Photovoltaic Prediction Method and System with Interpretable MLREP Multilevel Regression Integration

Abstract

The present invention discloses a photovoltaic prediction method and system for interpretable MLREP multi-level regression integration. S1. Collect multi-source data sets related to photovoltaic power generation; S2. Form a preprocessed multi-source data set; S3. Form each stratified data set; S4. Use a graph attention mechanism to model the spatio-temporal features in each of the stratified data sets and generate corresponding spatio-temporal feature weights; S5. Based on the spatio-temporal feature representation, construct a multi-level regression integration framework to generate sub-prediction results; S6. Integrate the sub-prediction results generated by the multiple regression sub-models using a weighted fusion method to generate the final predicted photovoltaic power generation result; S7. Perform interpretability analysis on the final predicted photovoltaic power generation result in combination with the aforementioned spatio-temporal feature weights and the prediction results of each regression sub-model, and generate an interpretability analysis report. The present invention can combine sensitivity index calculation to evaluate the source of prediction errors and give optimization suggestions.

Description

Technical Field

[0001] The present invention relates to the field of photovoltaic technology, and in particular to a photovoltaic prediction method and system that can interpret MLREP multi-level regression integration. Background Art

[0002] With the transformation of the global energy structure towards clean and renewable energy, as an important part of new energy, the installed capacity and application scope of photovoltaic power generation are constantly expanding. However, due to the influence of various factors such as meteorological environment, equipment status, and historical power generation laws on photovoltaic power generation, its power generation has obvious volatility and uncertainty.

[0003] Currently, photovoltaic power generation prediction mainly relies on physical models and data-driven models. The physical model method constructs a prediction model based on the electrical characteristics of photovoltaic cells, the solar radiation transmission model, and meteorological factors. In theory, it can better explain the operation mechanism of photovoltaic systems, but it has high requirements for the accuracy of input variables, is difficult to cope with complex and changing actual environments, resulting in large prediction errors. In addition, physical models often require a large number of parameter calibrations, with high data collection and processing costs, and are not easy to promote and apply.

[0004] In recent years, data-driven machine learning methods have been widely used in photovoltaic power generation prediction. Prediction models are trained based on historical data, including traditional statistical regression methods and deep learning methods. Although data-driven methods can adapt to different environmental conditions and improve prediction accuracy, there are still many problems in the existing technologies. First, a single regression model is difficult to simultaneously capture the long-term trend, short-term fluctuations, and instantaneous change characteristics of photovoltaic power generation, resulting in unstable prediction results. Second, the existing methods have limitations in spatio-temporal modeling and are difficult to fully utilize the complex relationships among meteorological data, equipment status data, and power generation data, resulting in limited prediction accuracy. In addition, most traditional data-driven methods lack interpretability and are difficult to effectively analyze the contributions of different influencing factors to prediction results, restricting their applications in actual photovoltaic scheduling and management.

[0005] Therefore, there is an urgent need for a new photovoltaic power generation prediction method that can comprehensively consider multi-level features, spatio-temporal correlation, and has high prediction accuracy and interpretability to improve the reliability and practical value of photovoltaic power generation prediction. Summary of the Invention

[0006] An object of the present invention is to propose a photovoltaic prediction method and system that can interpret MLREP multi-level regression integration. The present invention can calculate sensitivity indicators, evaluate the sources of prediction errors, and give optimization suggestions.

[0007] A photovoltaic prediction method that can interpret MLREP multi-level regression integration according to an embodiment of the present invention includes the following steps:

[0008] S1. Collect multi-source datasets related to photovoltaic power generation;

[0009] S2. Preprocess the collected multi-source datasets to form preprocessed multi-source datasets;

[0010] S3. Extract hierarchical features from the preprocessed multi-source datasets according to the operating characteristics of the photovoltaic power generation system to form each hierarchical dataset;

[0011] S4. Use the graph attention mechanism to model the spatio-temporal features in each hierarchical dataset, extract the spatio-temporal feature representations of each hierarchical dataset, and generate corresponding spatio-temporal feature weights;

[0012] S5. Construct a multi-level regression integration framework based on the spatio-temporal feature representations. The multi-level regression integration framework includes multiple regression sub-models, and each regression sub-model is trained using the spatio-temporal feature representations of the corresponding hierarchical dataset to generate sub-prediction results;

[0013] S6. Integrate the sub-prediction results generated by the multiple regression sub-models using a weighted fusion method to generate the final predicted photovoltaic power generation result;

[0014] S7. Perform interpretability analysis on the final predicted photovoltaic power generation result in combination with the aforementioned spatio-temporal feature weights and the prediction results of each regression sub-model, and generate an interpretability analysis report.

[0015] Optionally, S1 includes collecting multi-source datasets related to photovoltaic power generation. The multi-source datasets include meteorological data, historical power generation data, equipment status data, and environmental impact data. Set the collection time window to construct a multi-source dataset for photovoltaic power generation :

[0016] ;

[0017] wherein, represents the th photovoltaic power generation data record, is the number of photovoltaic power generation data collected within the time window , is the collection timestamp, is the numerical feature vector related to photovoltaic power generation, is the device identification of the source of the photovoltaic power generation data, is the category label of the photovoltaic power generation data.

[0018] Optionally, S2 includes the following steps:

[0019] S21. Based on the multi-source dataset for photovoltaic power generation Perform data screening, remove samples with missing values exceeding the set threshold in the data, and remove duplicate data to form a screened photovoltaic power generation dataset;

[0020] S22. Standardize the data format of the screened photovoltaic power generation dataset and set the normalization interval , and perform min-max normalization on the numerical feature vectors related to photovoltaic power generation to obtain a standardized photovoltaic power generation dataset;

[0021] S23. According to the time characteristics of the photovoltaic power generation data, align the time of the standardized photovoltaic power generation dataset, and use the sliding time window method to adjust the data timestamps so that the photovoltaic power generation data is aligned according to the time dimension to form a time-aligned photovoltaic power generation dataset;

[0022] S24. Fill in the missing values in the time-aligned photovoltaic power generation dataset based on the operating status of the photovoltaic power generation equipment, environmental impact factors, and meteorological characteristics, and complete some missing feature values to form a complete photovoltaic power generation dataset .

[0023] Optionally, the above S3 includes the following steps:

[0024] S31. According to the operating characteristics of the photovoltaic power generation system, perform feature extraction on the complete photovoltaic power generation dataset , and divide the features into long-term trend features, short-term fluctuation features, and instantaneous change features according to the time distribution and physical meaning of the data to form a hierarchical feature dataset of photovoltaic power generation , and the hierarchical feature dataset of photovoltaic power generation includes a long-term trend feature dataset , a short-term fluctuation feature dataset and an instantaneous change feature dataset , the long-term trend feature dataset contains the power generation change trend information on a long time scale, the short-term fluctuation feature dataset contains the power generation fluctuation characteristics within a day and short cycles, and the instantaneous change feature dataset contains the immediate impact of environmental conditions and equipment status changes on power generation;

[0025] S32. Perform feature calculation on the long-term trend feature dataset , and use historical time series data to extract the long-term trend feature vector of the photovoltaic power generation system ;

[0026] S33. Perform feature calculation on the short-term fluctuation feature dataset , analyze the photovoltaic power generation fluctuation pattern within a short time period, and extract the short-term fluctuation feature vector ;

[0027] S34. Calculate the features of the instantaneous change feature dataset to analyze the instantaneous impact of environmental variables and device status on photovoltaic power generation, and extract the instantaneous change feature vector ;

[0028] S35. According to the importance of feature data at different levels, fuse the long-term trend feature vector , short-term fluctuation feature vector and instantaneous change feature vector to form a complete photovoltaic power generation feature dataset :

[0029] .

[0030] Optionally, the S4 includes the following steps:

[0031] S41. According to the photovoltaic power generation hierarchical feature dataset , construct a multi-view spatio-temporal feature modeling graph for photovoltaic power generation , including the long-term trend view , short-term fluctuation view and instantaneous change view In each view, is the node set, and the node represents the corresponding feature vector , where , is the edge set, is the adjacency matrix;

[0032] S42. For each view , according to the spatio-temporal correlation of photovoltaic power generation data, use the time decay similarity function to calculate the adjacency matrix :

[0033] ;

[0034] Among them, represents the feature vector of node in view , represents the feature vector of node in view , is the scale parameter of view , represents the time difference between the data collection times of node and node , is the time decay factor of view ;

[0035] S43. For each view , the spatio-temporal associations between nodes are modeled using a multi-head graph attention mechanism, and the number of multi-heads is set , for each attention head , , the attention weights are calculated :

[0036] ;

[0037] Among them, is the feature transformation matrix of the th attention head in view is the corresponding attention coefficient vector, represents the feature concatenation operation, is the neighbor set of node in view ;

[0038] S44. For each node in each view , the multi-head attention results are fused to generate an updated spatio-temporal feature representation , and the updated spatio-temporal feature representation reflects the spatio-temporal feature dynamics of photovoltaic power generation under the corresponding view:

[0039] ;

[0040] Among them, is the non-linear activation function;

[0041] S45. The meteorological feedback data at the current moment is collected , and the average feature vector of view is calculated based on the updated feature representations of all nodes in each view:

[0042] ;

[0043] Among them, represents the number of nodes in view ;

[0044] S46. Based on the average feature vector and the meteorological feedback data , the adaptive fusion weights of each view are calculated using the mapping function :

[0045] ;

[0046] Generate the final set of spatio-temporal feature representations for photovoltaic power generation based on the updated feature representations and adaptive fusion weights of each view :

[0047] ;

[0048] Among them, represents the updated feature representations of all nodes in view ;

[0049] S48. Calculate the photovoltaic power generation feature weight vector based on the final set of spatio-temporal feature representations for photovoltaic power generation :

[0050] ;

[0051] Among them, is the feature fusion parameter matrix.

[0052] Optionally, the above S5 includes the following steps:

[0053] S51. Construct a multi-level regression integration framework based on the set of spatio-temporal feature representations for photovoltaic power generation , and define the set of regression sub-models , where , and are the long-term trend regression sub-model, short-term fluctuation regression sub-model, and instantaneous change regression sub-model respectively. Each regression sub-model corresponds to the long-term trend feature data set , short-term fluctuation feature data set and instantaneous change feature data set ;

[0054] S52. Use the supervised learning method to train the set of regression sub-models , set the target variable as the photovoltaic power generation at a future time . For each regression sub-model , optimize its parameter to minimize the prediction error. The loss function is defined as:

[0055] ;

[0056] Among them, is the number of training data samples of the regression sub-model , is the spatio-temporal feature representation of the sample in view , is the true value of the corresponding photovoltaic power generation, is the trainable parameter of the regression sub-model;

[0057] S53. Calculate the prediction confidence weights of each regression sub-model in the regression sub-model set based on the historical photovoltaic power generation data and the model training error : :

[0058] ;

[0059] wherein, reflects the prediction accuracy of the regression sub-model on the historical training data;

[0060] S54. Adopt a weighted regression fusion strategy, combine the prediction results of each regression sub-model, and generate a multi-level regression integrated prediction result :

[0061] ;

[0062] wherein, represents the predicted value of the future photovoltaic power generation at time , is the spatio-temporal feature representation of time in view , is the adaptive fusion weight of the regression sub-model.

[0063] Optionally, the S6 includes the following steps:

[0064] S61. Obtain the sub-prediction results of the regression sub-model set according to the multi-level regression integration framework, and define the sub-prediction result set :

[0065] ;

[0066] wherein, is the sub-prediction result generated by the long-term trend regression sub-model , is the sub-prediction result generated by the short-term fluctuation regression sub-model , is the sub-prediction result generated by the instantaneous change regression sub-model ;

[0067] S62. Calculate the final fusion weight according to the spatio-temporal feature weight vector and the regression sub-model prediction confidence weight ;

[0068] S63. Integrate all sub-prediction results by using a weighted fusion method to generate the final photovoltaic power generation prediction result :

[0069] 。

[0070] Optionally, the S7 includes the following steps:

[0071] S71. Based on the final photovoltaic power generation prediction result , combined with the spatio-temporal feature weight vector and the regression sub-model prediction confidence weight , construct an interpretable analysis framework for photovoltaic prediction to analyze the influence degree of each feature variable on the prediction result;

[0072] S72. Extract long-term trend features, short-term fluctuation features, and instantaneous change features from the photovoltaic power generation hierarchical feature dataset, calculate the contribution degree of each feature variable in the final prediction result, define the feature importance scoring rule, and assign weights according to the feature contribution degree, where:

[0073] Long-term trend features: The feature contribution degree is positively correlated with the fitting degree of the historical trend. The variables with higher feature contribution degrees include historical average power generation, standard deviation of power generation, and long-term trend fitting parameters;

[0074] Short-term fluctuation features: The feature contribution degree is positively correlated with the matching degree of short-term fluctuation amplitude and periodic change. The variables with higher feature contribution degrees include power change rate, main frequency component, and autocorrelation function result;

[0075] Instantaneous change features: The feature contribution degree is positively correlated with the instantaneous influence degree of environmental variables. The variables with higher feature contribution degrees include current irradiance, component temperature, and conversion efficiency;

[0076] S73. Trace the generation path of the prediction result according to the regression sub-model set, analyze the weights of each regression sub-model in the prediction process, and define the decision path tracing rule:

[0077] When the confidence weight of the long-term trend regression sub-model is the largest, the long-term trend regression sub-model has the highest contribution degree, indicating that the current prediction mainly depends on long-term trend information;

[0078] When the confidence weight of the short-term fluctuation regression sub-model is the largest, the short-term fluctuation regression sub-model has the highest contribution degree, indicating that the current prediction is significantly affected by short-term fluctuations;

[0079] When the confidence weight of the instantaneous change regression sub-model is the largest, the instantaneous change regression sub-model has the highest contribution degree, indicating that environmental variables have a great influence on the current prediction;

[0080] S74. Calculate the sensitivity index of photovoltaic prediction, analyze the response of the prediction result to the change of input variables, and define the sensitivity calculation rule:

[0081] If , then the prediction result is highly sensitive to the change of irradiance, and the meteorological fluctuation needs to be focused on. Among them, represents the partial derivative of the final photovoltaic power generation prediction result with respect to irradiance, that is, the change rate of the photovoltaic power generation prediction value when the irradiance changes, is the irradiance sensitivity threshold;

[0082] If , then the prediction result is sensitive to the change of component temperature, and the temperature compensation strategy needs to be evaluated. Among them, represents the partial derivative of the final photovoltaic power generation prediction result with respect to the surface temperature of the photovoltaic module, that is, the change rate of the photovoltaic power generation prediction value when the component temperature changes, is the component temperature sensitivity threshold;

[0083] If , then the prediction result is greatly affected by the photovoltaic conversion efficiency, and the component aging and maintenance conditions need to be concerned. Among them, represents the partial derivative of the final photovoltaic power generation prediction result with respect to the conversion efficiency of the photovoltaic module, that is, the change rate of the photovoltaic power generation prediction value when the conversion efficiency of the photovoltaic module changes, is the photovoltaic conversion efficiency sensitivity threshold;

[0084] S75. Generate an interpretability analysis report based on the feature importance evaluation, decision path tracking, and sensitivity index calculation results. The report content includes:

[0085] The final photovoltaic power generation prediction result and the prediction confidence interval;

[0086] The contribution degree ranking of each feature variable to the prediction result;

[0087] The regression sub-model and weight allocation on which the prediction result depends;

[0088] The sensitivity analysis of the prediction result to the change of the main input variables;

[0089] S76. Dynamically adjust the weights of the regression sub-model and feature weights .

[0090] An interpretable photovoltaic prediction system based on MLREP multi-level regression integration includes a computer program. When the computer program is executed by a processor, it realizes the steps of an interpretable photovoltaic prediction method based on MLREP multi-level regression integration.

[0091] The beneficial effects of the present invention are as follows:

[0092] (1) According to the operating characteristics of the photovoltaic power generation system, the input data is divided into three levels: long-term trend, short-term fluctuation, and instantaneous change, and the corresponding feature vectors are extracted respectively to more accurately describe the time-series dynamic characteristics of photovoltaic power generation. Through the hierarchical feature division, the long-term trend feature can reflect the overall power generation trend, the short-term fluctuation feature can capture the periodic changes, and the instantaneous change feature can cope with the fluctuations of the environment and equipment status within a short time.

[0093] (2) By constructing a multi-view spatio-temporal feature modeling graph, the long-term trend, short-term fluctuation, and instantaneous change of photovoltaic power generation are regarded as different views respectively, and the spatio-temporal features under each view are learned by using the graph attention mechanism. Specifically, within different time windows, the association weights of each data node are calculated by using multi-head graph attention to fully capture the spatio-temporal correlation of photovoltaic power generation data, and the weights of the adjacency matrix are dynamically adjusted by using the time decay similarity function to enhance the model's adaptive ability to time dynamics.

[0094] (3) An interpretability analysis method is introduced in the photovoltaic prediction process to make the prediction results transparent. Specifically, the feature importance evaluation method is used to analyze the contribution degree of different features in the final prediction results; through the decision path tracking technology, the main influencing factors adopted by different regression sub-models in the prediction process are revealed. In addition, combined with the calculation of sensitivity indicators, the sources of prediction errors are evaluated and optimization suggestions are given. Description of the Drawings

[0095] The drawings are used to provide a further understanding of the present invention, and constitute a part of the specification. Together with the embodiments of the present invention, they are used to explain the present invention and do not constitute a limitation to the present invention. In the drawings:

[0096] Figure 1 is a flowchart of a photovoltaic prediction method and system based on interpretable MLREP multi-level regression integration proposed by the present invention. Detailed Embodiments

[0097] Now, the present invention will be further described in detail with reference to the drawings. These drawings are all simplified schematic diagrams, only showing the basic structure of the present invention in a schematic way, so they only show the components related to the present invention.

[0098] Refer to Figure 1 , a photovoltaic prediction method based on interpretable MLREP multi-level regression integration, includes the following steps:

[0099] S1. Collect multi-source data sets related to photovoltaic power generation;

[0100] S2. Preprocess the collected multi-source data set to form a preprocessed multi-source data set;

[0101] S3. Extract hierarchical features from the preprocessed multi-source data set according to the operating characteristics of the photovoltaic power generation system to form each hierarchical data set;

[0102] S4. Use the graph attention mechanism to model the spatio-temporal features in each hierarchical data set, extract the spatio-temporal feature representations of each hierarchical data set, and generate corresponding spatio-temporal feature weights;

[0103] S5. Construct a multi-level regression integration framework based on the spatio-temporal feature representation. The multi-level regression integration framework includes multiple regression sub-models. Each regression sub-model is trained using the spatio-temporal feature representation of the corresponding hierarchical data set to generate sub-prediction results;

[0104] S6. Integrate the sub-prediction results generated by multiple regression sub-models using a weighted fusion method to generate the final photovoltaic power generation prediction result;

[0105] S7. Perform interpretability analysis on the final photovoltaic power generation prediction result in combination with the aforementioned spatio-temporal feature weights and the prediction results of each regression sub-model, and generate an interpretability analysis report.

[0106] In this embodiment, S1 includes collecting a multi-source data set related to photovoltaic power generation. The multi-source data set includes meteorological data, historical power generation data, equipment status data, and environmental impact data. Set a collection time window to construct a multi-source data set for photovoltaic power generation :

[0107] ;

[0108] Among them, represents the th photovoltaic power generation data record, is the number of photovoltaic power generation data collected within the time window , is the collection timestamp, is the numerical feature vector related to photovoltaic power generation, is the device identification of the source of photovoltaic power generation data, is the category label of photovoltaic power generation data.

[0109] In this embodiment, S2 includes the following steps:

[0110] S21. Perform data screening based on the multi-source data set for photovoltaic power generation , remove samples with missing values exceeding the set threshold in the data, and remove duplicate data to form a screened data set for photovoltaic power generation;

[0111] S22. Standardize the data format of the filtered photovoltaic power generation dataset and set the normalization interval , and perform min-max normalization on the numerical feature vectors related to photovoltaic power generation to obtain the standardized photovoltaic power generation dataset;

[0112] S23. According to the time characteristics of the photovoltaic power generation data, align the time of the standardized photovoltaic power generation dataset, and use the sliding time window method to adjust the data timestamps so that the photovoltaic power generation data is aligned according to the time dimension, forming the time-aligned photovoltaic power generation dataset;

[0113] S24. According to the operating status of the photovoltaic power generation equipment, environmental impact factors and meteorological characteristics, fill in the missing values in the time-aligned photovoltaic power generation dataset, and complete the partially missing feature values to form a complete photovoltaic power generation dataset .

[0114] In this embodiment, S3 includes the following steps:

[0115] S31. According to the operating characteristics of the photovoltaic power generation system, extract features from the complete photovoltaic power generation dataset . The features are divided into long-term trend features, short-term fluctuation features and instantaneous change features according to the time distribution and physical meaning of the data, forming a hierarchical feature dataset of photovoltaic power generation . The hierarchical feature dataset of photovoltaic power generation includes a long-term trend feature dataset , a short-term fluctuation feature dataset and an instantaneous change feature dataset . The long-term trend feature dataset contains the power generation change trend information on a long time scale, the short-term fluctuation feature dataset contains the power generation fluctuation characteristics within a day and short cycles, and the instantaneous change feature dataset contains the immediate impact of environmental conditions and equipment state changes on power generation;

[0116] S32. Calculate the features of the long-term trend feature dataset , and extract the long-term trend feature vector of the photovoltaic power generation system using historical time series data ;

[0117] S33. Calculate the features of the short-term fluctuation feature dataset , analyze the photovoltaic power generation fluctuation pattern within a short time period, and extract the short-term fluctuation feature vector ;

[0118] S34. Calculate the features of the instantaneous change feature dataset , analyze the instantaneous impact of environmental variables and equipment state on photovoltaic power generation, and extract the instantaneous change feature vector ;

[0119] S35. Combine the long-term trend feature vector , short-term fluctuation feature vector and instantaneous change feature vector to form a complete photovoltaic power generation feature data set :

[0120] .

[0121] In this embodiment, S4 includes the following steps:

[0122] S41. Based on the photovoltaic power generation hierarchical feature data set , construct a multi-view spatio-temporal feature modeling graph for photovoltaic power generation , including the long-term trend view , short-term fluctuation view and instantaneous change view In each view, is the node set, and the node represents the corresponding feature vector , where , is the edge set, is the adjacency matrix;

[0123] S42. For each view , calculate the adjacency matrix using a time-decaying similarity function based on the spatio-temporal correlation of photovoltaic power generation data:

[0124] ;

[0125] where, represents the feature vector of node in view , represents the feature vector of node in view , is the scale parameter of view , represents the difference in data collection time between node and node , is the time decay factor of view ;

[0126] S43. For each view , use the multi-head graph attention mechanism to model the spatio-temporal correlation between nodes, set the number of multi-heads , and for each attention head , ​​​​, calculate the attention weights :

[0127] ;

[0128] Among them, is the feature transformation matrix of the th attention head in the view, is the corresponding attention coefficient vector, represents the feature concatenation operation, is the node in the view neighbor set;

[0129] S44. For each node in each view , fuse the multi-head attention results to generate the updated spatio-temporal feature representation , and the updated spatio-temporal feature representation reflects the spatio-temporal feature dynamics of photovoltaic power generation under the corresponding view:

[0130] ;

[0131] Among them, is the non-linear activation function;

[0132] S45. Collect the meteorological feedback data at the current moment , and calculate the average feature vector of the view based on the updated feature representations of all nodes in each view:

[0133] ;

[0134] Among them, represents the number of nodes in the view ;

[0135] S46. Based on the average feature vector and the meteorological feedback data , use the mapping function to calculate the adaptive fusion weights of each view:

[0136] ;

[0137] S47. Based on the updated feature representations and adaptive fusion weights of each view, generate the final set of spatio-temporal feature representations of photovoltaic power generation :

[0138] ;

[0139] Among them, represents the updated feature representation of all nodes in the view ;

[0140] S48. According to the final set of spatio-temporal feature representations of photovoltaic power generation , calculate the photovoltaic power generation feature weight vector :

[0141] ;

[0142] Among them, is the feature fusion parameter matrix.

[0143] In this embodiment, S5 includes the following steps:

[0144] S51. According to the set of spatio-temporal feature representations of photovoltaic power generation construct a multi-level regression integration framework and define the set of regression sub-models , among which, , and are the long-term trend regression sub-model, short-term fluctuation regression sub-model and instantaneous change regression sub-model respectively. Each regression sub-model corresponds to the long-term trend feature data set , short-term fluctuation feature data set and instantaneous change feature data set ;

[0145] S52. Use the supervised learning method to train the set of regression sub-models , set the target variable as the photovoltaic power generation at a future moment , for each regression sub-model , optimize its parameters to minimize the prediction error. The loss function is defined as:

[0146] ;

[0147] Among them, is the number of training data samples of the regression sub-model , is the sample in the view spatio-temporal feature representation, is the true value of the corresponding photovoltaic power generation, is the trainable parameter of the regression sub-model;

[0148] S53. Calculate the prediction confidence weights of each regression sub-model in the set of regression sub-models according to the historical data of photovoltaic power generation and the model training error :

[0149] ;

[0150] Among them, reflects the prediction accuracy of the regression sub-model on historical training data;

[0151] S54. Adopt a weighted regression fusion strategy, combine the prediction results of each regression sub-model, and generate a multi-level regression integrated prediction result :

[0152] ;

[0153] Among them, represents the predicted value of the future photovoltaic power generation at time , is the spatio-temporal feature representation of time in view , is the adaptive fusion weight of the regression sub-model.

[0154] In this embodiment, S6 includes the following steps:

[0155] S61. According to the multi-level regression integration framework, obtain the sub-prediction results of the regression sub-model set, and define the sub-prediction result set :

[0156] ;

[0157] Among them, is the sub-prediction result generated by the long-term trend regression sub-model , is the sub-prediction result generated by the short-term fluctuation regression sub-model , is the sub-prediction result generated by the instantaneous change regression sub-model ;

[0158] S62. Calculate the final fusion weight according to the spatio-temporal feature weight vector and the regression sub-model prediction confidence weight ;

[0159] S63. Integrate all sub-prediction results using a weighted fusion method to generate the final photovoltaic power generation prediction result :

[0160] .

[0161] In this embodiment, S7 includes the following steps:

[0162] S71. Based on the final predicted photovoltaic power generation results , combined with the spatio-temporal feature weight vector and the confidence weight of the regression sub-model prediction , construct an interpretability analysis framework for photovoltaic prediction to analyze the influence degree of each feature variable on the prediction result;

[0163] S72. Extract long-term trend features, short-term fluctuation features and instantaneous change features from the photovoltaic power generation hierarchical feature dataset, calculate the contribution degree of each feature variable in the final prediction result, define the feature importance scoring rule, and allocate weights according to the feature contribution degree, where:

[0164] Long-term trend features: The feature contribution degree is positively correlated with the fitting degree of the historical trend. Variables with higher feature contribution degrees include historical average power generation, standard deviation of power generation, and long-term trend fitting parameters;

[0165] Short-term fluctuation features: The feature contribution degree is positively correlated with the matching degree of the short-term fluctuation amplitude and periodic change. Variables with higher feature contribution degrees include power change rate, main frequency component, and autocorrelation function result;

[0166] Instantaneous change features: The feature contribution degree is positively correlated with the instantaneous influence degree of environmental variables. Variables with higher feature contribution degrees include current irradiance, module temperature, and conversion efficiency;

[0167] S73. Trace the generation path of the prediction result according to the regression sub-model set, analyze the weight of each regression sub-model in the prediction process, and define the decision path tracing rule:

[0168] When the confidence weight of the long-term trend regression sub-model is the largest, the long-term trend regression sub-model has the highest contribution degree, indicating that the current prediction mainly depends on long-term trend information;

[0169] When the confidence weight of the short-term fluctuation regression sub-model is the largest, the short-term fluctuation regression sub-model has the highest contribution degree, indicating that the current prediction is significantly affected by short-term fluctuations;

[0170] When the confidence weight of the instantaneous change regression sub-model is the largest, the instantaneous change regression sub-model has the highest contribution degree, indicating that environmental variables have a great impact on the current prediction;

[0171] S74. Calculate the photovoltaic prediction sensitivity index, analyze the response of the prediction result to the change of input variables, and define the sensitivity calculation rule:

[0172] If , the prediction result is highly sensitive to irradiance changes, and meteorological fluctuations need to be focused on. Among them, represents the partial derivative of the final photovoltaic power generation prediction result with respect to irradiance, that is, the change rate of the photovoltaic power generation prediction value when the irradiance changes. is the irradiance sensitivity threshold;

[0173] If , the prediction result is sensitive to the change of component temperature, and the temperature compensation strategy needs to be evaluated. Among them, represents the partial derivative of the final photovoltaic power generation prediction result with respect to the surface temperature of the photovoltaic module, that is, the change rate of the photovoltaic power generation prediction value when the component temperature changes. is the component temperature sensitivity threshold;

[0174] If , the prediction result is greatly affected by the photovoltaic conversion efficiency, and the component aging and maintenance conditions need to be concerned. Among them, represents the partial derivative of the final photovoltaic power generation prediction result with respect to the conversion efficiency of the photovoltaic module, that is, the change rate of the photovoltaic power generation prediction value when the conversion efficiency of the photovoltaic module changes. is the photovoltaic conversion efficiency sensitivity threshold;

[0175] S75. According to the feature importance evaluation, decision path tracing and sensitivity index calculation results, generate an interpretability analysis report. The report content includes:

[0176] The final photovoltaic power generation prediction result and the prediction confidence interval;

[0177] The contribution degree ranking of each feature variable to the prediction result;

[0178] The regression sub-model and weight allocation on which the prediction result depends;

[0179] The sensitivity analysis of the prediction result to the change of the main input variables;

[0180] S76. Dynamically adjust the regression sub-model weights and feature weights .

[0181] An interpretable MLREP multi-level regression integrated photovoltaic prediction system includes a computer program. When the computer program is executed by a processor, it implements the steps of an interpretable MLREP multi-level regression integrated photovoltaic prediction method.

[0182] Example 1: On August 12, 2023, during the scheduling process, the operator found that the local weather forecast showed that there would be a short-term heavy rainfall in the area between 14:00 and 16:00 that day, accompanied by a rapid increase in cloud cover. However, the traditional photovoltaic prediction system of this power station failed to accurately reflect the risk of reduced photovoltaic power generation brought about by this weather change in the prediction report generated at 12:00. It still predicted that the photovoltaic power generation at 14:00 would be around 320 kW. Based on experience, the operators judged that this prediction was too optimistic and there was a large degree of uncertainty.

[0183] To further verify the accuracy of the prediction, the operator decided to introduce the method of the present invention, namely, the interpretable MLREP multi-level regression integrated photovoltaic prediction method, and test its prediction effect in actual application. They immediately invoked this method and input the meteorological data, historical power generation data, equipment status data, and environmental impact data during the period from 06:00 to 12:00 that day for calculation. Adopting a hierarchical feature extraction strategy, the impacts of long-term trends, short-term fluctuations, and instantaneous changes were calculated respectively, and the spatio-temporal characteristics of the photovoltaic power generation system were modeled using the graph attention mechanism, and finally the predicted value of photovoltaic power generation at 14:00 was generated.

[0184] At 12:15, this method obtained a new prediction result: the predicted photovoltaic power generation at 14:00 was about 245.3 kW, which was nearly 23.3% lower than the 320 kW predicted by the traditional method. This gap alerted the operator. They immediately decided to use the new prediction result to optimize the power grid scheduling and notified the dispatching center to adjust the operation plan of the standby thermal power units to prevent grid load imbalance caused by a sudden drop in photovoltaic power generation.

[0185] At 14:00, the actual monitoring data showed that the actual power generation of this photovoltaic power station was 241.6 kW. The prediction error of the method of the present invention was only 1.5%, while the error of the traditional method was as high as 32.8%. The operator compared and found that the method of the present invention could capture the short-term power fluctuations brought about by weather changes more accurately and avoid misjudgments caused by the lack of short-term meteorological change characteristics in the traditional method.

[0186] During the same period, the dispatching center received feedback from another photovoltaic power station. This power station experienced a sudden sandstorm on July 22. The prediction report generated by the traditional prediction system of this site at 07:00 showed that the photovoltaic power generation at 11:00 that day was about 410 kW. However, the site monitoring equipment found that the air quality index (AQI) increased sharply after 08:30, soaring from 55 (good) to 185 (moderate pollution) rapidly, and the visibility dropped to 1.2 km.

[0187] The operation team of this site decided to use the method of the present invention for correction prediction. At 08:45, the operation team input the meteorological data between 04:00 and 08:30 of that day into the method of the present invention, including wind speed, air particulate matter concentration, cloud thickness and temperature data. This method quickly calculated the impact of environmental changes on photovoltaic power generation, especially the impact of AQI changes on the light reception of the photovoltaic power station, and combined with historical data under similar meteorological conditions in the past to obtain a corrected prediction value of 358.7 kW at 11:00.

[0188] When the actual power generation data at 11:00 was recorded, it was found that the actual power generation was 352.9 kW. The prediction error of the method of the present invention was only 1.6%, while the prediction error of the traditional method reached 14.5%. The management personnel in the dispatching center realized that the traditional method had limited prediction ability in the face of sudden environmental changes, and the method of the present invention could more accurately predict the change trend of photovoltaic power generation through hierarchical feature extraction and spatio-temporal modeling, providing a more reliable tool for the intelligent management of future photovoltaic power stations.

[0189] In another case, on May 15, 2023, during routine maintenance, the operation team of a photovoltaic park found that a group of inverters (equipment ID: INV-20230515-008) had output fluctuations and unstable power since 10:30, and the feedback data from the photovoltaic module cleanliness detection sensor (equipment ID: SENSOR-CLN-101) showed that the cleanliness coefficient of this photovoltaic module had dropped to 0.72, lower than the normal operation threshold (0.85).

[0190] To judge the impact of this factor on photovoltaic power generation, the operation personnel used the method of the present invention to recalculate the predicted value of photovoltaic power generation at 13:00 on that day at 10:45. The prediction result of the traditional method was 510.2 kW, while the method of the present invention, considering the impact of the cleanliness of the photovoltaic module, gave a new predicted value of 468.9 kW and clearly pointed out in the prediction report that:

[0191] A 15.3% reduction in cleanliness has an impact on power generation efficiency of approximately 7.2%; the impact of short-term meteorological effects (cloud changes) is approximately 2.3%; the impact of equipment status (inverter fluctuations) is approximately 1.6%;

[0192] Based on the prediction data provided by the method of the present invention, the operation team decided to arrange a cleaning operation in advance and dispatched maintenance personnel to clean this group of photovoltaic modules at 11:30. The monitoring data at 13:00 showed that after the cleaning was completed, the actual power generation of this group of photovoltaic modules recovered to 497.1 kW, and the error decreased from 10.5% of the traditional method to 1.3%, verifying the effectiveness of the method of the present invention.

[0193] Table 1 Data comparison between the method of the present invention and the traditional method

[0194]

[0195] As can be seen from the above cases, compared with the traditional method, the method of the present invention has stronger short-term fluctuation prediction ability and environmental adaptability. Especially when dealing with sudden weather events, air pollution impacts, and equipment status fluctuations, the prediction error is significantly reduced, and the prediction reliability is significantly improved, providing more accurate decision-making support for the intelligent management of photovoltaic power plants.

[0196] Experimental data show that when the influencing factors of sudden weather and equipment failures are relatively strong, the prediction error of the method of the present invention is reduced by 60.2% - 95.4% compared with the traditional method, providing a more reliable basis for optimizing photovoltaic power generation scheduling and having wide engineering application value.

[0197] Based on the operating characteristics of the photovoltaic power generation system, the present invention divides the input data into three levels: long-term trend, short-term fluctuation, and instantaneous change, and extracts the corresponding feature vectors respectively to more accurately describe the time-series dynamic characteristics of photovoltaic power generation. Through hierarchical feature division, the long-term trend feature can reflect the overall power generation trend, the short-term fluctuation feature can capture periodic changes, and the instantaneous change feature can cope with the fluctuations of the environment and equipment status within a short period of time.

[0198] The present invention constructs a multi-view spatio-temporal feature modeling graph, takes the long-term trend, short-term fluctuation, and instantaneous change of photovoltaic power generation as different views respectively, and uses the graph attention mechanism to learn the spatio-temporal features under each view. Specifically, within different time windows, the multi-head graph attention is used to calculate the correlation weights of each data node, fully capturing the spatio-temporal correlation of photovoltaic power generation data, and dynamically adjusting the weights of the adjacency matrix through a time decay similarity function to enhance the model's adaptive ability to time dynamics.

[0199] The present invention introduces an interpretability analysis method in the photovoltaic prediction process to make the prediction results transparent. Specifically, the feature importance evaluation method is used to analyze the contribution degree of different features to the final prediction result; through the decision path tracing technology, the main influencing factors adopted by different regression sub-models in the prediction process are revealed; in addition, combined with the calculation of sensitivity indicators, the source of the prediction error is evaluated, and optimization suggestions are given.

[0200] The above is only a preferred specific embodiment of the present invention, but the protection scope of the present invention is not limited thereto. Any person skilled in the art within the technical scope disclosed by the present invention, according to the technical solution and inventive concept of the present invention, makes equivalent substitutions or changes, and should be covered by the protection scope of the present invention.

Claims

1. A photovoltaic prediction method that can interpret MLREP multi-level regression integration, characterized in that: The steps include: S1. Collect multi-source data sets related to photovoltaic power generation; S2. preprocessing the collected multi-source data set to form a preprocessed multi-source data set; S3. extract hierarchical features from the preprocessed multi-source data set according to the operating characteristics of the photovoltaic power generation system to form hierarchical data sets; S4. Modeling the spatiotemporal features in each hierarchical data set using a graph attention mechanism, extracting the spatiotemporal feature representations of each hierarchical data set, and generating corresponding spatiotemporal feature weights; S5. construct a multi-level regression integration framework based on the spatiotemporal feature representation, wherein the multi-level regression integration framework includes multiple regression sub-models, each of which is trained using the spatiotemporal feature representation of the corresponding hierarchical data set to generate a sub-prediction result; S6. Integrate the sub-prediction results generated by the multiple regression sub-models using a weighted fusion method to generate a final photovoltaic power generation prediction result; S7. Perform interpretability analysis on the final photovoltaic power generation prediction result in combination with the aforementioned spatiotemporal feature weights and the prediction results of each regression sub-model, and generate an interpretability analysis report; The S5 comprises the following steps: S51. Representing the set based on the spatiotemporal characteristics of photovoltaic power generation Build a multi-level regression integration framework and define a set of regression sub-models ,in, , and They are long-term trend regression sub-model, short-term fluctuation regression sub-model and instantaneous change regression sub-model. Each regression sub-model corresponds to a long-term trend feature data set. , short-term volatility characteristics dataset and instantaneous change feature datasets ; S52. Use supervised learning methods to ensemble regression sub-models Perform training and set the target variable as the photovoltaic power generation at the future moment , for each regression submodel , optimize its parameters , so that it minimizes the prediction error, the loss function is defined as: ； in, For the regression submodel The number of training data samples, For sample In view The spatiotemporal features in is the corresponding real value of photovoltaic power generation, is the trainable parameter of the regression sub-model; S53. Calculate the regression sub-model set based on the historical data of photovoltaic power generation and the model training error The prediction confidence weights of each regression sub-model : ； in, Reflection regression submodel Prediction accuracy on historical training data; S54. Use weighted regression fusion strategy to combine the prediction results of each regression sub-model to generate multi-level regression integrated prediction results : ； in, Indicates the moment The forecast value of future photovoltaic power generation, For the moment In view The spatiotemporal features in is the adaptive fusion weight of the regression sub-model.

2. A photovoltaic prediction method based on interpretable MLREP multi-level regression integration according to claim 1, characterized in that: S1 includes collecting multi-source data sets related to photovoltaic power generation, the multi-source data sets include meteorological data, historical power generation data, equipment status data and environmental impact data, setting a collection time window, and constructing a photovoltaic power generation multi-source data set : ； in, Indicates Photovoltaic power generation data records, For the time window The number of photovoltaic power generation data collected in is the acquisition timestamp, is the numerical feature vector related to photovoltaic power generation, It is the source device identifier of photovoltaic power generation data. is the category label of photovoltaic power generation data.

3. A photovoltaic prediction method based on interpretable MLREP multi-level regression integration according to claim 1, characterized in that: The S2 comprises the following steps: S21. Based on the photovoltaic power generation multi-source dataset Perform data screening, remove samples with missing values ​​exceeding the set threshold, and remove duplicate data to form a screened photovoltaic power generation data set; S22. Standardize the data format of the screened photovoltaic power generation data set and set the normalization interval , the numerical characteristic vector related to photovoltaic power generation Perform minimum-maximum normalization to obtain a standardized photovoltaic power generation data set; S23. Based on the time characteristics of the photovoltaic power generation data, the standardized photovoltaic power generation data set is time-aligned, and the data timestamp is adjusted using a sliding time window method so that the photovoltaic power generation data is aligned according to the time dimension to form a time-aligned photovoltaic power generation data set; S24. Based on the operating status of photovoltaic power generation equipment, environmental influencing factors and meteorological characteristics, the missing values ​​of the photovoltaic power generation data set after time alignment are filled, and some missing feature values ​​are completed to form a complete photovoltaic power generation data set .

4. A photovoltaic prediction method based on interpretable MLREP multi-level regression integration according to claim 1, characterized in that: The S3 comprises the following steps: S31. Based on the operation characteristics of the photovoltaic power generation system, the complete photovoltaic power generation data set Perform feature extraction. Feature extraction divides the features into long-term trend features, short-term fluctuation features and instantaneous change features according to the time distribution and physical meaning of the data, forming a stratified feature data set for photovoltaic power generation. The photovoltaic power generation hierarchical characteristic data set includes a long-term trend characteristic data set , short-term volatility characteristics dataset and instantaneous change feature datasets The long-term trend characteristic data set contains the power generation change trend information within a long time scale, the short-term fluctuation characteristic data set contains the power generation fluctuation characteristics within a day and a short period, and the instantaneous change characteristic data set contains the immediate impact of environmental conditions and equipment status changes on power generation; S32. Long-term trend feature data set Perform feature calculations and use historical time series data to extract the long-term trend feature vectors of photovoltaic power generation systems ; S33. Short-term volatility characteristics dataset Perform feature calculations, analyze the photovoltaic power generation fluctuation pattern in a short period of time, and extract short-term fluctuation feature vectors ; S34. For instantaneous change feature data set Perform feature calculations, analyze the instantaneous impact of environmental variables and equipment status on photovoltaic power generation, and extract instantaneous change feature vectors ; S35. According to the importance of feature data at different levels, the long-term trend feature vector , short-term volatility characteristic vector and the instantaneous change eigenvector Fusion to form a complete photovoltaic power generation characteristic data set : 。 5. The photovoltaic prediction method of the interpretable MLREP multi-level regression integration according to claim 1 is characterized in that: The S4 comprises the following steps: S41. Photovoltaic power generation layered feature dataset , constructing a multi-view spatiotemporal feature modeling diagram for photovoltaic power generation , including long-term trend views , Short-term volatility view and transient change views In each view, is a node set, each node represents the corresponding feature vector ,in , is the edge set, is the adjacency matrix; S42. For each view According to the spatiotemporal correlation of photovoltaic power generation data, the adjacency matrix is ​​calculated using the time-attenuated similarity function : ； in, Representation View Midpoint The characteristic vector of Representation View Midpoint The characteristic vector of For View The scale parameter, Representation Node With Node The difference in data collection time, For View The time decay factor of S43. For each view , a multi-head graph attention mechanism is used to model the spatiotemporal correlation between nodes, and the number of multi-heads is set , for each attention head , , calculate the attention weight : ； in, For View Middle The feature transformation matrix of the attention head is is the corresponding attention coefficient vector, represents the feature concatenation operation, For Node In view The set of neighbors in ; S44. For each view Each node in , fusion of multi-head attention results to generate updated spatiotemporal feature representation , the updated spatiotemporal feature representation Reflects the spatiotemporal dynamic characteristics of photovoltaic power generation in the corresponding view: ； in, is a nonlinear activation function; S45. Collect current weather feedback data , and the calculation view is represented based on the updated features of all nodes in each view The average eigenvector of : ； in, Representation View The number of nodes in the S46. Based on the average eigenvector With weather feedback data , using the mapping function Calculate the adaptive fusion weights of each view : ； S47. Generate the final photovoltaic power generation spatiotemporal feature representation set based on the updated feature representation and adaptive fusion weight of each view : ； in, Representation View Updated feature representation of all nodes in ; S48. Based on the final photovoltaic power generation spatiotemporal characteristics, the set , calculate the photovoltaic power generation characteristic weight vector : ； in, is the feature fusion parameter matrix.

6. A photovoltaic prediction method based on interpretable MLREP multi-level regression integration according to claim 1, characterized in that: The S6 comprises the following steps: S61.According to the multi-level regression integration framework, obtain the sub-prediction results of the regression sub-model set and define the sub-prediction result set : ； in, The long-term trend regression submodel The generated sub-prediction results, Regression submodel for short-term volatility The generated sub-prediction results, The regression submodel for instantaneous changes Generated sub-prediction results; S62. Based on spatiotemporal feature weight vector and regression sub-model prediction confidence weights Calculate the final fusion weights ; S63. Use weighted fusion method to integrate all sub-forecast results and generate the final photovoltaic power generation forecast result : 。 7. A photovoltaic prediction method based on interpretable MLREP multi-level regression integration according to claim 1, characterized in that: The S7 comprises the following steps: S71. Based on the final photovoltaic power generation prediction results , combined with the spatiotemporal feature weight vector and regression sub-model prediction confidence weights , construct the photovoltaic prediction interpretability analysis framework and analyze the influence of each characteristic variable on the prediction results; S72. Extract long-term trend characteristics, short-term fluctuation characteristics and instantaneous change characteristics based on the photovoltaic power generation layered feature data set, calculate the contribution of each feature variable in the final prediction result, define the feature importance scoring rule, and assign weights based on the feature contribution, where: Long-term trend characteristics: The characteristic contribution is positively correlated with the fitting degree of the historical trend. Variables with higher characteristic contribution include historical average power generation, power generation standard deviation, and long-term trend fitting parameters; Short-term fluctuation characteristics: The characteristic contribution is positively correlated with the matching degree of short-term fluctuation amplitude and periodic changes. Variables with higher characteristic contribution include power change rate, main frequency component, and autocorrelation function results; Instantaneous change characteristics: The characteristic contribution is positively correlated with the instantaneous impact of environmental variables. Variables with higher characteristic contributions include current irradiance, component temperature, and conversion efficiency. S73. Track the generation path of the prediction results based on the regression sub-model set, analyze the weight of each regression sub-model in the prediction process, and define the decision path tracking rules: When the confidence weight of the long-term trend regression submodel When the maximum, the long-term trend regression submodel The contribution is the highest, indicating that the current forecast mainly relies on long-term trend information; When the confidence weight of the short-term volatility regression submodel When the maximum value is reached, the short-term volatility regression submodel The contribution is the highest, indicating that the current forecast is significantly affected by short-term fluctuations; When the confidence weight of the regression sub-model changes instantaneously Maximum, instantaneous change regression submodel The contribution is the highest, indicating that environmental variables have a great influence on the current prediction; S74. Calculate the sensitivity index of photovoltaic prediction, analyze the response of prediction results to changes in input variables, and define sensitivity calculation rules: like , the forecast results are highly sensitive to irradiance changes, and special attention should be paid to meteorological fluctuations, among which, It represents the partial derivative of the final photovoltaic power generation prediction result with respect to the irradiance, that is, the rate of change of the photovoltaic power generation prediction value when the irradiance changes. is the irradiance sensitivity threshold; like , the prediction results are sensitive to component temperature changes, and the temperature compensation strategy needs to be evaluated, where It represents the partial derivative of the final photovoltaic power generation prediction result with respect to the surface temperature of the photovoltaic module, that is, the rate of change of the photovoltaic power generation prediction value when the module temperature changes. is the component temperature sensitivity threshold; like , the prediction results have a greater impact on the photovoltaic conversion efficiency, and attention should be paid to the aging and maintenance of components. It represents the partial derivative of the final photovoltaic power generation prediction result with respect to the conversion efficiency of the photovoltaic module, that is, the rate of change of the photovoltaic power generation prediction value when the conversion efficiency of the photovoltaic module changes. is the sensitivity threshold of photovoltaic conversion efficiency; S75. Generate an interpretability analysis report based on the feature importance assessment, decision path tracing and sensitivity index calculation results. The report content includes: Final photovoltaic power generation forecast results and forecast confidence intervals; Ranking of the contribution of each characteristic variable to the prediction results; The regression sub-models and weight distribution that the prediction results depend on; Sensitivity analysis of forecast results to changes in key input variables; S76. Dynamically adjust the weight of the regression sub-model based on the interpretability analysis report and feature weights .

8. A photovoltaic prediction system capable of interpreting MLREP multi-level regression integration, comprising a computer program, characterized in that: When the computer program is executed by a processor, the steps of a photovoltaic prediction method capable of interpreting MLREP multi-level regression integration as described in any one of claims 1 to 7 are implemented.

Images