Photovoltaic power generation power prediction method and system based on multi-source heterogeneous feature fusion

Abstract

The invention discloses a photovoltaic power generation power prediction method and system based on multi-source heterogeneous feature fusion. The method comprises the steps of obtaining historical time sequence data and historical image data of a photovoltaic power plant; performing feature extraction on the historical time series data to obtain a two-dimensional feature map corresponding to the historical time series data; performing image preprocessing on the historical image data; performing feature extraction on the historical image data after image preprocessing to obtain a corresponding three-dimensional feature map, finally converting the three-dimensional feature map into a two-dimensional feature map, and splicing the two-dimensional feature map corresponding to the historical image data and the two-dimensional feature map corresponding to the historical time series data to obtain a heterogeneous feature fused two-dimensional feature map; and inputting the heterogeneous feature fused two-dimensional feature map into the trained Bayesian neural network to obtain a predicted value and probability of the photovoltaic power generation power. According to the method, the limitation of a traditional prediction model in the aspects of feature selection and prediction results is overcome, so that the prediction accuracy and reliability are improved.

Description

Technical Field

[0001] The present invention relates to photovoltaic power prediction technology, and in particular to a photovoltaic power prediction method and system based on multi-source heterogeneous feature fusion. Background Art

[0002] In the context of today's global energy transformation and the rapid development of renewable energy, photovoltaic power generation has received widespread attention as an important clean energy source. Photovoltaic power generation power prediction is an important link to ensure the normal operation and optimized dispatch of the power grid, and its prediction accuracy is crucial to the safety and economic operation of the power grid. Traditional photovoltaic power generation power prediction technology mainly relies on deterministic single-point prediction methods, which provide power output predictions at a certain moment by real-time monitoring and analysis of meteorological conditions. However, the inherent defect of single-point prediction is that it is difficult to accurately capture the randomness and uncertainty of photovoltaic power generation, and the results often cannot meet the needs of practical applications. Since photovoltaic power generation is affected by a variety of meteorological factors, such as light intensity, temperature, wind speed, and cloud changes, this deterministic prediction method can easily lead to a large deviation between the prediction results and the actual power generation capacity, which brings certain difficulties to the dispatch and management of the power grid.

[0003] In recent years, probabilistic prediction methods for ultra-short-term power forecasting of photovoltaic power generation have gradually emerged. The core of these methods is to provide an uncertainty description of the output power, provide more information and support for decision makers, and to a certain extent help evaluate the distribution of power generation. However, most of the methods currently used for ultra-short-term probabilistic prediction of photovoltaic power generation rely on specific meteorological data from a single source. For example, many studies only use ground meteorological observation data or forecast results of specific meteorological models. The use of this single data source limits the diversity of feature extraction and cannot fully consider the complexity and variability of the photovoltaic power generation process, resulting in limited room for improvement in prediction accuracy. At the same time, it limits the model's ability to adapt to different meteorological conditions and environmental changes, and also restricts the reliability of the prediction results, making it impossible to fully reflect the complexity of photovoltaic power generation. In addition, in recent years, some studies have attempted to use only satellite cloud images for photovoltaic prediction. Although satellite cloud images can provide a wide range of cloud information, due to their limited temporal and spatial resolution, the details and local features of the data are difficult to be accurately captured, which to a certain extent affects the prediction effect of photovoltaic power generation. Similarly, relying solely on ground cloud maps for photovoltaic forecasts also faces shortcomings. Although ground cloud maps captured by all-sky imagers can provide real-time local weather conditions and cloud dynamics, data may be missing in bad weather, limiting its effectiveness in photovoltaic power generation forecasts. Summary of the invention

[0004] Technical problem to be solved by the present invention: In view of the above-mentioned problems in the prior art, a photovoltaic power generation prediction method and system based on multi-source heterogeneous feature fusion is provided, which can effectively combine and extract potential patterns and features in multi-source heterogeneous meteorological data and historical data, and obtain the probability distribution interval of photovoltaic ultra-short-term power generation, thereby overcoming the limitations of traditional prediction models in feature selection and prediction results, thereby improving the accuracy and reliability of prediction.

[0005] In order to solve the above technical problems, the technical solution adopted by the present invention is:

[0006] A photovoltaic power generation power prediction method based on multi-source heterogeneous feature fusion includes the following steps:

[0007] Obtain historical time series data and historical image data of photovoltaic power plants;

[0008] Extracting features from the historical time series data to obtain a two-dimensional feature graph corresponding to the historical time series data;

[0009] Performing image preprocessing on the historical image data;

[0010] Perform feature extraction on the historical image data after image preprocessing to obtain a corresponding three-dimensional feature map, convert the three-dimensional feature map of the historical image data into a corresponding two-dimensional feature map, and splice the two-dimensional feature map corresponding to the historical image data with the two-dimensional feature map corresponding to the historical time series data to obtain a two-dimensional feature map of heterogeneous feature fusion;

[0011] The two-dimensional feature map of heterogeneous feature fusion is input into the trained Bayesian neural network to obtain the predicted value and probability of photovoltaic power generation.

[0012] Furthermore, the historical time series data and historical image data specifically refer to all historical time series data and historical image data in a preset time period before the prediction moment. Before obtaining the historical time series data and historical image data of the photovoltaic power plant, it also includes: obtaining the time series data and image data in real time according to the specified sampling interval.

[0013] Furthermore, the historical time series data includes historical measured power generation data and historical measured weather data. When feature extraction is performed on the historical time series data, it specifically includes:

[0014] The historical measured power generation data and the historical measured weather data are combined into a data matrix, wherein the number of rows of the data matrix is ​​the number of sampling times in a preset time period before the prediction time, and the number of columns of the data matrix is ​​the number of features in the historical measured power generation data and the historical measured weather data;

[0015] The trained TS2Vec representation model is used to perform feature representation on the data matrix to obtain a data matrix after the feature dimension is upgraded as a two-dimensional feature map corresponding to the historical time series data.

[0016] Furthermore, the historical image data includes historical satellite cloud images and historical ground cloud images. When the historical image data is subjected to image preprocessing, the images of the historical satellite cloud images and historical ground cloud images are specifically subjected to grayscale processing, and then the image sizes are adjusted to a uniform size.

[0017] Further, when performing feature extraction on the historical image data after image preprocessing, specifically, feature extraction is performed on the historical satellite cloud images and historical ground cloud images after image preprocessing, specifically including:

[0018] The preprocessed current image data is used as the initial feature map. Figure 1 The first path performs global average pooling and global maximum pooling in turn, and then calculates the channel attention map through a shared multi-layer perceptron. The channel attention map is used to multiply the initial feature map in another path to obtain the intermediate feature map.

[0019] The intermediate features Figure 1 The global average pooling and global maximum pooling are performed in the first and second paths respectively, and then the results of the global average pooling and the global maximum pooling are concatenated and passed through the convolution layer to obtain the spatial attention map. The spatial attention map is used to multiply the intermediate feature map in another path to obtain the final feature map.

[0020] Furthermore, when the two-dimensional feature map of historical image data is spliced ​​with the two-dimensional feature map corresponding to the historical time series data, specifically, the two-dimensional feature map corresponding to the historical image data and the two-dimensional feature map corresponding to the historical time series data are both added with a channel dimension of 1 as the third dimension, and data are spliced ​​on the third dimension to obtain a three-dimensional feature map of heterogeneous feature fusion, and then the three-dimensional feature map of heterogeneous feature fusion is converted into the corresponding two-dimensional feature map.

[0021] Furthermore, when the two-dimensional feature map of heterogeneous feature fusion is input into the trained Bayesian neural network, specifically, the vector after the two-dimensional feature map of heterogeneous feature fusion is flattened is input into the trained Bayesian neural network.

[0022] The present invention also proposes a photovoltaic power prediction system based on multi-source heterogeneous feature fusion, comprising a memory, a processor and a computer program stored in the memory, wherein the processor executes the computer program to implement any step of the photovoltaic power prediction method based on multi-source heterogeneous feature fusion.

[0023] The present invention also proposes a computer-readable storage medium having a computer program stored thereon, which, when executed by a processor, implements the steps of any one of the photovoltaic power prediction methods based on multi-source heterogeneous feature fusion.

[0024] The present invention also proposes a computer program product, comprising a computer program, which, when executed by a processor, implements the steps of any one of the photovoltaic power prediction methods based on multi-source heterogeneous feature fusion.

[0025] Compared with the prior art, the advantages of the present invention are:

[0026] The present invention combines multi-source heterogeneous time series data and image data such as historical measured weather data, historical measured power generation, historical satellite cloud images, and historical ground cloud images. After extracting features from the time series data and image data respectively, the feature maps of the time series data and image data are spliced ​​to achieve feature extraction and fusion of different types of data. A Bayesian neural network (BNN) model is used to perform probabilistic prediction of distributed photovoltaic power generation. The BNN model's good uncertainty modeling capability is used to provide a reliable confidence interval for power prediction. BRIEF DESCRIPTION OF THE DRAWINGS

[0027] Figure 1 The figure is a flow chart of a method according to an embodiment of the present invention.

[0028] Figure 2 This is a schematic diagram of TS2Vec representation of historical measured data and preprocessing of historical satellite cloud images and historical ground cloud images according to an embodiment of the present invention.

[0029] Figure 3 Schematic diagram of a multi-source heterogeneous feature fusion module according to an embodiment of the present invention.

[0030] Figure 4 This is a schematic diagram of the structure of the Bayesian neural network model. DETAILED DESCRIPTION

[0031] The present invention is further described below in conjunction with the accompanying drawings and specific preferred embodiments, but the protection scope of the present invention is not limited thereby.

[0032] Embodiment 1

[0033] This embodiment proposes a photovoltaic power generation prediction method based on multi-source heterogeneous feature fusion, which comprehensively utilizes various heterogeneous data such as historical measured weather data, historical measured power generation, historical satellite cloud images and historical ground cloud images, and constructs a multi-source heterogeneous feature fusion module to fully explore the complementary advantages of different types of data, improve the accuracy and robustness of the prediction model, and overcome the shortcomings of relying on a single meteorological data source. Figure 1As shown, the method of this embodiment includes the following steps:

[0034] S1) acquiring historical time series data and historical image data of the photovoltaic power plant. In this embodiment, the historical time series data and historical image data specifically refer to all historical time series data and historical image data of a preset time period before the prediction moment, wherein the historical time series data includes historical measured weather data and historical measured power generation, and the historical image data includes historical satellite cloud images and historical ground cloud images;

[0035] S2) extracting features from the historical time series data to obtain a two-dimensional feature graph corresponding to the historical time series data. In this embodiment, when extracting features from the historical time series data, a TS2Vec characterization model is used to characterize the data matrix containing the historical measured weather data and the historical measured power generation;

[0036] S3) performing image preprocessing on the historical image data, specifically performing image preprocessing on the historical satellite cloud images and the historical ground cloud images, and uniformly processing them into single-channel images of a preset size;

[0037] S4) extracting features from the historical image data after image preprocessing to obtain a corresponding three-dimensional feature map, converting the three-dimensional feature map of the historical image data into a corresponding two-dimensional feature map, and splicing the two-dimensional feature map corresponding to the historical image data with the two-dimensional feature map corresponding to the historical time series data to obtain a two-dimensional feature map of heterogeneous feature fusion. In this embodiment, this step is achieved by inputting the feature characterization result of the data matrix containing the historical measured weather data and the historical measured power generation together with the historical satellite cloud map and the historical ground cloud map after image preprocessing into a multi-source heterogeneous feature fusion module to form a data matrix;

[0038] S5) The two-dimensional feature map of heterogeneous feature fusion is input into the trained Bayesian neural network (BNN). In this embodiment, the BNN module is used to process the data matrix output by the multi-source heterogeneous feature fusion module and output the photovoltaic power generation power and its probability interval distribution within the ultra-short-term prediction time period to obtain the predicted value and probability of photovoltaic power generation power.

[0039] Through the above steps, the probability prediction model established in this embodiment can replace the traditional deterministic prediction method, calculate and output the probability distribution of photovoltaic power generation, quantitatively describe the uncertainty of power prediction, establish a probability prediction interval, and provide more abundant decision-making basis for power dispatching, effectively reducing the impact of photovoltaic power generation fluctuations on the power system.

[0040] Each step is described in detail below.

[0041] Since the historical time series data and historical image data in this embodiment specifically refer to all historical time series data and historical image data in a preset time period before the prediction moment, before step S1, it also includes: acquiring time series data and image data in real time according to a specified sampling interval. In this embodiment, the sampling interval is preferably 15 minutes, and the preset time period is preferably 6 hours before the prediction moment.

[0042] In step S2 of this embodiment, the TS2Vec characterization model is used to characterize the data matrix containing the historical measured weather data and the historical measured power generation. This embodiment combines the feature characterization of the historical measured weather data and the historical measured power generation to provide more comprehensive information support for photovoltaic power probability prediction. Feature characterization can extract potential patterns and features in time series data, thereby improving the performance of the prediction model. Furthermore, the historical time series data includes historical measured power generation data and historical measured weather data. When extracting features from the historical time series data, Figure 2 As shown, specifically including:

[0043] S21) combining the historical measured power generation data and the historical measured weather data into a data matrix, wherein the number of rows of the data matrix is ​​the number of sampling times in a preset time period before the prediction time, and the number of columns of the data matrix is ​​the number of features in the historical measured power generation data and the historical measured weather data;

[0044] In this embodiment, the sampling interval is preferably 15 minutes, the preset time period is preferably 6 hours before the prediction time, and the historical measured weather data includes historical measured irradiance, historical measured temperature, historical measured humidity, historical measured wind speed and historical measured wind direction data. Therefore, the data matrix obtained by combining the historical measured power generation data with the historical measured weather data is a 24×6 data matrix, and the number of rows (24) indicates that 24 samples are taken within 6 hours, and the number of columns (6) indicates that each sampling records the data of 5 characteristic sources in the historical measured weather data and 1 characteristic source of the historical measured power generation data, a total of 6 characteristic sources;

[0045] S22) Use the trained TS2Vec representation model to perform feature representation on the data matrix to obtain a data matrix after the feature dimension is upgraded as a two-dimensional feature map corresponding to the historical time series data.

[0046] In this embodiment, the feature dimension of the data of size 24×6 is increased to 224 after being represented by the TS2Vec representation module, and a two-dimensional feature map of size 24×224 is obtained. TS2Vec (Time Series to Vector) is a model for learning time series representation. Its goal is to convert time series data into a vector representation of a fixed length to facilitate various subsequent tasks. Convolutional Neural Networks (CNN) or Recurrent Neural Networks (RNN) are usually used to extract time series features.

[0047] In this embodiment, the TS2Vec characterization model includes an input layer, a hidden layer, and an output layer. The input layer accepts a 24×6 input matrix. The hidden layer uses a multi-layer convolutional network to process the input data, gradually extracts complex features in the time series, and generates a feature representation for each time step. The output layer maps the output of the hidden layer to the target feature dimension, that is, 224 dimensions. The output layer dimension is 24×224, that is, a 224-dimensional feature vector is generated for each time step.

[0048] The TS2Vec characterization model in this embodiment can perform self-supervised learning by constructing positive and negative sample pairs, where the positive sample pair refers to adjacent time steps in the same time series, and the negative sample pair refers to the time steps in different time series. At the same time, the TS2Vec characterization model uses a loss function such as InfoNCE to ensure that the features of adjacent time steps in the same time series are more similar. Adam or similar optimization algorithms are used for model training. Monitor the training and validation losses during training and adjust the model hyperparameters when necessary. The trained model can convert a 24×6 input matrix into a 24×224 two-dimensional feature map. After the input of each time step passes through the model, a 224-dimensional feature vector is output.

[0049] In step S3 of this embodiment, when the historical image data is preprocessed, specifically, the image preprocessing module is used to grayscale the images of the historical satellite cloud images and the historical ground cloud images, and then the image sizes are adjusted to a uniform size.

[0050] like Figure 2 As shown, in this embodiment, the sampling interval is preferably 15 minutes, and the preset time period is preferably 6 hours before the prediction time. Therefore, the historical satellite cloud map and the historical ground cloud map are 24 three-channel color pictures with a size of H×W with a sampling interval of 15 minutes within the past 6 hours from the current time. The image preprocessing module first performs image grayscale processing to transform the original color three-channel image into a single-channel image, and then transforms the image size to 24×224×224.

[0051] like Figure 3As shown, in step S4 of this embodiment, when extracting features from the historical image data after image preprocessing, a parallel neural network is specifically designed to extract features from the historical satellite cloud images and historical ground cloud images after image preprocessing. In this embodiment, the parallel neural network includes a CBAM (Convolutional Block Attention Module) convolutional attention module composed of a channel attention module and a spatial attention module, wherein the channel attention module performs global average pooling and global maximum pooling on the input, compresses the features of each channel into a real number, and then obtains a channel attention map through a shared multi-layer perceptron and a sigmoid function, and multiplies the channel attention map and the input of the channel attention module to obtain the output of the channel attention module. The output of the channel attention module will be used as the input of the spatial attention module, and then global average pooling and global maximum pooling will be performed. The results of the two poolings will be spliced ​​and passed through a convolution layer and a sigmoid function to generate a spatial attention map. The spatial attention map and the output of the channel attention module will be multiplied to obtain the output of the CBAM convolutional attention module. The size of the feature map after the CBAM convolutional attention module is consistent with the size of the input historical satellite cloud map and historical ground cloud map after image preprocessing. Therefore, when extracting features from the historical satellite cloud map and historical ground cloud map after image preprocessing, it specifically includes:

[0052] S41) using the preprocessed current image data as an initial feature map. In this embodiment, the size of the initial feature map is an image of 24×224×224;

[0053] S42) Initial feature Figure 1 The path performs global average pooling and global maximum pooling in turn, and then the channel attention map is calculated through a shared multi-layer perceptron;

[0054] In this embodiment, for the initial feature map, global average pooling and maximum pooling are performed on each channel in the spatial dimension, and the expression is as follows:

[0055] V avg =GAP(F),V max =GMP(F)

[0056] Among them, Vavg represents the result of global average pooling, and Vmax represents the result of maximum pooling. Both are input into a shared multi-layer perceptron (MLP), and a channel attention weight scalar M is generated through nonlinear activation (such as ReLU) and Sigmoid function. c , the expression is as follows:

[0057] M c =σ(MLP(V avg ))+MLP(Vmax )

[0058] S43) Use the channel attention map to multiply the initial feature map in another way to obtain the intermediate feature map, which is expressed as follows:

[0059] F ′ =M c ·F

[0060] Among them, F represents the initial feature map, F ′ Represents the intermediate feature map, with a size of 24×224×224;

[0061] S44) The intermediate feature Figure 1 The global average pooling and global maximum pooling are performed on the first and second paths respectively, and then the results of global average pooling and global maximum pooling are concatenated and passed through the convolution layer to obtain the spatial attention map;

[0062] In this embodiment, average pooling and maximum pooling are performed on the channel dimension to obtain two feature maps of size 24×224×224, which are the processing results of the complete image itself for a single-channel image. The two feature maps are concatenated on the channel dimension, and then a 7×7 convolution kernel is used to convolve the concatenated feature maps to generate a spatial attention map M. s , the expression is as follows:

[0063] M s =σ(f 7×7 [A avg ; A max ])

[0064] Among them, [A avg ; A max ] is the concatenation of the average pooling result and the maximum pooling result, A avg is the average pooling result, A max is the maximum pooling result;

[0065] S45) Use the spatial attention map to multiply the intermediate feature map in another way to obtain the final feature map, which is expressed as follows:

[0066] F″=M s ⊙F′

[0067] Among them, ⊙ represents element-by-element multiplication, and F″ represents the final feature map, which is still of size 24×224×224.

[0068] like Figure 3As shown, in step S4 of this embodiment, when converting the three-dimensional feature map of the historical image data into the corresponding two-dimensional feature map, specifically, the feature map after passing through the CBAM convolutional attention module is transformed into a three-dimensional feature map of size 24×224×224 by using a Linear linear fully connected layer through a linear connection transformation module, and then the feature map is transformed into a two-dimensional feature map of size 24×224 by removing the channel dimension of size 1.

[0069] Correspondingly, in step S4 of this embodiment, when the two-dimensional feature map of the historical image data is spliced ​​with the two-dimensional feature map corresponding to the historical time series data, specifically, the two-dimensional feature map (size is 24×224) corresponding to the historical satellite cloud map and the historical ground cloud map in the historical image data after the linear connection transformation module and the two-dimensional feature map (size is 24×224) corresponding to the historical time series data represented by the TS2Vec characterization model are both increased by a channel dimension of size 1 as the third dimension, and data are spliced ​​on the third dimension to obtain a three-dimensional feature map (size is 24×224×3) of heterogeneous feature fusion, and then the three-dimensional feature map of heterogeneous feature fusion is converted into the corresponding two-dimensional feature map, specifically, the three-dimensional feature map of size 24×224×3 is transformed into 24×224×1 through the linear connection transformation module using the Linear linear fully connected layer, and then the feature map is transformed into a two-dimensional feature map of size 24×224 by removing the channel dimension of size 1.

[0070] In step S5 of this embodiment, the BNN module is used to process the data matrix of size 24×224 output by the multi-source heterogeneous feature fusion module and output the photovoltaic power generation power and its probability interval distribution within the ultra-short-term prediction period. When the two-dimensional feature map of the heterogeneous feature fusion is input into the trained Bayesian neural network, the vector after the two-dimensional feature map of the heterogeneous feature fusion is flattened is specifically input into the trained Bayesian neural network.

[0071] In this embodiment, the data matrix of size 24×224 from the multi-source heterogeneous feature fusion module is flattened to obtain a one-dimensional vector of length 5376. The structure of the Bayesian neural network is as follows: Figure 4 As shown, the input layer obtains a vector of length 5376 and directly passes it as the input of the hidden layer. The length of the vector output by the output layer is the number of photovoltaic power samples in the next hour, that is, the length of 4 sampling intervals.

[0072] The number of neurons in the hidden layer of the Bayesian neural network of this embodiment is set to 256. After the hidden layer performs forward propagation calculation on the vector of length 5376 transmitted from the input layer, the output vector of length 256 is expressed as follows:

[0073]

[0074] Where f() is the activation function of the hidden layer (e.g. ReLU), x i is the input feature, b j is the bias, W ij is the weight, i represents the input unit, and j represents the hidden layer unit.

[0075] The output layer of the Bayesian neural network of this embodiment obtains a vector of length 256 transmitted from the hidden layer. The number of neurons in the output layer is set to 4, corresponding to 4 sampling intervals in the next hour. The output length of the vector is 4, which represents the predicted value of photovoltaic power generation. The expression is as follows:

[0076]

[0077] Among them, g() is the activation function of the output layer, which can be a linear function, and b k is the bias, W jk is the weight, j represents the hidden layer unit, and k represents the output unit.

[0078] Each weight between the neurons in the hidden layer and the output layer of the Bayesian neural network of this embodiment is regarded as obeying a Gaussian distribution with a mean of μ and a variance of σ, but obeying different Gaussian distributions, multiple prediction results can be obtained by sampling multiple times from each Gaussian distribution to obtain different weight values ​​when performing probability prediction, thereby obtaining prediction probability intervals under different confidence levels. Specifically, the following steps are included:

[0079] S51) obtaining input features and corresponding true labels as training data, then using variational inference or Monte Carlo method to update μ and σ parameters of each weight, and finally outputting optimized weight distribution parameters;

[0080] S52) obtaining input features as test data, sampling multiple times from the Gaussian distribution of each weight, and performing forward propagation on the weight set obtained by each sampling, obtaining multiple prediction results and forming a prediction value distribution;

[0081] S53) According to the multiple prediction results, the prediction mean and standard deviation are calculated, and the confidence interval of the prediction value under different confidence levels is calculated based on this, and the expression is as follows:

[0082]

[0083] in, represents the mean of the predicted values, z α / 2 represents the critical value of the standard normal distribution, corresponding to the confidence level 1-α, σ y Represents the standard deviation of the predicted values.

[0084] Embodiment 2

[0085] This embodiment proposes a photovoltaic power prediction system based on multi-source heterogeneous feature fusion, including a memory, a processor and a computer program stored in the memory, and the processor executes the computer program to implement the steps of the photovoltaic power prediction method based on multi-source heterogeneous feature fusion of embodiment one.

[0086] This embodiment also proposes a computer-readable storage medium on which a computer program is stored. When the computer program is executed by a processor, the steps of the photovoltaic power prediction method based on multi-source heterogeneous feature fusion of the first embodiment are implemented.

[0087] This embodiment also proposes a computer program product, including a computer program, which, when executed by a processor, implements the steps of the photovoltaic power prediction method based on multi-source heterogeneous feature fusion in the first embodiment.

[0088] In summary, the present invention discloses a photovoltaic power prediction method and system based on multi-source heterogeneous feature fusion, which combines multi-source heterogeneous time series data and image data such as historical measured weather data, historical measured power generation, historical satellite cloud images, and historical ground cloud images to construct a multi-source heterogeneous feature fusion module to realize feature extraction and fusion of different types of data, and adopts a Bayesian neural network (BNN) model to perform probability prediction of ultra-short-term power of distributed photovoltaic power generation. Relying on the good uncertainty modeling ability of the BNN model, a reliable confidence interval is provided for power prediction. While effectively combining and extracting potential patterns and features in multi-source heterogeneous meteorological data and historical data, the probability distribution interval of photovoltaic ultra-short-term power generation is obtained, which overcomes the limitations of traditional prediction models in feature selection and prediction results, thereby improving the accuracy and reliability of prediction.

[0089] The above is only a preferred embodiment of the present invention, and the protection scope of the present invention is not limited to the above embodiments. All technical solutions under the concept of the present invention belong to the protection scope of the present invention. It should be pointed out that for ordinary technicians in this technical field, some improvements and modifications without departing from the principle of the present invention should also be regarded as the protection scope of the present invention.

Claims

1. A photovoltaic power generation prediction method based on multi-source heterogeneous feature fusion, characterized in that: The following steps are involved: Obtain historical time series data and historical image data of photovoltaic power plants; Extracting features from the historical time series data to obtain a two-dimensional feature graph corresponding to the historical time series data; Performing image preprocessing on the historical image data; Perform feature extraction on the historical image data after image preprocessing to obtain a corresponding three-dimensional feature map, convert the three-dimensional feature map of the historical image data into a corresponding two-dimensional feature map, and splice the two-dimensional feature map corresponding to the historical image data with the two-dimensional feature map corresponding to the historical time series data to obtain a two-dimensional feature map of heterogeneous feature fusion; The two-dimensional feature map of heterogeneous feature fusion is input into the trained Bayesian neural network to obtain the predicted value and probability of photovoltaic power generation.

2. The photovoltaic power generation prediction method based on multi-source heterogeneous feature fusion according to claim 1 is characterized in that: The historical time series data and historical image data specifically refer to all historical time series data and historical image data in a preset time period before the prediction moment. Before obtaining the historical time series data and historical image data of the photovoltaic power plant, it also includes: obtaining the time series data and image data in real time according to the specified sampling interval.

3. The photovoltaic power generation prediction method based on multi-source heterogeneous feature fusion according to claim 1 is characterized in that: The historical time series data includes historical measured power generation data and historical measured weather data. When feature extraction is performed on the historical time series data, it specifically includes: The historical measured power generation data and the historical measured weather data are combined into a data matrix, wherein the number of rows of the data matrix is ​​the number of sampling times in a preset time period before the prediction time, and the number of columns of the data matrix is ​​the number of features in the historical measured power generation data and the historical measured weather data; The trained TS2Vec representation model is used to perform feature representation on the data matrix to obtain a data matrix after the feature dimension is upgraded as a two-dimensional feature map corresponding to the historical time series data.

4. The photovoltaic power generation prediction method based on multi-source heterogeneous feature fusion according to claim 1 is characterized in that: The historical image data includes historical satellite cloud images and historical ground cloud images. When the historical image data is subjected to image preprocessing, the images of the historical satellite cloud images and historical ground cloud images are subjected to grayscale processing, and then the image sizes are adjusted to a uniform size.

5. The photovoltaic power generation prediction method based on multi-source heterogeneous feature fusion according to claim 4 is characterized in that: When performing feature extraction on the historical image data after image preprocessing, the feature extraction is performed on the historical satellite cloud images and historical ground cloud images after image preprocessing respectively, which specifically includes: The preprocessed current image data is used as the initial feature map, and the initial feature map is subjected to global average pooling and global maximum pooling in sequence. Then, the channel attention map is calculated through a shared multi-layer perceptron, and the channel attention map is used to multiply the initial feature map in another way to obtain an intermediate feature map. The intermediate feature map is subjected to global average pooling and global maximum pooling respectively, and then the results of global average pooling and global maximum pooling are concatenated and passed through a convolutional layer to obtain a spatial attention map. The spatial attention map is used to multiply the intermediate feature map in another way to obtain the final feature map.

6. The photovoltaic power generation prediction method based on multi-source heterogeneous feature fusion according to claim 1 is characterized in that: When splicing the two-dimensional feature map of historical image data with the two-dimensional feature map corresponding to the historical time series data, specifically, the two-dimensional feature map corresponding to the historical image data and the two-dimensional feature map corresponding to the historical time series data are both increased with a channel dimension of 1 as the third dimension, and data are spliced ​​on the third dimension to obtain a three-dimensional feature map of heterogeneous feature fusion, and then the three-dimensional feature map of heterogeneous feature fusion is converted into the corresponding two-dimensional feature map.

7. The photovoltaic power generation prediction method based on multi-source heterogeneous feature fusion according to claim 1 is characterized in that: When the two-dimensional feature map of heterogeneous feature fusion is input into the trained Bayesian neural network, specifically, the vector after flattening the two-dimensional feature map of heterogeneous feature fusion is input into the trained Bayesian neural network.

8. A photovoltaic power generation prediction system based on multi-source heterogeneous feature fusion, comprising a memory, a processor and a computer program stored in the memory, characterized in that: The processor executes the computer program to implement the steps of the photovoltaic power prediction method based on multi-source heterogeneous feature fusion as described in any one of claims 1 to 7.

9. A computer-readable storage medium having a computer program stored thereon, characterized in that: When the computer program is executed by a processor, the steps of the photovoltaic power prediction method based on multi-source heterogeneous feature fusion as described in any one of claims 1 to 7 are implemented.

10. A computer program product, comprising a computer program, characterized in that When the computer program is executed by a processor, the steps of the photovoltaic power prediction method based on multi-source heterogeneous feature fusion as described in any one of claims 1 to 7 are implemented.

Images