The Electricity Load Prediction Model for Residential Buildings: A Critical Review of Output Types, Prediction Methods and Driving Factors

Abstract

An electrification revolution in the Chinese building energy field has been promoted by China’s carbon peak and carbon neutrality goals. An accurate electricity load prediction was essential to resolve the conflict between substations which was caused by the current increase in energy demand, on both the generation and consumption sides. This review provided an in-depth study of prediction models for residential building electricity load and inspected various output types, prediction methods and driving factors. The prediction types were divided into three categories: (i) time scale, (ii) geographical scale and (iii) regional scale. Predictive model building methods were classified as classical, algorithms based on Machine Learning (ML) or Deep Learning (DL) and hybrid methods. Driving factors included single and multiple features. By summarizing the driving factors, the influence of improving the prediction accuracy according to the characteristics of output types on selecting the driving factors correctly was discussed. The review provided a key perspective for future studies in electricity load prediction by analyzing the regional variations in electricity load characteristics. It suggested that the regional electricity load prediction model for residential buildings based on diverse driving factors in each region was established to offer valuable solutions for future residential planning and energy distribution.

1. Introduction

In recent decades, China’s demand for electricity has increased due to economic and technological advancements, population growth, and improved lifestyle standards, which has created both challenges and opportunities for sustainable development’s strategic positioning. Coal-fired electricity generation will account for 61% of total electricity generation in 2020 [1]. As of 2021, the electricity industry continued to be the primary supply sector for coal-fired power plants, accounting for 86% of China’s energy consumption [2]. As a result, China’s electricity industry’s high energy consumption and emissions have created ecological and environmental issues [3,4]. In the context of such high carbon emissions, urban residential buildings account for approximately 40% of emissions [5], and emissions continue to rise with the trend of increased urbanization [6]. In other countries as well, one study found that residential electricity consumption in Portugal has grown faster than GDP per capita since 1990, so it is necessary to study the electricity load of residential buildings [7]. The analysis and prediction of the electricity load of residential buildings can not only realize the effective allocation of electricity resources by evaluating the electricity load in a future time period and reduce the risk of supply and demand conflicts, but also adjust the national electricity supply side structural reform and systemize the planning of renewable energy systems in the long run [8].

To address energy consumption and emissions, governments around the world have implemented low-carbon policies. The proposed Peak Carbon and Carbon Neutral targets in China have helped to accelerate the transition to electrification of building energy use [9]. At the moment, both the electricity supply and the customer sides are beginning to shift from fossil fuels to clean electricity, with the electricity supply side focusing on new energy generation through the efficient use of renewable energy sources. Renewable energy generation from wind and solar photovoltaic can be electrically clean, but it can increase grid load during peak electricity times in summer and winter, resulting in a supply-demand imbalance [10]. Especially affected by seasonality, the seasonal electricity demand for cooling in the south and heating in the north surged. The seasonal capacity of renewable energy is precisely reduced [11,12], thus making it more difficult for the grid system to operate stably [13].

There are a variety of flexible loads in residential buildings, as shown in Figure 1, including adjustable loads, transferable loads and breakable loads. Flexible electricity consumption can be achieved by integrating both the subjective behavior of occupants and the objective factors of the environment, thereby enhancing the convenience of grid operations [14]. With the widespread adoption of distributed power supply and power storage technology, residential buildings are no longer just “payment” consumers, but have an identity based on the integration of “production-regulation-consumption”, which can not only account for energy consumption, but also for independent energy production [15]. Simultaneously, the flexibility of the use of personnel behavior for internal electricity regulation and control, to achieve energy adjustment, thus realizing the effect of flexible electricity [16]. However, “production-regulation-consumption” is prone to two-way feedback between each other, such as dynamic change [17] and time delay [18], which has a negative effect on the stable operation and flexible regulation of the electricity system to some extent. Therefore, considering the influence of energy users on the power load of residential buildings, the diversity and flexibility of the behavior of analysts to predict the change law of the power load of residential buildings is crucial for alleviating the contradiction between the substation, the generation side, and the consumption side while realizing the electricity company’s rational planning.

The widespread use of sensors, computer networks and smart grid technologies has greatly helped the collection of building energy information data. Smart grids, which are installed in a home or building’s primary or secondary circuits, measure, store and transmit real-time electricity consumption data with high temporal resolution [19]. These data can be analyzed and processed to benefit consumers and the industry [20]. At present, the application of smart grid in urban residential buildings is beginning to spread, and most young residents are positive about smart meters, stating that they can adjust the use of electrical appliances to achieve energy savings by recognizing the information fed back from the smart meter: real-time energy use and dynamic electricity prices [21]. As shown in Figure 2, power related departments and companies have improved the microgrid [22] and smart grid architecture to enable electricity meters to interact with different residential buildings. A large amount of information data are collected through smart meter feedback [23], which can then be analyzed for effective prediction, regulation and planning. Residential electricity consumption prediction, demand prediction and load prediction are all important topics dealing with the current increase in energy consumption in the electricity industry. In this paper, the above forecasts are expressed in terms of electricity load prediction. The electricity load prediction of residential buildings is realized by analyzing the relationship between different driving factor combinations and load data as well as the change rule [24]. However, the randomness of renewable energy, the volatility of electricity, the diversity of residential buildings and the flexibility of human behavior have brought great challenges to the power load prediction of residential buildings.

At present, the majority of existing electricity load prediction research focuses on prediction methods, including data processing, feature extraction and the optimization of prediction methods, with the goal of improving model accuracy. Power-related departments and companies need to quantify the balance between capacity and energy use to ensure the safety of power grid transmission and distribution. This is because electricity-related departments and companies must quantify the balance between electricity supply and consumption to ensure grid transmission and distribution security [25]. However, there is a lack of discussion on the relationship between the predicting target types and driving factors of a residential building electricity load prediction model in the context of the energy user’s integrated “production-regulation-consumption” identities.

Therefore, this paper makes a comprehensive summary of the current work of electricity load prediction model for residential buildings. Based on the published research results, this paper deeply studies the prediction types, driving factors and prediction methods used in the prediction model of residential building electricity loads, and makes the following conclusions:

• This paper analyzes the target types of the electricity load prediction model, and di-vides the output target types into three scales: (i) time scale; (ii) geographical scale; (iii) regional scale. The differences and similarities between urban and rural areas in the geographical scale of the prediction model of residential building electricity loads, as well as the characteristics of electricity load prediction in different regional scales, are discussed. The analysis lays the groundwork for future investigations into the diversification of the electricity load prediction model, as well as potential future research directions;
• This paper provides a systematic review of advanced modeling techniques for predicting electricity loads. We introduce and categorize electrical load prediction methods into three types: classical prediction methods, ML and DL prediction algorithms, and hybrid prediction methods;
• By comparing the relationship between the prediction types and the driving factors, the combination types of driving factors of the residential building prediction model are summarized. This study emphasizes the importance of correctly selecting driver combinations as input variables to achieve accurate prediction based on different output target types. It provides a valuable design basis for the study of future power system operation and planning. The influence of different driver combinations as input features on the accuracy of the prediction model is also discussed.

The remainder of the paper is arranged as follows. Section 2 describes the research review methodology and the literature’s data statistics. Section 3 describes the output types for the electricity load prediction model for residential buildings. Section 4 compares three aspects of model data analysis, prediction methods and model optimization to existing electricity load prediction model methods. Section 5 discusses the importance of selecting appropriate single- and multi-feature inputs for accurately predicting electricity loads with different types, as well as summarizing the input feature grouping categories of existing electricity load prediction models. Section 6 contains the main conclusions. Figure 3 shows a summary of the key processes used in this paper to establish an electricity load prediction model for residential buildings.

2. Review Methodology

In the literature review methodology, the PRISMA (Preferred Reporting Items for Systematic Reviews and Meta-Analyses) model can improve the transparency, integrity and quality of systematic reviews and meta-analysis studies [26]. Consequently, this paper employs the PRISMA model to conduct a systematic review of the literature on electricity load prediction for existing residential buildings, showing the review process from one stage to another stage, and gives the reasons for inclusion and exclusion of the literature screening.

2.1. Literature Search

A content-analysis-based literature review methodology was used. Databases such as the American Institute of Physics (AIP), Web of Science and Science Direct were used for the literature review. Considering the relevance and high quality of the retrieved documents, this paper adopts the literature search strategy of combining keywords, synonyms and Boolean operators “AND” and “OR” to conduct a comprehensive search on the prediction types, methods and driving factors in the electricity load prediction process of residential buildings. In addition, we also extended the search of the relevant literature published by the relevant author’s institution and the citations in the search results. There are relatively few studies in this part of extended search, but after the preliminary screening and full text reading, the studies with strong correlations were selectively included. The keywords of the literature search are shown in the

Table 1. Three key words of predictive model building process.

Prediction Targets (2010–2024)	Prediction Methods (1990–2024)	Driving Factors (2010–2024)
Residential building Short-term electricity load prediction Medium-term electricity load prediction Long-term electricity load prediction Urban and rural housing Climatic province Regional power load Electricity consumption	Data analysis Classical prediction algorithm Time series Regression analysis Machine learning algorithm Deep learning algorithm Combinatorial algorithm Hybrid algorithm Model evaluation Model optimization	Energy-using behavior Meteorological parameter Electrovalence Building parameter Historical data Electrical equipment Geographical location characteristics Economic level Family information

:

2.2. Selection Criteria

In the literature search process based on the PRISMA model, the first phase of the literature search was completed through clearly defined inclusion AND exclusion criteria for the research questions and objectives of the systematic review, and the second phase of the literature search strategy using a combination of keywords, synonyms and Boolean operators “and” and “OR”. The follow-up is literature screening and extraction. Figure 4 depicts literature retrieval, screening and inclusion. After excluding studies that were non-English, content biased, incomplete and of non-required age, the literature was screened using the following content criteria:

• The electricity load prediction field only considers residential buildings, excluding the electricity load prediction of other fields such as industry and commerce;
• In the establishment of prediction model, a variety of machine learning algorithms or deep learning algorithms were considered for the comparative analysis of prediction accuracy;
• Combining subjective and objective driving factors for residential building electricity load prediction objects;
• For the data collection stage, smart meters, questionnaires and a variety of data collection devices can be combined to fully collect subjective and objective driving information such as climate parameters, schedules and energy consumption habits of electricity consumers;
• For the establishment of the residential building electricity load prediction model, consider the characteristics of prediction targets of different granularity, including time range and spatial granularity.

As shown in the Figure 4, according to the PRISMA model process, 148 papers were searched from academic journals, and 29 papers were searched from extended sources, and a total of 177 results were searched. Of the 177 records, 11 were duplicates. Of the 166 results, 6 results published before 2010 excluding the classical literature, and five studies whose primary language of publication was not English were excluded. Among the remaining 155 results, 9 results unrelated to electricity load and 5 incomplete results were excluded. A qualitative analysis was performed on the remaining 143 results, and 19 results were excluded by comparing content criteria. The remaining 124 results are discussed in depth to summarize.

Table 2. Multi-scale power load forecasting models based on forecasting methods and driving factors.

Ref.	Output Types	Prediction Methods	Driving Factors
[27]	Ultra-short-term load (one-minute)	Prophet model with Variational Mode Decomposition, Recurrent Neural Network, Back Propagation Neural Network, Long Short-Term Memory, Gated Recurrent Unit and Time Convolution Network	Time series data
[28]	Ultra-short-term load	Convolutional Neural Network—Long Short-Term Memory models	Meteorological factor data
[29]	The daily load on winter and summer peak days	Multiple Linear Regression, Stochastic Time Series, General Exponential Smoothing, State Space Method and Knowledge-Based Approach.	Dry bulb temperature, dew point temperature and wind speed.
[30]	Short-term household load (hour level)	A novel multiple cycles self-boosted neural network (MultiCycleNet) framework	Household electricity consumption pattern
[31]	Mid-term electricity demand (months)	Multiple Regression models	Climatic and economic factors
[32]	Medium-term of load (a week ahead)	Ensemble Machine Learning models	Time series data
[33]	The annual load	Artificial Neural Networks	Calendrical information, annual peak loads and weather data
[34]	Long-term electricity load at hourly frequencies	Multiple Linear Regression models	Economic, environmental and weather conditions
[35]	Short-term electrical energy load	Convolutional Neural Network and Long Short-Term Memory Network	The residential building dataset
[36]	Mid and long-term energy demand in distribution grids	Dynamic Mode Decomposition algorithm	Diverse time series energy consumption data
[37]	Medium-Term Regional Electricity Load	Support Vector Machine, Random Forest, Non-Linear Auto-Regressive Exogenous Neural Network and Long Short-Term Memory	Historical load, temperature and wind speed
[38]	Short-term peak regulation demand of rural electricity	Long Short-Term Memory Network	Air condition system, indoor temperature, peak load reduction, and revenue
[39]	The seasonal electricity consumption, hourly electricity load, and peak and average loads for individual and regional rural residences	Stochastic model	Electricity usage behavior
[40]	Monthly electricity consumption	Multiple Linear Regression, Machine Learning method including Support Vector Machine, Random Forest and Deep Learning method including Long Short-Term Memory Network-Gated Recurrent Unit	Climatic and historical electricity datasets
[41]	Energy use of air condition system in residential buildings	Artificial Neural Networks and Gradient Boosting Decision Trees	Environment parameters and air condition behavior data of residents in eight cities across three different climate zones in China
[42]	Daily and annual residential electricity consumption during the non-heating period	Nonlinear Response Functions	Awareness of poor and senior, citizens different house layouts and outdoor air average temperature
[43]	Hourly residential electric load	Compressive spatio-temporal load forecasting approach	Historical load data, house size, occupancy level and usage behavior of appliances
[44]	Urban-scale per capita comprehensive electricity consumption and per capita residential electricity consumption	The grey prediction model	The 18 macroeconomic factors and 2 energy consumption indicators for 30 cities
[45]	hourly electricity consumption	Principal Components method	Electricity consumption data
[46]	Long-term electricity consumption of a region	Collaborative Principal Component Analysis and Fuzzy Feed-Forward Neural Network	The historical annual energy consumption in Taiwan
[47]	Short-term load on the user-side of micro-grid	A hybrid load model of Empirical Mode Decomposition, Extended Kalman Filter, Extreme Learning Machine with Kernel and Particle Swarm Optimization	The load of small residential areas
[48]	Pattern-based short-term loa	Linear Regression models	Time series data with multiple seasonal cycles
[49]	Monthly electricity demand	Multiple Regression model	Weather variables, gross domestic product, and population growth.
[50]	Long-term electricity consumption	Different Regression models based on co-integrated or stationary data	Historical electricity consumption, gross domestic product, gross domestic product per capita and population.
[51]	Electric load on multiple time scales (e.g., daily, weekly, quarterly, annually)	Auto-Regressive and Moving-Average Components	Historical load data
[52]	Annual and daily household electricity demand	A combination of statistical and engineering modelling approaches	Demographic characteristics, occupancy patterns, and the features, ownership, and utilization patterns of electric appliances
[53]	Electric load on multiple time scales (e.g., daily, monthly)	Linear and Polynomial Regression, Support Vector Regression and Random Forest	The 28 days electricity consumption of residential unit with a solar PV array
[54]	The next month and time-series household electricity consumption	Support Vector Regression model	Energy behaviors, personality trait, demographic/building features, weather indicators and the last month consumption
[55]	Regional electricity load	Support Vector Machines with Genetic Algorithms	The 20 load data for Taiwan regional electricity load
[56]	The 1 h- and 24 h-ahead electric load	Artificial Neural Network	Electric load and temperature data
[57]	The 1 h -ahead electric load	Deep Neural Network Convolutional, Neural Network and Recurrent Neural Network	Weather, timing and holiday information of the targeting hour and historical load of 24 h before
[58]	Individual household electric load	Pooling-Based Deep Recurrent Neural Network and Long Short-Term Memory network	Half-hourly sampled electricity consumption, customer types, allocation of tariff scheme and Demand Side Management
[59]	Short- and medium-term monthly electric load for a wider metropolitan area	Recurrent Neural Networks, Long Short-Term Memory network and genetic algorithm	Historical load data, holidays, weather and weekday features
[60]	The electricity consumption for residential buildings for the next 24 h	Feed-Forward Artificial Neural Network algorithm	The data from smart meters and weather sensors
[61]	Electric energy consumption in real-time, short-term, medium-term and long-term timespans	Combination of Convolutional Neural Network and Bi-directional Long Short-Term Memory	Several variables in the individual household electric power consumption and three variables collected from energy consumption sensors
[62]	Monthly residential electricity consumption	Multiple regression technique	Ambient temperature, rainfall, relative humidity, wind speed, economic and social factors
[63]	Electricity–temperature sensitivities	Segmented Linear Regression approach	Smart meter data records of electricity use for 1245 households, hourly ambient temperature records

shows a comprehensive review of the relevant literature on the establishment of the multi-scale electricity load prediction model based on prediction methods and driving factors.

3. Output Types

With the rise in residential building electricity demand and carbon emissions, the electrical clean-up transition has resulted in a gradual diversification of electricity load prediction applications into time and space. Therefore, the reasonable comparison and analysis of the characteristics of different residential building electricity load types have become an important direction for the development of today’s electricity industry. Analyzing the differences in output types is critical for focusing on the implementation of response policies to achieve effective grid capacity allocation and ensure stable operation, and this process can provide the electricity sector and the company with references for operating and dispatching modes for the development of future electricity trends [64]. This section focuses on the various output types for electricity load prediction, and three scales of output types are discussed using a summary of existing studies: time scale (Section 3.1), geographical scale (Section 3.2) and regional scale (Section 3.3).

3.1. Time Scale

Accurate electricity load prediction can help electricity companies to arrange the operation and scheduling of electricity supply, transmission and distribution for a future time period, but the time scale of electricity load prediction needs to be clarified before the electricity companies carry out the planning and control, because the prediction of the electricity load value in different time periods, it is necessary to select historical data at different times and add the driving factors that affect the electricity load for reference. As a result, this section discusses and analyzes electricity load prediction time scales, as well as the characteristics of various time scale predictions, using existing time scale prediction models from the literature. Existing electricity load prediction in advance of the prediction time length can be divided into four categories: very-short-term load prediction [28,29], short-term load prediction [30,31], medium-term load prediction [32,33] and long-term load prediction [34,35], with the differences shown in

Table 3. Time scale characteristics.

Time Scale	Output Targets	Function
Very short-term	Electricity loads for the next few minutes or tens of minutes	Real-time and emergency scheduling of electricity
Short-term	Electricity loads for the next day or multiple days in a row	Distribution and coordination of electricity Planning of electricity unit combinations
Medium-term	Future monthly or annual electricity loads	Electric electricity facilities repairs Large-scale electricity dispatching Financial supply plan
Long-term	Future electricity loads ranging from five to ten years	Future electricity unit construction investment and planning

. This section may be divided by subheadings. It should provide a concise and precise description of the experimental results, their interpretation, as well as the experimental conclusions that can be drawn.

Existing research has found that the majority of scheduling and planning for electricity system operation is based on short-term electricity load prediction [65]. Because electricity supply and demand imbalances are common during peak periods of consumption, accurate short-term load prediction can help to mitigate this risk [35]. Currently, the majority of the variables influencing the accuracy of short-term electricity load prediction are user behavior, variations in the quality and quantity of appliance types, and price fluctuations [66]. It should be noted that the time period for short-term future prediction is short, and there are special days in the prediction [36], such as holidays, as well as time variables with long regularity periods, which can affect the accuracy of short-term prediction.

With climate change, the widespread use of renewable energy, and the emergence of producer-consumer identity, medium- and long-term load prediction is gradually becoming a key technology that electricity utility departments must master [37]. In modeling, medium- and long-term predictions differ from short-term predictions in three ways. The first is that there are different data sensitivity requirements. Medium- and long-term load requirements are lower than short-term predictions, which require high temporal resolution data due to the volatility of residential activity [67]. The second is that the amount of training data varies. Because medium- and long-term predictions are made over a longer period of time [68], a large amount of data must be analyzed to extract the law of change in electricity load. Finally, the driving factors differ. Medium- and long-term prediction focuses on major events that occur in larger cycles, as well as seasonal variations, changes in demand for large equipment and economic changes [69]. Accurate long-term prediction is difficult to achieve due to the high level of uncertainty and unpredictability in the prediction process [70]. Furthermore, data for long time periods, such as meteorological conditions and economic data, are difficult to obtain, which can result in low accuracy in long-term prediction models [71]. The excessive overestimation of long-term prediction can result in a surplus of electricity equipment built with large investments, whereas underestimation can easily lead to an imbalance between supply and demand, resulting in customer dissatisfaction, such as the occurrence of extreme weather, which increases electricity consumption [72]. However, the underestimation of electricity transmission and distribution capacity causes grid failures and grid paralysis on the consumer side. As a result, the appropriate time of resolution for scientific analysis must be chosen based on the expected trend of load changes in the coming period.

3.2. Geographical Scale

To provide an informed perspective by delving deeper into the changing laws of electricity loads in the residential building sector, it is necessary to investigate the geographic variations of electricity loads in residential buildings [73]. At present, due to the influence of climatic conditions and geographical location, the electricity load of residential buildings varies greatly in aspects such as lifestyle, building type, economic development, population density and policy adjustment. In this section, the geographic scale of the output target for electricity load prediction is divided into two categories: urban–rural and climate.

In China, urban and rural areas, as living places for regular human activities and energy consumption, analyzing the characteristics of electricity loads in urban and rural residential buildings, plays an important role for socioeconomic development [74]. In recent decades, as income levels and living standards have risen, urban households have produced more energy demand and carbon emissions than rural households [75]. Nie et al. [76] investigated the gap between urban and rural residents’ electricity consumption under the influence of the electricity price reform policy and discovered that the current trend of daily electricity consumption of rural and urban residents is the same, but the daily electricity consumption of rural households is larger and more volatile than that of urban households, making electricity load prediction more difficult. This phenomenon can be understood from two perspectives.

The objective reasons are primarily energy type and policy changes. With the recent increase in public awareness of energy conservation [77] and the rapid development of clean heating [78], rural areas in the north began to respond to the policy of coal-to-electricity conversion; the use of clean energy instead of coal electricity generation is efficiently utilized, and thus the amount of electricity in rural residential buildings has increased significantly in comparison to previous years. Furthermore, in the field of energy, the use of distributed energy resources in northern rural areas has become increasingly common [79], such as photovoltaic electricity generation and distributed charging piles, but the distributed electricity supply is unstable due to changes in environmental parameters, resulting in unpredictable power generation performance.

Subjective reasons primarily include human behaviors and life concepts. It is estimated that China will eventually achieve a two-thirds urbanization rate, but more than 500 million people will remain in rural areas over the next two decades [80]. At the moment, rural areas are aging and have young children, and the majority of young and middle-aged laborers have moved to cities to work and study, resulting in the rural economy’s relative backwardness and consciousness. There are inherent differences between urban and rural areas in terms of building attributes, household incomes, infrastructure conditions, and material resource gaps [38], which result in different feedbacks of the electricity load of residential buildings between urban and rural residents with different changes in these driving factors. Furthermore, the flexibility of rural residents’ activities, such as living and sleeping, working arrangements, and energy use behaviors [39], will result in unstable electricity load changes, making it difficult to investigate the laws in the data.

Because rural electricity facilities are not comprehensive, buildings are aging, and other factors [81], which make it difficult to obtain rural electricity data and make it difficult for the electricity company to investigate the rule of change of the rural electricity load, research on rural residential building electricity load prediction is currently relatively less. As shown in Figure 5, with the widespread use of smart grid in urban residential buildings, electricity companies will be able to monitor electricity data via smart meters, allowing for more accurate prediction and analysis of electricity load in order to regulate and control effectively. As a result, it is recommended that future investment in rural electricity infrastructure, in addition to rural areas where daily electricity demand is high, make use of distributed energy resources to fully tap the resource potential of rural areas [82], in order to increase rural energy self-sufficiency.

China has a vast geographical area and a variety of climates. Because building thermal design must be tailored to the regional climate, China’s geography is divided into five building thermal design climate regions: severe cold region, cold region, hot summer and cold winter region, hot summer and warm winter region, and mild region.

In terms of electricity energy, the degree of variation in energy consumption of residential buildings varies by climate region [83]. This difference is mainly caused by the differences in climate conditions and geographical locations, which lead to the differences in residents’ building needs [84], energy use behavior and daily life. Residents in different regions have different comfort requirements for residential buildings, so there are variations in energy consumption behaviors, such as equipment parameter settings [40] and terminal equipment types [85]. Liu et al. [41] collected data on AC electricity usage from households in eight cities located in three climate zones in China and established a prediction model for AC electricity on/off behaviors based on an artificial neural network and gradient boosting decision tree algorithm, and found that there are large differences in energy use behaviors among households in the three climate zones, and the differences are mainly due to the geo-graphical characteristics and living habits. There are significant differences in energy use behaviors among households in three climate zones, but air conditioning use patterns are somewhat similar. Furthermore, differences in energy generation and consumption types and ways of energy use exist between regions of China due to differences in geography, lifestyle and climate [86].

3.3. Regional Scale

With increasing urbanization, the demand for electrical loads from residential buildings may exceed the capacity of the grid itself in the future, which could result in electricity system breakdowns, higher social crime rates, an increase in residential health problems, and economic paralysis [42], particularly for energy customers who rely heavily on electricity supply. Households are the unit of electricity consumption in society, so analyzing household electricity consumption benefits regional, urban, and social electricity planning [87], and the living environment of the family has a significant effect on the variation of electricity load. In particular, residential buildings of different sizes form regional electricity loads, which at the same time maintain their diversity. This section investigates and evaluates the regional scale of electricity load prediction. As shown in Figure 6, based on the summary and analysis of existing studies on electricity, load prediction for residential buildings at various regional levels is summarized and analyzed, and the regional scales of electricity load prediction output types are classified as household level, building level, city level and national level.

More research has been conducted on electricity load prediction for residential buildings at the household level. The term “household level prediction” refers to the creation of a residential electricity load prediction model using historical data from single-family or multi-family electricity loads in a residential building. The term “household level prediction” refers to the creation of a residential electricity load prediction model for a residential building using historical data on single-family or multi-family electricity loads in that residential building. The single-family residential electricity load prediction focuses on analyzing the complex relationship between electricity consumption and driving factors (for example, temperature, time variables, and occupancy). The multi-family residential electricity load prediction focuses on analyzing the differences in electricity load variations between different customers, and the study discovered that the differences are primarily due to household, building, and behavioral attributes [43]. Furthermore, Du et al. [88] examined the consumption of users on different floors of a residential building during the summer and transition seasons, and discovered that the building’s different indoor environments caused the electricity costs of households on lower floors to be significantly higher than the electricity costs of households on higher floors. As a result, when users live in the same building, the electricity load changes may appear to be large. Therefore, it is necessary to cluster and group the analysis according to the multiple influencing factors affecting these differences [89], and the prediction model is established separately for the characteristics of household electricity loads of different classes, which can reduce the prediction error.

The term “building level residential electricity load prediction” refers to determining the changing law of electricity load of various types of buildings by analyzing the physical characteristics of the buildings [90], as well as developing a more accurate residential electricity load prediction model for the entire building, which is critical for planning the transmission and distribution of the building complex’s electricity capacity in a region. There is heterogeneity in residential electricity load prediction at the city level [91], which results in different driver types of cities with different sizes and types having different impacts on energy consumption. It is difficult to analyze the electricity load variation law of residential buildings in different cities. It is mainly caused by two aspects, one is the difference of climate, geographical location and living habits among cities; Second, cities are inclusive. In the context of economic development and urbanization, they will accept residents from all over the world. Therefore, it is not easy to collect enough differentiated data [44] to build a model. Similarly, residential electricity load prediction at the national level faces the same challenges as city level prediction, but there are few predictions of residential electricity loads for different countries, and most studies have compared the relationship of income levels on national energy allocations, particularly in low-income countries [92].

Relatively speaking, most of the current research focuses on using the relevant data of single-family residential buildings to analyze the electricity load variation rule or the electricity load data of different users combined with different combinations of driving factors. Some studies predict the electricity load change of a specific city, but few analyze the characteristics of driving factors in different regions to establish representative feature sets at regional scale. Therefore, by establishing a regional residential electricity load prediction model to determine the level of electricity consumption in a specific region, capacity planning can be realized for transmission and supply grids that transmit electricity from a power supply or regional grid to a city or region.

4. Model of Prediction Methods

This section presents a systematic review of the relevant literature on prediction model methods, as well as a comparative analysis of existing model building methods, divided into three sections: model data analysis, prediction method and model optimization. This section is organized as follows: Section 4.1 describes and compares the most common current input feature analysis methods. Section 4.2 summarizes the methods currently used for mainstream prediction. Section 4.3 discusses model optimization methods.

4.1. Analysis of Model Data

In the process of electricity load prediction, many load driving factors need to be input to improve the prediction effect of the model. However, the number of input driving factors needs to be combined according to the type of predicted output types. Having too many driving factors tends to increase the amount of model computation and reduce model execution efficiency. On the contrary, using too few driving factors can reduce the accuracy of model calculations and result in an incorrect analysis of the relationship between electricity consumption and driving factors. Therefore, this section summarizes and analyzes the data analysis prior to building the electricity load model, focusing on the Correlation Analysis (CA), Principal Component Analysis (PCA) and Sensitivity Analysis (SA).

The CA is a statistical method that measures the relationship between factors. It calculates the correlation coefficient to determine the linear relationship between factors, such as the relationship between social, economic, and climatic factors in predicting electricity load [93]. Pearson, Spearman and Kendall are the most commonly used correlation coefficients [94]. The computational process is not dependent on specific data and can be applied to tasks such as factors selection and feature engineering.

The PCA transforms the original variable into a new set of mutually orthogonal principal components using linear transformation, which is a dimensionality reduction method used to solve the collinearity problem among multiple factors [95]. To address the issue of long-term load uncertainty [45], the load profile of total electricity consumption can be modeled using PCA, which takes into account driving factors in a more effective way to establish a collaborative mechanism, thereby improving the precision and accuracy of residential electricity load prediction [46]. The PCA method is primarily used to reduce data dimensionality while retaining the original data’s key features.

The SA examines the sensitivity of the output changes of the model are to changes in the system parameters or surrounding conditions composed of the combination of driving factors [96]. It is frequently used in optimization methods in which the SA can determine which parameters have a significant impact on the system or model, such as in electricity load prediction, where the effects of household size, occupancy, thermal resistance of wall insulation, and airflow rate on energy changes are investigated [97].

In the electricity load prediction, it is necessary to correctly analyze the relationship between the driving factors in the data or between the output target and the driving factors, so as to achieve the optimal prediction performance of the established model in the appropriate data through reasonable selection of data.

4.2. Summary of Prediction Methods

In recent years, the subject of residential building electricity load prediction has developed rapidly, which can be combined with smart meters and other technologies to monitor the generation, transmission and consumption of electricity. The “producer and marketer” can balance the relationship between supply and demand by predicting the law of electricity load in advance. But the core technology of building prediction model is prediction method, or the method of building prediction mathematical model. In this section, we categorize electricity load prediction methods into classical methods, algorithms based on ML and DL, and hybrid methods, and summarize the advantages, disadvantages, and scope of application of each prediction method to provide references for future load prediction.

4.2.1. Classical Prediction Method

Classical prediction methods primarily include regression analysis, time series methods and Karman filtering [98]. The Kalman filter algorithm has a simple numerical implementation and recursive properties, which can be used to estimate the model state and unknown parameter values in real time [47]. Linear modeling of residential short-term electricity load prediction with a Kalman filter based on smart meters. The Extended Kalman Filter (EKF) was then investigated as a nonlinear prediction model.

Regression analysis is a statistical method that explains changes in the dependent variable Y using changes in the explanatory variable X. The goal is to find an appropriate mathematical model and determine the model’s best fit coefficient using the provided data. Unitary regression [48] and multiple regression [49] are two types of regression analysis methods that are distinguished by the relationship between the explanatory and dependent variables. Different models can be used for regression analysis depending on the accuracy requirements of electricity load prediction. Simple regression prediction models with different input variables (historical electricity consumption, GDP, climate and population) are established using parametric and sensitivity analysis, and the prediction error is controlled within 10%. Different multivariate regression models should be chosen for different climates and building scenarios [99,100]. The calculation theory for the electricity load prediction regression analysis method is simple, and the arithmetic speed is quick, but it is difficult to clearly analyze the relationship between various load influencing factors and electricity load during the calculation process, resulting in low accuracy of the results.

The time series method was first proposed by American scholars Box et al. [101]. The theoretical method of mathematical modeling, which describes the relationship between time and load values based on historical load data, is used to forecast future trends and development of electricity system loads. The method necessitates a continuous time series array for historical load data. There are currently four types of time series methods: autoregressive (AR) model, moving average (MA) model, autoregressive moving average (ARMA) model, and autoregressive integrated moving average (ARIMA) [51], which are suitable for modeling historical load data in different seasonal cycles (daily, weekly, quarterly, and yearly) and effectively reduce prediction error. The time series method uses historical load data with a seasonal cycle for modeling and does not require any additional inputs [102]. The time series method, like regression analysis, is simple, but it only considers time variables and ignores other factors influencing electricity load, such as climate change, resulting in a large margin of error in the final prediction.

Although classical prediction model methods are simple and easy to estimate, their application in real-time changing environments may result in unsatisfactory prediction accuracy because electricity loads are affected by a number of nonlinear factors and the changes in electricity loads are stochastic.

4.2.2. Prediction Algorithms Based on ML and DL

With the advancement of computer technology, numerous structural networks and algorithms have emerged. These algorithms outperform classical statistical methods in terms of learning capabilities, can handle complex nonlinear functions, and can analyze the relationship between electricity loads and linear and nonlinear decision variables [103]. ML and DL algorithms, such as Support Vector Machine (SVM), Random Forests (RF), K-Nearest Neighbor (KNN), Artificial Neural Networks (ANN), Convolutional Neural Networks (CNN), Recurrent Neural Networks (RNN), Long-Short Term Memory (LSTM) and Gated Recurrent Unit (GRU), are commonly used for predicting electricity loads [52]. Jose Mari M. Arce et al. [53] used Linear Regression, Polynomial Regression, SVR, and RF algorithms to predict total residential end-of-month electricity consumption. The study found that nonlinear data cause the linear and polynomial regression prediction models to have large errors, while the SVM and RF prediction models have small errors, and RF is not suitable for long-term prediction.

The SVM was originally used for data classification, but due to its excellent nonlinear data processing capabilities, it can also be used for electricity load prediction [104]. The SVM convergence speed is fast, and there is no problem adjusting the number of network layers or the local optimization solution, but the computational difficulty is high due to the complex structure of the dataset. Support Vector Regression (SVR) is an application of SVM to regression problems. It combines various behaviors of household electricity consumption to predict consumption under multiple intervention strategies [54]. SVR has good performance in solving nonlinear regression and time series prediction problems [55]. ANNs are nonlinear systems that mimic the neural network of the human brain in order to learn and solve problems. An ANN is made up of three layers: input, one or more hidden layers, and output. The signal enters the input layer, travels through the hidden layers, and finally reaches the output layer. The training time required for the data is determined by a variety of factors, including the complexity of the problem, the amount of data, the network structure, and the training parameters used [56]. Muhammad Waseem Ahmad and Tawfiq Al-Saba et al. compared Back Propagation Neural Networks (BPNN) and Random Forest (RF) for long-term electricity load prediction and discovered that ANN produces results that are close to the actual data when dealing with long-term data and multiple parameters [105,106].

The DL algorithm, which attempts to extract features from a dataset and automatically learns the extracted ones, has a better method of learning the hidden layers and parameter initialization techniques [57] to learn the features of the data in depth than traditional neural networks. CNN, RNN, and LSTM are the most commonly used deep learning algorithms for electricity load prediction. The RNN is a type of neural network model for processing time-series data [107]. The basic structure of RNN is particularly simple, primarily by storing the network’s output in a memory unit, which enters the neural network with the next input, and the difference from the structure of other prediction models is the cyclic connection. However, while RNN can partially overcome the overfitting problem caused by some deep learning methods [58], as time passes, the growth of sequences causes pattern memory of previous moments to fade. Hochreiter et al. [108] proposed the LSTM, which can combine short-term and long-term memory and, to some extent, solves the problem of disappearing gradients. Salah Bouktif et al. [59] analyzed a city’s electricity consumption data by optimizing various LSTM parameters and feature selections, which improved prediction accuracy even more. The LSTM excels at handling large datasets with long time series. However, RNN and LSTM only consider the temporal correlation with the sequences, ignoring the sequences’ spatial features. The CNN uses a deep convolutional layer architecture to extract spatial features from historical load data in order to perform regression prediction and classification [109], which improves model accuracy. Nadjib Mohamed Mehdi Bendaoud et al. [110] proposed a 2D input novel CNN model for analyzing the degree of change of historical load data and driving factors affecting electricity consumption, which has important applications in electricity load prediction and time series studies.

4.2.3. Hybrid Prediction Methods

To improve prediction accuracy, some studies have attempted to integrate different models of hybrid models, which, when compared to a single model, can not only capture more electricity load variation laws, but also compensate for the single model’s shortcomings [111]. The first approach predicts the electricity load using different models, and the final predicted value is the sum of each model’s predicted value multiplied by its weights. For example, a single model such as BPNN, Genetic Algorithm Back Propagation Neural Network (GABPNN), Generalized Regression Neural Network (GRNN) and SVM are used to build an electricity load model for prediction, and the weights of each model are optimized. The predicted value of each model is multiplied by the weight to obtain the final predicted value [112].

The second approach is to divide the process of predicting electricity load into several constituent blocks, select appropriate models to compute each block separately, and then combine the final prediction results in multiple models to complete the realization in collaboration [113]. Hybrid models are commonly used to combine classical models with ML or DL models, as well as multiple ML and DL models. Multiple regression models are combined with LSTM models to predict short-, medium- and long-term national electricity loads [114]. Analyzing the relationship between electricity load and multiple driving factors, the (multiple linear regression) MLR and ANN models were combined to create a hybrid model to predict daily electricity consumption, and the prediction accuracy was found to be higher than that of a single prediction model through optimization [115]. Because of the large number of input features and historical load data in the prediction model, multiple ML and DL models can be used to complete the data processing, feature process, model building and model optimization stages of the electricity load prediction process. Oprea et al. [60] developed a prediction framework consisting of seven algorithms for data with various consumption categories and weather variables. The best performing algorithm can be chosen automatically for determining the load profile and day-ahead prediction based on various scenarios. Le et al. [61] proposed a hybrid CNN-Bi-LSTM model in which the CNN extracts variable-important information from individual household electricity consumption data, and the Bi-LSTM module uses the feature information of the electricity load prediction model, demonstrating that prediction accuracy is superior to existing methods. As a result, selecting appropriate prediction methods to achieve accurate electricity consumption predictions is critical in developing-country energy planning.

4.3. Model Optimization

Although the hybrid model has higher prediction accuracy than the single model, both the single and hybrid models require some network structure parameters and hyperparameters to be adjusted based on multiple analysis and calculations to adapt to the characteristics of the training data. And because such computational adjustments have a significant impact on the performance of the prediction model, optimizing the parameters to achieve model optimization is a critical component of the prediction process. Common prediction optimization models include Grid Search Algorithm (GSA), Particle Swarm Optimization (PSO), Genetic Algorithm (GA) and other optimization algorithms.

The GSA adjusts the parameters sequentially in steps within a specified range of parameters, uses the adjusted parameters to learn and train, and selects the parameter with the highest accuracy on the validation set among all parameters. This is actually a training and comparison process. The goal is to find the highest value in the array. GSA has the advantage of randomly exploring the hyperparameter space to avoid local optimal solutions, which is useful in cases where the parameter space is very large. The GSA method, despite having a simple computational principle [116], is inefficient and takes a long time when dealing with a large parameter space.

The PSO is an evolutionary computing technique that arose from the behavioral analysis of bird flock feeding [117]. Its basic principle is that each particle searches for the best solution in the parameter space and marks it as the current individual extreme value. Based on information sharing, the optimal individual extreme value is identified as the current global best solution for the entire particle population. In the face of nonlinearity and randomness in electricity peak load prediction, the PSO is used to accurately capture peak electricity load by optimizing parameters [118]. The PSO is simple, fast, and requires fewer setup parameters, but it has a limited ability to search locally.

The GAs are a collection of search algorithms inspired by natural evolution theory [119]. The GA can provide high-quality solutions to a wide range of search, optimization, and learning problems by simulating the natural selection and reproduction processes. The GA can be used in ML algorithms to optimize parameters. Li et al. [120] combined the ANN and GA models to find the optimal hyperparameter sets in hyperparameter space, completing the model optimization, and the accuracy of building electricity consumption prediction was higher than that of the ANN model.

In addition to the prediction optimization models mentioned above, some studies use multiple optimization algorithms to optimize parameters. Cui Hua et al. [121] used Bayesian and genetic algorithms to optimize hyperparameters on a multiscale grid. Jiang et al. [122] proposed a two-step hybrid global optimization algorithm with a designed Grid Traverse Algorithm (GTA) and a PSO to determine the optimal prediction parameters for the determination of the two-step hybrid global optimization algorithm for the best prediction parameters of SVR, which has high performance when compared to other optimization algorithms.

The application of these advanced technologies in the residential building electricity load prediction model can significantly improve the accuracy of predicting, optimize load scheduling, enhance the adaptability of the system, and improve the efficiency of energy management [123]. In the future, through continuous optimization and improvement, these technologies can achieve more intelligent and efficient electricity load management.

5. Selection of Driving Factors

Future residential electricity load varies with randomness and complexity, so the development of an electricity load prediction model must not only analyze the intrinsic characteristics of the load historical data, but also consider the impact of the load change on the extrinsic characteristics of the dialectical analysis. Only by understanding the law of change in electricity load can a specific space and time in the future electricity load prediction value be more reliable. There are numerous factors in the electricity system that influence the change in electricity load, and the information such as house age, number of equipment, family, employment and number of bedrooms can provide different accuracy characteristics for different algorithms [124]. Therefore, it is necessary to select different variables for different prediction targets when developing an electricity load prediction model. According to literature research, the electricity load prediction task can be divided into two categories: single-feature prediction and multi-feature prediction. Single-feature prediction uses historical load data to predict electricity load data, whereas multi-feature prediction uses factors influencing both historical load and electricity load data to predict electricity load data.

5.1. Single-Factor Input

In the long run, electricity load fluctuates with regularity over a long period of time, so it is clear that time is an important factor in determining the error between the predicted and actual electricity load values. The single-factor electricity load prediction model refers to learning useful information from electricity load data so that the model’s final prediction result is close to the true value. In the single-factor load prediction task, the electricity load data are the input factor X, and also the prediction Y target. The single-factor electricity load prediction model accepts only electricity load data as input factors. The prediction process focuses primarily on analyzing the cyclical and seasonal patterns of change in the electricity load prediction data [125]. When analyzing electricity load prediction types with different time scales, the change laws have different cycles, such as the work system cycle for short-term prediction, and special holidays and other characteristics for medium-term prediction.

5.2. Multi-Factor Input

Although long-term historical data on electricity load can be used to predict electricity consumption within a certain accuracy range, the appropriate addition of some factors that affect the change of electricity load in residential buildings can significantly reduce the error between the actual and predicted values. When the prediction time period is short, single-variable time series are likely to provide a simpler and more obvious computational process for more accurate prediction results. However, multivariate prediction accurately analyzes a sufficient number of influencing factors to predict the future state of electricity load [126]. The cycle of change law varies, particularly when analyzing the target of electricity load prediction at different time scales, such as the short-term prediction of the work system cycle, the medium-term prediction of special holidays, and other characteristics.

The majority of early researchers concentrated on determining the impact of climate change on the electricity load of residential buildings. According to one study [62], an increase in temperature by 1 °C will result in a 6.79% increase in electricity consumption. However, average rainfall has an impact on electricity consumption, but it is less than the impact of temperature increase. As a result, it is clear that temperature is one of the most important constraints on electricity load changes. As a result, it is clear that temperature is one of the most important constraints on electricity load changes. In response to the current lack of research on the correlation of electricity data forecasts with spatial and temporal resolution, Mo Chen et al. [63] calculated electricity-temperature sensitivities using hourly ambient temperature records and discovered that the strongest relationship was found using cumulative daily electricity consumption and average daily temperature and temperature. According to the study [127], higher prices reduce electricity consumption, so the government should fluctuate electricity prices to the optimum level in order to minimize electricity consumption waste. There are dynamic fluctuations in loads and electricity prices, so it is important to accurately analyze the relationship between electricity prices and loads to formulate the electricity pricing policy. Shayeghi et al. [128] proposed a multiple-input multiple-output (MIMO) model that takes into account the correlation between electricity price and load, which is used to predict both load and price signals in departments and enterprises that deal with electricity.

Changes in personnel behavior in residential buildings have the potential to save significant amounts of energy, and there is a self-reinforcing feedback relationship between attitudes and behaviors [129]. Yohanis et al. [130] investigated 27 representative residential electricity consumption patterns and discovered that building type, geographical location, appliance ownership and scale, and occupant attributes (including number of occupants, income, and residence pattern) all have distinct but significant effects on electricity consumption. Furthermore, based on various application scenarios, historical load data should be combined with appropriate input factors to accurately predict future load change trends. Heating equipment usage [131] and building attributes [132] have a significant impact on residential electricity load during the winter heating season in northern regions. The difference in air-conditioning electricity consumption during the summer cooling period is largely due to differences in economic income and psychological awareness [133], which is reflected in the fact that households living in public rental housing or with lower household income tend to be more frugal in their use of air-conditioning and other refrigeration equipment. The majority of studies focus on differences in household information, living habits, and climatic conditions that affect the variation of electricity loads in residential buildings. The impact of different equipment usage patterns in residential buildings is also significant, particularly in distinguishing between the characteristics of electricity load variations of different users that have a positive effect.

As a result, based on the analysis and summary of the three scales of input factors, prediction methods and selection of the output types of residential building electricity load prediction, this paper can summarize the above factors affecting the change of electricity load, as shown in Figure 7, and divide the various input factors into the following aspects: environmental factors, personnel factors, behavioral factors, residential factors, and energy factors.

These characteristics are relatively different in different geographical areas or regions, and some of them are significantly different from other regions, such as the ownership of heating terminal equipment in the north and south, as well as significantly different climatic parameters in cold regions, hot summers, and warm winters. These characteristics of significant differences in the corresponding residential building electricity load changes in the trend have a large difference. In the same moment or time period, similar changes in the trend may appear, but the amount of change is different, and there may be a direct opposite trend of change. As a result, depending on the spatial and temporal characteristics of the target of electricity load prediction for residential buildings, different characteristics of these five factors can be chosen to reduce the error between the prediction and error values. It is also possible to use some of the survey’s significant differences to collect data and factors influencing the electricity load of residential buildings in different regions or geographic areas, as well as to develop a framework for predicting the electricity load of residential buildings at the regional level.

6. Conclusions

Prediction is a random and targeted process that is difficult to achieve with accuracy. This is because the process of electricity load prediction is to analyze and calculate the laws between the historical data of electricity load and the driving factors from monitoring or investigation, and then to predict the electricity load in a future period by combining some real values (household characteristics, building attributes, energy use behavior, etc.) and non-occurring variables (climate data, economic data, etc.) through this law. Some of the data used in this process are random, diverse and volatile. In other words, the prediction topic can be described as a random-to-random problem.

In the analysis of residential building electricity load prediction, the characteristics of electricity load in different households may be very different. A study conducted a research and analysis on the related driving factors of social economy, construction and equipment of residential construction groups, and found that 20 of the 62 factors related to the electricity load had a significant positive impact, while the rest of the factors had a relatively low impact [134]. Therefore, the purpose of this paper was to summarize the power factors, family factors, behavior factors, residential factors and energy factors, so as to be able to apply to the analysis of residential electricity load changes in different time and space scales to obtain more accurate results. On the whole, there are commonalities in the variation rules of residential electricity load at different scales, but there are significant qualitative and quantitative differences in the relative details of driving factors for improving the prediction accuracy. From the power supply perspective, effective electricity supply and distribution capacity planning can only be achieved through a targeted analysis of the temporal and spatial variations in the electricity load of residential buildings. This analysis should consider the combined effects of time, geography and region. For electricity load managers and users, strategies such as balanced scheduling, flexible scheduling and distributed energy scheduling should be selected based on the characteristics of the time scale, geographic scale and regional scale of residential building electricity loads. Therefore, only by accurately predicting the electricity load of residential buildings can power grid companies reasonably formulate power supply scheduling, meet the needs of residents, and improve the reliability and economy of the system.