A photoelectric power prediction method based on deep learning fusion model

Abstract

The present invention relates to the field of photovoltaic power generation technology, and discloses a photovoltaic power prediction method based on a deep learning fusion model, including: performing variable correlation analysis on historical environmental data to determine a cluster variable set and a cluster result data set; inputting the cluster result data set into a first prediction model, and extracting photovoltaic data spatial feature data based on the cluster result data set using the first prediction model; inputting the photovoltaic data spatial feature data into a second prediction model to obtain photovoltaic data cross-time period correlation data; inputting photovoltaic data cross-time period correlation data into a third prediction model to generate a photovoltaic power prediction interval of a target photovoltaic power station. By implementing the present invention, a deep learning fusion model, combined with historical environment and power generation data, deeply analyzes variable correlation and cross-time period characteristics, accurately extracts photovoltaic data features, effectively predicts photovoltaic power intervals, significantly improves prediction accuracy, and helps photovoltaic power stations to operate and dispatch efficiently.

Description

Technical Field

[0001] The present invention relates to the technical field of photovoltaic power generation, and in particular to a photovoltaic power prediction method based on a deep learning fusion model. Background Art

[0002] Photovoltaic power, as the core indicator of solar power generation, directly reflects the power generation capacity of photovoltaic power generation systems under different environmental conditions. With the growing global demand for renewable energy, accurate prediction of photovoltaic power has become the key to improving the operating efficiency of power systems, optimizing energy allocation, and ensuring stable operation of power grids. However, the prediction of photovoltaic power is affected by many factors, including environmental parameters such as light intensity, temperature, humidity, wind speed, as well as the geographical location, equipment status, and operating history of photovoltaic power generation systems. There are often complex nonlinear relationships between these factors, making the prediction of photovoltaic power a challenging task.

[0003] The current photovoltaic power prediction methods mainly include statistical methods, physical model methods and machine learning methods. Statistical methods make predictions by constructing statistical relationships between historical data and photovoltaic power, but it is difficult to capture complex nonlinear relationships and dynamic changes in environmental factors. The physical model method establishes a prediction model based on physical laws and the operating mechanism of the photovoltaic power generation system. Although it has high accuracy, the model is highly complex and has strict requirements on input data. Machine learning methods, especially deep learning methods, have shown strong capabilities in processing nonlinear relationships, high-dimensional data and large-scale data sets, but they still face some challenges in practical applications. Due to the diversity of environmental factors and photovoltaic power generation system characteristics, a single prediction model is difficult to adapt to the prediction needs in different scenarios, resulting in weak generalization of prediction results in different scenarios. Summary of the invention

[0004] In view of this, the present invention provides a photovoltaic power prediction method based on a deep learning fusion model to solve the problem in the related technology that due to the diversity of environmental factors and photovoltaic power generation system characteristics, a single prediction model is difficult to adapt to the prediction needs in different scenarios, resulting in weak generalization ability of the prediction results in different scenarios.

[0005] The present invention provides a photovoltaic power prediction method based on a deep learning fusion model, the method comprising: acquiring historical environmental data and historical power generation data of a target photovoltaic power station, performing variable correlation analysis on the historical environmental data, and determining a cluster variable set; determining a cluster result data set based on the cluster variable set, the historical environmental data, and the historical power generation data; inputting the cluster result data set into a first prediction model, and extracting photovoltaic data spatial feature data based on the cluster result data set using the first prediction model; inputting the photovoltaic data spatial feature data into a second prediction model, and parsing the photovoltaic data spatial feature data using the second prediction model to obtain photovoltaic data cross-time period correlation data; inputting the photovoltaic data cross-time period correlation data into a third prediction model, and using the third prediction model to generate a photovoltaic power prediction interval of the target photovoltaic power station based on a preset loss function, photovoltaic data spatial feature data, and photovoltaic data cross-time period correlation data.

[0006] The photovoltaic power prediction method based on the deep learning fusion model provided in this embodiment, first, by obtaining the historical environmental data and historical power generation data of the target photovoltaic power station, and performing variable correlation analysis to determine the cluster variable set, effectively screens out environmental variables that have a significant impact on photovoltaic power generation, such as light intensity, temperature, humidity, etc. Through correlation analysis, redundant or weakly correlated variables are removed, thereby improving the efficiency and accuracy of subsequent model training. The determination of the cluster variable set lays a solid foundation for subsequent data processing and model construction. Secondly, by performing cluster analysis on historical data, the power generation mode under different time periods or different environmental conditions can be identified, which is helpful to understand the dynamic characteristics of the photovoltaic power generation system, and the internal structure and laws between the data can be reflected through the cluster result data set, providing more refined input data for subsequent photovoltaic power prediction. Then, by inputting the clustering result data set into the first prediction model, extracting the spatial feature data of photovoltaic data, and using the powerful feature extraction ability of the deep learning algorithm, the key spatial features of photovoltaic data are automatically learned and extracted from the clustering result data set. These feature data can more accurately describe the relationship between photovoltaic power generation and environmental variables, providing a rich source of information for subsequent photovoltaic power prediction, and significantly improving the accuracy of the prediction. Afterwards, by inputting the spatial feature data of photovoltaic data into the second prediction model, the cross-time correlation data of photovoltaic data is parsed, and the spatial feature data of photovoltaic data is further parsed to mine the correlation between data in different time periods, which is of great significance for understanding the periodicity and seasonal changes of photovoltaic power generation. The acquisition of cross-time correlation data helps the model better grasp the long-term trend of photovoltaic power generation, so as to make more reasonable predictions. Finally, by inputting the cross-time correlation data of photovoltaic data into the third prediction model, the photovoltaic power prediction interval of the target photovoltaic power station is generated. Based on the preset loss function, the spatial characteristic data of photovoltaic data and the cross-time correlation data are comprehensively considered to generate the photovoltaic power prediction interval of the target photovoltaic power station. The specific power prediction value and the prediction uncertainty range are obtained, which provides a more comprehensive and reliable reference basis for the scheduling and planning of the power system. By introducing the prediction interval, it is possible to better cope with the volatility and uncertainty of photovoltaic power generation, improve the stability and economy of the power system, and solve the problem in related technologies that due to environmental factors and the diversity of photovoltaic power generation system characteristics, a single prediction model is difficult to adapt to the prediction needs in different scenarios, resulting in weak generalization ability of the prediction results in different scenarios. BRIEF DESCRIPTION OF THE DRAWINGS

[0007] In order to more clearly illustrate the specific implementation methods of the present invention or the technical solutions in the prior art, the drawings required for use in the specific implementation methods or the description of the prior art will be briefly introduced below. Obviously, the drawings described below are some implementation methods of the present invention. For ordinary technicians in this field, other drawings can be obtained based on these drawings without paying creative work.

[0008] Figure 1 It is a flowchart of a photoelectric power prediction method based on a deep learning fusion model according to an embodiment of the present invention. DETAILED DESCRIPTION

[0009] In order to make the purpose, technical solution and advantages of the embodiments of the present invention clearer, the technical solution in the embodiments of the present invention will be clearly and completely described below in conjunction with the drawings in the embodiments of the present invention. Obviously, the described embodiments are part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative work are within the scope of protection of the present invention.

[0010] Photovoltaic power, as the core indicator of solar power generation, directly reflects the power generation capacity of photovoltaic power generation systems under different environmental conditions. With the growing global demand for renewable energy, accurate prediction of photovoltaic power has become the key to improving the operating efficiency of power systems, optimizing energy allocation, and ensuring stable operation of power grids. However, the prediction of photovoltaic power is affected by many factors, including environmental parameters such as light intensity, temperature, humidity, wind speed, as well as the geographical location, equipment status, and operating history of photovoltaic power generation systems. There are often complex nonlinear relationships between these factors, making the prediction of photovoltaic power a challenging task.

[0011] The current photovoltaic power prediction methods mainly include statistical methods, physical model methods and machine learning methods. Statistical methods make predictions by constructing statistical relationships between historical data and photovoltaic power, but it is difficult to capture complex nonlinear relationships and dynamic changes in environmental factors. The physical model method establishes a prediction model based on physical laws and the operating mechanism of the photovoltaic power generation system. Although it has high accuracy, the model is highly complex and has strict requirements on input data. Machine learning methods, especially deep learning methods, have shown strong capabilities in processing nonlinear relationships, high-dimensional data and large-scale data sets, but they still face some challenges in practical applications. Due to the diversity of environmental factors and photovoltaic power generation system characteristics, a single prediction model is difficult to adapt to the prediction needs in different scenarios, resulting in weak generalization of prediction results in different scenarios.

[0012] The photovoltaic power prediction method based on the deep learning fusion model provided in this embodiment, first, by obtaining the historical environmental data and historical power generation data of the target photovoltaic power station, and performing variable correlation analysis to determine the cluster variable set, effectively screens out environmental variables that have a significant impact on photovoltaic power generation, such as light intensity, temperature, humidity, etc. Through correlation analysis, redundant or weakly correlated variables are removed, thereby improving the efficiency and accuracy of subsequent model training. The determination of the cluster variable set lays a solid foundation for subsequent data processing and model construction. Secondly, by performing cluster analysis on historical data, the power generation mode under different time periods or different environmental conditions can be identified, which is helpful to understand the dynamic characteristics of the photovoltaic power generation system, and the internal structure and laws between the data can be reflected through the cluster result data set, providing more refined input data for subsequent photovoltaic power prediction. Then, by inputting the clustering result data set into the first prediction model, extracting the spatial feature data of photovoltaic data, and using the powerful feature extraction ability of the deep learning algorithm, the key spatial features of photovoltaic data are automatically learned and extracted from the clustering result data set. These feature data can more accurately describe the relationship between photovoltaic power generation and environmental variables, providing a rich source of information for subsequent photovoltaic power prediction, and significantly improving the accuracy of the prediction. Afterwards, by inputting the spatial feature data of photovoltaic data into the second prediction model, the cross-time correlation data of photovoltaic data is parsed, and the spatial feature data of photovoltaic data is further parsed to mine the correlation between data in different time periods, which is of great significance for understanding the periodicity and seasonal changes of photovoltaic power generation. The acquisition of cross-time correlation data helps the model better grasp the long-term trend of photovoltaic power generation, so as to make more reasonable predictions. Finally, by inputting the cross-time correlation data of photovoltaic data into the third prediction model, the photovoltaic power prediction interval of the target photovoltaic power station is generated. Based on the preset loss function, the spatial characteristic data of photovoltaic data and the cross-time correlation data are comprehensively considered to generate the photovoltaic power prediction interval of the target photovoltaic power station. The specific power prediction value and the prediction uncertainty range are obtained, which provides a more comprehensive and reliable reference basis for the scheduling and planning of the power system. By introducing the prediction interval, it is possible to better cope with the volatility and uncertainty of photovoltaic power generation, improve the stability and economy of the power system, and solve the problem in related technologies that due to environmental factors and the diversity of photovoltaic power generation system characteristics, a single prediction model is difficult to adapt to the prediction needs in different scenarios, resulting in weak generalization ability of the prediction results in different scenarios.

[0013] According to an embodiment of the present invention, an embodiment of a photovoltaic power prediction method based on a deep learning fusion model is provided. It should be noted that the steps shown in the flowchart of the accompanying drawings can be executed in a computer system such as a set of computer executable instructions, and although a logical order is shown in the flowchart, in some cases, the steps shown or described may be executed in an order different from that shown here.

[0014] In this embodiment, a method for predicting photoelectric power based on a deep learning fusion model is provided. Figure 1 is a flow chart of a method for predicting photoelectric power based on a deep learning fusion model according to an embodiment of the present invention. Figure 1 As shown, the process includes the following steps:

[0015] Step S101, obtaining historical environmental data and historical power generation data of a target photovoltaic power station, performing variable correlation analysis on the historical environmental data, and determining a clustering variable set.

[0016] Specifically, historical environmental data of photovoltaic power stations, such as irradiance, temperature, wind speed, etc., and power generation data are collected. The collected data are cleaned, denoised, and normalized to improve the efficiency of model training. The main environmental factors affecting photovoltaic power are determined through statistical analysis methods, and clustering variables, such as date type, season, weather conditions, etc., are selected based on the results of variable correlation analysis.

[0017] Step S102: determining a clustering result data set based on a clustering variable set, historical environmental data, and historical power generation data.

[0018] Specifically, the above step S102 includes:

[0019] Step S1021, initializing the position of each cluster center in the cluster variable set, and assigning an initial membership degree to each data point in the historical environmental data.

[0020] Specifically, the clustering variable set is a set of variables that have a significant impact on photovoltaic power generation, which is screened from the original environmental data through variable correlation analysis. The cluster center is the point used to represent the center of each category in the clustering algorithm. Before the clustering algorithm starts, the initial cluster center position needs to be selected randomly or according to a certain strategy (such as the K-means++ algorithm). These positions will serve as the starting point for subsequent iterative updates. Membership refers to the degree or probability that a data point belongs to a cluster. In fuzzy clustering, data points can belong to multiple clusters at the same time, but the membership of each cluster is different. After the cluster center is initialized, the distance from each data point to each cluster center is calculated based on the Euclidean distance, and the initial membership is assigned based on these distances. In hard clustering, data points are explicitly assigned to the nearest cluster center, and the membership can be regarded as 0 or 1.

[0021] Step S1022, updating the position of each cluster center according to the initial membership degree of each data point until a preset iteration condition is reached to obtain a clustering result cluster set.

[0022] Specifically, the cluster center is updated based on the membership of the current data point, with the goal of making the cluster center reflect the characteristics of the data point to which it belongs more accurately. In each iteration, the location of the cluster center is recalculated based on the membership of the data point. Then, the membership of the data point is redistributed based on the new location of the cluster center until the location of the cluster center no longer changes significantly.

[0023] Step S1023: determining a clustering result data set based on the clustering result cluster set and the historical power generation data.

[0024] Specifically, the clustering result data set, i.e., the similar day data set, refers to a set of data sets with similar characteristics obtained through clustering analysis. The data points in each set, i.e., date or time period, have similar performance in environmental conditions and power generation. After clustering is completed, each cluster center and its corresponding data point set constitute a clustering result set. Since these data points have a high degree of similarity in environmental conditions and power generation, they can be regarded as similar day data sets, which reflect the power generation characteristics of photovoltaic power stations under different environmental conditions.

[0025] Step S103: input the clustering result data set into the first prediction model, and use the first prediction model to extract photovoltaic data spatial feature data based on the clustering result data set.

[0026] Specifically, the first prediction model is a convolutional neural network model, and step S103 includes:

[0027] Step S1031, normalizing the clustering result data set.

[0028] Specifically, the Convolutional Neural Network (CNN) is a deep learning model that is particularly good at processing data with a grid structure. It extracts local features of the data through the convolution layer and reduces the spatial size of the data through the pooling layer to increase the abstract level of the features. Using methods such as minimum-maximum normalization or Z-score normalization, each feature value in the clustering result data set is converted to a specified numerical range.

[0029] Step S1032, performing sliding window processing on the normalized clustering result data to obtain a photovoltaic data spatial feature map.

[0030] Specifically, the one-dimensional time series data is converted into a two-dimensional image form so that the spatial features can be extracted using the convolution and pooling operations of CNN. A fixed-size window is defined and slid over the normalized data. Each time it slides, the data segment within the window is used as a local area of ​​an "image" to construct a complete spatial feature map of the photovoltaic data.

[0031] Step S1033, in the first prediction model, a convolution operation, a pooling operation and a feature vector extraction operation are sequentially performed on the photovoltaic data spatial feature map to obtain photovoltaic data spatial feature data.

[0032] Specifically, multiple convolution kernels are slid on the spatial feature map of photovoltaic data to extract local features of the data. Each convolution kernel generates a new feature map, and these feature maps are combined to form a higher-level feature representation. The feature map obtained after the convolution operation is downsampled to reduce the spatial size of the feature map while retaining important features and reducing computational complexity. In the last few layers of CNN, the feature map obtained after convolution and pooling operations is usually converted into a one-dimensional feature vector through a fully connected layer, that is, the spatial feature data of photovoltaic data, which contains an abstract feature representation of the input data for subsequent photovoltaic power prediction tasks.

[0033] Step S104, inputting the spatial characteristic data of photovoltaic data into the second prediction model, and using the second prediction model to analyze the spatial characteristic data of photovoltaic data to obtain the photovoltaic data cross-time period correlation data.

[0034] Specifically, the second prediction model is a long short-term memory neural network, and step S104 includes:

[0035] Step S1041 , converting the photovoltaic data spatial characteristic data into a time series form to obtain a photovoltaic data spatial characteristic data sequence.

[0036] Specifically, the Long Short-Term Memory (LSTM) neural network is a special recurrent neural network (RNN). By introducing a gating mechanism to solve the gradient vanishing or gradient exploding problems that are prone to occur when traditional RNNs are trained on long sequence data, LSTM can effectively capture long-term dependencies in sequence data. Since LSTM is designed to process time series data, it is necessary to convert the spatial feature data of photovoltaic data into a one-dimensional time series form so that the LSTM model can process it. The rows or columns of the two-dimensional feature map are regarded as different time points of the time series, and the feature map is converted into a feature vector sequence.

[0037] Step S1042, using the forward layer of the second prediction model to sequentially process the photovoltaic data spatial feature data sequence to obtain a forward hidden state.

[0038] Specifically, the forward LSTM layer can capture the information flow from the past to the present in the data sequence and generate a hidden state containing past information. The forward layer of the LSTM processes the input of each time step step by step according to the order of the input sequence, updates the hidden state and cell state, and generates the corresponding output.

[0039] Step S1043, using the backward layer of the second prediction model to perform reverse processing on the photovoltaic data spatial feature data sequence to obtain a backward hidden state.

[0040] Specifically, through the backward LSTM layer, the model is able to capture information from the future to the present, providing another perspective on the data sequence. This is very useful for understanding the context of sequence data. Similar to the forward layer, but the input sequence is in reverse order. The backward layer also updates the hidden state and cell state, and generates a hidden state after reverse processing.

[0041] Step S1044, concatenating the forward hidden state and the backward hidden state to obtain the photovoltaic data cross-time correlation data.

[0042] Specifically, by concatenating the forward hidden state and the backward hidden state, the model is able to simultaneously utilize past and future information to generate a more comprehensive data representation that includes cross-period correlations of the data.

[0043] At each time step, the forward hidden state and the backward hidden state are concatenated in the feature dimension to form a new hidden state vector with doubled dimension. This vector, as the output of the current time step, contains the cross-period correlation information extracted from the entire sequence. Ultimately, these output vectors constitute a collection of cross-period correlation data of photovoltaic data.

[0044] Step S105, inputting the photovoltaic data cross-time correlation data into a third prediction model, and using the third prediction model to generate a photovoltaic power prediction interval of a target photovoltaic power station based on a preset loss function, photovoltaic data spatial feature data and photovoltaic data cross-time correlation data.

[0045] Specifically, the preset loss function includes a prediction error loss function and a prediction interval width loss function, the third prediction model is a quantile regression model, and step S105 includes:

[0046] Step S1051, obtaining each preset quantile in the preset loss function.

[0047] Specifically, determine the quantile points that need to be estimated in quantile regression, which are used to construct prediction intervals. According to actual needs, select a series of representative quantile points, such as 0.05, 0.25, 0.5 (median), 0.75, 0.95, etc.

[0048] Step S1052, using a third prediction model to generate photovoltaic power prediction values ​​corresponding to each preset quantile based on each preset quantile, photovoltaic data spatial characteristic data and photovoltaic data cross-time period correlation data.

[0049] Specifically, the Quantile Regression Model is a statistical method used to estimate the conditional quantile, that is, the quantile of the conditional distribution of the dependent variable under a given independent variable value. Unlike traditional least squares regression, quantile regression can give the predicted value of the dependent variable at any quantile, thereby constructing a prediction interval. The spatial characteristic data of photovoltaic data and the cross-period correlation data of photovoltaic data are used as input, and the quantile regression model is used for training to obtain the photovoltaic power prediction value at each preset quantile.

[0050] Step S1053, using the prediction error loss function to calculate the prediction error loss corresponding to each preset quantile based on the photovoltaic power prediction value corresponding to each preset quantile, to obtain the prediction error loss.

[0051] Specifically, the prediction error loss function is a function used to measure the difference between the model's predicted value and the actual value. In quantile regression, an asymmetric loss function is used to calculate the prediction error. Using the preset prediction error loss function, the prediction error loss is calculated for each preset quantile.

[0052] Step S1054, using the prediction interval width loss function to calculate the prediction interval width loss based on the photovoltaic power prediction values ​​corresponding to each preset quantile.

[0053] Specifically, prediction interval width loss function: a function used to control the width of the prediction interval. A wider prediction interval may contain more actual values, but will reduce the accuracy of the prediction; a narrower prediction interval may miss some actual values, but can improve the accuracy of the prediction. The prediction interval width loss function is used to find a balance between the two. Using the preset prediction interval width loss function, the width loss of the prediction interval is calculated based on the photovoltaic power prediction values ​​corresponding to each preset quantile.

[0054] Step S1055, weighted summing up the prediction error loss and the prediction interval width loss is performed using preset weights to obtain a total loss function value.

[0055] Specifically, preset weights are assigned to the prediction error loss and the prediction interval width loss, and their weighted sum is used to obtain the total loss function value, and the preset weights are obtained based on data training.

[0056] Step S1056, using the third prediction model to generate a photovoltaic power prediction value interval of the target photovoltaic power station based on the total loss function value, the photovoltaic data spatial characteristic data and the photovoltaic data cross-time period correlation data.

[0057] Specifically, step S1056 includes:

[0058] Step a1: Execute a back propagation algorithm to iteratively adjust the parameters in the third prediction model according to the total loss function value.

[0059] Specifically, the backpropagation algorithm is a method for training a neural network by calculating the gradient of the loss function with respect to the network parameters and updating the parameters in the opposite direction of the gradient to minimize the loss function. In the quantile regression model, the backpropagation algorithm is executed to iteratively adjust the parameters in the third prediction model.

[0060] Step a2: During each iteration, the predicted photoelectric power value at each preset quantile under the current iteration number is calculated until the preset iteration number is reached.

[0061] Specifically, the model parameters at the current iteration number are used, combined with the spatial feature data of photovoltaic data and the cross-period correlation data of photovoltaic data, to calculate the photovoltaic power prediction value at each preset quantile. These prediction values ​​will be used to calculate the loss function value at the current iteration number and serve as the input of the back propagation algorithm.

[0062] Step a3: after the iteration is completed, the photovoltaic power prediction value at each preset quantile under the current iteration number is output to obtain the photovoltaic power prediction value interval of the target photovoltaic power station.

[0063] Specifically, after the iteration process is completed, the photovoltaic power prediction values ​​at each preset quantile calculated at the last iteration are used as output to form the upper and lower bounds of the photovoltaic power prediction value interval, for example, the 5th percentile is used as the lower bound and the 95th percentile is used as the upper bound. The photovoltaic power prediction value interval finally obtained will provide the target photovoltaic power station with reliable information about the possible range of photovoltaic power values ​​in the future.

[0064] In an optional manner of some embodiments, before performing variable correlation analysis on the historical environmental data, the method further includes: performing interpolation processing on the historical environmental data.

[0065] Specifically, in actual data collection, there may be missing data due to various reasons such as sensor failure and communication interruption. Interpolation processing can estimate and fill these missing data points based on the information of existing data points, avoiding deviations caused by missing values ​​during the analysis process.

[0066] The photovoltaic power prediction method based on the deep learning fusion model provided in this embodiment, first, by obtaining the historical environmental data and historical power generation data of the target photovoltaic power station, and performing variable correlation analysis to determine the cluster variable set, effectively screens out environmental variables that have a significant impact on photovoltaic power generation, such as light intensity, temperature, humidity, etc. Through correlation analysis, redundant or weakly correlated variables are removed, thereby improving the efficiency and accuracy of subsequent model training. The determination of the cluster variable set lays a solid foundation for subsequent data processing and model construction. Secondly, by performing cluster analysis on historical data, the power generation mode under different time periods or different environmental conditions can be identified, which is helpful to understand the dynamic characteristics of the photovoltaic power generation system, and the internal structure and laws between the data can be reflected through the cluster result data set, providing more refined input data for subsequent photovoltaic power prediction. Then, by inputting the clustering result data set into the first prediction model, extracting the spatial feature data of photovoltaic data, and using the powerful feature extraction ability of the deep learning algorithm, the key spatial features of photovoltaic data are automatically learned and extracted from the clustering result data set. These feature data can more accurately describe the relationship between photovoltaic power generation and environmental variables, providing a rich source of information for subsequent photovoltaic power prediction, and significantly improving the accuracy of the prediction. Afterwards, by inputting the spatial feature data of photovoltaic data into the second prediction model, the cross-time correlation data of photovoltaic data is parsed, and the spatial feature data of photovoltaic data is further parsed to mine the correlation between data in different time periods, which is of great significance for understanding the periodicity and seasonal changes of photovoltaic power generation. The acquisition of cross-time correlation data helps the model better grasp the long-term trend of photovoltaic power generation, so as to make more reasonable predictions. Finally, by inputting the cross-time correlation data of photovoltaic data into the third prediction model, the photovoltaic power prediction interval of the target photovoltaic power station is generated. Based on the preset loss function, the spatial characteristic data of photovoltaic data and the cross-time correlation data are comprehensively considered to generate the photovoltaic power prediction interval of the target photovoltaic power station. The specific power prediction value and the prediction uncertainty range are obtained, which provides a more comprehensive and reliable reference basis for the scheduling and planning of the power system. By introducing the prediction interval, it is possible to better cope with the volatility and uncertainty of photovoltaic power generation, improve the stability and economy of the power system, and solve the problem in related technologies that due to environmental factors and the diversity of photovoltaic power generation system characteristics, a single prediction model is difficult to adapt to the prediction needs in different scenarios, resulting in weak generalization ability of the prediction results in different scenarios.

[0067] Although the embodiments of the present invention have been described in conjunction with the accompanying drawings, those skilled in the art may make various modifications and variations without departing from the spirit and scope of the present invention, and such modifications and variations are all within the scope defined by the appended claims.

Claims

1. A photoelectric power prediction method based on a deep learning fusion model, characterized in that: The method comprises: Acquire historical environmental data and historical power generation data of a target photovoltaic power station, perform variable correlation analysis on the historical environmental data, and determine a cluster variable set; Determining a clustering result data set based on the clustering variable set, the historical environmental data and the historical power generation data; Inputting the clustering result data set into a first prediction model, and using the first prediction model to extract photovoltaic data spatial feature data based on the clustering result data set; Inputting the photovoltaic data spatial characteristic data into a second prediction model, and using the second prediction model to analyze the photovoltaic data spatial characteristic data to obtain photovoltaic data cross-time period correlation data; Inputting the photovoltaic data cross-time correlation data into a third prediction model, and using the third prediction model to generate a photovoltaic power prediction interval of a target photovoltaic power station based on a preset loss function, photovoltaic data spatial feature data, and photovoltaic data cross-time correlation data; The step of determining a clustering result data set based on the clustering variable set, the historical environmental data, and the historical power generation data includes: Initializing the position of each cluster center in the cluster variable set, and assigning an initial membership degree to each data point in the historical environmental data; According to the initial membership of each data point, the position of each cluster center is updated until the preset iteration condition is reached to obtain the clustering result cluster set; Determining a clustering result data set based on the clustering result cluster set and the historical power generation data; The first prediction model is a convolutional neural network model, the clustering result data set is input into the first prediction model, and the photovoltaic data spatial feature data is extracted based on the clustering result data set using the first prediction model, including: Normalizing the clustering result data set; The normalized clustering result data is processed by sliding window to obtain the spatial characteristic map of photovoltaic data; In the first prediction model, a convolution operation, a pooling operation and a feature vector extraction operation are sequentially performed on the photovoltaic data spatial feature map to obtain photovoltaic data spatial feature data; The step of inputting the photovoltaic data spatial characteristic data into a second prediction model, and using the second prediction model to analyze the photovoltaic data spatial characteristic data to obtain photovoltaic data cross-time period correlation data includes: Converting the photovoltaic data spatial characteristic data into a time series form to obtain a photovoltaic data spatial characteristic data sequence; sequentially processing the photovoltaic data spatial feature data sequence using the forward layer of the second prediction model to obtain a forward hidden state; Using the backward layer of the second prediction model, the photovoltaic data spatial feature data sequence is processed in reverse order to obtain a backward hidden state; The forward hidden state and the backward hidden state are concatenated to obtain photovoltaic data cross-time correlation data; The preset loss function includes a prediction error loss function and a prediction interval width loss function, the third prediction model is a quantile regression model, the photovoltaic data cross-time period correlation data is input into the third prediction model, and the third prediction model is used to generate a photovoltaic power prediction interval of a target photovoltaic power station based on the preset loss function, the photovoltaic data spatial feature data and the photovoltaic data cross-time period correlation data, including: Get each preset quantile in the preset loss function; Generate photovoltaic power prediction values ​​corresponding to each preset quantile using a third prediction model based on each preset quantile, photovoltaic data spatial characteristic data, and photovoltaic data cross-time period correlation data; Using the prediction error loss function, based on the photovoltaic power prediction value corresponding to each preset quantile, the prediction error loss corresponding to each preset quantile is calculated to obtain the prediction error loss; The prediction interval width loss function is used to calculate the prediction interval width loss based on the photovoltaic power prediction values ​​corresponding to each preset quantile; The prediction error loss and the prediction interval width loss are weighted and summed by preset weights to obtain the total loss function value; The third prediction model is used to generate a photovoltaic power prediction value interval of a target photovoltaic power station based on the total loss function value, photovoltaic data spatial characteristic data and photovoltaic data cross-time period correlation data.

2. The method according to claim 1, characterized in that The method of using the third prediction model to generate a photovoltaic power prediction value interval of a target photovoltaic power station based on the total loss function value, photovoltaic data spatial characteristic data, and photovoltaic data cross-time period correlation data includes: According to the total loss function value, executing a back propagation algorithm to iteratively adjust the parameters in the third prediction model; In each iteration process, the predicted value of the photoelectric power at each preset quantile under the current number of iterations is calculated until the preset number of iterations is reached; After the iteration is completed, the photovoltaic power prediction value at each preset quantile under the current iteration number is output to obtain the photovoltaic power prediction value interval of the target photovoltaic power station.

3. The method according to claim 1, characterized in that Before performing variable correlation analysis on the historical environmental data, the method further includes: performing interpolation processing on the historical environmental data.

Images