A Multimodal Fusion Photovoltaic Power Prediction Method for All-Sky Imager Data

Abstract

The present invention discloses a multi-modal fusion photovoltaic power prediction method for all-sky imager data, comprising the following steps: obtaining the time-series sky image data in the all-sky imager and the historical photovoltaic power data of the photovoltaic array; respectively extracting features from the time-series sky image data and the historical photovoltaic power data to obtain the modal feature information of the time-series sky image data and the modal feature information of the historical photovoltaic power data; fusing the modal feature information of the time-series sky image data and the modal feature information of the historical photovoltaic power data to obtain a feature matrix; obtaining the final latent features according to the feature matrix; training a photovoltaic power prediction model using the final latent features; and predicting the photovoltaic power using the photovoltaic power prediction model. The present invention performs feature extraction processing on the time-series sky image data and the historical photovoltaic power data, effectively reducing the prediction error caused by the intermittency and uncertainty of solar power generation.

Description

Technical Field

[0001] The present invention relates to the technical field of photovoltaic power prediction, and more specifically, to a multi-modal fusion photovoltaic power prediction method for all-sky imager data. Background Technique

[0002] With the wide application of solar energy, the power generation of renewable energy represented by photovoltaics has developed rapidly. In particular, the popularization of rooftop photovoltaic panels has become an important solution to cope with destructive climate change and promote sustainable energy development. With the continuous deepening of this trend, the penetration rate of solar energy in the power system is expected to increase steadily. The cumulative installed capacity of photovoltaics in China has increased from 43.18 GW in 2015 to 609.49 GW in 2023, with an average annual compound growth rate of about 39%. The annual newly added photovoltaic installed capacity and the cumulative photovoltaic installed capacity both rank first in the world. Promoting the development of photovoltaic power generation has become one of the important ways to promote the transformation of China's power system.

[0003] For photovoltaic power prediction, there is a multi-step photovoltaic power prediction method based on multi-view feature extraction and multi-task learning in the prior art. First, in order to obtain rich and comprehensive feature information, LSTM, MLP, and CNN are used to extract the temporal features, global features, and local features in the original photovoltaic power data respectively. Secondly, the multi-step photovoltaic power prediction task is transformed into multiple single-step photovoltaic power prediction subtasks, and a multi-task learning model based on a lightweight attention mechanism and an expert network (lightweightattention-based mixture of experts, LAMoE) is used for multi-step prediction to fully utilize the correlation of multi-step prediction values. Finally, an improved dynamic weight averaging method is used to adaptively optimize and adjust the loss weights for photovoltaic prediction. However, due to the limitation of the feature information of the original photovoltaic power data, the consideration of spatial and cloud feature information factors is lacking, and there is a disadvantage of deviation in photovoltaic prediction accuracy. Summary of the Invention

[0004] The present invention provides a multi-modal fusion photovoltaic power prediction method for all-sky imager data, which solves the technical problem in the prior art that due to the limitation of the feature information of the original photovoltaic power data, the consideration of spatial and cloud feature information factors is lacking, and there is a deviation in photovoltaic prediction accuracy.

[0005] To solve the above technical problems, the technical solution of the present invention is as follows:

[0006] A multi-modal fusion photovoltaic power prediction method for all-sky imager data, comprising the following steps:

[0007] Obtain the temporal sky image data in the all-sky imager and the historical photovoltaic power data of the photovoltaic array;

[0008] Extract features from the time-series sky image data and historical photovoltaic power data respectively to obtain the modal feature information of the time-series sky image data and the modal feature information of the historical photovoltaic power data;

[0009] Fuse the modal feature information of the time-series sky image data and the modal feature information of the historical photovoltaic power data to obtain a feature matrix;

[0010] Obtain the final latent feature according to the feature matrix;

[0011] Train a photovoltaic power prediction model using the final latent feature;

[0012] Use the photovoltaic power prediction model to predict the photovoltaic power.

[0013] In the above technical means, feature extraction processing is performed on the time-series sky image data collected by the all-sky imager and the historical photovoltaic power data of the photovoltaic array, fully considering the time-series, space, and cloud dynamic feature information in different modalities, effectively reducing the prediction error caused by the intermittency and uncertainty of solar power generation, improving the accuracy of multi-step prediction of photovoltaic power in solar power stations, and solving the technical problem of the deviation of photovoltaic prediction accuracy in the existing technology due to the limitation of the feature information of the original photovoltaic power data and the lack of consideration of spatial and cloud feature information factors.

[0014] Further, extracting features from the time-series sky image data to obtain the modal feature information of the time-series sky image data includes:

[0015] Reduce the dimension of the time-series sky image data to obtain a grayscale image, and convert the grayscale image into a grayscale pixel grid map;

[0016] Define a circular domain with the center of the grayscale pixel grid map as the center and a radius of , where the circular domain does not exceed the range of the grayscale pixel grid map, is an integer and , ;

[0017] For the circular domain with a radius of , take the average value of the pixel grayscale values of the grayscale pixel grid map within the circular domain to obtain the local mean corresponding to the circular domain with a radius of ; When it is greater than 1, for the circular domain with a radius of , for the circular domain with a radius of and less than The pixel gray values are averaged to obtain the local mean corresponding to the circular neighborhood with a radius of ;

[0018] When the circular neighborhood touches the edge of the grayscale pixel grid map, the local means are aggregated to obtain the modal feature information of the time-series sky image data.

[0019] Furthermore, extracting features from the time-series sky image data to obtain the modal feature information of the time-series sky image data further includes:

[0020] Using the median filtering method, for the eigenvalue points near the sun in the modal feature information of the time-series sky image data, calculate the median value of a 3x3 fixed window, and replace the original feature points with the median value.

[0021] Furthermore, extracting features from the historical photovoltaic power data to obtain the modal feature information of the historical photovoltaic power data includes:

[0022] Calculating the mean value of the historical photovoltaic power data and the standard deviation ;

[0023] Calculating the abnormal data points in the historical photovoltaic power data according to the Z-score method;

[0024] Interpolating and resampling the missing points and abnormal data points in the historical photovoltaic power data to obtain a photovoltaic power sequence;

[0025] After processing the photovoltaic power sequence using BiLSTM, splicing the forward and reverse outputs to obtain the modal feature information of the historical photovoltaic power data.

[0026] Furthermore, calculating the abnormal data points in the historical photovoltaic power data according to the Z-score method includes:

[0027]

[0028] In the formula, represents a photovoltaic power data point. When Z is greater than a preset threshold, it is determined that the photovoltaic power data point is an abnormal data point.

[0029] Furthermore, fusing the modal feature information of the time-series sky image data and the modal feature information of the historical photovoltaic power data to obtain a feature matrix includes:

[0030] Aligning the time stamps of the modal feature information of the time-series sky image data and the modal feature information of the historical photovoltaic power data and then superimposing them to obtain a feature matrix .

[0031] Furthermore, according to the feature matrix, the final potential features are obtained, including:

[0032] The feature matrix Expanding to obtain a feature vector, and filtering from the feature vector to obtain regularized feature information;

[0033] The regularized feature information is processed through a secondary feature extraction module and a pooling layer with L1 norm penalty to obtain the final potential feature.

[0034] Furthermore, the regularized feature information is obtained by screening the feature vector, including:

[0035] A feature selection module based on L1 norm regularization is constructed, and a group of mean-effective features are selected from the high-dimensional input space by the feature selection module for the feature vector to obtain the regularized feature information.

[0036] Furthermore, the secondary feature extraction module is a multi-layer stacked encoder and decoder structure.

[0037] Furthermore, the photovoltaic power prediction model is a linear regression prediction model.

[0038] Compared with the prior art, the technical solution of the present invention has the following beneficial effects:

[0039] The present invention proposes a multi-modal fusion photovoltaic power prediction method for sky imager data. First, the time series sky image data of the sky imager is obtained, and the sky image is reduced to a grayscale image. A new method is created, in which the center coordinates of the grayscale pixel matrix are taken as the circle point, the radius parameter is R, and R is gradually accumulated to extract the average grayscale pixel value in the circular neighborhood. This process can effectively extract the brightness information, cloud features and other visual features in the image, and provide rich basic data for subsequent analysis. In order to further improve the quality of the data, the median filter method is used to process the image to remove the noise caused by the mirror reflection effect of sunlight, so as to obtain more accurate image features. Secondly, the historical photovoltaic power time series data of the photovoltaic array of the solar panel is collected, and the BiLSTM can be used to simultaneously capture the forward and backward time dependencies in the data and mine its inherent nonlinear characteristics. By aligning the timestamps, the time series features of different modes are fused into a feature matrix, and then the useful feature information is screened by L1 norm regularization regression. Finally, the final potential feature vector matrix output by the secondary feature processing module is input into the linear regression prediction model to improve the prediction accuracy of photovoltaic power based on the sky image. BRIEF DESCRIPTION OF THE DRAWINGS

[0040] Figure 1Schematic flowchart of a multimodal fusion photovoltaic power prediction method for all-sky imager data provided by an embodiment of the present invention;

[0041] Figure 2 Schematic diagram of the principle of a multimodal fusion photovoltaic power prediction method for all-sky imager data provided by an embodiment of the present invention;

[0042] Figure 3 Photovoltaic power prediction effect diagram for a certain day obtained by using the multimodal fusion photovoltaic power prediction method for all-sky imager data proposed by the present invention provided by an embodiment of the present invention. Detailed implementation manners

[0043] The accompanying drawings are only for illustrative purposes and should not be construed as limitations on this patent;

[0044] To better illustrate this embodiment, some components in the accompanying drawings are omitted, enlarged or reduced, and do not represent the dimensions of actual products;

[0045] For those skilled in the art, it is understandable that some well-known structures and their descriptions in the accompanying drawings may be omitted.

[0046] The technical solutions of the present invention will be further described below with reference to the accompanying drawings and embodiments.

[0047] Embodiment 1

[0048] This embodiment provides a multimodal fusion photovoltaic power prediction method for all-sky imager data, as Figure 1 shown, including the following steps:

[0049] Obtain the sequential sky image data in the all-sky imager and the historical photovoltaic power data of the photovoltaic array;

[0050] Extract features from the sequential sky image data and the historical photovoltaic power data respectively to obtain the modal feature information of the sequential sky image data and the modal feature information of the historical photovoltaic power data;

[0051] Fuse the modal feature information of the sequential sky image data and the modal feature information of the historical photovoltaic power data to obtain a feature matrix;

[0052] Obtain the final latent features according to the feature matrix;

[0053] Train a photovoltaic power prediction model using the final latent features;

[0054] Use the photovoltaic power prediction model to predict the photovoltaic power.

[0055] As Figure 2As shown in the figure, it is a schematic diagram of the principle of a multi-modal fusion photovoltaic power prediction method for all-sky imager data provided by this embodiment, specifically as follows:

[0056] The time-series sky image data in the all-sky imager includes sky image data with a resolution of one minute;

[0057] The historical photovoltaic power data of the photovoltaic array includes historical photovoltaic power data with a size of 2048x2048 and a sampling frequency of one minute.

[0058] In a further embodiment, features are extracted from the time-series sky image data to obtain the modal feature information of the time-series sky image data, including:

[0059] The time-series sky image data is reduced in dimension to obtain a grayscale image, and the grayscale image is converted into a grayscale pixel grid map;

[0060] In this embodiment, the range of grayscale levels of the grayscale pixel grid map is 0-255; by reducing the dimension of the time-series sky image data, the texture and edge features of the time-series grayscale image can be regularly extracted, and the dynamic change trend of the cloud layer can also be obtained from the time-series image;

[0061] Define a circular domain with the center of the grayscale pixel grid map as the center and a radius of , where the circular domain does not exceed the range of the grayscale pixel grid map, is an integer and , ;

[0062] In this embodiment, the pixel points within the circular domain with a radius of are determined by the following formula:

[0063]

[0064] In the formula, is the pixel coordinate within the circular neighborhood, is the center coordinate;

[0065] For the circular domain with a radius of , the average value of the pixel grayscale values of the grayscale pixel grid map within the circular domain is taken to obtain the local mean corresponding to the circular domain with a radius of ; When it is greater than 1, for the circular domain with a radius of , the average value of the pixel grayscale values of the grayscale pixel grid map whose distance from the center is greater than and less than is taken to obtain the local mean corresponding to the circular domain with a radius of ;

[0066] In this embodiment, the formula for the circular field mean is as follows:

[0067]

[0068] In the formula, is the local mean accumulated within the k-th radius for point ; is centered on the central pixel point and is the set of all pixels used to calculate the mean within a circular neighborhood with a radius of ; is the number of pixel points within the neighborhood;

[0069] When the circular field touches the edge of the grayscale pixel grid map, the local means are aggregated to obtain the modal feature information of the time-series sky image data.

[0070] In a further embodiment, extracting features from the time-series sky image data to obtain the modal feature information of the time-series sky image data further includes:

[0071] Using the median filtering method to calculate the median value of a 3x3 fixed window for the eigenvalue points near the sun in the modal feature information of the time-series sky image data, and replacing the original feature points with the median value.

[0072] In this embodiment, the median filtering method is used to reduce sunlight reflection.

[0073] In a further embodiment, extracting features from the historical photovoltaic power data to obtain the modal feature information of the historical photovoltaic power data includes:

[0074] Due to uncontrollable factors, the historical power data of the device often has missing values and outliers, and the missing values and outliers need to be processed. First, calculate the mean and the standard deviation of the historical photovoltaic power data;

[0075] Calculate the outlier data points in the historical photovoltaic power data according to the Z-score method;

[0076] On a straight line, every position between two points can be predicted. The goal of linear interpolation is to find the value of (x,y) for x at a certain point between and . Therefore, interpolation resampling can be performed on the missing points and outlier data points in the historical photovoltaic power data to obtain a photovoltaic power sequence;

[0077] After processing the photovoltaic power sequence using BiLSTM, the forward and reverse outputs are concatenated to obtain the modal feature information of the historical photovoltaic power data.

[0078] In a specific embodiment, the BiLSTM is composed of two independent LSTM units, which are responsible for processing the input sequence in two directions (forward and reverse). The main formula of the LSTM unit is as follows:

[0079]

[0080]

[0081]

[0082]

[0083]

[0084]

[0085] In the formula, 、 、 、 、 、 、 and are the weight matrices and biases corresponding to the inputs of the network activation function; 、 、 、 、 are the forget gate, input gate, cell state, output gate, and hidden state respectively; represents the sigmoid activation function, is the hyperbolic tangent function.

[0086] After the historical photovoltaic power data is processed by BiLSTM, the forward and reverse outputs are concatenated to capture the long-term temporal dependencies and non-linear feature information in the sequence. Through complementary feature information, redundant information is eliminated to generate new time series modal features.

[0087] In a further embodiment, calculating the abnormal data points in the historical photovoltaic power data according to the Z-score method includes:

[0088]

[0089] In the formula, represents the photovoltaic power data point. When Z is greater than the preset threshold, the photovoltaic power data point is determined to be an abnormal data point.

[0090] In a specific embodiment, the threshold is set to 3. If Z is greater than 3, it is identified as a value deviating from the normal range, and then the abnormal value is marked.

[0091] In a further embodiment, the modal feature information of the temporal sky image data and the modal feature information of the historical photovoltaic power data are fused to obtain a feature matrix, including:

[0092] The modal feature information of the temporal sky image data and the modal feature information of the historical photovoltaic power data are aligned in time stamp and then superimposed to obtain a feature matrix 。

[0093] In a further embodiment, according to the feature matrix, the final potential features are obtained, including:

[0094] The feature matrix is unfolded to obtain a feature vector, and the regularized feature information is screened from the feature vector;

[0095] The regularized feature information is processed by a quadratic feature extraction module and a pooling layer with L1 norm penalty to obtain the final potential features.

[0096] In a further embodiment, screening the regularized feature information from the feature vector includes:

[0097] A feature selection module based on L1 norm regularization is constructed, and the feature vector is passed through the feature selection module to select a set of mean-effective features from the high-dimensional input space to obtain the regularized feature information.

[0098] In a further embodiment, the quadratic feature extraction module is a multi-layer stacked encoder and decoder structure.

[0099] In this embodiment, the quadratic feature extraction module based on the Transformer architecture can capture long-range temporal dependencies through the self-attention mechanism in the Transformer, effectively model the global information of the temporal image features, and gradually extract different temporal scales and higher-level abstract features through the multi-layer stacked encoder and decoder structure.

[0100] In a further embodiment, a pooling operation based on L1 norm penalty is used to reduce the dimension to obtain the final potential features.

[0101] In a further embodiment, it is characterized in that the photovoltaic power prediction model is a linear regression prediction model.

[0102] Embodiment 2

[0103] In this embodiment, to verify the effectiveness of a multi-modal fusion photovoltaic power prediction method for all-sky imager data proposed by the present invention, first, sky images and historical photovoltaic power generation data from 6:00 to 20:00 within a month in a certain place are obtained, and preprocessing is performed on the two different modal data. Then, a feature matrix is fused and screened according to the data characteristics after preprocessing. Finally, through a secondary feature processing module based on the transformer architecture, the final potential features are output, input into a linear regression model, and the data of a certain day is used for prediction to obtain as Figure 3 shown in the prediction effect diagram of the photovoltaic power generation, in Figure 3 , the solid trend line at the center of the circle represents the true value of the photovoltaic power generation, and the solid trend line at the center of the triangle represents the predicted value of the photovoltaic power generation. From Figure 3 the curves corresponding to the two values shown, it can be seen that the error between the true value of the photovoltaic power generation and the predicted value of the photovoltaic power generation is very small. Using the method proposed by the present invention can effectively improve the accuracy of the photovoltaic power generation prediction.

[0104] Embodiment 3

[0105] This embodiment proposes a multi-modal fusion photovoltaic power prediction system for all-sky imager data. The system includes:

[0106] A sky image and historical photovoltaic power data acquisition and processing unit, which is used to acquire the image modality of the all-sky imager and the historical photovoltaic power of the photovoltaic array, and perform preliminary processing on the relevant data;

[0107] A time-series feature processing unit, which is used to extract features and perform smoothing processing on the sky image data, and screen out the feature information of brightness, cloud layer, and vision through the feature module; the time-series photovoltaic power data is used to obtain time-series feature information through a BiLSTM network.

[0108] A modal feature fusion unit, which aligns the time stamps of the image modality and the time series modality and superimposes them to form new features, flattens them into long vectors, and inputs them into an L1-norm regularized regression feature selection module to screen out useful feature information;

[0109] A model construction unit, which is used to construct a secondary feature processing model and a linear regression prediction model. The multi-modal fusion data passes through the secondary feature processing model to extract abstract features of different time scales and higher levels. Finally, the features input are pooled through an L1-norm penalty pooling layer to form the potential features of the final input prediction model;

[0110] A prediction unit, which is used to input the final potential features obtained by the deep learning network feature processing model into the final linear regression prediction model to predict the photovoltaic output power of the power station.

[0111] The same or similar reference numerals correspond to the same or similar components;

[0112] The terms used to describe the positional relationship in the drawings are for illustrative purposes only and should not be construed as limiting the patent;

[0113] Obviously, the above embodiments of the present invention are merely examples for clearly illustrating the present invention and are not limitations on the implementation manners of the present invention. For those of ordinary skill in the art, other different forms of changes or modifications can be made based on the above description. It is not necessary and impossible to list all the implementation manners here. Any modifications, equivalent replacements, and improvements made within the spirit and principles of the present invention shall be included within the protection scope of the claims of the present invention.

Claims

1. A multi-modal fusion photovoltaic power prediction method for all-sky imager data, characterized in that Including the following steps: Obtain the time-series sky image data in the all-sky imager and the historical photovoltaic power data of the photovoltaic array; Extract features from the time-series sky image data and the historical photovoltaic power data respectively to obtain the modal feature information of the time-series sky image data and the modal feature information of the historical photovoltaic power data; Fuse the modal feature information of the time-series sky image data and the modal feature information of the historical photovoltaic power data to obtain a feature matrix; Obtain the final latent features according to the feature matrix; Train a photovoltaic power prediction model using the final latent features; Predict the photovoltaic power using the photovoltaic power prediction model; Extract features from the time-series sky image data to obtain the modal feature information of the time-series sky image data, including: Reduce the dimension of the time-series sky image data to obtain a grayscale image, and convert the grayscale image into a grayscale pixel grid map; Define a circular area centered at the center of the grayscale pixel grid map with a radius of , where the circular area does not exceed the range of the grayscale pixel grid map, is an integer and , ; For a circular area with a radius of , the average value of the pixel gray values of the grayscale pixel grid map within the circular area is taken to obtain the local mean corresponding to the circular area with a radius of ; When it is greater than 1, for a circular area with a radius of , the average value of the pixel gray values in the grayscale pixel grid map whose distance from the center of the circle is greater than and less than is taken to obtain the local mean corresponding to the circular area with a radius of ; When the circular domain touches the edge of the grayscale pixel grid map, summarize the local means to obtain the modal feature information of the time-series sky image data.

2. The multi-modal fusion photovoltaic power prediction method for all-sky imager data according to claim 1, wherein Extracting features from the time-series sky image data to obtain the modal feature information of the time-series sky image data further includes: Using the median filtering method to calculate the median value of a 3x3 fixed window for the eigenvalue points near the sun in the modal feature information of the time-series sky image data, and replacing the original feature points with the median value.

3. The multi-modal fusion photovoltaic power prediction method for all-sky imager data according to claim 1, characterized in that Extract features from the historical photovoltaic power data to obtain the modal feature information of the historical photovoltaic power data, including: Calculate the mean of the historical PV power data and the standard deviation ; Calculate the abnormal data points in the historical photovoltaic power data according to the Z-score method; Interpolate and resample the missing points and abnormal data points in the historical photovoltaic power data to obtain a photovoltaic power sequence; After processing the photovoltaic power sequence using BiLSTM, splice the forward and reverse outputs to obtain the modal feature information of the historical photovoltaic power data.

4. The multi-modal fusion photovoltaic power prediction method for all-sky imager data according to claim 3, wherein Calculating the abnormal data points in the historical photovoltaic power data according to the Z-score method includes: In the formula, represents a photovoltaic power data point. When Z is greater than a preset threshold value, it is determined that the photovoltaic power data point is an abnormal data point.

5. The multi-modal fusion photovoltaic power prediction method for all-sky imager data according to claim 1, characterized in that Fusing the modal feature information of the time-series sky image data and the modal feature information of the historical photovoltaic power data to obtain a feature matrix, including: Align the modal feature information of the time-series sky image data and the modal feature information of the historical photovoltaic power data by time stamps and then superimpose them to obtain a feature matrix .

6. The multi-modal fusion photovoltaic power prediction method for all-sky imager data according to claim 5, wherein Obtaining the final latent features according to the feature matrix, including: Unfold the feature matrix to obtain a feature vector, and screen out regularization feature information from the feature vector; Processing the regularization feature information through a secondary feature extraction module and a pooling layer with L1 norm penalty to obtain the final latent features.

7. The multimodal fusion photovoltaic power prediction method for all-sky imager data according to claim 6, characterized in that, Selecting the regularization feature information from the feature vectors, including: Construct a feature selection module based on L1 norm regularization, and pass the feature vectors through the feature selection module to select a set of mean-effective features from the high-dimensional input space to obtain the regularization feature information.

8. The multi-modal fusion photovoltaic power prediction method for all-sky imager data according to claim 6, characterized in that The secondary feature extraction module is a multi-layer stacked encoder and decoder structure.

9. The multimodal fusion photovoltaic power prediction method for all-sky imager data according to any one of claims 1 to 8, characterized in that The photovoltaic power prediction model is a linear regression prediction model.

Images