Machine Learning- and Artificial Intelligence-Derived Prediction for Home Smart Energy Systems with PV Installation and Battery Energy Storage

Abstract

Overview: Photovoltaic (PV) systems are widely used in residential applications in Poland and Europe due to increasing environmental concerns and fossil fuel energy prices. Energy management strategies for residential systems (1.2 million prosumer PV installations in Poland) play an important role in reducing energy bills and maximizing profits. Problem: This article aims to check how predictable the operation of a household PV system is in the short term—such predictions are usually made 24 h in advance. Methods: We made a comparative study of different energy management strategies based on a real household profile (selected energy storage installation) based on both traditional methods and various artificial intelligence (AI) tools, which is a new approach, so far rarely used and underutilized, and may inspire further research, including those based on the paradigm of Industry 4.0 and, increasingly, Industry 5.0. Results: This paper discusses the results for different operational scenarios, considering two prosumer billing systems in Poland (net metering and net billing). Conclusions: Insights into future research directions and their limitations due to legal status, etc., are presented. The novelty and contribution lies in the demonstration that, in the case of domestic PV grids, even simple AI solutions can prove effective in inference and forecasting to support energy flow management and make it more predictable and efficient.

1. Introduction

Photovoltaic (PV) energy is one of the most important and widely available renewable energies, and with the energy crisis and the need to protect the environment, investment in it by states and companies is increasing every year, especially in the area of artificial intelligence (AI) applications in PV systems [1,2,3]. PV systems are widely used in residential applications in Poland and Europe due to growing environmental concerns and the price of fossil fuels. Energy management strategies for home systems in Poland play an important role in lowering energy bills and maximizing profits.

Uncertainty in PV power generation refers to the variability and unpredictability of the amount of electricity generated by a PV system over time. Several factors contribute to this uncertainty:

• Weather conditions: The generation of solar energy is highly dependent on weather conditions, primarily the intensity and timing of sunshine. Thus, cloud cover, atmospheric haze and other weather phenomena can cause fluctuations in solar irradiance, leading to variations in output;
• Day–night cycle: PV systems only generate electricity during daylight hours—the length of the day varies throughout the year depending on location and season, which affects the total energy generated;
• Seasonal variability: Solar energy production changes with the changing seasons; moreover, in some regions, the availability of sunlight can be significantly reduced during cloudy or rainy seasons;
• Latitude and location: The amount of sunlight received by a PV system depends on its geographical location. Areas closer to the equator tend to receive more consistent and intense sunlight, leading to a more predictable output;
• Shading and obstacles: Shading caused by nearby buildings, trees or other obstacles can create “hot spots” on PV panels, leading to a reduction in output. Shading can vary throughout the day and year, introducing uncertainty;
• Dust and dirt: The accumulation of dust, dirt or other debris on PV panels can reduce their efficiency and overall output.

Various strategies can be used to manage and reduce the uncertainty associated with PV generation:

• Forecasting: Weather and solar intensity forecasting models can provide short- and medium-term forecasts of solar availability, helping system operators to anticipate and plan for fluctuations;
• Energy storage: Integrating energy storage systems (e.g., batteries) with PV installations can help store excess energy during periods of high production and discharge it during periods of low production, thus smoothing output;
• Diversification: Combining solar power with other renewable energy sources, such as wind or hydropower, can help to balance fluctuations in generation from different sources;
• Smart grids: Smart grid technologies enable real-time monitoring and control of electricity generation and consumption, allowing variable energy sources such as solar power to be managed more efficiently;
• Advanced inverters: Modern inverters have features that enable them to manage fluctuations in power generation more efficiently, such as maximum power point tracking (MPPT) algorithms;
• Maintenance and cleaning: Regular maintenance and cleaning of PV panels can help maximize their efficiency and reduce performance degradation caused by dirt or dust accumulation.

Uncertainty in PV power generation is inherent due to the variability of natural conditions, but advances in technology and energy management strategies have enabled more efficient use of solar energy in the overall energy mix.

Prosumers account for almost 80% of the installed PV capacity in Poland. At the end of last year, the number of prosumers exceeded 1.2 million, and the total capacity of their installations amounted to more than 9.3 GW. This means that the average power of a micro-installation is 7.6 kW. The power forecast assumes that the cumulative PV capacity in the country will stand at 20 GW in 2025. By 2030, this figure should reach 29 GW, including 15 GW of power in prosumer sources, including 4 GW of PV systems owned by businesses [4,5].

From a technical point of view, a PV microgrid offers better energy efficiency, higher power quality and greater service reliability by integrating multiple micro-sources, regulating the impact of distributed generators and improving compatibility with the power grid. From the point of view of users, the microgrid allows for improvement in the quality and continuity of services, at low voltages and in grid and off-grid mode (islands). To date, a number of solutions have been used to improve the system’s performance by reducing the maximum power points (MPPs) of PV systems, such as the quasi-Z-source tripod inverter (QZSI) and the improved control strategy developed for it [6], and in selecting data-driven algorithms for predicting the behavior of lithium-ion batteries (LIBs), their charging and discharging process, data processing and functioning [7].

Energy management strategies for residential systems are constantly being refined, and the optimal disposition of microgrids has been and continues to be a hotspot of research and practice. It uses a wide variety of approaches, including

• An optimizer strategy for mud-ring feeding of bottlenose dolphins to solve a dynamic economic emissions problem;
• Load following dispatch strategy optimized in Bangladesh conditions;
• Moth–flame optimization algorithm based on position disturbance updating strategy (MFO_PDU) [8,9,10,11].

Despite many attempts, the problem of a low-cost, simple, adaptable, scalable solution to spread the concept of PV-based renewable energy in most developed and developing countries has not been solved. Bans and subsidies alone will not lift the restrictions—the use of PV energy simply needs to become more cost-effective than existing solutions, including on a mass scale. Advanced AI-based solutions can and should help with this.

AI has proven its effectiveness in data analysis and management of processes of natural origin (biological, medical) [12,13,14] as well as industrial [15,16,17]. The first two papers [12,13] concern the application of AI-based filtering of electroencephalographic signals for control, while the next ones concern the AI-based analysis of the gait of stroke patients [14] and industrial applications of AI: optimization of 3D printing and water supply as part of Industry 4.0 [15,16,17].

The rationale for AI-based research and tools in PV grids is the need to improve the efficiency, reliability and overall performance of solar energy systems. Machine learning (ML) offers the ability to analyze complex data patterns, optimize system performance, predict energy production and consumption and improve maintenance practices.

In the area of PV systems management, AI methods and techniques often perform better than previously used traditional approaches, and the high involvement of scientists and engineers has encouraged the finding of new solutions or improvement of previous ones, so that AI algorithms can improve the efficiency and accuracy of the solutions used. This is not altered by the limitations of AI, including the significant amounts of data and computational time required to learn solutions based on a data-driven approach (ML) [1]. ML is indeed a subset of AI. Nowadays, AI is increasingly being used to analyze, infer and predict PV system performance, including prediction of the amount of energy produced (point, interval, probabilistic), detection of anomalies and faults in PV modules/cells, monitoring the maximum power point of PV modules, extraction, modelling and optimization of PV module design parameters, or the estimation of PV module surrogate model parameters, and also the forecasting of energy demands, generation costs and associated prices, also for PV/Thermal (PV/T) and Concentrating PV (CPV) systems [1,2,3]. This applies to both PV generation and PV energy storage and conversion systems.

New AI-based features of the PV grid include two-way digital communication (including reporting, alerting, command confirmation), self-monitoring and self-healing (if necessary), adaptation and switching of operating modes, intuitive control, mechanisms of protection and security, power quality, voltage and frequency control and maintaining system stability (Figure 1). At all levels (local, network), data-driven (ML) models play a key role, mainly ANN, because it allows the logic to be separated from the physical structure (i.e., there is no need for such accurate physical models). It seems that the wider use of AI will simplify communications and reduce the complexity of the model. This will be supported by the increase in the autonomy of the PV grid. AI in PV grids is used in four main areas:

• Smart management (including PV grid autonomy): AI can be used to optimize the operation and control of PV grids. ML algorithms can analyze real-time data from PV panels, weather forecasts, energy demands and other relevant factors to make intelligent decisions about when and how electricity is generated, stored and distributed. This allows the grid to dynamically respond to changing conditions, increase energy efficiency and effectively balance supply and demand;
• Reliability and resilience: AI can improve the reliability and resilience of PV grids by predicting and preventing faults or failures. Predictive maintenance algorithms can analyze sensor and device data to detect anomalies and potential problems before they lead to system downtime. Artificial intelligence-based fault detection and diagnosis can contribute to faster response times and reduced downtime, ensuring a more reliable power supply;
• The human-centric approach to designing AI solutions involves developing AI solutions that are user-friendly, easy to understand and compatible with human needs and behavior. In the context of PV grids, this could include creating user interfaces that allow grid operators or consumers to interact intuitively with AI systems. In addition, AI can help with demand forecasting, helping grid operators make informed decisions based on human behavior and energy consumption patterns;
• Privacy and data security: AI in PV grids must ensure the privacy and security of sensitive data. AI solutions that handle data from PV panels, energy consumption and other sources must comply with data protection regulations and implement robust security measures. Encryption, secure data transmission and access control mechanisms are key to protecting data privacy.

Integrating AI into PV networks can lead to more efficient and resilient energy systems, better decision making and an improved user experience. However, it is important to address potential challenges such as algorithm bias, data quality and the need for continuous monitoring and maintenance of AI systems to ensure their effectiveness and reliability.

Accurate forecasting remains a key task for the effective integration of PV output power into the grid, and is necessary to be taken into account when planning, managing and optimizing microgrids (MG), integrating them with smart buildings or for the efficient use of electric vehicles (including their batteries as grid elements). This is especially difficult in multi-stage forecasting and must take into account a number of factors, including weather [2,3]. Traditional numerical and probabilistic methods or physical models have also been replaced here by AI techniques, in particular by a subset of AI techniques called ML, deep learning (DL) and hybrid methods [2,3]. The ML system has the ability to learn from the data to model the system and, during deployment and normal operation, it learns from further real data, gaining the ability to adapt the operation to new situations and better fit the specifics of the system [2].

ML methods and data-driven control (DDC) techniques can support monitoring, control, optimization and fault detection in power generation systems, including assessing, countering or withstanding the effects of associated uncertainties. These benefits are mainly seen in the areas of visibility, maneuverability, flexibility, profitability and safety (the so-called “5-TYs”) [18].

The originality of AI lies in its ability to unlock insights into the complex data of the solar PV network, while its importance lies in optimizing energy production, management and maintenance. Performance metrics are essential to objectively assess the effectiveness of AI solutions in achieving these goals and contributing to the development of solar energy technologies.

Global annual potential savings from applying AI to PV systems were USD 7–8 billion in 2021 and are expected to reach USD 15 billion by 2024. With the rapid development of digital technologies such as 5G and the cloud, more than 90% of PV plants are expected to be fully digitized by 2025, making PV plants “smart” and efficient. The widespread integration of AI and PV will facilitate cross-discovery and interconnection of PV plant equipment and improve the efficiency of power generation and O&M through joint optimization. AI techniques can offer promising new capabilities for PV systems, including proactive identification and protection of PV modules, and a reduction in equipment failures using AI diagnostic algorithms. It will also be possible to optimize the demand-supply tracking algorithm for micro-installations (smart homes) and large PV farms operating in domestic and global energy markets (smart grids) [1,2,3].

This article aims to check how predictable the operation of a household PV system is in the short term—such predictions are usually made 24 h in advance.

Its novelty and contribution lies in demonstrating that, for domestic PV grids, even simple AI solutions can prove effective in inference and prediction to support energy flow management and make it more predictable and efficient. Accessible, effective AI solutions do not have to be complex or expensive, facilitating transparency for users. This fills a research gap—the field of AI and renewable energy is developing rapidly; hence, it is necessary not only to make advances in the analysis of domestic solar PV grids, but also to disseminate good practices and AI-based solutions cheaply and quickly. There are still many potential research gaps in the field of simple AI analysis of domestic solar PV networks—some of which are as follows:

• Real-time data integration: While efforts have been made to analyze national PV grids using artificial intelligence, there may still be a gap in effectively integrating real-time data from a variety of sources, such as weather forecasts, energy consumption patterns and PV system performance. The development of AI algorithms that can seamlessly incorporate these data streams could lead to more accurate and flexible grid analysis;
• User-friendly decision support: Most AI analysis for domestic PV grids is for experts or researchers. There may be a research gap in creating user-friendly interfaces or tools that allow homeowners and small businesses to easily interpret AI-based insights and make informed decisions about optimizing their PV systems;
• Predictive maintenance and fault detection: AI can predict maintenance needs and identify faults in PV systems. However, there may be a research gap in developing algorithms that can accurately predict specific maintenance requirements or diagnose system faults in a simplified way, helping non-experts to take appropriate action;
• Localized energy storage optimization: AI can optimize the use of PV energy storage, but there may be a research gap in developing algorithms that take into account local energy policies, incentives and market conditions. Such algorithms could help homeowners decide when to store excess energy, when to sell it back to the grid or when to draw from the grid;
• Load management and demand response: AI analysis often focuses on PV system performance and optimization. However, a research gap may exist in incorporating load management and demand response strategies. This could include AI algorithms that dynamically adjust energy consumption patterns based on grid conditions, time-of-day tariffs and user preferences;
• Integration with smart home systems: Many homes are becoming “smart” through the integration of different devices and systems. There may be a research gap in exploring how AI analysis of home solar PV networks can be seamlessly integrated into broader smart home management platforms, providing a holistic approach to optimizing energy consumption;
• Uncertainty and robustness: AI models for PV grid analysis may not always account for uncertainties such as fluctuations in weather forecasts or changes in user behavior. Addressing this research gap may involve developing AI approaches that are robust to uncertainty and can provide reliable recommendations even in less predictable scenarios;
• Behavioral aspects and user adoption: it is important to understand how users interact with and adopt AI-based recommendations for their PV networks. Research can focus on exploring user behavior, motivations and barriers to adopting AI suggestions, with the aim of improving the overall effectiveness of AI-based network analytics;
• Privacy and security: As AI analysis involves the collection and processing of data from national PV networks, there may be a research gap in exploring privacy techniques and ensuring the security of data and AI models to protect user information;
• Long-term performance evaluation: Many AI analyses focus on short- to medium-term optimization. However, there may be a research gap in assessing the long-term performance of NGAs under AI-guided strategies, taking into account factors such as system degradation, technological advances and changing energy landscapes.

This article’s introduction reviewed the state of the art of home energy management, including commonly used objectives, constraints and solutions for PV and energy storage applications. This is followed by a comparative study of different energy management strategies based on a real household profile (selected energy storage installation) based on both traditional methods and various AI tools, which is a new approach, so far rarely used and underutilized, and may inspire further research, including those based on the Industry 4.0 paradigm and, increasingly, Industry 5.0. This article discusses the results for different operational scenarios, considering two prosumer billing systems in Poland (net metering and net billing). Finally, conclusions and insights on future research directions and their limitations due to the legal status, etc., are also presented.

2. Material and Methods

2.1. Dataset

A dataset from the Polish home PV network (Figure 2), downloaded hourly for 30 consecutive days (720 records) in April/May 2023, was analyzed. The data were read from the system, stored and analyzed with full accuracy, and were rounded only for the purposes of presentation.

Fourteen following data were analyzed as inputs of ANN (

Table 1. Presentation of dataset (selected parameters).

Parameter	Mean	SD	Min	Q1	Median	Q3	Max
Current air temperature (°C)	15.42	5.45	7	11	14	17	22
Current wind speed (m/s)	4.49	1.03	0	2.25	4.50	5.24	6.32
Forecasted air temperature (°C)	16.38	6.12	7	12	16	18	25
Forecasted wind speed (m/s)	5.00	1.11	0	2.11	5.00	7.33	11.11
Power (Watt)	17.55	4.46	12.50	14.78	16.45	19.11	22.71
Total CO2 reduction (tons)	10.28

, Figure 3):

• Current weather (air temperature (°C), wind speed (m/s), cloudiness (five classes represented by numbers));
• Weather forecast (air temperature (°C), wind speed (m/s), cloudiness (five classes, e.g., cloudy represented by numbers));
• Power (Watts);
• Current power direction (four classes represented by numbers);
• PV use (Watts);
• Power network use (Watts);
• Energy storage status (%);
• Energy consumption by receivers (Watts);
• Standardized power (%);
• Energy balance (Watt-hour);
• Income (PLN) (daily, monthly, yearly, total);
• Total CO₂ reduction (tons), including equivalent number of trees or car use (km).

Outputs were:

• Predicted (24 h after) energy balance (Watt-hour);
• Predicted (24 h after) total CO₂ reduction (tons).

The third result should be the daily income (PLN), for the whole data set, this value was always 0 in the tested system.

The choice of input signals was based on their availability (ability to be simply downloaded from the PV system interface) and their ability to be used directly in the AI system by simply converting them to numbers. Normalization and scaling of the input and output signals resulted in an equal interpretation of the values of all signals processed by the ANN/CNN.

Data were collected in an .xls spreadsheet (Microsoft Excel, Microsoft, Redmond, WA, USA). In total, 720 data records were obtained. Data were prepared for AI processing by being checked for gaps, errors and outliers and then normalized. The data were randomly divided into a teaching set (70%, i.e., 504 samples/records) and a testing set (30%, i.e., 216 samples/records).

2.2. Methods

In this paper, four approaches to data analysis and model building were tested, including two traditional, simpler ones (linear regression and non-linear polynomial regression) and two based on artificial neural networks (hand-written and automated). The first two methods were chosen for their simplicity and easy and widespread use.

The main differences between the “hand-written dedicated AI-based solution” and the “semi-automated solution based on pre-prepared scripts” lie in the level of customization, accuracy and effort required during development and maintenance. The differences between these two concepts are described below:

• Hand-Written Dedicated AI-Based Solution:
  • Involves creating custom AI solutions from scratch, tailored to solve a specific problem or task;
  • Developers and data analysts build an AI model, often using ML or DL techniques, and train it on relevant data;
  • The solution is designed to handle complex tasks and adapt to different scenarios;
  • Requires an in-depth understanding of AI algorithms, data pre-processing, feature engineering and model training;
  • Development process can be time and resource intensive;
  • The resulting solution is usually more accurate and flexible as it is optimized for the specific problem;
  • Maintenance and updates may require ongoing efforts to ensure the model remains effective as conditions change;
• Semi-Automated Solution Based on Pre-Prepared Scripts:
  • It involves using pre-existing scripts or code templates to develop an AI solution;
  • The scripts often contain predefined algorithms, configurations and processing steps;
  • Developers may need to adapt and customize the scripts to suit a specific use case;
  • Quicker to implement compared to building a solution from scratch;
  • It may be suitable for simpler tasks that do not require highly specialized models;
  • The level of customization and flexibility may be limited by the available scripts;
  • Updates and modifications may be simpler, but may still require an understanding of the underlying code;
  • May be useful for rapid prototyping or when time and resources are limited.

The main features are as follows:

• Customization: A hand-written, dedicated AI-based solution offers a higher level of customization to specific requirements, whereas a semi-automated solution relies on existing scripts, potentially limiting customization options;
• Accuracy and complexity: Manually written solutions tend to be more accurate and can handle complex tasks due to their customized nature. Semi-automated solutions may be less accurate and suitable for simpler tasks;
• Development effort: Hand-written solutions require more time and expertise during development, while semi-automated solutions can be implemented more quickly, but may not have the same degree of optimization;
• Maintenance and upgrades: Both solutions require maintenance, but manually written solutions may require more ongoing efforts to adapt to changing conditions, while updates for semi-automated solutions may be simpler to implement;
• The choice between these approaches depends on factors such as the complexity of the task, available resources, time constraints and the desired level of customization and accuracy.

The linear regression model is based on the assumption that there is a linear relationship between the dependent (output) variables and a vector of independent (input) variables. This relationship is modelled by including a random (error) component. The classical method of least squares is most commonly used for this purpose. This method is the oldest and easiest to apply, although it has drawbacks, including little robustness to outliers. The assumption of the model used here was that it is possible to create such rules described by a system of equations (or an equivalent matrix), which will allow, on the basis of the parameters present, the prediction of predicted values in 24 h.

In the case of polynomial regression, the higher the degree of the polynomial, the better the fit of the model to sample non-linearity, but the higher the computational complexity. In the case studied, it was assumed that the correlation coefficient should be at least 0.8 to achieve a good representation of the function. It is worth noting, however, that in polynomial regression, when selecting the parameters of the approximating function, it should be estimated what function is to be expected, and this is very difficult for the current case. The similarity of the individual variables can only be estimated after 24 h based on the autocorrelation values.

In this study, a data-driven approach was also used, i.e., ML. In this study, two different approaches were compared for solving the same problem: a hand-written dedicated AI-based solution (Figure 3, MATLAB R2023a software with Neural Networks and Deep Learning toolboxes, MathWorks, Tulsa, OK, USA) and a semi-automated solution based on pre-prepared scripts (ML.NET in Visual Studio 2022, Microsoft, Redmond, MA, USA). The primary criteria for evaluating the effectiveness of the solutions were the RMSE value and accuracy (separately for teaching and testing).

Various neural networks, both traditional and deep, were tested in this study, but, with the research problem formulated in this way and the number and characteristics of the data (mainly numerical), the best results with the simplest structure were given by a three-layer neural network (multilayer perceptron (MLP)) with a back-propagation (BP) algorithm, a minimum mean square error (RMSE) optimization and a naive initialization technique. BP is widely used due to its ease of implementation, fast convergence, efficiency and lack of need for prior knowledge. MLP operation is described by the following equation:

y = f(x,Ѳ)

(1)

where:

• f—activation function,
• x—network inputs,
• y—network outputs,
• Ѳ—set of parameters mapping input x to an output class y.

The network with three layers (1–3) is described using three functions:

f(x) = f₃(f₂(f₁(x)))

(2)

• f(x)—output of network,
• f₁(x)—output of 1st layer,
• f₂(x)—output of 2nd layer,
• f₃(x)—output of 3rd layer.

Each layer is described as a function of

y = f(Wx + b)

(3)

where:

• y—output,
• W—weights,
• x—inputs,
• b—bias.

Many models were tested. The best solutions were selected for presentation: a traditional three-layer MSE (implemented traditionally) and a four-layer convolutional network with two hidden layers (implemented in ML.NET).

The use of a sigmoidal (non-linear) transition function

$$y=\frac{1}{1-e^{-x}}$$

(4)

in the networks is required to make an appropriate choice of bias to ensure a more accurate fit to the data (through a more accurate predictive function in the output). The selection of a steeper transition function resulted in a better transfer of differences with more categories.

The number of neurons in the input layer depends on the amount of input data, and the number of neurons in the output layer on the amount of output data. The number of neurons in the hidden layer was selected experimentally on the basis of the authors’ experience and the method of successive approximations to the global optimum (samples were concentrated in subsequent search rounds). This approach requires learning about 100–150 neural networks on average, but with automation and the use of dedicated software, it can be the solution that gives the best results. The ML.NET (Visual Studio 2022, Microsoft, Redmond, MA, USA) solution, pre-configured, was faster, but less accurate and with a higher RMSE.

Accuracy is traditionally calculated as the fraction of the predictions for which the model is correct and is based on calculations based on a set of normalized training data. In this case, an accuracy of 0.87, or 87%, means 87 correct predictions for every 100 test examples.

The estimated learning time (central processing unit (CPU) local, AMD Ryzen 5 5600U 64-bit processor, Radeon Graphics 2.30 GHz card and 16 GB RAM) did not exceed 10 min in either case. This confirms that this type of data can be analyzed locally or in the cloud.

The advantage of the ML.NET environment is the simultaneous checking of many algorithms (up to 50) and the identification of the best one, and the relatively short setup (for a person who has the environment and design prepared in advance, e.g., class catalogues with individual data in the case of classification) takes about 20 min. In the case of the traditional approach, it takes much longer and must be performed separately for each network configuration, but it allows solutions tailored to the form of data to be chosen. This is because some of the knowledge comes from the designer’s experience rather than being coded into the system, making manual network tuning unique and difficult to replicate in an automated system. The automatic approach is faster, but it is not known whether the speed advantage is enough to achieve a better model performance. In this case, a dedicated AI solution simply allows for better personalization of the offer, e.g., towards cost reduction or increasing the efficiency of the PV grid at certain times of the day. On the other hand, the AI algorithm allows you to capture more dependencies within the processes and user behavior that are not obvious and noticeable by humans, and then include them in the control algorithm. Therefore, the choice of a specific solution may depend on the preferences of users.

Results were exported to .xlsx and .csv files, and in the case of ML.NET, C# code was also available in the form of two projects (ML.ConsoleApp and ML.Model in MS Visual Studio 2022 (Microsoft, Redmond, WA, USA)) for further processing.

3. Results

The term “hyperparameters” refers to settings that are not learned during training but determine how the model learns. The tuning of hyperparameters is a key step in the development of an ML model, as it involves choosing the best configuration for the different parameters that affect the learning process. Balancing hyperparameter tuning and computational cost is a trade-off that depends on available resources, project goals and time constraints: it is important to find a balance between accurate exploration of the hyperparameter space and efficient use of computational resources.

Tuning hyperparameters and selecting input features are iterative processes and do not have a one-size-fits-all solution. Correct implementation requires experimentation and fine-tuning to achieve the best results for a particular problem (

Table 2. Selected hyperparameters taken into account during MLP model tuning.

Hyperparameter	Description
Number of hidden layers	Experimenting with different architectures to find the one that works best.
Activation function(s)	Select the appropriate activation functions for each layer (from basic, e.g., rectified linear unit (ReLU), sigmoid, tanh or advanced, e.g., Leaky ReLU or scaled exponential linear unit (SELU)) in a way that maximizes the model’s performance.
Weight initialization	Experiments with different weight initialization methods (random initialization, Xavier, He et al.), as the initial values of the weights can affect network learning.
Learning rate value	Finding the right learning rate (step size during optimization) to ensure convergence without overshooting or getting stuck in local minima.
Learning rate schedule	Adaptation of the learning rate of the network using schedules (step fading, exponential fading, cosine annealing, etc.).
Batch size	Finding the right batch size is important as the batch size controls the number of training samples used in each update of the model weights, with too small batches likely to result in more noise, while too large batches can lead to slower convergence.
Number of epochs	Deciding at what number of epochs to stop training to prevent over-fitting.
Regularization techniques	Tuning the strength of the regularization is essential as techniques such as dropout, L1/L2 regularization prevent over-tuning.
Optimization	Experimenting with different optimizers (stochastic gradient descent (SGD), Adam, root mean square propagation (RMSprop), etc.) to find the one that best suits the problem at hand, as they affect training speed and network convergence differently.

).

The selection of appropriate input functions or variables during tuning procedure is crucial for successful MLP learning (

Table 3. Input function or variables taken into account during MLP model tuning procedure.

Stage of the Procedure	Description
Feature analysis	Analysis of the dataset, understanding the nature of the features and their impact on the final result, correlation analysis, assessing the importance of features that may be relevant to the problem.
Feature engineering (optional)	Creating new features or transforming existing features in a way that improves the model’s ability to capture patterns in the data.
Feature selection	Feature selection to reduce dimensionality through forward selection, backward elimination or recursive feature elimination.
Cross-validation	Evaluate the performance of different feature sets and hyperparameter configurations, helping to prevent over-fitting and providing a more accurate estimate of model performance.
Iterative tuning	Continuous iteration through the stages of feature selection and hyperparameter tuning with evaluation of different combinations until the best performing model is found on the validation data.

).

The hyperparameters for our MLP models selected by the procedure described in Table 3 included the following:

• Number of hidden layers and units (neurons) in each layer (

  Table 4. The best MLP network models.

  ANN Structure	Activation Function in the Hidden Layer	Activation Function in the Output Layer	Accuracy (Learning) [%]	Accuracy (Testing) [%]	RMSE
  MLP 14-22-2	Sigmoid	Sigmoid	86.11	87.32	0.05
  MLP 14-24-2	Sigmoid	Sigmoid	87.92	88.07	0.02
  MLP 14-26-2	Sigmoid	Sigmoid	88.43	88.92	0.01
  MLP 14-28-2	Sigmoid	Sigmoid	87.56	87.16	0.02
  MLP 14-30-2	Sigmoid	Sigmoid	86.99	86.53	0.03

  );
• Activation functions: ReLU, sigmoid, tanh;
• Learning rate;
• Number of training examples used in each iteration of the gradient descent;
• Epochs: The number of iterations of the entire training data set during the learning of the network;
• L2 or L1 regularization to prevent over-fitting;
• Dropout Rate: percentage of neurons randomly dropped during learning to prevent over-fitting;
• Optimization algorithm used to update model weights: Adam, SGD, RMSProp;
• Initialization of network weights (Xavier, He, etc.);
• Schedule of learning rate changes over time for faster convergence;
• Architecture modifications: addition of skipped links.

Computational costs proved to be an important factor when tuning the hyperparameters of the model. To find the best combination of hyperparameters that led to optimal model performance, we found that Bayesian optimization was more efficient here than grid or random search and faster than more advanced methods such as genetic algorithms.

Computational modelling, including that based on artificial neural networks, is useful and cost-effective (including considering their computational cost) in those cases where traditional statistical methods fail or are insufficient. For the purpose of finding a single good solution, we created about 150 models based on ANNs, while only the results for the best ones are presented in this paper.

The results obtained using linear regression were not statistically significant, i.e., for all selected options the software showed a message that for such a dataset it was not possible to perform the full process, i.e., to create and analyze the regression equations. This means that no linear approximation was found for the hypothesis that a forecast can be made one day in advance based on the input data set. Linear multivariate regression models should not be the preferred forecasting method due to the lack of a normal distribution of the data. It follows that the relationship is non-linear, so perhaps a non-linear model is the solution to the problem posed in this way.

In the case of polynomial regression, analyses were performed using both custom settings and the MATLAB software’s built-in Spline Smoothing, Moving Average or Weighted Average functions that automate this process. Unfortunately, in this case, the highest correlation value obtained was 0.657, while the expected correlation value was no less than 0.8.

As we have already mentioned, each ANN layer contained neurons with the same activation function (sigmoidal), primarily due to their high flexibility (Table 4).

AR results after using other activation functions were significantly worse. When evaluating and comparing neural networks, we relied on the RMSE, Accuracy (learning) and Accuracy (training) values. The learning process of the selected networks involved repeating the learning patterns and modifying the network weights accordingly until the target RMSE was reached after a maximum of 1000 epochs (Figure 4).

The best accuracy results in this study were obtained for the MLP structure: 14 neurons in the input layer, 26 neurons in the hidden layer and 2 neurons in the output layer (i.e., MLP 14-26-2), with the lowest RMSE for the data from the training set of 0.01 (Table 4). Competing convolutional network CNN 14-28-28-2 achieved a lower maximal accuracy for the same data set (accuracy (learning) 85.31%, accuracy (testing) 86.41%) with similar RMSE values (0.01) (

Table 5. A comparison of the selected different models used to solve this problem (own studies, we did not use hybrid approaches, combining different methods and techniques).

Model	Accuracy (Learning) [%]	Accuracy (Testing) [%]	RMSE
MLP 14-26-2	88.43	88.92	0.01
MLP 14-24-2	87.92	88.07	0.02
CNN 14-28-28-2	85.31	86.41	0.01
Support Vector Regression (SVR) *	85.19	87.89	0.02
eXtreme Gradient Boosting (XGBoost)	69.32	70.81	0.02
Polynomial regression	to low highest correlation value obtained (Spearmann’s Rho 0.657)

* penalty factor, insensitive loss function, and the kernel function parameter selected experimentally based on the literature [19].

).

MLP-based optimization was more accurate. As a data-driven approach (ML), it provided a simple and quick solution to the problem on the basis of properly selected input and output data without full knowledge of the rules and mechanisms. This enables quick adaptation of the solution to a different set of training data (including a more extensive one), or even a different PV system. It should be noted, however, that a change in the amount of input or output data will entail the need to change the network structure, and the continuous use of the network in a PV installation requires training the network on current data, so it is not entirely maintenance-free. The inclusion of new data from new devices (i.e., with different characteristics) requires the initial testing of the network to check the accuracy of the prediction.

Cross-validation is a statistical method used to estimate the skill of ML models. It is commonly used in ML to compare and select a model for a given prediction problem: it is easy to understand, easy to implement, and results in skill estimates that generally have less variance than other methods. For cross-validation, scikit-learn was used (Figure 5).

The concept presented is in line with the recently expressed view that highly complex, labor-intensive and costly AI solutions are not the only, mainstream or safest direction for AI development. Simple, usable AI algorithms, currently implemented and tuned by hand, which can also be developed in the future as low-code or no-code AI solutions, are pointed out as an alternative. This approach provides a way to control commonly used AI solutions, especially those used en masse, including on mobile devices.

The need for large-scale energy storage at a grid level is often emphasized. According to the authors, appropriate control of home energy storage, heat pumps and residential electric car chargers can also very effectively improve energy balancing in the grid (this is illustrated by Figure 6).

4. Discussion

A review of the literature (original and review articles and meta-analyses) conducted in three leading databases (Web of Science, Scopus, DBLP) based on specified keywords (AI, ML, PV and similar in English) showed that AI-based improvements in PV covered three basic developmental periods of computational support so far:

• Using artificial neural networks (1950s–1970s);
• Via data-driven approaches (ML, 1980s–2010s);
• By means of DL (present).

From a technological point of view, it is important to use systems that are effective but as simple as possible in everyday practice, avoiding the creation of large AI systems whose operation is not fully transparent to users. Simpler and handier mobile applications are becoming increasingly popular. Essential AI technologies used to increase the competitiveness and safety of PV include Expert Systems (ES), Fuzzy Logic (FL), traditional ANNs, DL and genetic algorithms (GA). There is a lack of standards in this area [21,22].

Regression, i.e., the reduction of an issue of interdependence of random variables to a functional dependence, does not always work in the case of PV network analysis. The complexity of the dependence of the variables means that, increasingly, only complex multivariate non-linear modelling will be able to cope with this task.

The results obtained and their interpretation indicate that, from the perspective of previous research and working hypotheses, the use of AI in PV systems is a technology expected by designers, manufacturers and users (both private and institutional), and the current trends are in line with the Industry 4.0 paradigm, as well as Industry 5.0, in which human and natural needs are placed at the center of the production process, also emphasizing the sustainability of the supply chain [23,24,25]. However, the concept of Industry 5.0 is still vaguely defined in the specialist literature and there is still a lot of debate in the area of applications of the Industry 4.0 paradigm. Despite this, this is how the EU, governments, companies and other stakeholders want to improve employee wellbeing, reduce work-related illnesses and raise awareness of the importance of motor fitness, fatigue, strain and effort. To date, such approaches have typically been developed in laboratories and only sometimes translated into field applications [23,24,25,26], which is an important aspect of the research to date. Effective AI-based modelling and optimization, appropriate task allocation, ergonomics and safety, based on data from motion and biological sensors, interfaces, Human Digital Twins and other simulation and visualization tools are becoming crucial [23,24,25,26]. The key is harmonistic and synergistic cooperation of people and machines. The technological competitiveness of Industry 4.0 and Industry 5.0 are analyzed holistically. In developed economies, the importance of engineering competencies and skills, supply chain and efficient implementation of emerging technologies: AI, the Internet of Things and big data sets is growing. In emerging economies, training and skills assessment, organizational sustainability and structure as well as new technologies (digitization, internet of things) are key elements [23,24,25]. Currently, the prevailing view, supported by scientific research, is that an energy sector based on fossil fuels has a devastating impact on the natural environment, which makes it necessary to shift the main energy burden to alternative energy sources (including PV microgrids) [27].

AI is becoming particularly important in the management of PV micro-installations in Poland. Under the old net-metering rules, owners of PV systems under 10 kW could feed up to 80% of their power into the grid, while PV systems between 10 kW and 50 kW could feed up to 70% of their electricity into the grid. Under the new net-billing rules, prosumers must prepare a bill that takes into account the energy they produce. The price is calculated according to a special model referring to the price of a kilowatt-hour in so-called day-ahead trading.

An even bigger change in Poland will be the introduction of dynamic tariffs in 2024. Dynamic electricity tariffs, or cheap electricity at certain times, will allow the most energy-intensive activities to be shifted to times of day when electricity will be cheapest. Smart electricity meters will make this possible. Contrary to appearances, the cheapest laundry, cooking or home heating will not be at night, but on the contrary—at noon. Managing such systems will be very demanding.

Figure 7 shows the need to store the energy produced from PV installations at a time when the price of energy is low. Additionally, this will help to reduce the load on the power grids.

The energy market around the world is preparing for the implementation of energy storage, particularly at the consumer and user level. This is illustrated in Figure 8. It shows that, soon, we will not be able to imagine investing in a PV installation on our roof without investing in home energy storage. It is expected that by 2030, half of all installed PV installations will be equipped with energy storage [28].

There are many factors to consider when choosing energy storage for the home. First and foremost, technical parameters are taken into account: the power of the system, the amount of energy stored, full-cycle efficiency, the area occupied by the installation, energy density, service life, method of connection to the energy system, level of reliability. No less important are the investment and operating costs (unit and total). Investors often take into account the sophistication of the energy storage technology and its impact on the environment. The following parameters were selected as part of the analysis:

• Rated power of energy storage—in the case of energy storage, the power rating is a parameter that defines the amount of energy that the most commonly used lithium-ion or lead–acid batteries are capable of delivering at any given time. It is expressed in kilowatts (kW). This parameter is important, because, if the power rating of the energy storage is too low, the number of devices it will be able to power at the same time is significantly reduced. It is therefore important to select it according to the individual needs of future energy storage users;
• Energy storage capacity—the capacity of the PV energy storage device, also referred to as capacitive capacity or simply storage capacity. This is expressed in kilowatt hours (kWh) and is the product of the voltage at which the storage device operates and the ampere hours of the battery. In turn, this parameter indicates how much a particular storage unit will be able to store. It is worth mentioning that the energy storage capacity can be freely expanded by adding another battery connected in series to its installation. In the case of capacity, it is also worth paying attention to the power rating. These two parameters should be matched in relation to each other, as an energy store based on lithium-ion batteries with a high capacity and low power rating will be able to power a small number of devices, but for a long time. In contrast, one with a low capacity and high power rating will power a large number of devices at the same time, but only for a short period. The standard storage capacity should be selected on the basis of the average daily electricity demand, taking into account the useful capacity of the storage rather than its nominal capacity;
• Battery efficiency—when deciding to install energy storage, it is also worth bearing in mind that the amount of electricity contained by charging the battery via a PV installation is never 100%. It all depends on the efficiency of the batteries used in the device. This is why energy storage units based on lithium-ion batteries, which have an efficiency of around 95%, have become increasingly popular in recent years. This means that for every 10 kWh supplied to them, 9.5 kWh can be successfully collected. In comparison, lead–acid batteries have an efficiency of 75%;
• Level of discharge—when it comes to being able to use the amount of energy produced and stored in the batteries, the level of discharge specified by the manufacturer is also important. It indicates the level to which lithium-ion or lead–acid batteries can be safely discharged without affecting their capacity and subsequent performance. Therefore, if the installed energy storage has a capacity of 10 kWh and its manufacturer has indicated that the level of discharge is 90%, this means that the battery should not be discharged below 2 kWh. Thus, its users only have 8 kWh of energy available, despite the 10 kWh in storage. Furthermore, the level of discharge also indicates the class of the battery being used. Therefore, the higher its value, the more expensive it is;
• Safety of energy storage operation—the topic of the safety of domestic energy storage was an important issue a few years ago. It was influenced by the use of lithium-ion batteries in energy storage. Lithium batteries are a relatively new technology—they only started to be used in portable devices in the early 1990s. By comparison, lead–acid batteries made their market debut around a century earlier. Overcharging lithium batteries poses a risk of fire, and their transport is subject to hazardous material conditions. All of this means that when lithium-ion batteries began to be used in domestic energy storage, no longer as small units in phones, cameras or laptops, but as storage units with a much larger capacity, doubts were raised about whether this posed a risk to the occupants of the home.

4.1. Reference to Results of Earlier Studies

Microgrids of renewable energy sources (including PV) use computerization, automation and robotization to increase efficiency and quality of services, and to optimize the use of environmentally friendly distributed energy resources. Research to date confirms that, despite imperfections such as outages, they are vulnerable to AI-based performance management and performance optimization to anticipate and fix outages as part of preventive maintenance. The application of AI in PV includes optimization, improvement of energy quality and identification of failure points using artificial neural networks (ANN), genetic algorithms (GA), fuzzy logic (FL), particle swarm optimization (PSO), heuristic optimization, artificial bee colony (ABC), as well as simpler solutions such as regression and classification [27]. Integration of the PV grid with the power grid requires the prediction of energy generation using PV panels. So far, weather data (temperature, humidity, cloud cover, wind speed, etc.), solar radiation and insolation have been collected in hourly intervals, and analyses have been made using deep neural networks (LSTM, GRU and transformers), with the latter proven to be the most effective (RMSE in the range of 0.21–0.24) [29]. In the case of a risk of PV panels being covered with dust, an AI-based system for detecting the level of dust on PV panels was developed, combined with dust cleaning units, and tested in real field conditions in various weather conditions [30]. AI-based models can recognize location/region specifics, long-term spatial and temporal variables, and anomalies in insolation patterns. This supports the increasing of the accuracy of the estimation of PV panel configuration parameters and generation capacity for 24 h renewable energy production planning. The existing conventional measurements and models based on partial differential equation solutions are still used in forecasts for large areas and at least medium-term forecasts, but differential learning is also being developed, combining the advantages of traditional and AI-based approaches, treating binomial nodes of the network as the output sum, without reducing the complexity of the network. The accuracy achieved is 87.8–88.1% for PV panels, with much lower values for wind power plants (36.3–46.7%) [31]. Charging electric vehicles becomes much easier by coordinating EV charging stations with PV-powered arrays using an AI model [32]. The inclusion of renewable energy sources, including PV, in the energy mix improves the cumulative power generated by the power system and allows for a better response to changes in energy demand. The intermittent nature of the supply of energy from PV panels requires both an efficient maximum power tracking scheme, synchronization and power system oscillations to maintain its stability. Previous technological models in this area also included phase-locked loops, virtual synchronous generators, power system stabilizers and flexible AC transmission systems, but they are increasingly being supplemented or even replaced by AI-based coordinated control systems [33]. Changes in temperature and solar radiation can have the greatest impact on a sharp increase or decrease in power output. A hybrid model consisting of a convolutional neural network (CNN, classifies weather conditions) and long short-term memory (LSTM, which learns power generation patterns based on weather conditions) allows for stable forecasting of power generation. Forecasting accuracy was 92.94–95.42% depending on the day (sunny, cloudy) [34]. AI tools in the cloud are already able to automatically optimize the aggregate performance of the microgrid towards the creation of an automatically controlled microgrid integrating active technologies (including PV). This can be used not only for effective energy management and flexibility forecasting, but also for optimizing the efficiency of buildings and microgrids towards near-zero energy consumption (like the PLUG-N-HARVEST ICT platform) [35]. Differences in performance, robustness, accuracy, and generalizability are now the basis for which AI technique are used. More and more often, especially for large data sets and newly created (sometimes dedicated to a given system PV) hybrid methods—combining AI with other optimization methods have better results than a single method [36]. AI-based approaches used to study the impact of climate change on solar energy and predict the generated power showed that the DL-based model was the best with the minimum set of features describing the operating conditions of the PV panels (about 7 features, MSE = 0.15), and the polynomial regression model was best with larger trait sets (more than 10). The linear regression model gave the worst results [37]. The hybrid maximum power point tracking method with zero steady-state oscillations, based on a combination of a genetic algorithm and a perturbation and observation method, can quickly track the global maximum power point for a PV system even under almost any atmospheric conditions, including partial shade [38]. Furthermore, the currently used AI techniques affect the PV value chain and their advantage over systems using conventional solutions. This is a factor worth considering when studying building business models [39]. Even devices such as the Romanian off-grid PV system for irrigation are subject to AI modelling to improve the efficiency of its mechanical and electrical components (monthly energy efficiency, monthly pumping capacity (flow rate), monthly total efficiency). The model was based on two main weather parameters: solar irradiance and ambient temperature based on the maximum power point tracking method [40]. The intelligence of the Grasshopper Optimization Algorithm is also used to control proportional integral controller parameters in island PV systems [41]. Reliable forecasting of solar radiation intensity two days ahead is possible thanks to simple solutions such as UTSA SkyImager for imaging the entire sky and (after supplementing with AI algorithms: DL and gradient boosted trees (GBT)) for forecasting radiation intensity from the sub-image surrounding the sun [42]. AI-based models allow the classification of the technical condition of PV cells based on electroluminescence images at the level of PV cells and current–voltage curves of PV cells to classify cells based on their production efficiency [43].

4.2. Limitations of Own Studies

The further development of renewable energy sources is facing increasing barriers. For example, in Poland, there has recently been an increasing number of refusals to connect domestic PV investments to the grid. Already, nearly half of all applications submitted are rejected, with refusals so far mainly affecting high-power installations, and now also Poles wishing to install panels for their own use. This means that, despite EU subsidies, not everyone will be able to benefit from them. Under Polish regulations, such a refusal to connect to the grid excludes the installation of PV. Importantly, the operator should state when this will be possible (e.g., when the grid will be upgraded). An opportunity to change this situation is to invest in the development of the grid on the basis of its dynamic predictive models that take into account changes expected in the coming years, such as the increase in the number of electric cars being charged in homes and blocks of flats (in the case of flats). AI also offers an opportunity here to support research, monitoring, inference and prediction, but this requires its integrated industrial-scale application in energy companies. The costs of adapting the electricity system to distributed energy (including PV installations) will be enormous, and the process of upgrading the grid itself will be spread over many years and be adaptive, as the number of PV installations will increase, but the population itself will tend to decrease and the population distribution will change—towards migration to larger centers [4,5].

In this study, the short-term effects were considered (up to 30 days), but battery ageing and changes in system performance are long-term changes that have yet to be studied. However, the planned long-term models are subject to much larger errors. Electricity production by PV cells takes place under conditions of adequate sunlight. Hence, it may be reasonable to assume that similar forecasts for the winter period and months with worse weather (less or highly variable insolation) will have a much higher error rate than forecasts for summer months (as in this study). This will, of course, also affect the accuracy of long-term forecasts covering the winter months.

4.3. Theoretical Implications

The integration of AI in PV networks has several theoretical implications for both the research community and practitioners. These implications span a variety of disciplines, from renewable energy and engineering to computer science and economics:

• Advanced energy management paradigms: the use of AI introduces advanced energy management paradigms that use ML, optimization and control theories. Researchers can explore new ways to model, analyze and optimize the performance of the PV grid, leading to the development of innovative algorithms and techniques.
• Interdisciplinary collaboration: the application of AI in PV requires collaboration between experts in renewable energy, computer science, data analysis and other fields. This interdisciplinary approach encourages researchers from different fields to collaborate, leading to a mutual exchange of ideas and the development of holistic solutions.
• Modelling complex systems: AI enables the modelling and simulation of complex interactions in PV networks, taking into account factors such as weather conditions, energy demand patterns, grid constraints and user behavior. Researchers can delve into the modelling of complex systems to better understand the dynamic behavior of PV networks and assess the impact of AI-based strategies.
• Data-driven insight: The use of AI facilitates data-driven decision making by analyzing large amounts of data from PV systems. Researchers can explore new data pre-processing, feature extraction and analysis techniques to uncover valuable insights that were previously difficult to discern using traditional methods.
• Algorithm development and evaluation: AI in PV grids requires the design, development and evaluation of AI algorithms tailored to specific optimization and energy management tasks. Researchers can contribute by creating novel algorithms that address challenges such as real-time energy forecasting, load scheduling, fault detection and adaptive control.
• Uncertainty and robustness: Theoretical research can focus on resolving uncertainties and increasing the reliability of PV grid solutions based on artificial intelligence. This includes developing algorithms that take into account uncertain weather forecasts, fluctuations in PV system performance and other sources of uncertainty to ensure reliable and efficient grid management.
• Human-centric design: Researchers can explore the human-centric design of AI-powered PV grid systems by considering how users interact with and respond to AI recommendations. This includes understanding users’ preferences, behavior and decision-making processes to design user-friendly interfaces and algorithms.
• Economic and market dynamics: The integration of AI into PV grids introduces economic and market dynamics that affect energy trading, prices and consumption patterns. Researchers can study how AI-based strategies influence energy markets, pricing mechanisms and incentives for renewable energy deployment.
• Policy and regulatory implications: The use of AI in PV grids could have implications for energy policy, regulations and grid integration standards. Researchers can analyze the potential impact of AI on policy implementation, grid stability and integration of distributed energy resources.
• Sustainability and environmental impact: Theoretical research can explore the broader implications of AI-guided PV grid management for sustainability and environmental goals. This includes assessing how AI-based strategies contribute to reducing carbon emissions, optimizing resource use and increasing overall energy efficiency.
• Long-term system evolution: Researchers can study the long-term evolution of AI-enabled PV grid systems by considering factors such as technological advances, system scalability, and integration of emerging energy storage technologies.
• Education and training: As AI becomes a key component of PV grid management, researchers can contribute to educational initiatives by developing curricula, training materials, and resources that enable practitioners to successfully apply AI techniques in real-world scenarios.

Theoretical postulates formulated in this way may be a facilitation for young engineers in AI applications in PV, who are just entering the profession and need to gain experience.

Making connections between policy aspects of energy and experimental results is crucial to grounding research and showing its real-world implications. Such a connection helps bridge the gap between theoretical work and practical applications, ensuring that research findings are relevant and influential in addressing energy challenges. Here is how this connection can be established:

• Policy context: Start by outlining the relevant energy policies or regulations that govern the research area and shape the energy landscape and influence decision making;
• Research objectives: Identify research objectives, key issues, the planned way forward for the field;
• Experimental design, such as an experimental set-up, methodology and data collection process to achieve the research objectives;
• Policy relevance: Linking research objectives to the policy context and their impact on energy policy, incentive schemes, and grid integration policy (in the case of PV);
• Quantitative analysis: The results of the experiments shown using quantitative data and relevant visualizations, and how the above results are in line with or diverge from theoretical expectations;
• Policy implications: The results of the experiment in the context of energy policy, whether the results are in line with or challenge existing policies and possible policy changes that could make use of the research results;
• Cost–benefit analysis: The cost-effectiveness of implementing the research results in real-world scenarios, their economic feasibility, environmental impacts and potential energy savings;
• Stakeholder engagement: Identifying key stakeholders (policy makers, industry representatives, support groups)—how research findings can address their concerns or objectives, fostering collaboration and engagement;
• Case studies: Scenarios illustrating how experimental results can translate into practical applications for visualizing the real-world implications of the research;
• Policy recommendations: Suggestions for improving existing policies (increasing energy efficiency, sustainability, moving closer to achieving other relevant goals);
• Long-term impact of implementing the research results: How they can contribute to wider energy-related goals (reducing carbon emissions, increasing energy security, etc.);
• Communication channels: Ways to reach policy makers, industry experts and the public (policy briefs, presentations, workshops and engagement through relevant conferences, associations, platforms).

4.4. Direction for Further Studies

Directions for further research include multi-domain approaches, based on the potential of combined analyses (including instrumental ones), offering research results and new AI methods, techniques and tools to wider user groups of their potential users to ensure the translation of pilot studies into real world practice in PV grids and cooperating with the power systems. Only universality, integration and cooperation of solutions will allow the intended effect to be achieved [23,24,25].

An essential direction for further research is to test the three best AI solutions on the data for at least an entire year (greater variety of climatic conditions) and to determine, for the best solution, the rarest data acquisition interval that allows the system to operate effectively (analysis, inference, prediction) in real time.

From the implementation point of view, standardization and validation of ready-made AI solutions for PV grids should be a priority—this will avoid incompatibilities at the technical level.

It seems that the most important future research direction is the use of AI to protect against periodic energy shortages, and globally against an energy crisis.

Further research is influenced by changes in Polish law and differences between net metering and net billing compensatory programs. The first is net metering, which compensates for the retail rate, and the second is net billing, which compensates for the lower supply or wholesale rate. The choice of net billing results in a decrease in the value of the surplus electricity of the PV system, which may reduce the financial return and extend the duration of the return on investment in the PV system.

The use of AI in the PV industry should be approached globally. This field can be used not only in the area indicated in this article, but also in other stages [4,5,44,45,46]:

• Production stage—The need to use AI systems appears already at the stage of the production of cells and entire PV modules. To compete with production facilities in China, Europe needs to produce more technologically advanced and more reliable products. European research institutes indicate the need to build self-learning production management systems in production plants that use process data to continuously optimize production. The production of technologically advanced cells and entire PV modules currently requires many complex production processes and materials, and the amount of recorded data is huge. Thanks to the “self-learning factories”, we have new, innovative tools at our disposal, thanks to which we can save time and money in the development of the production of new PV modules [47,48,49,50].
• Installation design stage—selection of installation and modelling of meteorological data.
• With the development of investments in PV farms, it becomes more and more necessary to optimize the decision of where to place them. Tools based on artificial neural networks are needed to analyze the potential of plots of land for PV farms and the potential of house roofs for investments in PV installations. To summarize, we already have widely known applications such as PVSol and many others, but we need a system for multi-criteria analyses for entire districts or areas. The production of energy from PV depends on the weather conditions. When the weather is sunny, the amount of power generated by the PV system is high. When the weather is partly cloudy or cloudy, then there is less power. Therefore, power generation from PV systems is unreliable, intermittent and variable. To manage the uncertain and variable power generation of PV systems, grid operators use classic PV power forecasts and, increasingly, digital methods based on the so-called methods of intelligent management of the PV generation network.
• Operation stage—Faults in PV modules affect the energy yield of a PV installation more than any other factor. These errors vary greatly depending on the stage at which they appear. Face-to-face inspection and traditional supervisory control and data acquisition (SCADA) are often unable to pinpoint the root cause of errors in a short period of time. A global study by TÜV Rhineland on 12 GW PV plants found that 30% of the plants had major defects, 50% of which were due to defects in the PV modules. Failures of PV modules directly affect the energy yields and benefits of PV installations. Accurate and quick locating of faults in PV strings (for large farms, tens of thousands of PV modules!) is an urgent problem for power plant owners. However, conventional methods of inspecting solar modules are both time-consuming and labor-intensive. Additionally, they increase operating costs. In a large PV power plant, there are many PV modules and PV strings scattered over a large area. For example, a typical 100 MW PV power plant contains approximately 400,000 PV modules grouped in 20,000 PV strings and covers an area of 72.5 km². It is impossible to make a 100% meticulous comparison of voltages and currents, or images of thermography photos, due to the huge number of PV modules and the complex test environment, even with the use of drones. This means that only spot checks are possible.
• Security—Safety is also a very important aspect for PV installations. Therefore, Huawei inverters and power optimizers are equipped with the AFCI (Arc Fault Circuit Interrupter) function to protect against the formation of an electric arc, directly affecting the safety of the entire PV installation. Thanks to the use of artificial intelligence, after detecting an arc, the inverter operation is interrupted within 2.5 s. AFCI is a system that uses elements of artificial intelligence, analyzes DC harmonics and allows the flow of this current to be interrupted if patterns are detected in the spectrum that may indicate an arc in the DC link of the inverter (the spectrum of the arc discharge current is similar to the white noise spectrum, and its energy is concentrated around the frequency range of 10–100 kHz). In this situation, the inverter disconnects the DC circuit in less than 2 s and generates an alarm, indicating the location of the arc with the accuracy of the module (in the case of fully optimized installations). The task of the installer is then to easily and efficiently remove the problem. The function protecting against the formation of an electric arc is also provided by power optimizers that have the ability to detect and interrupt the electric arc at the level of the PV module. An electric arc arising in a DC circuit is a very dangerous phenomenon. The temperature at the point of arc formation can reach up to 3000 °C, and as long as solar energy is supplied to the generator and the arc is not automatically interrupted, it can cause a serious fire hazard.

5. Conclusions

The achieved low RMSE values and relatively high accuracy values confirm that ANN can constitute an important extension to the capabilities of the PV grid software. The widespread deployment of PV systems is a step forward for countries, businesses of all sizes and households to achieve global green energy goals. AI can effectively support this process thanks to faster data processing, optimization of parameters, more accurate inference and earlier prediction of damage. PV systems, thanks to the continuous development of technology and high reliability at low costs, are successfully implemented on a large scale, as well as—in the distributed generation model—on a small scale, contributing to the creation of smart city and Industry 5.0 infrastructure. We are also increasingly convinced that business models that are effective on the market should contain some element of AI as a carrier of their modernity.

According to the authors, AI will be most relevant for intelligent forecasting and management of energy in our homes and across entire markets. This is particularly relevant in the case of increasing periodic surpluses of RES energy production in Europe. It is all about so-called smart forecasting.

Systems of this type will be needed to manage the balancing of energy generation: the relationship between the level of output from PV or wind power and generation from conventional or nuclear power plants. The sudden increase in renewable energy production is already leading to so-called negative energy prices in Europe. This could result in RES sources being disconnected in the absence of management. In this context, the use of energy storage at grid level and in our homes seems crucial.

Even when it comes to renewables, forecasting is widely used to accurately determine energy production in specific geographical areas. DL algorithms have more predictive power than all industry experts combined. Forecasting in this sense can take many forms, from predicting demand and price trends to identifying potential growth areas.