Photovoltaic power generation prediction method and device, and storage medium

Abstract

The invention relates to the technical field of photovoltaic power generation prediction, in particular to a photovoltaic power generation prediction method and device and a storage medium. The photovoltaic power generation prediction method comprises the following steps: inputting historical real weather features and historical real power generation into a trained teacher model to obtain teacher model feature vectors and teacher model predicted power generation; inputting the historical predicted weather features and the historical real generating capacity into a to-be-trained student model to obtain a student model feature vector and a student model predicted generating capacity; training a student model according to the teacher model feature vector, the teacher model predicted generating capacity, the student model feature vector, the student model predicted generating capacity and the historical real generating capacity to obtain a trained student model; and predicting photovoltaic power generation according to the trained student model. According to the method provided by the invention, when the relationship between the predicted weather characteristics and the power generation amount is learned, a real weather assisted learning method is added, and the robustness and accuracy of the model are improved.

Description

Technical Field

[0001] The present application relates to the technical field of photovoltaic power generation prediction, and in particular to a photovoltaic power generation prediction method, device, and storage medium. Background Art

[0002] As a clean and renewable energy source, solar energy has received increasing attention and emphasis worldwide. With the continuous development and application of photovoltaic technology, photovoltaic power generation has become an important part of the new energy field. To maintain the balance and stability of power consumption and energy storage, accurate prediction of photovoltaic power generation is crucial. The accuracy of photovoltaic power generation prediction directly depends on the accuracy of weather forecasts (radiation / temperature / humidity / wind speed / cloud cover, etc.). Due to the inherent errors in weather forecasts, the prediction of photovoltaic power generation based on these weather characteristics is also inaccurate, with relatively high errors.

[0003] With the development of artificial intelligence, traditional machine learning and deep learning have become a common method for photovoltaic power generation prediction. In recent years, the mainstream approach has been to perform (without considering predicted weather) based on historical data of photovoltaic power generation and relatively accurate weather data corrected by a weather platform during the model design and optimization phase. However, in actual implementation, modeling and prediction are carried out using predicted weather data. This approach fails to take into account the significant differences between predicted weather and actual weather, resulting in relatively large prediction errors during implementation. Summary of the Invention

[0004] Embodiments of the present application aim to provide a photovoltaic power generation prediction method, device, and storage medium to solve the problem in the prior art that the photovoltaic power generation prediction model fails to consider the significant differences between predicted weather and actual weather, resulting in errors in prediction results.

[0005] To solve the above technical problems, the embodiments of the present application provide the following technical solutions: According to a first aspect of the present application, there is provided a photovoltaic power generation prediction method, including: Obtaining historical predicted weather characteristics, historical actual weather characteristics, and historical actual power generation; Inputting the historical actual weather characteristics and the historical actual power generation into a trained teacher model to obtain a teacher model feature vector and a teacher model predicted power generation; Inputting the historical predicted weather characteristics and the historical actual power generation into a student model to be trained to obtain a student model feature vector and a student model predicted power generation; Calculating a multi - variable total loss based on the teacher model feature vector, the teacher model predicted power generation, the student model feature vector, the student model predicted power generation, and the historical actual power generation to train the student model, and obtaining a trained student model; Perform prediction on photovoltaic power generation according to the trained student model.

[0006] Optionally, the teacher model is trained based on historical real weather features.

[0007] Optionally, calculating the multi - variable total loss according to the teacher model feature vector, the predicted power generation of the teacher model, the student model feature vector, the predicted power generation of the student model, and the historical real power generation includes: Calculate the distillation loss according to the teacher model feature vector and the student model feature vector; Calculate the soft - label loss according to the predicted power generation of the student model and the predicted power generation of the teacher model; Calculate the real - label loss according to the predicted power generation of the student model and the historical real power generation; Calculate the multi - variable total loss based on the distillation loss, the soft - label loss, and the real - label loss.

[0008] Optionally, the calculation formula for the multi - variable total loss is:

[0009] Where, is the multi - variable total loss, is the distillation loss, is the soft - label loss, is the real - label loss, , and are weight coefficients, is the probability distribution of the teacher model feature vector, is the probability distribution of the student model feature vector, is the number of samples, is the th true value of the sample, is the predicted value of the teacher model for the th sample, is the predicted value of the student model for the th sample.

[0010] Optionally, the teacher model and the student model have the same network structure.

[0011] Optionally, the network structure includes a global feature module, a feature pyramid module, a time - series module, a multi - head attention module, and a high - level feature expression module connected in sequence.

[0012] Optionally, the global feature module uses dilated convolution and large convolution kernels to extract features.

[0013] Optionally, the feature pyramid module inputs the features extracted in the previous stage into multiple branches. After each branch extracts features using dilated convolutions with different dilation rates, the features of multiple branches are concatenated and superimposed by channel.

[0014] According to a second aspect of the present application, there is provided an electronic device, including at least one processor and a memory communicatively connected to the at least one processor. The memory stores instructions executable by the at least one processor. When the instructions are executed by the at least one processor, the at least one processor is enabled to execute the photovoltaic power generation prediction method described above.

[0015] According to a third aspect of the present application, there is provided a computer storage medium storing instructions or programs. When the instructions or programs are executed by at least one processor, the at least one processor is enabled to execute the photovoltaic power generation prediction method described above.

[0016] The beneficial effects of the embodiments of the present application are as follows: Different from the prior art, in the embodiments of the present application, a photovoltaic power generation prediction method is provided. First, historical predicted weather features, historical real weather features, and historical real power generation are obtained; then, the historical real weather features and historical real power generation are input into a trained teacher model to obtain a teacher model feature vector and a teacher model predicted power generation, and the historical predicted weather features and historical real power generation are input into a student model to be trained to obtain a student model feature vector and a student model predicted power generation; then, a multi-source total loss is calculated based on the teacher model feature vector, the teacher model predicted power generation, the student model feature vector, and the student model predicted power generation to train the student model, and a trained student model is obtained; finally, the photovoltaic power generation is predicted according to the trained student model. The method of the present application adds a learning method assisted by real weather when learning the relationship between predicted weather features and power generation, improves the robustness of the model, and reduces the problem of cross-domain accuracy. Description of the Drawings

[0017] One or more embodiments are exemplarily illustrated by pictures in the corresponding drawings. These exemplary illustrations do not limit the embodiments. Elements with the same reference numerals in the drawings are represented as similar elements, unless otherwise stated, and the drawings in the figures do not constitute a proportional limitation.

[0018] Figure 1 Exemplarily shows a training framework diagram of a teacher model; Figure 2 Exemplarily shows a training framework diagram of a student model; Figure 3 Exemplarily shows a module diagram of a feature extraction network structure; Figure 4 Exemplarily shows the structural diagram of the global feature module; Figure 5 Exemplarily shows the structural diagram of the feature pyramid module; Figure 6 Exemplarily shows the schematic flowchart of a photovoltaic power generation prediction method; Figure 7 Exemplarily shows the structural diagram of an electronic device. Detailed implementation manners

[0019] To make the objectives, technical solutions and advantages of the embodiments of the present application clearer, the technical solutions in the embodiments of the present application will be clearly and completely described below with reference to the accompanying drawings in the embodiments of the present application. Obviously, the described embodiments are some but not all of the embodiments of the present application. All other embodiments obtained by those of ordinary skill in the art based on the embodiments of the present application without creative efforts shall fall within the protection scope of the present application.

[0020] In addition, the technical features involved in the various embodiments of the present application described below can be combined with each other as long as they do not conflict with each other.

[0021] 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 the logical order is shown in the flowchart, in some cases, the steps shown or described can be executed in a different order than here.

[0022] The key terms in the present application are explained as follows: Historical true weather features: The weather features finally and truly recorded at a certain past time. For example: Assume that the current is day T, and the meteorological features that actually occurred on day T-2 have been truly recorded by various meteorological sensors. Then the records of these true meteorological conditions on day T-2 are called "historical true" weather.

[0023] Historical predicted weather features: The predicted values of future weather. For example: Assume that the current is day T, and on day T-3, the weather on day T-2 was predicted (weather forecast). Then standing on day T, both day T-3 and day T-2 are past history. Therefore, the predicted weather on day T-3 for day T-2 is called "historical predicted" weather.

[0024] Prelu (Parametric rectified linear unit) activation function: An improved ReLU activation function that automatically adjusts the slope of the negative interval by introducing a learnable parameter, solving the problem of the gradient being 0 in the negative interval of the traditional ReLU ("neuron death" problem).

[0025] An embodiment of the present application proposes a photovoltaic power generation prediction method based on a knowledge transfer model, and the knowledge transfer model includes a student model and a teacher model. Among them, the student model and the teacher model have the same network structure and different input features. By transferring the knowledge of the pre-trained teacher model to the student model, the student model can obtain more knowledge from the teacher model, thereby improving the robustness and accuracy of photovoltaic power generation prediction. Specifically, the teacher model is trained with relatively high-quality real weather features, and the accuracy and robustness of the model are relatively high. The student model is trained based on predicted weather features, and the feature vector and prediction result output by the teacher model are used as the training objectives of the student model, so that the student model can pay attention to the differences between the predicted weather features and the real weather features, obtain more knowledge, and thus have high robustness in future power generation predictions.

[0026] Please refer to Figure 1 , which shows a training framework diagram of a teacher model. As Figure 1 shown, the training process of the teacher model is as follows: (1) Input the real weather features into the teacher model; (2) After the real weather features are extracted by the feature extraction network structure, they are input into the high-level feature expression module for compression processing to obtain the teacher model feature vector; (3) Obtain the teacher model prediction result (i.e., the predicted power generation of the teacher model) based on the teacher model feature vector; (4) Calculate the loss based on the teacher model prediction result and the real label (i.e., the historical real power generation corresponding to the real weather features), and train the teacher model based on this loss to obtain the trained teacher model. Among them, the real weather features are relatively accurate weather features corrected by history. The teacher model feature vector is a feature vector with a preset fixed length, and the main purpose is to realize the knowledge transfer between the teacher model and the student model.

[0027] Please refer to Figure 2 , which shows a training framework diagram of a student model. As Figure 2 shown, the training process of the student model is as follows: (1) Load Figure 1The pre-trained teacher model; (2) Input the predicted weather features into the student model, and input the true weather features corresponding to the predicted weather features into the teacher model; (3) The teacher model obtains the teacher model feature vector and the teacher model prediction result based on the true weather features, and the student model obtains the student model feature vector and the student model prediction result based on the predicted weather features; (4) Calculate the multi-source total loss based on the distillation loss between the teacher model feature vector and the student model feature vector, the soft label loss between the teacher model prediction result and the student model prediction result, and the true label loss between the student model prediction result and the true label, and train the student model based on this multi-source total loss to obtain the trained student model. It should be noted that when training the student model, the teacher model only performs inference and does not perform training. In the process of training the student model, using a multi-source loss function can enable the student model to learn knowledge in different domains and improve the robustness of the model. It can be understood that the weather features used to train the teacher model can be in different time periods from the weather features used to train the student model. For example, assuming it is September currently, the weather features used to train the teacher model can be the historical true weather features from January to June, and the weather features used to train the student model are the historical predicted weather features and historical true weather features from May to August.

[0028] Please refer to Figure 3 , which shows a module diagram of the feature extraction network structure. As Figure 3 shown, the feature extraction network structure mainly includes a global feature module, a feature pyramid module, a time series module, and a multi-head attention module. Among them, the global feature module uses dilated convolution and large convolution kernels to extract features to increase the receptive field and capture multi-scale context information. The feature pyramid module uses dilated convolutional layers with different dilation rates to process the input feature map, which enables the network to capture combinations of weather features at different scales, effectively increasing the receptive field of the model without significantly increasing the computational cost.

[0029] Please refer to Figure 4 , which shows a structural diagram of the global feature module. As Figure 4 shown, the processing process of the global feature module is as follows: (1) Input the weather features, that is, the true weather features or the predicted weather features. (2) Perform dilated convolution, where the convolution kernel size is 15 and the dilation rate is 1; (3) Perform batch normalization layer processing; (4) Perform non-linear mapping through the Prelu activation function; (5) Perform dilated convolution again, where the convolution kernel size is 9 and the dilation rate is 1; (6) Perform batch normalization layer processing again; (4) Perform non-linear mapping through the Prelu activation function again.

[0030] Please refer to Figure 5 , which shows a structural diagram of the feature pyramid module. As Figure 5As shown in Figure 1, the processing process of the feature pyramid module is as follows: (1) input the features extracted in the previous stage; (2) input them to multiple branches in parallel, each of which contains a hole convolution layer, a batch normalization layer and a Prelu activation function. For example, Figure 5 The features extracted in the previous stage are input into four branches. The size of the dilated convolution kernel of the first branch is 1, and the dilation rate is 1; the size of the dilated convolution kernel of the second branch is 3, and the dilation rate is 12; the size of the dilated convolution kernel of the third branch is 3, and the dilation rate is 24; the size of the dilated convolution kernel of the fourth branch is 3, and the dilation rate is 36; (3) The features of the four branches are concatenated and superimposed by channel.

[0031] The time series module is used to extract time series features of historical weather features based on the long short-term memory network (LSTM) to capture the changing relationship between photovoltaic power generation and historical weather features in the time dimension. Through the time series module, the model can better focus on the following features: 1) the pattern information of photovoltaic power generation changing over time; 2) the periodic characteristics of photovoltaic power generation, such as the seasonal variation characteristics of photovoltaic power generation, and the long-term trend of photovoltaic power generation; 3) the continuity relationship between photovoltaic power generation in time; 4) reduce the problems of gradient hour or gradient explosion in calculation. In one embodiment, the time series module adopts a two-layer LSTM structure.

[0032] The multi-head attention module is used to focus on different parts of the input sequence from multiple different angles or representation spaces. The features and models learned by each attention head may be different, which helps the model to fully understand the data. By using multiple independent attention heads, the model can learn more diverse feature combinations, each focusing on different information, and finally merging this information, thereby enhancing the overall expressiveness of the model. Since the attention mechanism can directly connect any two positions in the sequence, it is particularly suitable for modeling long-distance dependencies. This is very useful for capturing long-term patterns such as seasonality and trends in time series. In one embodiment, the multi-head attention module uses 10 attention heads.

[0033] In an embodiment of the present application, introducing a combination of a time series module and a multi-head attention module into the feature extraction network structure can enable the model to better pay attention to the characteristics of time and space, thereby improving the robustness of the model.

[0034] The high-level feature expression module uses a fully connected layer to convert the input high-dimensional feature vector into a low-dimensional feature vector with a fixed length through a series of linear transformations. Since the fully connected layer linearly combines all the features extracted by the feature extraction network structure, it can play a role in feature compression, condensing complex feature information into a relatively low-dimensional space, thereby realizing the knowledge transfer between the teacher model and the student model.

[0035] In one embodiment, between the feature pyramid module and the time series module, there is also a point convolution module. The point convolution module performs point convolution on the feature map output by the feature pyramid module to reduce the number of channels of the feature map, thereby reducing the computational complexity of the subsequent layers. The 1*1 convolution can perform linear combination between different channels to achieve cross-channel information fusion, which is beneficial to enhancing the expression ability of features.

[0036] Taking the historical weather features of one day as an example below, the processing process of the historical weather features by the teacher model or the student model is described as follows: Step 1, input historical weather features with a dimension of (1, 48, 200).

[0037] Among them, 1 represents the input time unit of days, 48 represents that each hour in 12 hours of the day is calculated as 48 intervals according to 15 minutes, and 200 is the weather feature.

[0038] Step 2, extract features through the feature extraction network structure, including: a) Extract features through the global feature module, and output a feature map with a dimension of (1, 192, 50); b) Extract features through the feature pyramid module, and each branch outputs a feature map with a dimension of (1, 192, 50); c) Combine the four branches by channels, and output a feature map with a dimension of (1, 768, 50); d) Perform dimensionality reduction through point convolution, and output a feature map with a dimension of (1, 192, 50); e) Extract features through LSTM, and output a feature map with a dimension of (1, 192, 100); f) Extract features through multi-head attention, and output a feature map with a dimension of (1, 100); g) Extract features through the high-level feature expression module, and output a feature map with a dimension of (1, 96).

[0039] Step 3, output the prediction result, a feature map with a dimension of (1, 48).

[0040] Please refer to Figure 6 , Figure 6 which shows a schematic flowchart of a photovoltaic power generation prediction method. The method includes: S601, Obtain historical predicted weather features, historical actual weather features, and historical actual power generation.

[0041] Specifically, collect historical predicted meteorological data of a photovoltaic power station, historical actual meteorological data after correcting the historical predicted meteorological data, historical power generation power data, relevant time data, etc., to obtain historical predicted weather features with dimensions of A B C, historical actual weather features with dimensions of A B C, and historical actual power generation with dimensions of A B. Among them, A is the number of days, B is the number of sunshine time intervals per day, and C is the number of weather features affecting power generation. Weather features affecting power generation may include temperature, radiation, air pressure, wind speed, etc. For example, the historical predicted weather features are: {Date: October 20, 2024, Time interval: 1, Temperature: 28°C, Humidity: 65%, Air pressure: 1008 Pa, ……, Time interval: 2, Temperature: 28°C, Humidity: 65%, Air pressure: 1008 Pa, ……, …… Time interval: 48, Temperature: 25°C, Humidity: 55%, Air pressure: 1008 Pa, ……,}.

[0042] The historical actual power generation corresponding to the historical predicted weather features is: {2.5, ……, 10}, a total of 48 values. The historical actual weather features correspond to the historical predicted weather features and are relatively accurate weather features corrected according to the actual meteorological data based on the historical predicted weather features. The historical actual power generation is the actual power generation corresponding to the date and time interval of the historical predicted weather features / historical actual weather features. That is, a training sample for training the student model includes historical actual weather features, historical predicted weather features, and historical actual power generation (i.e., the true label).

[0043] S602, Input the historical actual weather features and historical actual power generation into the trained teacher model to obtain the teacher model feature vector and the teacher model predicted power generation.

[0044] Specifically, the teacher model includes Figures 3 to 5 the feature extraction network structure and the advanced feature expression module shown. The trained teacher model is trained based on the historical actual weather features, and the loss function used during training is the MAE function (Mean Absolute Error), and the formula is:

[0045] Among them, is the number of samples, is the true value of the th sample, is the predicted value of the

[0046] During the training process of the student model, the teacher model is only used for prediction based on historical true weather features and is not trained. After the historical true weather features are input into the trained teacher model, the teacher model feature vector and the teacher model predicted power generation are output.

[0047] S603. Input the historical predicted weather features and the historical true power generation into the student model to be trained to obtain the student model feature vector and the student model predicted power generation.

[0048] Specifically, the student model also includes Figures 3 to 5 the feature extraction network structure and the high-level feature expression module shown in. After the historical predicted weather features are successively extracted by the feature extraction network structure and the high-level feature expression module, the student model feature vector is obtained, and the student model predicted power generation is obtained based on the student model feature vector.

[0049] Step S604. Train the student model by calculating the multi-source total loss according to the teacher model feature vector, the teacher model predicted power generation, the student model feature vector, the student model predicted power generation, and the historical true power generation to obtain the trained student model.

[0050] In one embodiment, a multi-source loss function is constructed to train the student model, so that the student model can learn knowledge in different domains. The multi-source loss function is composed of a distillation loss function, a soft label loss function, and a true label loss function. Among them, the distillation loss adopts the KL divergence (Kullback-Leibler Divergence, also known as relative entropy) loss function, and the distillation loss is calculated by calculating the distribution difference between the teacher model feature vector and the student model feature vector. The formula is as follows:

[0051] Among them, is the probability distribution of the teacher model feature vector, is the probability distribution of the student model feature vector. Specifically, is the teacher model feature vector or the student model feature vector. Softmax normalization is performed on the teacher model feature vector or the student model feature vector, that is, the feature vector is converted into a probability distribution. The Softmax formula is:

[0052] Among them, = is the feature vector, is the probability of the th category. After Softmax normalization, the probabilities of all categories sum to 1, i.e., = 1.

[0053] The soft label loss uses the MAE function. Taking the predicted power generation of the student model as the prediction value and the predicted power generation of the teacher model as the label, the soft label loss is calculated as follows:

[0054] where is the prediction value of the teacher model for the th sample, is the prediction value of the student model for the th sample.

[0055] The true label loss also uses the MAE function. Taking the predicted power generation of the student model as the prediction value and the historical true power generation as the label, the true label loss is calculated as follows:

[0056] where is the true value of the th sample, is the prediction value of the student model for the th sample.

[0057] In one embodiment, the formula for the multi - variable total loss is:

[0058] where , and are weight coefficients, and + + = 1. Generally, > > .

[0059] Step S605, predict the photovoltaic power generation according to the trained student model.

[0060] The photovoltaic power generation prediction method provided by the embodiments of the present application first obtains historical predicted weather features, historical actual weather features, and historical actual power generation; then inputs the historical actual weather features and historical actual power generation into the trained teacher model to obtain the teacher model feature vector and the teacher model predicted power generation, and inputs the historical predicted weather features and historical actual power generation into the student model to be trained to obtain the student model feature vector and the student model predicted power generation; then, trains the student model by calculating the multivariate total loss according to the teacher model feature vector, the teacher model predicted power generation, the student model feature vector, and the student model predicted power generation to obtain the trained student model; finally, predicts the photovoltaic power generation according to the trained student model. The method of the present application, when learning the relationship between the predicted weather features and the power generation, adds a learning method assisted by the actual weather, improves the robustness of the model, and reduces the problem of cross-domain accuracy.

[0061] According to an embodiment of the present application, there is provided an electronic device, such as Figure 7 , which is a schematic structural diagram of an electronic device provided by an embodiment of the present application. The electronic device 100 may include a processor 10, a communication interface 30, a memory 20, and a communication bus. Among them, the processor 10, the communication interface 30, and the memory 20 complete communication with each other through the communication bus. The processor 10 may call the logical instructions in the memory 20 to execute the above-mentioned photovoltaic power generation prediction method.

[0062] In addition, when the logical instructions in the above-mentioned memory 20 are implemented in the form of a software functional unit and sold or used as an independent product, they may be stored in several computer-readable storage media. Based on such an understanding, the technical solution of the present application, in essence, or the part that contributes to the prior art or a part of the technical solution may be embodied in the form of a software product. The computer software product is stored in a storage medium and includes several instructions for causing a computer device (which may be a personal computer, a server, or a network device, etc.) to execute all or part of the steps of the photovoltaic power generation prediction method described above. The aforementioned storage media include: USB flash drives, mobile hard disks, read-only memories (ROM, Read-Only Memory), random access memories (RAM, Random Access Memory), magnetic disks, or optical discs, etc., which can store program codes.

[0063] According to an embodiment of the present application, there is provided a computer-readable storage medium of the type described above. The computer-readable storage medium stores a computer program. When the computer program is executed by a processor, the processor executes the steps of the photovoltaic power generation prediction method described above.

[0064] Through the description of the above embodiments, those skilled in the art can clearly understand that each embodiment can be implemented by means of software plus a general hardware platform, and of course, it can also be implemented by hardware. Based on such an understanding, the essence of the above technical solution, or the part that contributes to the related technology, can be embodied in the form of a software product. This computer software product can be stored in a computer-readable storage medium, such as ROM / RAM, magnetic disk, optical disk, etc., and includes several instructions to enable a computer device (which can be a personal computer, a server, or a network device, etc.) to execute the methods described in each embodiment or some parts of the embodiments.

[0065] The above is only the specific embodiment of the present application, but the protection scope of the present application is not limited thereto. Any person skilled in the art can easily think of changes or substitutions within the technical scope disclosed in the present application. Therefore, any modification, equivalent substitution, improvement, etc. made within the spirit and principle of the present application shall be included in the protection scope of the present application.

Claims

1. A photovoltaic power generation prediction method, characterized in that, The method includes: Obtaining historical predicted weather features, historical actual weather features, and historical actual power generation; Inputting the historical actual weather features and the historical actual power generation into a trained teacher model to obtain a teacher model feature vector and a teacher model predicted power generation; Inputting the historical predicted weather features and the historical actual power generation into a student model to be trained to obtain a student model feature vector and a student model predicted power generation; Calculating a multi - variable total loss based on the teacher model feature vector, the teacher model predicted power generation, the student model feature vector, the student model predicted power generation, and the historical actual power generation to train the student model, and obtaining a trained student model; Predicting photovoltaic power generation based on the trained student model.

2. The method according to claim 1, characterized in that, The teacher model is trained based on historical actual weather features.

3. The method according to claim 1, characterized in that, The calculating of the multi - variable total loss based on the teacher model feature vector, the teacher model predicted power generation, the student model feature vector, the student model predicted power generation, and the historical actual power generation includes: Calculating a distillation loss based on the teacher model feature vector and the student model feature vector; Calculating a soft - label loss based on the student model predicted power generation and the teacher model predicted power generation; Calculating a real - label loss based on the student model predicted power generation and the historical actual power generation; Calculating the multi - variable total loss based on the distillation loss, the soft - label loss, and the real - label loss.

4. The method according to claim 3, wherein The calculation formula of the multi - variable total loss is: Among them, is the total multi - loss, is the distillation loss, is the soft - label loss, is the true - label loss, 、 and are weight coefficients, is the probability distribution of the feature vectors of the teacher model, is the probability distribution of the feature vectors of the student model, is the number of samples, is the true value of the th sample, is the predicted value of the teacher model for the th sample, is the predicted value of the student model for the th sample.

5. The method according to any one of claims 1 to 4, characterized in that, The teacher model and the student model have the same network structure.

6. The method according to claim 5, characterized in that, The network structure includes a global feature module, a feature pyramid module, a time - series module, a multi - head attention module, and a high - level feature expression module connected in sequence.

7. The method according to claim 6, characterized in that, The global feature module extracts features using dilated convolution and large convolution kernels.

8. The method according to claim 6, wherein The feature pyramid module inputs the features extracted in the previous stage into multiple branches. After each branch extracts features using dilated convolution with different dilation rates, the features of multiple branches are concatenated and superimposed by channel.

9. An electronic device, characterized in that, Including at least one processor and a memory communicatively connected to the at least one processor, the memory stores instructions executable by the at least one processor, and the instructions are executed by the at least one processor so that the at least one processor can execute the photovoltaic power generation prediction method according to any one of claims 1 to 8.

10. A computer storage medium, characterized in that, The computer storage medium stores instructions or programs, and when the instructions or programs are executed by at least one processor, the at least one processor is enabled to execute the photovoltaic power generation prediction method according to any one of claims 1 to 8.

Images