A Review of State-of-the-Art and Short-Term Forecasting Models for Solar PV Power Generation

: Accurately predicting the power produced during solar power generation can greatly reduce the impact of the randomness and volatility of power generation on the stability of the power grid system, which is beneﬁcial for its balanced operation and optimized dispatch and reduces operating costs. Solar PV power generation depends on the weather conditions, such as temperature, relative humidity, rainfall (precipitation), global solar radiation, wind speed, etc., and it is prone to large ﬂuctuations under different weather conditions. Its power generation is characterized by randomness, volatility, and intermittency. Recently, the demand for further investigation into the uncertainty of short-term solar PV power generation prediction and its effective use in many applications in renewable energy sources has increased. In order to improve the predictive accuracy of the output power of solar PV power generation and develop a precise predictive model, the authors used predictive algorithms for the output power of a solar PV power generation system. Moreover, since short-term solar PV power forecasting is an important aspect of optimizing the operation and control of renewable energy systems and electricity markets, this review focuses on the predictive models of solar PV power generation, which can be veriﬁed in the daily planning and operation of a smart grid system. In addition, the predictive methods identiﬁed in the reviewed literature are classiﬁed according to the input data source, and the case studies and examples proposed are analyzed in detail. The contributions, advantages, and disadvantages of the predictive probabilistic methods are compared. Finally, future studies on short-term solar PV power forecasting are proposed.


Introduction
The energy crisis, air pollution, global warming, and other environmental issues have stimulated the development of renewable energy, which is expected to account for about 40% of energy consumption by 2030 [1].Solar PV power generation refers to a power generation device that uses a PV module to directly convert solar energy into electricity energy.This is a novel, highly promising, and comprehensive energy utilization method with the advantages of low environmental pollution, no pollution of air and water resources, no noise pollution, the ability to adapt to local conditions, low installation cost, and on-site consumption when connected to the power grid.It can achieve the coexistence of power generation and consumption and is currently one of the most promising PV technologies.According to Rethink Energy data, in the first three seasons of 2022, the global installed solar energy capacity increased by 54 GW, a year-on-year increase of 37.8%.The total installed capacity in the first nine months of this year was about 142.5 GW.The forecast shows that the annual installed capacity will reach 222 GW [2,3].According to the latest report from the European Photovoltaic Association SPE, the installed capacity of new devices in the 27 EU countries in 2022 was 41.4 GW, a net increase of 28.1 GW compared to last year, achieving a year-on-year increase of 47%.By 2022, the cumulative installed capacity was expected to reach 208.9 GW.According to the statistical data released by the National Energy Administration of China, the new installed power capacity in 2022 was 87.41 GW, and by 2022, the cumulative installed power capacity was 396.261GW.
The prediction of power generation was carried out very early due to the early establishment of a large number of solar observation stations in Europe and the United States, more assistance from advanced technology and equipment, and the accumulation of sufficient historical data.The main work involves the use of different predictive models to improve forecasting accuracy, and part of the work is to summarize existing methods or analyze their economic benefits.The methods for realizing PV power generation forecasting are mainly divided into traditional predictive methods in physics and statistics, novel forecasting methods using machine learning, optimization algorithms, and deep learning, as well as hybrid models.
More recently, in artificial intelligence (AI) or neural network (NN) approaches, a new short-term PV predictive method based on the artificial neural network (ANN) or recurrent neural network (RNN) was proposed.This method employs dynamic artificial neural networks to predict solar radiation and temperature, thereby achieving the prediction of the solar power energy output [4][5][6][7][8][9].Sudden changes in solar radiation near the surface are extracted from ground-based cloud image sampling technology and are combined with similar day-based and ANN-based approaches to ensure accuracy in solar radiation prediction [10][11][12][13].Lima et al. (2020) used AI methods in a new adaptive topology based on portfolio theory (PT) technology to make short-term predictions of effective solar PV power generation for global solar radiation [14].
Next, some solar PV power generation forecasting models based on machine learning or optimization algorithms, such as the support vector machine (SVM), support vector regression (SVR), extreme learning machine (ELM), gradient boosting decision tree (GBDT), and adaptive boosting learning (ABL), have been proposed [15][16][17][18][19][20][21][22][23][24][25][26][27][28][29][30][31][32][33].These use a large number of satellite images and a significant amount of data.When compared with traditional time series analysis, the forecasting accuracy is significantly improved.Ziyabari et al. (2022) used a novel multi-range attentive gated current residual network (ResAttGRU) model and meteorological data, the clear sky index, and solar Ireland to predict short-term solar radiation [34].This model also proposes the use of a strong multi-timescale in the proposed architecture, and the GRU can utilize temporal information at a lower computational cost than the popular long short-term memory (LSTM) method.Doubleday et al. (2021) established utility-scale photovoltaic (PV) plants at multiple time horizons based on the Bayesian model-averaging (BMA) algorithm and numerical weather forecasting (NWP) and obtained a probabilistic solar power forecasting model [35].
In addition, deep learning methods, such as the long short-term memory (LSTM) network model, the recursive short-term memory (Rec LSTM) network, convolutive long short-term memory (Conv LSTM), and the multi-step CNN stacked LSTM model , are used to predict the solar PV output power.Talat et al. (2021) proposed a new multi-layer feed-forward neural network (MFFNN) for solar PV power generation forecasting, considering thermal effects and environmental conditions [57].The results obtained from the MFFNN-MVO and MFFNN-GA models were studied through environmental temperature, wind speed, and solar irradiance.Jebli et al. (2021) established a multi-layer perceptron (MLP) model, which is a network composed of multi-layer interconnected nodes combined with the clear sky index to achieve the classification of environmental factors.They then optimized the weight of the multi-layer perceptron through the artificial bee colony algorithm to predict solar PV output power.This non-linear forecasting model has a better effect than the linear forecasting model since the output power is intermittent and random [58].
Moreover, some forecasting works have used hybrid and ensemble models.Ma et al. (2021Ma et al. ( -2022) ) proposed new forecasting models, such as VMD-LSTM-RVM, CNN-LSTM-MLP, MC-WT-CBiLSTM depth, NARX-CVM, wavelet-adversarial deep, GBRT-Med-KDE model, and TG-A-CNN-LSTM, and implemented interval forecasting for microgrids, providing a good solution for the energy management of microgrids [59][60][61][62][63]. Meng et al. (2021) proposed a new hybrid wavelet-adversarial deep model for power generation forecasting using satellite and global horizontal radiation (GHI) forecasting.This method integrates a wavelet neural network model with a three-stage adaptive modification solution to the DA to improve the algorithm's ability to modify local and global searches, and it provides relatively reliable forecasting results [64].Wang et al. (2022) proposed a hybrid LSTM-SVR-BO model that combines machine learning methods and statistical methods and conducted comparative tests on multiple time dimensions to better reflect the accuracy of the experimental results.They verified the advantages of the proposed method, which can achieve better forecasting results than a single model [65].Zhang et al. (2022) proposed the hybrid gradient boosting regression tree-median and kernel density estimation (GBRT-Med-KDE) models.This study proposes a short-term solar power interval prediction method for solar PV power generation, which effectively predicts global solar radiation.This method can obtain more reliable and stable interval forecasting results [66].Du et al. (2022) proposed a forecasting model based on the theory-guided and attention-based CNN-LSTM (TG-A-CNN-LSTM), which can ignore meteorological data such as temperature and wind speed.In the training process, data mismatch and boundary constraints are introduced into the loss function, and positive constraints are used to limit the output of the model.This model demonstrates better forecasting accuracy, stability, and robustness characteristics for solar PV power generation when compared to a single forecasting model [67].Furthermore, Ghasvarian Jahromi et al. (2020) conducted forecasting work using statistical methods such as the hidden Markov model (HMM), similarity-based forecasting models (SBFMs), and Kalman filtering (KF) and applied them to the probability forecasting of solar power generation [68,69].Mutavhatsindi et al. (2021) achieved good results when predicting the production of solar power plants using the quantitative regression average (QRA) regression model based on meteorological data [70][71][72][73][74].
To date, several review papers on solar PV power forecasting have been studied.Maciel, Rajagukguk, et al. (2021) outlined short-term methods for predicting solar PV power generation.In addition to using different forecasting methods to improve forecasting performance, another part of the work is to summarize and analyze the existing PV power generation forecasting methods developed in recent years based on time scales, forecasting models, and output data [75][76][77].Wu et al. (2022) summarized machine learning, deep learning, algorithm optimization, and hybrid forecasting models to achieve the modeling and forecasting of meteorological factors.Of these methods, the solar radiant intensity is a key parameter, and its forecasting results will directly affect the output power of PV power stations [78,79].Furthermore, Sudharshan and Mohamad Radzi summarized 161 and 306 related papers, respectively, and introduced various combinations, influencing factors, issues, limitations, and suggestions for achieving the solar PV power generation prediction of hybrid ANNs, machine learning methods, or algorithm optimization [80,81].
This review work intends to provide a clear and concise understanding of the different predictive models for solar radiation and solar PV power generation forecasting.In order to satisfy the requirements of large-scale solar PV power grid integration and further improve the forecasting accuracy of short-term solar PV power generation, it is necessary to develop a short-term solar PV power forecasting model based on state-of-the-art hybrid AI algorithms to accomplish accurate, robust, and efficient solar PV power forecasting.The main contribution of this paper is a review of the impacts of different irradiance forecasting techniques for solar PV power prediction, as follows: 1.
This paper discusses a systematic understanding of the selection and application scope of various prediction models, including Neural Networks (NNs), machine learning models or algorithm optimization, deep learning models, hybrid AI models, and probability models; 2.
This paper summarizes the current trends in solar PV power forecasting techniques, including their advantages and disadvantages, and the contributions of various solar PV power forecasting models.Some important metrics, such as the time resolution, model type, accuracy, and parameters, are presented; 3.
These models have different predictive capabilities, and the weights of each model are updated in real time to improve the comprehensive predictive capabilities of the models and have good application prospects for solar PV power forecasting; 4.
The paper reviews and analyzes case studies and examples in the literature that accurately predict short-term solar PV power forecasting with uncertainty and stochasticity.
Finally, the paper draws a conclusion and presents the existing issues in the methodologies.Future research directions are suggested.

Review of the Development of the Literature on Solar PV Power Forecasting Models
Improving the predictive accuracy of solar PV power generation is conducive to the optimal dispatching of microgrids.This paper analyzes the multi-time-scale optimal dispatching model of microgrids, which can effectively deal with the risks brought about by solar PV power prediction errors to system operation and achieve the optimal dispatching of solar PV microgrid systems.Then, starting from the necessity of improving the predictive accuracy of solar PV power generation, the impact of different predictive accuracies of solar PV output power on the optimal dispatch of microgrids is analyzed, and it is shown that the predictive accuracy of solar PV power generation can be achieved.Optimized scheduling that is more in line with the actual operation shows the practicability and necessity of improving the forecasting accuracy of power generation.

Forecasting Techniques
Previously, review articles with a wide scope (prediction techniques, sources of input databases, statistical metrics, temporal and spatial coverage, etc.) were produced.In recent years, relevant scholars have conducted theoretical research and practical simulations.This paper presents a comprehensive review of novel techniques for predicting solar PV power generation.Figure 1 shows a predictive model of solar PV power generation.The advantage of these methods (AI or neural networks (NNs), machine learning or optimization algorithms, deep learning, hybrid models, and other statistical analysis methods) is that the amount of training data can be greatly reduced, and they also avoid the excessive weighting of individual data.

Literature Classification Based on Methods
Modern solar PV power generation forecasting methods mainly include AI neural networks, the support vector machine, wavelet analysis, hybrid and ensemble model forecasting, etc. Neural networks have the characteristics of self-reasoning, selforganization, and information memory.They also have a strong fitting ability, complex mapping ability, fault tolerance, and learning ability and are suitable for dealing with a

Literature Classification Based on Methods
Modern solar PV power generation forecasting methods mainly include AI neural networks, the support vector machine, wavelet analysis, hybrid and ensemble model forecasting, etc. Neural networks have the characteristics of self-reasoning, self-organization, and information memory.They also have a strong fitting ability, complex mapping ability, fault tolerance, and learning ability and are suitable for dealing with a large number of unstructured and strongly dynamic regular problems.The relationship between solar PV power generation and time is usually random and non-linear because variations in solar radiation are affected by external conditions, such as temperature, relative humidity, rainfall, rainfall hours, sunshine hours, and full-day sunshine.Neural networks (ANNs) are the most frequently used machine learning techniques in short-term solar PV power forecasting.Hybrid predictive models are designed by combining two or three deep learning techniques or combining optimization algorithms with AI methods.They address the aforementioned shortcomings of a single predictive model by finding optimal features, hyperparameters, and training algorithms.The review works on solar PV power generation forecasting for time resolution, model type, accuracy, and the parameters used are presented in Table 1.

Summary of Forecasting Techniques
A literature review was conducted using (1) the Web of Science, (2) IEEE Xplore, (3) MDPI, (4) Engineering Village, and (5) Google Scholar databases from 2020 to 2023 for publications on short-term solar PV power prediction.In the past three years, the amount of research in this field has significantly increased, which is consistent with the global growth in solar power generation.This indicates that these predictive tech-Energies 2023, 16, 5436 9 of 30 nologies for solar PV power generation are becoming more important as their penetration rate in the power grid increases.In the initial search for this paper, a total of 217 papers were reviewed and identified using five academic literature databases.A total of 102 relevant articles were identified based on second-review keywords, titles, abstracts, article content, and the journal's main subject of interest.The final 74 papers were selected and analyzed based on reviewing the impact factor, review process, citation, exploration of issues and challenges, and future studies.Based on the temporal resolution, the number of AI methods used in the model, and the accuracy of the model, the performance level of short-term wind power prediction models is evaluated for the reviewed works, recommending prediction models with better performance.These models are mainly divided into five categories: artificial intelligence or neural networks (NNs), machine learning models (MLs) or algorithm optimization, deep learning models (DL), hybrid artificial intelligence models, and probability models.A list of all the papers is presented in the references.

Distribution of Input Data for the Reviewed Works
It was found from the reviewed literature that solar power generation can be predicted through different input source databases, as shown in Figure 1. Figure 2 presents the distribution of the five database input sources, of which the models using meteorological records [81][82][83][84] or numerical weather prediction (NWP) [85][86][87] are dominant, accounting for 49% and 25%, respectively.In several studies, 15% of the power generation information was shared from nearby PV power plants [56,59,88], 6% of the studies used satellite images as the input source data [89,90], and some studies combined with sky images have been very promising.Such studies account for 5% of all studies, although further work is needed to correctly identify cloud layers [72,[91][92][93].When considering their spatial resolution and the temporal level at which they are applied, NWP, satellite images, and sky images are plotted based on their spatial resolution, while the statistical methods are represented based on their spatial range.If inputs from NWP models or satellite or sky images are input into statistical prediction models, the spatial range of statistical methods will be expanded.It was found from the reviewed literature that solar power generation can be predicted through different input source databases, as shown in Figure 1. Figure 2 presents the distribution of the five database input sources, of which the models using meteorological records [81][82][83][84] or numerical weather prediction (NWP) [85][86][87] are dominant, accounting for 49% and 25%, respectively.In several studies, 15% of the power generation information was shared from nearby PV power plants [56,59,88], 6% of the studies used satellite images as the input source data [89,90], and some studies combined with sky images have been very promising.Such studies account for 5% of all studies, although further work is needed to correctly identify cloud layers [72,[91][92][93].When considering their spatial resolution and the temporal level at which they are applied, NWP, satellite images, and sky images are plotted based on their spatial resolution, while the statistical methods are represented based on their spatial range.If inputs from NWP models or satellite or sky images are input into statistical prediction models, the spatial range of statistical methods will be expanded.

Distribution of Forecasting Methods for the Reviewed Works
Figure 3 shows the distribution of studies analyzed regarding the techniques used.We found that 16% of all studies included artificial intelligence or neural network (NN) models, 31% included machine learning models or algorithm optimization, 34% included deep learning (DL) models, 13% included mixed artificial intelligence models, and probability models accounted for 6% of all studies.This selection is limited to publications produced in 2020 or later, as the purpose of this work was to focus on the latest trends and developments in solar power energy forecasting.The most common approaches among the papers reviewed were AI techniques, especially deep learning and machine learning or optimization algorithms, which accounted for 34% and 31% of the studies, respectively.

Distribution of Forecasting Methods for the Reviewed Works
Figure 3 shows the distribution of studies analyzed regarding the techniques used.We found that 16% of all studies included artificial intelligence or neural network (NN) models, 31% included machine learning models or algorithm optimization, 34% included deep learning (DL) models, 13% included mixed artificial intelligence models, and probability models accounted for 6% of all studies.This selection is limited to publications produced in 2020 or later, as the purpose of this work was to focus on the latest trends and developments in solar power energy forecasting.The most common approaches among the papers reviewed were AI techniques, especially deep learning and machine learning or optimization algorithms, which accounted for 34% and 31% of the studies, respectively.

Statistical Metrics for the Reviewed Works
There are many methods to determine errors in solar power generation prediction, and Table 1 uses various statistical metrics to describe the accuracy of different shortterm solar power generation prediction models produced in the past three years.In Figure 4, we develop and propose many methods for calculating errors, such as RMSE, MAE, MAPE, nRMSE, R 2 , MSE, MRE, nMAE, MBE, SMAPE, MASE, and WMAPE, and attempt to present the error values as completely as possible so that they can be used for the study of future short-term solar power generation prediction, which needs to be improved and evaluated.The most commonly used methods for counting errors in the literature on short-term solar power generation prediction are RMSE, MAE, and MAPE in the respective proportions of 25%, 22%, and 17%.The root mean square error (RMSE) is the most commonly used metric since it describes the measurement of the average distribution of errors.The RMSE is a good method for describing prediction errors because it does not consider the difficulty of the predictions made under different meteorological conditions.In addition, most predictive models tend to use some variants of the RMSE to evaluate the performance of their predictive models.
The research on the above short-term solar PV power generation shows that the accuracy of traditional single prediction models, such as BP neural networks [10], SVM [12,25], etc., is far from sufficient.It is easy to fall into local optimal solutions, thereby

Statistical Metrics for the Reviewed Works
There are many methods to determine errors in solar power generation prediction, and Table 1 uses various statistical metrics to describe the accuracy of different short-term solar power generation prediction models produced in the past three years.In Figure 4, we develop and propose many methods for calculating errors, such as RMSE, MAE, MAPE, nRMSE, R 2 , MSE, MRE, nMAE, MBE, SMAPE, MASE, and WMAPE, and attempt to present the error values as completely as possible so that they can be used for the study of future short-term solar power generation prediction, which needs to be improved and evaluated.The most commonly used methods for counting errors in the literature on short-term solar power generation prediction are RMSE, MAE, and MAPE in the respective proportions of 25%, 22%, and 17%.

Statistical Metrics for the Reviewed Works
There are many methods to determine errors in solar power generation prediction, and Table 1 uses various statistical metrics to describe the accuracy of different shortterm solar power generation prediction models produced in the past three years.In Figure 4, we develop and propose many methods for calculating errors, such as RMSE, MAE, MAPE, nRMSE, R 2 , MSE, MRE, nMAE, MBE, SMAPE, MASE, and WMAPE, and attempt to present the error values as completely as possible so that they can be used for the study of future short-term solar power generation prediction, which needs to be improved and evaluated.The most commonly used methods for counting errors in the literature on short-term solar power generation prediction are RMSE, MAE, and MAPE in the respective proportions of 25%, 22%, and 17%.The root mean square error (RMSE) is the most commonly used metric since it describes the measurement of the average distribution of errors.The RMSE is a good method for describing prediction errors because it does not consider the difficulty of the predictions made under different meteorological conditions.In addition, most predictive models tend to use some variants of the RMSE to evaluate the performance of their predictive models.
The research on the above short-term solar PV power generation shows that the accuracy of traditional single prediction models, such as BP neural networks [10], SVM [12,25], etc., is far from sufficient.It is easy to fall into local optimal solutions, thereby The root mean square error (RMSE) is the most commonly used metric since it describes the measurement of the average distribution of errors.The RMSE is a good method for describing prediction errors because it does not consider the difficulty of the predictions made under different meteorological conditions.In addition, most predictive models tend to use some variants of the RMSE to evaluate the performance of their predictive models.
The research on the above short-term solar PV power generation shows that the accuracy of traditional single prediction models, such as BP neural networks [10], SVM [12,25], etc., is far from sufficient.It is easy to fall into local optimal solutions, thereby reducing the prediction accuracy.Deep learning (DL) networks are neural networks with many hidden layers, which can actively and comprehensively grasp the abstract features of samples by using layer-by-layer training and learning methods to form a feature space [86,89].It overcomes the shortcomings of BP neural networks and SVM, thereby effectively improving the prediction accuracy.In addition, due to machine learning techniques, such as extreme learning machines, where the input weights and hidden layer thresholds can be randomly set, the calculated hidden layer output weights can have significant fluctuations, leading to unstable prediction results.In order to reduce prediction errors, the particle swarm optimization algorithm has a strong global search ability and simple optimization, overcoming the disadvantage of the extreme learning machine model, in which the output weights are prone to random fluctuations [17,19].A forgetting mechanism or adaptive extreme learning machine is employed to optimize the number of neurons in the hidden layer within a certain range to solve the problem of the poor generalization ability of extreme learning machines [21,87].Due to the advantages and disadvantages of different prediction models, hybrid prediction methods are used to optimize the data processing results of different models based on specific strategies to obtain better solar PV power generation prediction results and ultimately improve predictive accuracy [92,93].It was found that hybrid prediction methods have the optimization characteristics of the prediction results.These models fully leverage the advantages of various hybrid prediction models, effectively overcoming the poor adaptability and low prediction accuracy of individual models and providing a more practical reference for the optimization and dispatch of PV microgrids.

Scientific Contributions and Comparison of Reviewed Works
In the past decade, studies on solar PV power generation prediction have become more and more popular.This paper covers the contribution of the recent progressive solar PV power forecasting technology and explores the advantages and disadvantages of the various solar PV power forecasting models produced in the past three years, as shown in Table 2.These forecasting models have different forecasting capabilities, update the weights of each model in real time, have an improved comprehensive forecasting capability, and have good application prospects for solar PV power generation forecasting.Increasing dimensions of the input vector.
[5] August 2021 A daily clustering method based on statistical features, such as daily average, maximum, and standard deviation of solar PV power, is adopted in the datasets to address the impact of uncertain weather on the prediction model. -

Date of Publication and Location Main Contribution
Advantages Disadvantages [8] June 2020 An accurate prediction model and results can be obtained for specific regional meteorological data.
A promising alternative to accurate power prediction for practical PV power plants.
Lowing the prediction accuracy of ANN models due to the chaotic nature of meteorological parameters.
[ This database size limits the prediction horizon of the models.
[13] November 2020 Similar hour-based and hybrid methods have presented better performance than commonly deployed prediction techniques.
The outputs of both solar PV prediction methods are dynamically weighted based on weather types and the MAE.
Increasing dimensions of the input vector.
[14] January 2020 Creating a hybrid model of four different artificial intelligence prediction methods to obtain the optimal policy for each prediction technique to reduce predictability errors.It is difficult to adjust the LSTM parameters and determine whether it converges.
[16] October 2021 Forecasted photovoltaic power generation using historical weather data and different time resolutions.
Suggested a model that only requires a set of dates to specify a prediction period and more inputs.
The modeling would take a longer time due to the large amount of historical data.
[17] October 2022 The ELM method was employed to ensure faster computing time and more direct microcomputer realization.
FFNN, with the particle swarm optimization algorithm, is used to achieve the search when computing the optimal weight.
Using PSO to select the parameters of adaptive ELM will make the computing time longer.
[18] June 2020 Assisted power grid operators in better planning the economic dispatch of solar energy grid-connected electricity.
The accuracy of short-term predictions can further be improved by using a longer time period of earlier data.
It is necessary to do more work (usually for several years) on larger datasets to confirm this.
[19] June 2021 To avoid erroneous optimistic predictions, the predicted power generation should be reduced to avoid affecting the stability of virtual power plants.
Better accuracy and time resolution of irradiance prediction is achieved for the next hour interval.
It is required to acquire the percentage of uncovered sun and cloud images within the next hour.It is difficult to sharpen the accuracy of an individual model.
[21] November 2021 Provided time and space compensation, as well as comprehensive power regulation while assisting energy dispatch units in generating strategies, is crucial for the stability and security of the energy system and its continuous optimization.
The proposed method can reduce the training time while improving accuracy.
The degree of uncertainty in photovoltaic power generation is closely related to the chaotic nature of weather conditions.
[ The integrated prediction models are much more accurate than single prediction models.
Recalculating the weight of each new input sample to improve the accuracy of a single prediction model.
[23] October 2021 Trained Abha's solar photovoltaic system data using seven famous machine learning algorithms to predict photovoltaic power generation.
Obtaining relatively low prediction error of the algorithms.The MSE of RF was the worst. [24] November 2022 The STVAR model demonstrated good model performance by predicting at a time resolution of 5 min to 1 h.
The prediction system can reduce the cost without installing and maintaining the solar irradiance sensor.
The research limitations of the irradiance prediction model (STVAR model) affect the final PV prediction results.
[25] November 2021 Machine learning (ML) model is an efficient tool that can predict the power performance of any solar photovoltaic power generation.
The high reliability and accuracy of the GPR prediction model can be verified.
-Square exponential GPR shows worse performance due to the complicated relationship between input parameters and the dielectric coefficient.-Cubic SVM presents worse performance due to the complicated relationship between input parameters and PV module power.
[ Expediting computing time and lowering the complexity of the model.
The selection of parameters using MK-RVFLN affects the accuracy of the prediction model.
[27] September 2022 Clustering initial training set and day-ahead power forecasting using adaptive k-means.
Gru network has excellent prediction results, better robustness, and fewer errors.
Increasing dimensions of the input vector.
[28] May 2022 Had better global radiation prediction results.
Proposing seven machine learning models for PV power generation forecasting.
Increasing dimensions of the input vector.
[29] February 2022 Selecting the SVR-based model parameters using PSO-based algorithms to improve the model performance Reaching better performance of the forecast algorithm.
Using algorithm bar parameters will lead to longer operation time.
[30] April 2022 A novel solar PV power prediction method composed of a feature extraction, clustering method, and hybrid classification regression prediction engine.
-The forecasting computation is quicker.
-Individual training for each subset is achieved by a prediction engine.-Obtaining the final solar PV power forecasting by using a relevancy-based combination of these two predictions.
-Increasing dimensions of the input vector.- The internal parameters of the subset need to be well selected.In some cloudy or cloudy days conditions, the model prediction accuracy needs to be improved.The whole PV system uses the solar modules with the lowest power to calculate the worst-case power generation performance, and mismatch loss is a major problem.
[45] September 2022 With historical satellite images as input, the FPP model based on CNN is employed to predict the future PV power fluctuation mode.
Reaching better performance for the prediction algorithm.
-Complex computation process.-Use of cloud computing.
[46] January 2021 The predicted results can successfully close the expected output and well capture the intra-hour ramping.
Achieving good performance of the prediction algorithm.
It is difficult to adjust the LSTM parameters and determine whether it converges.
[ -The feature of 3-D CAE models could not find well.-Increasing dimensions of the input vector.
[49] October 2020 The significance of the proposed synthetic prediction is highlighted to promote the more effective use of public sky prediction types and achieve more reliable PV power generation predictions.
Studied the performance of the proposed model in different seasons with different intraday horizon lengths.
[50] November 2022 To deal with the scenarios of missing data, an integrated model is proposed for probabilistic PV power generation prediction.
-Increasing dimensions of the input vector.-Computing time is too long.
[51] January 2021 In order to predict the solar irradiance at several locations simultaneously, a prediction model trained with several artificial neural networks is proposed.
A family of flexible and robust deep learning models for solar irradiance prediction is proposed.
-Increasing dimensions of the input vector.
[52] September 2020 A CNN model is proposed to find out the non-linear characteristics and invariant structures in the previous output power data so as to promote the prediction of PV power.
-CNN was used to preprocess the data.-Reaching better performance of the forecast algorithm.
-Increasing dimensions of the input vector -Computing time is too long.It is difficult to adjust the LSTM parameters and determine whether it converges.
[57] August 2021 Optimizing the number of neurons in the hidden layer, weight, and bias of the proposed neural network using MVO and GA algorithms.
Multi-layer Feed-forward neural network (MFFNN) is used to study the accuracy of MFFNN-MVO and MFFNN-GA models.
-Increasing dimensions of the input vector.
[58] February 2021 The relevance of the studied model in real-time and short-term solar energy prediction was evaluated using appropriate prediction models to ensure optimized management and safety requirements.
Artificial neural networks (ANN) have demonstrated good performance in real-time and short-term predictions.
-Increasing dimensions of the input vector.
[59] June 2021 Proposed a forecasting model with higher prediction accuracy and relatively small overall fluctuations.
Decomposing the PV power sequence to reduce the complexity and instability of the raw data by the VMD decomposition technology.
Prediction errors and fluctuations are large. [

Date of Publication and Location Main Contribution
Advantages Disadvantages [65] September 2022 In order to better reflect the accuracy of the prediction model, comparative tests were conducted on multiple time dimensions to verify the advantage of the prediction model.
A new prediction model has been proposed, which has an average improvement of about 15% in prediction accuracy and stability as compared to other prediction models.
Adjusting parameters using the BO algorithm increases the time cost of training the model.
[66] September 2022 An ensemble interval prediction method was proposed for solar power generation prediction to obtain higher-quality prediction intervals than other AI methods Obtaining more reliable and stable interval prediction results.
The KDE method takes a longer total computing time as compared to other AI methods.It is difficult to adjust the LSTM parameters and determine whether it converges.
[68] July 2020 Providing better accuracy than other investigated methods for cost computation.
Proposed a novel prediction model that outperforms other investigated methods in terms of accuracy and computational time.
Prediction accuracy can be increased with other new effective techniques.
[69] June 2020 SBFM predicts PV power at high time resolution by using low time resolution weather variables.
PV power generation prediction with a five-minute time resolution can substantially obtain accurate results.
Increasing dimensions of the input vector.
[70] August 2021 Generalized a novel prediction model to find the optimal prediction by affine transformation mapping for a given available measurement.
Irradiance, temperature, relative humidity, and solar zenith angle are selected as highly correlated inputs of WRF prediction model.
[71] November 2020 Completed the prediction combination of machine learning models with convex combination and Quantile regression averaging (QRA).
The forecasting performance of Diebold Mariano and Giacomini White tests is remarkable.
-Increasing dimensions of the input vector.
[72] May 2023 The data set of the proposed new prediction model contains weather features, which is more cost-efficient and more suitable for scenarios where there is no dedicated hardware or hard-to-obtain input features.
The proposed RF and DNN prediction models utilize widely available weather features and operate quite well even in the event of sudden fluctuations in PV output.
Increasing dimensions of the input vector.
[73] July 2021 The accuracy of the proposed multi-graph model is superior to other benchmark models in the day-ahead prediction cases.
When compared to the deep learning benchmark models, the single-graph prediction model had lower cost regarding training time.
-Increasing dimensions of the input vector.-Longer computing time.
[74] March 2022 The proposed model and algorithm can lower the dimensionality of the model and improve its prediction accuracy.
Proposed a short-term PV power generation prediction model based on combined fuzzy clustering, a genetic algorithm programming system (GAPS), and radial basis function (RBF) for meteorological data to improve the prediction accuracy.
The parameter numbers of the search window size affect the accuracy of the proposed models.
In the process of predicting solar PV power, each prediction model has its own advantages and disadvantages.Due to the research limitations, it is difficult to achieve high-precision predictions or different types of predictions with a single prediction model.With the continuous increase in the solar PV power grid connection capacity and the increase in the solar PV power penetration power, the State Grid Corporation has implemented increasingly high requirements for the scheduling and prediction accuracy of solar PV power.Based on this, establishing a combined prediction model for solar PV power prediction by integrating the advantages of various prediction models is of great significance for improving the accuracy of solar PV power prediction.Therefore, conducting research on solar PV power prediction based on artificial intelligence algorithms and optimizing prediction models has practical value in engineering for improving the accuracy of solar PV power prediction and the reliability of grid connection scheduling.

State-of-the-Art Approaches for Short-Term Solar PV Power Forecasting
The short-term solar PV power forecasting model is discussed in depth, as shown in Figure 5.The latest approaches to short-term solar PV power forecasting developed in the past three years are reviewed to provide an important reference for solar PV power grid integration.In order to improve the accuracy of solar PV power forecasting, this paper gives a detailed overview of the contributions, advantages, and disadvantages of various delivered solar PV power forecasting models, as well as presenting future research work.These advanced forecasting models can be approximately classified into artificial intelligence/neural networks (NNs), machine learning or optimization algorithms, deep learning, hybrid and ensemble forecasting models, and other statistical analysis methods.The proposed novel short-term solar PV power forecasting models provide very useful information for power system operation and control with high renewable energy penetration.
power prediction by integrating the advantages of various prediction models is of great significance for improving the accuracy of solar PV power prediction.Therefore, conducting research on solar PV power prediction based on artificial intelligence algorithms and optimizing prediction models has practical value in engineering for improving the accuracy of solar PV power prediction and the reliability of grid connection scheduling.

State-of-the-Art Approaches for Short-Term Solar PV Power Forecasting
The short-term solar PV power forecasting model is discussed in depth, as shown in Figure 5.The latest approaches to short-term solar PV power forecasting developed in the past three years are reviewed to provide an important reference for solar PV power grid integration.In order to improve the accuracy of solar PV power forecasting, this paper gives a detailed overview of the contributions, advantages, and disadvantages of various delivered solar PV power forecasting models, as well as presenting future research work.These advanced forecasting models can be approximately classified into artificial intelligence/neural networks (NNs), machine learning or optimization algorithms, deep learning, hybrid and ensemble forecasting models, and other statistical analysis methods.The proposed novel short-term solar PV power forecasting models provide very useful information for power system operation and control with high renewable energy penetration.

Insolation Prediction for Solar PV Power Generation
A solar cell is a converter that directly converts solar light energy into electrical energy due to the PV effect.Photodiodes will convert the sun's light energy into electrical energy, which can be connected in series and parallel to form a battery array to increase the output.Equation ( 1) is given by using least squared curve fitting: The novel short-term solar PV power forecasting techniques

Insolation Prediction for Solar PV Power Generation
A solar cell is a converter that directly converts solar light energy into electrical energy due to the PV effect.Photodiodes will convert the sun's light energy into electrical energy, which can be connected in series and parallel to form a battery array to increase the output.Equation ( 1) is given by using least squared curve fitting: Of these, P S is the electrical energy obtained from solar energy (kW), which is the record of the solar power plant.P sb is the total capacity of the solar cell (kW), which is a constant value (units: kW), S t is the global solar radiation (MJ/m 2 ) obtained by the Central Weather Bureau, k is the design coefficient of solar module (parameters for curve fitting) and in solar PV power generation, and S t is the main factor affecting the power generation output and is also the main variable used to predict solar PV power generation, making Equation (1) more in line with actual solar power generation.Meteorological data, such as the air temperature, relative humidity, precipitation, precipitation hours, sunshine hours, and global solar radiation, provided by Central Weather Bureau (CWB) Observation Data Inquiry System, were used as input variables for the solar irradiance-related information database; the output variable is the global solar radiation for the solar PV power generation prediction techniques, as shown in Figure 5.

PV Array Model
The PV cell is a p-n junction semiconductor with characteristics similar to diodes.The parameters of the PV cell are modeled, as can be seen in Figure 6.The current source generates the photocurrent, I ph , which is proportional to the solar irradiation.The relation between the array terminal current and voltage is presented in reference [94].The maximum power point of the photovoltaic (PV) array is variational, so a search algorithm is needed according to the current-voltage (I-V) and power-voltage (P-V) characteristics of the solar cell.
where R s is the series resistance, R sh is the shunt resistance, I SC is the light-induced current, n is the diode ideality factor, I PVO is the diode saturation current, and V T is the thermal voltage.K is the Boltzmann constant (1.38 × 10 −23 J/ • k), and q is the electronic charge.I SC depends on the irradiance level S and on the array temperature T, while I PVO and V T depend on T only.The PV array current I PV is a non-linear function of the PV array voltage V PV of the irradiance level S and of the temperature [94,95].
tion between the array terminal current and voltage is presented in reference [94].The maximum power point of the photovoltaic (PV) array is variational, so a search algorithm is needed according to the current-voltage (I-V) and power-voltage (P-V) characteristics of the solar cell.
where Rs is the series resistance, Rsh is the shunt resistance, ISC is the light-induced current, n is the diode ideality factor, IPVO is the diode saturation current, and VT is the thermal voltage.K is the Boltzmann constant (1.38 × 10 −23 J/°k), and q is the electronic charge.ISC depends on the irradiance level S and on the array temperature T, while IPVO and VT depend on T only.The PV array current IPV is a non-linear function of the PV array voltage VPV of the irradiance level S and of the temperature [94,95].

Data Mining Technique
The data mining technique is used for data processing, and more meaningful data are selected from the database as modeling data, as shown in Figure 7.The problem dealt with by data mining is finding meaningful hidden information in a big database.Power generation forecasts are similar to solar energy.

Data Mining Technique
The data mining technique is used for data processing, and more meaningful data are selected from the database as modeling data, as shown in Figure 7.The problem dealt with by data mining is finding meaningful hidden information in a big database.Power generation forecasts are similar to solar energy.

Hourly Similarity (HS)-Based Method
The reference data selection method based on the hourly similarity (HS) forecasting method introduces the concept of the horizontal axis and the vertical axis of time.The hour of the prediction day to be forecasted is called the prediction hour.Firstly, the prediction day is used to find weather information for the reference day, which is the day before and the next day (the day after).The reference hours are selected from the prediction hour and the reference day.The reference hours are the hours before and after the prediction hour.These reference hours are used as reference data.The reference hours of the hourly similarity prediction method were selected from the hypothetical case demonstration, as shown in Figure 8. Grey is historical data.Yellow represents future data.

Hourly Similarity (HS)-Based Method
The reference data selection method based on the hourly similarity (HS) forecasting method introduces the concept of the horizontal axis and the vertical axis of time.The hour of the prediction day to be forecasted is called the prediction hour.Firstly, the prediction day is used to find weather information for the reference day, which is the day before and the next day (the day after).The reference hours are selected from the prediction hour and the reference day.The reference hours are the hours before and after the prediction hour.These reference hours are used as reference data.The reference hours of the hourly similarity prediction method were selected from the hypothetical case demonstration, as shown in Figure 8. Grey is historical data.Yellow represents future data.
Demonstration of how to select reference data; assuming that 12:00 on the prediction day is the prediction hour, the reference hours include 11:00 on the current day, 11:00-13:00 on the previous day, and 12:00-13:00 on the prediction day.A total of six pieces of data are selected for the reference hours (collectively referred to as the reference data).
diction day is used to find weather information for the reference day, which is the day before and the next day (the day after).The reference hours are selected from the prediction hour and the reference day.The reference hours are the hours before and after the prediction hour.These reference hours are used as reference data.The reference hours of the hourly similarity prediction method were selected from the hypothetical case demonstration, as shown in Figure 8. Grey is historical data.Yellow represents future data.Demonstration of how to select reference data; assuming that 12:00 on the prediction day is the prediction hour, the reference hours include 11:00 on the current day, 11:00-13:00 on the previous day, and 12:00-13:00 on the prediction day.A total of six pieces of data are selected for the reference hours (collectively referred to as the reference data).
Figure 8 shows the data types used in data mining for the similar day prediction method, and the data mining steps are as follows: Step 1: Select the database range and reference day from the prediction hour.
Step 2: Determine the reference data from the prediction hour and reference day.
Step 3: Normalize the data first and then perform sequence similarity searching for each layer based on the reference hours of each layer.Each reference hour has its own set of sorted data.
Step 4: Integrate a set of data from the same layer, and all the integrated data are modeling data.Figure 8 shows the data types used in data mining for the similar day prediction method, and the data mining steps are as follows: Step 1: Select the database range and reference day from the prediction hour.
Step 2: Determine the reference data from the prediction hour and reference day.
Step 3: Normalize the data first and then perform sequence similarity searching for each layer based on the reference hours of each layer.Each reference hour has its own set of sorted data.
Step 4: Integrate a set of data from the same layer, and all the integrated data are modeling data.
As an example, two layers with Layers 1 and 2 can be used for the reference data.The normalization for a particular {m, r} is shown in Equation (4): where Nr d−m,r is a similar hour, j ∈ [1, 2, . . ., f ] is the input space dimension number, d is the reference data for day, and m = [0, 1] as well as r ∈ {t + 1, t, t − 1} are the concepts of the horizontal axis and vertical axis of time at a similar hour.N j d−i,k is the original data for hour k in the database, and i ∈ [0, 1, 2, . . ., v] is the number of the days with the total number set to v; k ∈ [1, 2, . . ., u] is the hour number with u = 24.N mr i,k is sequenced for each reference hour and is used to figure out the degree of similarity in the data.L is the number of selections at a similar hour, and H DMT is the training data selected at a similar hour, as shown in Equations ( 5) and ( 6): After data mining, the modeling data are selected by the hourly similarity (HS)-based prediction method.The modeling data include the training data and test data.The training data are the integrated data obtained after sequencing the data (the sequencing data do not include reference data), and the reference data are used as the test material.The modeling data selected by data mining can be used to train the models of various state-of-the-art approaches for short-term solar PV power forecasting.

Internet of Things (IOT) Technology
Data obtained from solar PV power generation and several environmental sensors were collected to store in the Raspberry Pi database and corresponding data tables using Internet of Things technology.Through the Raspberry Pi environment, a Python crawler program can be developed to grab the weather forecast information from the local environmental observatory of the Central Meteorological Bureau and store the weather forecast information in the database.Raspberry Pi is also applied to set up the human-machine interface and display it in website form while viewing it remotely via the internet.Furthermore, the collection progress is checked to confirm the hardware operation status and collect the data stably [96][97][98].
After long-term data collection, the amount of data required for the input layer parameters of the neural network is obtained.Data tables for variables, such as solar PV power generation data and environmental sensor data, are exported from the database management system and are first brought into the model to train the input parameters of the fuzzy neural network while performing data preprocessing.After the data preprocessing is completed, the data are divided into a training group and a test group.The training group data are used to continuously train the internal parameters of the neural network, and then the proposed method is verified by the test group.The feasibility and accuracy of the data collection framework are shown in Figure 9.

Sky-Image-Based Methods
The automatic identification of clouds, cloud matching, and cloud area corrections based on ground cloud images and the estimation of the cloud movement direction are carried out to allow accurate judgments to be made on clouds that are about to cover the sun and to improve the accuracy and speed of big data feature prediction for solar PV power generation.Next, efficient pixel-sensitive prediction models can be developed based on satellite imagery to track the cloud shape and motion and study satellite measurements and high-resolution cloud images (e.g., images from ground-based sky cameras).In addition, these cloud image information-correlation features have been comprehensively used for classification and prediction while verifying the feasibility of the model using different datasets [99].
Based on dynamic sky images, the characteristics of the cloud layer are extracted to estimate the future cloud movement path by using the object tracking algorithm, and then the cloud cover of the sun is calculated according to the cloud movement path.Finally, the change in insolation is estimated through the long short-term memory (LSTM) network.The paper had the aim of finding a method for predicting the movement path of cloud cover and, at the same time, estimating the sun's shading of cloud piles and forecasting power variation due to the changes in insolation through the long short-term memory (LSTM) network.This information can be provided to power dispatchers or the EMS (energy management system) in advance to allow them to effectively respond to

Sky-Image-Based Methods
The automatic identification of clouds, cloud matching, and cloud area corrections based on ground cloud images and the estimation of the cloud movement direction are carried out to allow accurate judgments to be made on clouds that are about to cover the sun and to improve the accuracy and speed of big data feature prediction for solar PV power generation.Next, efficient pixel-sensitive prediction models can be developed based on satellite imagery to track the cloud shape and motion and study satellite measurements and high-resolution cloud images (e.g., images from ground-based sky cameras).In addition, these cloud image information-correlation features have been comprehensively used for classification and prediction while verifying the feasibility of the model using different datasets [99].
Based on dynamic sky images, the characteristics of the cloud layer are extracted to estimate the future cloud movement path by using the object tracking algorithm, and then the cloud cover of the sun is calculated according to the cloud movement path.Finally, the change in insolation is estimated through the long short-term memory (LSTM) network.The paper had the aim of finding a method for predicting the movement path of cloud cover and, at the same time, estimating the sun's shading of cloud piles and forecasting power variation due to the changes in insolation through the long short-term memory (LSTM) network.This information can be provided to power dispatchers or the EMS (energy management system) in advance to allow them to effectively respond to the impact of cloud clusters shading the sun on the grid [100,101].
A schematic diagram of the sky-image-based methods is shown in Figure 10.The system configuration can be divided into three parts, among which D1 is the part used for analyzing the characteristics of all-sky clouds covering the sun and predicting the movement path of cloud clusters; D2 is the part used for extracting the characteristics of ground-based all-sky pyranometers, and D3 is the part used for predicting the solar irradiance and the solar PV power generation.Part D1 in Figure 10 shows the design of a method to predict the moving path of cloud layers through the whole sky image and the moving path of the sun and deduce the moving path of the cloud layers.The moving path of the cloud layer takes into account the moving path of the sun.For sun shading conditions, the predictive path of cloud layer movement is regarded as future information, and the real-time value for the insolation observation of the ground-based all-sky pyranometer is determined in part D2.The input of the intelligent learning network is used to deduce the change in insolation, and the variation in solar PV power generation can be obtained according to the power and insolation curve (PV power curve) of the solar PV module.In recent years, various research institutions and scholars have adopted different cutting-edge methods to reduce power fluctuations and randomness in the power prediction of solar power generation, as well as to prevent possible errors and omissions in the original data and allow certain results to be achieved.However, there are still some problems to be solved.First of all, in the future, the sample space can be further expanded, and the diurnal insolation and the dimension of data samples can be increased to predict the diversity of solar data.According to solar power generation data with different characteristics, the prediction model was further optimized to increase the applicability of the model.Secondly, according to the characteristics of the existing hybrid model, the parameter optimization method was further improved to ensure that the prediction model has high prediction accuracy at different time sampling rates, making it suitable for different prediction situations.An overview of future solar PV forecasting studies is given in the next section.

Future Studies and Development
As an important subsystem of smart management systems for microgrids, solar PV power generation prediction systems play a vital role in the development of solar energy.Due to the close relationship between solar radiation and meteorological conditions, such as seasons, cloudy and sunny days, and day and night, novel solar PV power output predictive methods have been developed in the past few years to allow for the balanced operation and optimized dispatch of the power grid system.These methods have been used in experiments, and some results have been achieved to solve the intermittent and random power problems associated with solar PV power generation prediction as In recent years, various research institutions and scholars have adopted different cutting-edge methods to reduce power fluctuations and randomness in the power prediction of solar power generation, as well as to prevent possible errors and omissions in the original data and allow certain results to be achieved.However, there are still some problems to be solved.First of all, in the future, the sample space can be further expanded, and the diurnal insolation and the dimension of data samples can be increased to predict the diversity of solar data.According to solar power generation data with different characteristics, the prediction model was further optimized to increase the applicability of the model.Secondly, according to the characteristics of the existing hybrid model, the parameter optimization method was further improved to ensure that the prediction model has high prediction accuracy at different time sampling rates, making it suitable for different prediction situations.An overview of future solar PV forecasting studies is given in the next section.

Future Studies and Development
As an important subsystem of smart management systems for microgrids, solar PV power generation prediction systems play a vital role in the development of solar energy.Due to the close relationship between solar radiation and meteorological conditions, such as seasons, cloudy and sunny days, and day and night, novel solar PV power output predictive methods have been developed in the past few years to allow for the balanced operation and optimized dispatch of the power grid system.These methods have been used in experiments, and some results have been achieved to solve the intermittent and random power problems associated with solar PV power generation prediction as well as to reduce possible errors and omissions in the original input data.Based on the latest advances in AI neural networks, machine learning, and deep learning methods, this paper examined the temporal resolution, the parameters used, the accuracy, and the research limitations and reviewed the contributions, advantages, and disadvantages of the latest hybrid prediction models for the development of solar PV power generation.However, there are still some issues that need to be improved.The following points describe the main aspects that can be studied further: (1) Weather variable predictions: Recent investigations only selected meteorological stations based on historical survey data.However, the meteorological information from different regions is inevitably different.Therefore, considering the impacts of the geographical environment, weather, or climate-related factors at the location of the meteorological station can definitely improve the accuracy of solar radiation predictions.In addition, in terms of other meteorological and site determination factors used for solar radiation forecasting, such as the temperature, humidity, precision, pressure, and solar radiation, etc., the impacts of these factors on the prediction results need to be explored, as these could be included as input factors for future meteorological data from the Meteorological Bureau to improve prediction accuracy; (2) Modeling the prediction algorithms through cloud images: Cloud areas based on ground cloud images are automatically identified, matched, and corrected to estimate the direction of cloud movement and make accurate judgments about clouds that are about to cover the sun.It is necessary to improve the accuracy and speed of feature prediction for big data used for solar PV power generation.Efficient pixel-sensitive prediction models were developed based on satellite images to track the shape and motion of clouds and study satellite measurements and high-resolution cloud images (such as images from ground sky cameras).These correlation features of cloud image information are comprehensively utilized for classification and prediction, for which different datasets are applied to verify the feasibility of the model.New hybrid models or multiple optimization algorithms, including cloud information for predictive models, are also integrated to improve the models and their prediction accuracy; (3) Solar PV power generation forecasting: Weather forecasting is selected based on data characteristics, and machine learning or optimization algorithms are added to the solar PV power generation prediction model, for example, optimization algorithms with RNN-LSTM, to optimize the superparameters and enhance the prediction accuracy.These deep learning (DL) models or ensemble models (EMs) are implemented for solar PV power generation forecasting to provide more stable power to the grid; (4) Data preprocessing or data feature analysis: Through data preprocessing and the clustering analysis of initial training sets to predict solar PV power generation, the accuracy of the prediction model is significantly improved.Secondly, the computing cost is reduced, the regression accuracy is significantly improved, and the model's own features are effectively found for predictions through the preprocessing and correlation analysis of input data.When compared with general data preprocessing methods, data preprocessing is further optimized, improving the applicability of FFT methods; (5) Improvement of inaccurate or missing data: In order to expand the ability of irradiance prediction methods to predict the power capacity of new solar power plants without data, we explored prediction methods that can handle repeated and frequent continuous multi-point data loss, for example, extracting data suitable for the target domain from different data domains or using data from other regions as a supplement when the training data for the target location are insufficient.Therefore, it is of practical significance to improve short-term solar PV predictions of inaccurate or missing data; (6) Integration with the power system: Accurate PV power generation forecasting is very important for the scheduling and regulation of power systems after the grid connection, and its results can be integrated into the entire energy management system or utilities to improve grid performance and achieve a higher level of renewable energy integration.Secondly, variation in power generation can have an impact on the voltage and frequency of the power system at any time, solving the problems of economic dispatch, grid integration, and the mismanagement of power management systems caused by the variability of solar energy.Furthermore, based on the basic viewpoint of large-scale or distributed solar PV systems, load forecasting, demand response applications, aggregate capacity prediction, and the dispatch of a large number of distributed solar PV systems can be obtained.When combined with pumped storage power stations, adjustable biomass power stations, or PV battery systems, they can stably transmit solar PV power generation and improve the flexibility of power dispatch.

Conclusions
This paper first presented the significance of using solar PV power for energy conservation and emission reduction issues, as well as the technical challenges faced when predicting solar PV power generation.The necessity of developing prediction systems for solar PV power generation and improving the model's accuracy was clarified.Some existing physical and statistical learning methods have deficiencies, such as high modeling costs and large input data requirements when performing predictions, while traditional machine learning methods have trouble processing missing data, are prone to overfitting, and ignore the correlations between attributes in the dataset.This paper further reported on many novel prediction models for PV power generation based on deep learning or hybrid models that integrate multiple meteorological factors such as temperature, relative humidity, rainfall (precipitation), global solar radiation, wind speed, etc.By analyzing the mean square error (MSE) value and the determination coefficient (R-Squared) value, we proved that the proposed method further improves prediction accuracy when compared to previous prediction methods.Secondly, this paper introduced the current situation of solar PV power generation forecasting from a global perspective.Most of these efforts cover the field of short-term PV power generation forecasting, which has grown significantly in the past few years.These advanced solar short-term PV power generation prediction models were classified and compared in terms of temporal resolution, the parameters used, accuracy, and research limitations.In addition, this paper reviewed the latest progress in short-term solar PV power generation based on artificial intelligence methods, emphasizing their contributions to model development, their advantages and disadvantages, and the areas of future study and development.The contributions of this review work are as follows: (1) The most advanced algorithms for short-term solar PV power generation forecasting were evaluated; (2) The accuracy, advantages, and disadvantages of various new AI hybrid models were evaluated; (3) Existing challenges and issues, such as short-term solar PV power generation data diversity, algorithm structure, hyperparametric adjustment, optimization integration, and AI hybrid issues, were explored; (4) The development and future possibilities of efficient short-term solar PV power generation prediction methods based on artificial intelligence were proposed.Future research directions and challenges for existing short-term solar PV power generation prediction methods were provided; (5) The impacts of meteorological information and cloud image information in terms of improving data preprocessing or data feature selection and analysis and data inaccuracies or loss were explored.The distribution of the database input sources, forecasting methods, and predictive error metrics was analyzed, and the effective use of machine learning or optimization algorithms and deep learning models to improve the accuracy of existing models was discussed to increase forecasting accuracy; (6) It was shown that improving the prediction accuracy of short-term solar PV power generation is beneficial to the optimal scheduling of microgrids and integration with the optimization of power systems.

Figure 1 .
Figure 1.Short-term solar PV power generation prediction model.

Figure 2 .
Figure 2. Ratio of input data of the reviewed works.

Figure 2 .
Figure 2. Ratio of input data of the reviewed works.

Figure 3 .
Figure 3. Distribution of forecasting methods used in the reviewed works.

Figure 4 .
Figure 4. Proportion of statistical metrics for the reviewed works.

Figure 3 .
Figure 3. Distribution of forecasting methods used in the reviewed works.

32 Figure 3 .
Figure 3. Distribution of forecasting methods used in the reviewed works.

Figure 4 .
Figure 4. Proportion of statistical metrics for the reviewed works.

Figure 4 .
Figure 4. Proportion of statistical metrics for the reviewed works.
and robustness of the TG-A-CNN-LSTM model by testing the performance of sparse data prediction models.

Figure 5 .
Figure 5. Classification of the novel short-term solar PV power forecasting techniques.

Figure 5 .
Figure 5. Classification of the novel short-term solar PV power forecasting techniques.

Figure 8 .
Figure 8. Schematic diagram of the selection of reference data for the similar day prediction method.

Figure 8 .
Figure 8. Schematic diagram of the selection of reference data for the similar day prediction method.

32 Figure 9 .
Figure 9. Configuration diagram of the IOT technology prediction system.

Figure 9 .
Figure 9. Configuration diagram of the IOT technology prediction system.

Table 1 .
The model type, accuracy, and parameters for the reviewed works.

Table 2 .
Main contributions, advantages, and disadvantages of reviewed works in terms of solar PV power forecasting.