNMSE 0.143
[63]Short-termGBoostRMSEp: 17%YYNY
Table 4. Green Energy Storage Analysis and Prediction.
Table 4. Green Energy Storage Analysis and Prediction.
IDPurposeBattery TypeModelsModel Performance
[65]SoC EstimationLithium-ion battery packsLSTM-RNN0.007 RMSE
[66]SoC EstimationNickel-cobalt-aluminum oxidePhysics Constrained NN0.0036 RMSE
[67]SoC Estimation based on VoltageLithium-ion batteries connected in parallelMEA-BP Algo.0.51 MAPE%
[68]SoCBAK 18,650 batteryGRU0.0177 RMSE
EstimationCNN-GRU0.0167 RMSE
[69]SoC Estimation18,650 lithium-ion battery batteriesED1.3913 RMSE
DA-LSTM0.4918 RMSE
DA-GRU0.3869 RMSE
[70]Monitoring SoCCommercial Lithium iron phosphateRandom ForestAccuracy 0.9759
MLPAccuracy 0.9738
Monitoring SoHCommercial Lithium iron phosphateLinear Reg.1.07 RMSE
(With Savgol Filter)PLS Reg.1.54 RMSE
MLP1.96 RMSE
Random Forest1.53 RMSE
[71]SoH EstimationLithium-ionLSTM0.0249 RMSE
BiLSTM0.0099 RMSE
[72]SoH PredictionNickel cobalt Manganese;LSTM with Transfer Learning0.42 RMSE
Lithium cobalt oxide;
Lithium iron phosphate;
[73]Long Horizon SoH PredictionLithium-ionRNN0.0157 RMSE
[74]SoH EvaluationLithium iron phosphate;Linear Reg.0.0179 RMSE
Lithium cobalt oxide;SVM0.0132 RMSE
Nickel cobalt;RVM0.0155 RMSE
ManganeseGPR0.0127 RMSE
[75]SoH time series predictionLithium-ionGradient Boosting0.26 RMSE
[76]Lifetime ForecastingNickel Cobalt ManganeseGPR0.0159 RMSE
[77]Battery Monitoring System—RULLithium-ionLSTM0.058 RMSE
[78]RUL PredictionLithium-ionADNN0.0666 RMSE
Bayesian Reg.0.1192 RMSE
Linear Reg.0.1200 RMSE
SVM0.1066 RMSE
[79]Early Lifetime PredictionLithium-ionGPR119 RMSE
SVM115 RMSE
RF152 RMSE
Table 5. Load forecasting using Machine Learning.
Table 5. Load forecasting using Machine Learning.
IDPrediction HorizonModelsModel PerformanceParameters
Weather
Historical Load DataTemperatureHumidityWind SpeedElectricity Price CurrencyDay/TimeSmart Meters
[84]Short-term and Mid-termLSTMR2: 0.94 RMSE: 21,561.57 MAE: 15,502.61YYNNYYNN
ANNR2: 0.73 RMSE: 45,782.79 MAE: 34,045.72
[89]Short-termHybrid modelMAE: 38.61 MAPE: 0.60%YYNNNNNN
[85]Short-termHybrid modelMAPE: 2670.6NYYYNNNN
[86]Short-term and Mid-termSVRMAPE: 3.60NYNNNNYY
ANNMAPE: 3.65
MLRMAPE: 4.53
[88]Short-termMLPMSE: 320.07 MAE: 109.3 RMSE: 146.07
MAPE: −45.41
YYYYNNNN
SVMMSE: 207 MAE: 83.03 RMSE: 117.7 MAPE: 31.28
LRMSE: 294.7 MAE: 112.9 RMSE: 140 MAPE: 32.52
[87]Mid-termIFOA-w-LSSVMMAPE: 1.42%YNNNNNNN
[90]Short-term and Long-term Day AheadWeek AheadYYNNNNNN
ANNMAPE 15.6115.3
SVRMAPE 11.8819.45
RFMAPE 16.4111.02
XGBMAPE 11.8412.33
FNTMAPE 24.1718.07
Table 6. Usage Analysis and Prediction.
Table 6. Usage Analysis and Prediction.
IDPrediction HorizonModelsModel PerformanceParameters
Weather
Historical Electricity Consumption DataTemperatureHumidityWind Speed
[92]Short-termSVRMAE: 45.94 MSE: 10,454.92YYNN
ANNMAE: 52.68 MSE: 4918.03
RFMAE: 33.84 MSE: 8231.56
[91]Mid-termHybridMAPE: 4.29YYNN
[93]Short-termRFRMSE: 3754.7 MAE: 2663.5YYYY
DTRMSE: 4613.9 MAE: 3962.3
SVRRMSE: 3898.7 MAE: 3046.0
[94]Long-termSVMRMSE: 2.39YYNN
Table 7. Forecasting Prices.
Table 7. Forecasting Prices.
IDPrediction HorizonModelsModel PerformanceParameters
Historical Data Weather
Electricity ConsumptionElectricity GenerationElectricity PriceLoad DataDemand DataTemperatureHumidity
[95]Short-termMLPMAE: 6 layers—1.84YYNNNYY
MAE: 2 layers—1.88
[96]Short-termANNMAPE: 10.23%NYYNNYY
[97]Short-termARXMDE: 2.06 MWE: 3.04NNYYNNN
ARMAXMDE: 2.8 MWE: 2.95
[98]Mid-termHybridMAE: 2.67 MSRE: 0.14YNYNYNN
[99]Long-termS-ARIMAMSE: 0.02 MAE: 0.09 RMSE: 0.12NYYNNNN
[100]Mid-termMNNMAPE: 3.25% MSE: 9.11NNYNNYN
Disclaimer/Publisher’s Note: The statements, opinions and data contained in all publications are solely those of the individual author(s) and contributor(s) and not of MDPI and/or the editor(s). MDPI and/or the editor(s) disclaim responsibility for any injury to people or property resulting from any ideas, methods, instructions or products referred to in the content.

Share and Cite

MDPI and ACS Style

Mehta, Y.; Xu, R.; Lim, B.; Wu, J.; Gao, J. A Review for Green Energy Machine Learning and AI Services. Energies 2023, 16, 5718. https://doi.org/10.3390/en16155718

AMA Style

Mehta Y, Xu R, Lim B, Wu J, Gao J. A Review for Green Energy Machine Learning and AI Services. Energies. 2023; 16(15):5718. https://doi.org/10.3390/en16155718

Chicago/Turabian Style

Mehta, Yukta, Rui Xu, Benjamin Lim, Jane Wu, and Jerry Gao. 2023. "A Review for Green Energy Machine Learning and AI Services" Energies 16, no. 15: 5718. https://doi.org/10.3390/en16155718

Note that from the first issue of 2016, this journal uses article numbers instead of page numbers. See further details here.

Article Metrics

Back to TopTop