Short-term photovoltaic power generation power prediction method based on multi-source data fusion

Abstract

The invention relates to a short-term photovoltaic power generation power prediction method based on multi-source data fusion. According to the method, an end-to-end multi-mode deep learning framework is constructed, satellite images, meteorological time sequences and historical power generation data are effectively fused, and high-precision short-term prediction of photovoltaic power generation power is achieved. The method comprises the following steps: firstly, processing a satellite image by using a convolutional neural network, processing a time sequence by using a bidirectional LSTM model, and then splicing multi-source features; and an attention mechanism is introduced to perform feature selection so as to improve the prediction accuracy. Experimental verification is carried out, and the result shows that compared with a prediction method of a single data source, the prediction precision is improved, the predictability of photovoltaic power generation is enhanced, economic dispatching of a power grid is facilitated, and application and popularization potential is achieved.

Description

Technical Field

[0001] The present invention belongs to the technical field of power generation of new energy power systems, and specifically relates to a short-term photovoltaic power generation power prediction method based on multi-source data fusion. Background Art

[0002] As renewable energy continues to gain popularity in electricity production, photovoltaic power generation, as a clean and renewable energy source, is playing an increasingly important role. However, the efficiency and reliability of photovoltaic power generation are affected by many factors, such as daily and seasonal meteorological conditions, light intensity, cloud cover, etc. In order to achieve sustainable operation of photovoltaic power generation systems and optimize power production, short-term photovoltaic power generation power forecasting becomes essential. This forecast not only improves the efficiency of power production, but also helps power system planners and operators better manage power supply, reduce energy waste, and reduce carbon emissions, thereby promoting sustainable development.

[0003] Although the importance of short-term PV power forecasting is widely recognized, existing technologies still face a series of bottlenecks. Traditional forecasting methods mainly rely on meteorological data and time series analysis, which are constrained by data inaccuracies and limitations. In particular, traditional models have difficulty effectively processing unstructured data, such as satellite imagery, which provides more comprehensive spatial range information and is crucial to improving forecast accuracy. In addition, the time series modeling capabilities of existing models are limited and it is difficult to capture the complex associations between multiple data sources, resulting in poor forecasting performance. Therefore, a new and innovative approach is urgently needed to address these issues to achieve more accurate and reliable short-term PV power forecasting.

[0004] In order to address the above challenges, the present invention proposes a multimodal deep learning framework, which aims to improve the performance and efficiency of short-term photovoltaic power generation prediction. The framework can receive multi-source data from different sensors and data sources, including satellite images, time series data, meteorological data, etc. through a flexible multimodal data input layer. Using convolutional neural networks and recurrent neural networks (BiLSTM), the present invention can effectively process image and time series data and capture spatial and temporal correlations. Through a multimodal fusion layer and an attention mechanism, the present invention fuses the feature vectors generated by different data sources and dynamically adjusts the weights to improve the prediction performance. Most importantly, the present invention allows end-to-end training, thereby maximizing the performance and prediction accuracy of the model. This innovative multimodal deep learning framework is expected to bring revolutionary progress to the field of short-term photovoltaic power generation prediction, solve the bottleneck of the existing technology, provide a more accurate and reliable prediction method, and promote the sustainable use of renewable energy. Summary of the invention

[0005] The purpose of the present invention is to overcome the shortcomings of the prior art and to provide a short-term photovoltaic power generation prediction method based on multi-source data fusion, which can realize multi-source data fusion and utilize more exogenous information, thereby realizing high-precision photovoltaic system power generation prediction, providing a reference basis for power system operation decision-making, and improving the economy and stability of new power systems.

[0006] The technical solution adopted by the present invention is: a short-term photovoltaic power generation prediction method based on multi-source data fusion, comprising the following steps:

[0007] S1: Build a multimodal data input layer: Design a multimodal data input layer that allows receiving multi-source data from different sensors and data sources, including satellite images, time series data, and meteorological data. Ensure that the architecture of the input layer is flexible and can seamlessly integrate various data sources and convert them into a format that can be processed by the model.

[0008] S2: Convolutional Neural Networks for Image Data: Use convolutional neural networks to process satellite image data. Implement convolutional neural network layers to extract spatial features from satellite images and encode them into useful feature vectors.

[0009] S3: BiLSTM to process time series data; Use BiLSTM to process time series data. Use BiLSTM to capture time correlation and dynamic features to adapt to data changes in different time periods.

[0010] S4: Build a multimodal fusion layer: Develop a multimodal fusion layer that can effectively fuse the feature vectors generated by the convolutional neural network and BiLSTM. The fusion layer needs to solve the time difference problem between different data sources and generate a more comprehensive feature representation.

[0011] S5: Introduce the attention mechanism: The introduction of the attention mechanism enables the model to dynamically adjust the weights of different data sources. According to the needs of the current task, it automatically selects which data sources are more important for prediction to improve the prediction performance.

[0012] S6: End-to-end training: Allows the entire multimodal deep learning framework to be trained end-to-end, ensuring that the different components of the model work together to maximize performance and perform unified parameter optimization.

[0013] S7: Performance evaluation and optimization: Conduct a comprehensive performance evaluation of the developed multimodal deep learning framework. Optimize and improve the model based on the experimental results to ensure its excellent performance in time series forecasting tasks.

[0014] Specifically, the main goal of the multimodal data input layer is to process different types of data, including satellite images, time series data, and meteorological data, and convert them into feature vectors for model processing. The following is the specific architecture for each data type:

[0015] 1) Satellite image data processing: For satellite image data, convolutional neural network is used to process it to extract spatial features.

[0016] Input image size: Assume that the input size of the satellite image is (W, H, C), where W represents the width of the image, H represents the height of the image, and C represents the number of channels.

[0017] Convolutional layer architecture: The specific architecture is as follows Figure 2 As shown, multiple convolutional layers and pooling layers are used to gradually extract features. Convolution kernel size: Usually a convolution kernel of size (3, 3) is used. Number of convolution kernels: Different numbers of convolution kernels are set according to the task requirements. Pooling layer: Use maximum pooling to reduce the size of the feature map. Output feature vector: The output after the convolution layer will be a three-dimensional tensor, which is converted into a fixed-size feature vector of size (F,) through global average pooling or a fully connected layer, where F is the dimension of the feature vector.

[0018] 2) Time series data processing: In order to model the time series of the serialized image feature sequence, the BiLSTM model is introduced. BiLSTM is a recurrent neural network that can capture sequence information. The model has two directions in time series: forward (positive) and backward (reverse), which process the sequence data respectively. Suppose the time series of the forward serialization is The output of BiLSTM is The calculation process of forward and backward LSTM is as follows:

[0019] Forward LSTM calculation:

[0020]

[0021] Backward LSTM calculation:

[0022]

[0023] in, represents the forward input sequence, represents the reverse input sequence, is the output of the forward LSTM, is the output of the reverse LSTM.

[0024] In order to fuse the outputs of the forward and backward LSTMs, the BiLSTM model concatenates the outputs of the two directions to obtain the final output H, as shown below:

[0025]

[0026] Among them, [] represents the concatenation operation, and H is the final output feature sequence of BiLSTM. The hidden states of the forward and backward LSTM are concatenated together to form a feature vector with a size of 2H, where H is the hidden state dimension of the LSTM unit.

[0027] 3) Multimodal fusion layer: The goal of the multimodal fusion layer is to effectively fuse the feature vectors from the convolutional neural network and BiLSTM and solve the problem of temporal differences in the data.

[0028] Input feature vector: The feature vectors from the convolutional neural network and BiLSTM are (F,) and (2H,), respectively.

[0029] Fusion mechanism: Use a simple concatenation operation to connect the two feature vectors to form a comprehensive multimodal feature vector with size (F+2H,).

[0030] The ReLU activation function is used to activate the fused feature vector to increase nonlinearity.

[0031] Specifically, the purpose of the attention mechanism is to allow the model to dynamically adjust the weights of different data sources so that the model automatically selects which data sources are more important in the current task. This helps improve prediction performance because different data sources may have different information values ​​at different times or under different conditions. In the multimodal deep learning framework proposed in the present invention, the attention mechanism is used to dynamically adjust the weights of satellite image data and time series data so as to perform targeted information fusion in different situations.

[0032] The attention weights are used to weight the feature vectors of different data sources to generate weighted feature vectors. Consider two data sources, namely the feature vectors of satellite image data (convolutional neural network output) and denoted as F img , and the feature vector of the time series data (BiLSTM output) is denoted as F seq .

[0033] The calculation of attention weight α is achieved by the following formula:

[0034] α=softmax(w T [F img , F seq ])

[0035] Among them, [F img , F seq ] indicates that F img and F seqsplice together operations. seq is the learned weight vector, whose dimension is the same as [F img , F seq ] are consistent in dimension. The softmax function will ensure that the sum of the weights α is 1, making them normalized weights.

[0036] According to the calculated attention weight α, calculate the weighted feature vector F weighted as follows:

[0037] F weighted =α·[F img , F seq ]

[0038] This weighted feature vector F weighted It will contain information from different data sources and be used in subsequent model prediction steps, such as time series prediction tasks.

[0039] The optimization of the attention mechanism usually needs to consider factors such as regularization, loss function, and training data. For example, regularization is used to prevent the attention weight from being overly concentrated on a certain data source. In addition, through supervised learning, a suitable loss function is used to train the network to obtain the optimal attention weight. The specific implementation method and parameter settings will be adjusted according to the specific task and dataset.

[0040] Specifically, the end-to-end training described in step S6 means that the entire model is trained as a unified system, rather than splitting it into multiple independent parts. In the multimodal deep learning framework, end-to-end training means that the entire model (including components such as multimodal data input layer, convolutional neural network, BiLSTM, multimodal fusion layer, attention mechanism, etc.) will be trained as a whole, rather than training each component separately. This helps ensure that different components adapt to each other, learn and adjust parameters together to maximize the performance of the model.

[0041] S6.1 Forward propagation: On the training data, the input data is forward propagated through the entire multimodal deep learning framework. This includes inputting data from different data sources into the corresponding components and performing operations such as feature extraction and attention weight calculation.

[0042] S6.2 Loss function: Define a loss function to measure the performance of the model. The loss function is usually related to the nature of the task. For time series prediction tasks, the mean square error is used. The specific calculation method is as follows:

[0043]

[0044] Where N is the number of samples; y i is the actual target value. is the predicted value of the model.

[0045] S6.3 Back propagation: According to the loss function, the gradient of each parameter is calculated by the back propagation algorithm. These gradients will be used for parameter update.

[0046] S6.4 Parameter Update: Use gradient descent or its variants to update the model parameters. The updated parameters will be adjusted according to the gradient direction to minimize the loss function.

[0047] Specifically, the performance evaluation and optimization in step S7 includes the following steps:

[0048] S7.1 Dataset division: Divide the available dataset into training set, validation set and test set. Use 80% of the data for training and 20% for validation and testing.

[0049] S7.2 Model training: Perform end-to-end training of a multimodal deep learning framework using the training set. Use an optimization algorithm, such as gradient descent, to minimize a loss function (such as mean squared error).

[0050] S7.3 Validation set evaluation: Use the validation set to evaluate the performance of the model. Calculate various performance indicators such as root mean square error (RMSE), mean absolute error (MAE), etc. to understand the initial performance of the model.

[0051] S7.4 Model optimization: Adjust and optimize the model based on the evaluation results of the validation set. Try different hyperparameter settings, network architecture modifications, or increase regularization to improve the model.

[0052] S7.5 Test set evaluation: Use the test set for final performance evaluation. This is an independent dataset that evaluates the generalization ability of the model and ensures that the model performs well on unseen data.

[0053] In time series forecasting tasks, commonly used performance evaluation indicators include the root mean square error (RMSE): used to measure the standard deviation of the model's forecast error. Its calculation formula is as follows:

[0054]

[0055] Mean absolute error (MAE): The average value of the prediction error used to measure the model. Its calculation formula is as follows:

[0056]

[0057] Coefficient of determination (R-squared): measures the correlation between the model's predicted value and the true value. Its calculation formula is as follows:

[0058]

[0059] in, is the average of the true values.

[0060] Model optimization steps:

[0061] Hyperparameter Tuning: Use methods like grid search, random search, etc. to tune the model’s hyperparameters like learning rate, batch size, number of layers, etc. to find the best settings.

[0062] Network architecture improvement: Based on the performance of the validation set, try to improve the network architecture, such as increasing or decreasing the number of layers, adjusting the number of neurons, etc.

[0063] Regularization: Use regularization techniques, such as L1 regularization or L2 regularization, to reduce overfitting and improve generalization.

[0064] Ensemble methods: Consider using ensemble learning methods such as bagging or boosting to further improve model performance.

[0065] Data augmentation: Apply data augmentation techniques such as rotation, scaling, or adding noise to improve the robustness of the model.

[0066] Iterative training: Repeatedly train and validate the model until performance requirements are met.

[0067] Beneficial effects of the present invention: The present invention can fuse multi-source data from different sensors and data sources, including satellite images, time series data, and meteorological data, thereby providing more comprehensive information and further improving the accuracy and precision of time series prediction. Secondly, the introduced attention mechanism allows the model to dynamically adjust the weights of different data sources, automatically selecting which data sources are more important for prediction according to task requirements, thereby improving the adaptability and generalization ability of the model. End-to-end training ensures that the different components of the model work together and optimize parameters uniformly, further improving the performance of the model and making it perform well in time series prediction tasks. This combined effect makes the present invention a powerful tool that can be applied to a variety of fields, such as renewable energy management and meteorological forecasting, providing new possibilities for future research and applications. BRIEF DESCRIPTION OF THE DRAWINGS

[0068] Figure 1 It is a main step flow chart of the present invention;

[0069] Figure 2 is a network architecture diagram of the present invention;

[0070] Figure 3 is a detailed sub-step flow chart of step S6 of the present invention;

[0071] Figure 4is a detailed sub-step flow chart of step S7 of the present invention;

[0072] Figure 5 It is a comparison curve between the photovoltaic power generation power prediction result and the actual value in the embodiment of the present invention. DETAILED DESCRIPTION

[0073] The following will be combined with the drawings in the embodiments of the present invention to clearly and completely describe the technical solutions in the embodiments of the present invention. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without creative work are within the scope of protection of the present invention, and are specifically described in combination with the embodiments below.

[0074] like Figure 1 and 2 As shown, the present invention comprises the following steps:

[0075] S1: Build a multimodal data input layer; Design a multimodal data input layer that allows receiving multi-source data from different sensors and data sources.

[0076] S2: Convolutional neural network processing image data; Use convolutional neural network to process satellite image data and encode it into useful feature vectors. Satellite image data processing:

[0077] Assume that the input size of the satellite image is (W, H, C), where W represents the width of the image, H represents the height of the image, and C represents the number of channels. The specific architecture is as follows Figure 2 As shown, multiple convolutional layers and pooling layers are used to gradually extract features. The output after the convolutional layer will be a three-dimensional tensor, which is converted into a fixed-size feature vector with a size of (F,) through global average pooling or a fully connected layer, where F is the dimension of the feature vector.

[0078] S3: BiLSTM processes time series data; BiLSTM is used to process time series data. BiLSTM is used to capture time correlation and dynamic features to adapt to data changes in different time periods. The model has two directions in time series: forward (positive) and backward (reverse), which process the sequence data respectively. Suppose the time series of forward serialization is The output of BiLSTM is The calculation process of forward and backward LSTM is as follows:

[0079] Forward LSTM calculation:

[0080]

[0081] Backward LSTM calculation:

[0082]

[0083] in, represents the forward input sequence, represents the reverse input sequence, is the output of the forward LSTM, is the output of the reverse LSTM. In order to fuse the outputs of the forward and backward LSTMs, the BiLSTM model connects the outputs of the two directions to obtain the final output H, as shown below:

[0084]

[0085] Among them, [] represents the concatenation operation, and H is the final output feature sequence of BiLSTM. The hidden states of the forward and backward LSTM are concatenated together to form a feature vector with a size of 2H, where H is the hidden state dimension of the LSTM unit.

[0086] S4: Build a multimodal fusion layer: Develop a multimodal fusion layer that can effectively fuse the feature vectors generated by the convolutional neural network and BiLSTM. The fusion layer needs to solve the time difference problem between different data sources, generate a more comprehensive feature representation, and solve the time difference problem of the data.

[0087] Input feature vector: The feature vectors from the convolutional neural network and BiLSTM are (F,) and (2H,) respectively. The two feature vectors are concatenated using a concatenation operation to form a comprehensive multimodal feature vector of size (F+2H,).

[0088] The ReLU activation function is used to activate the fused feature vector to increase nonlinearity.

[0089] S5: Introduce the attention mechanism; according to the requirements of the current task, automatically select which data sources are more important for prediction to improve prediction performance.

[0090] Consider two data sources, namely the feature vector of satellite image data (convolutional neural network output) denoted as F img , and the feature vector of the time series data (BiLSTM output) is denoted as F seq .

[0091] The calculation of attention weight α is achieved by the following formula:

[0092] α=softmax(w T [F img , F seq ])

[0093] Among them, [F img , F seq ] indicates that Fimg and F seq splice together operations. seq is the learned weight vector, whose dimension is the same as [F img , F seq ] are consistent in dimension. The softmax function will ensure that the sum of the weights α is 1, making them normalized weights.

[0094] According to the calculated attention weight α, calculate the weighted feature vector F weighted as follows:

[0095] F weighted =α·[F img , F seq ]

[0096] Weighted eigenvector F weighted Will contain information from different data sources.

[0097] S6: End-to-end training: Allows the entire multimodal deep learning framework to be trained end-to-end, ensuring that the different components of the model work together to maximize performance and perform unified parameter optimization.

[0098] S6.1 Forward propagation: On the training data, the input data is forward propagated through the entire multimodal deep learning framework.

[0099] S6.2 Loss function: Define a loss function to measure the performance of the model. For time series prediction tasks, use mean square error. The specific calculation method is as follows:

[0100]

[0101] Where N is the number of samples; y i is the actual target value. is the predicted value of the model.

[0102] S6.3 Back propagation: According to the loss function, the gradient of each parameter is calculated by the back propagation algorithm. These gradients will be used for parameter update.

[0103] S6.4 Parameter Update: Use gradient descent or its variants to update the model parameters. The updated parameters will be adjusted according to the gradient direction to minimize the loss function.

[0104] S7: Performance evaluation and optimization: Conduct a comprehensive performance evaluation of the developed multimodal deep learning framework.

[0105] S7.1 Dataset division: Divide the available dataset into training set, validation set and test set. Use 80% of the data for training and 20% for validation and testing.

[0106] S7.2 Model training: Perform end-to-end training of a multimodal deep learning framework using the training set. Use an optimization algorithm, such as gradient descent, to minimize a loss function (such as mean squared error).

[0107] S7.3 Validation set evaluation: Use the validation set to evaluate the performance of the model. Calculate various performance indicators such as root mean square error (RMSE), mean absolute error (MAE), etc. to understand the initial performance of the model.

[0108] S7.3 Model optimization: Adjust and optimize the model based on the evaluation results of the validation set. Try different hyperparameter settings, network architecture modifications, or increase regularization to improve the model.

[0109] S7.4 Test set evaluation: Use the test set for final performance evaluation. This is an independent dataset that evaluates the generalization ability of the model and ensures that the model performs well on unseen data.

[0110] The effectiveness of the present invention is verified as follows:

[0111] A photovoltaic power station was selected as the verification data source. The installed capacity of the power station is 10MW, and 20,000 monocrystalline silicon photovoltaic modules are installed, each with a power of 500W. The photovoltaic modules are connected in series in groups of 100, with a total of 200 groups. Each group is connected to a 100kW inverter. The photovoltaic power station collected power generation data and meteorological data for the past two years with a time resolution of 15 minutes. In addition, the matching satellite image data of the photovoltaic power station taken by the Himawari-8 geostationary satellite in the past two years was obtained. The model was built after cleaning and normalizing the data.

[0112] Model parameters:

[0113] (1) Convolutional neural network: input image size: 256×256×3; convolution layer: use 3 convolution layers, convolution kernel size 3×3; the number of convolution kernels is 32, 64, and 128 respectively; pooling layer: maximum pooling, pooling window 2×2; output dimension: 256;

[0114] (2) BiLSTM: LSTM hidden unit: 128; sequence length: 96 (data from the past 24 hours); output dimension: 256;

[0115] (3) Multimodal fusion layer: Input dimension: convolutional neural network output 256 + BiLSTM output 256 = 512; Output dimension: 300

[0116] (4) Attention layer: input dimension: 300; output dimension: 300

[0117] The Adam optimization algorithm is used for model training, and the main hyperparameters are set as follows: number of training set samples: 100,000; number of validation set samples: 20,000; number of test set samples: 20,000; number of training rounds: 100; batch size: 64; initial learning rate: 0.001; learning rate decay: decay once every 10 rounds, with a decay coefficient of 0.8; the Early Stopping method is used during training, and training is stopped when the validation set loss has not decreased for 5 consecutive rounds to prevent overfitting. The loss function uses the mean square error MSE. To alleviate overfitting, L2 regularization is added to the convolutional layer and fully connected layer of the model, and the regularization coefficient is 0.0001. By adjusting the hyperparameters, the best MSE LOSS of the model on the validation set is 0.042, and the best MAE is 98.3. The trained model is used to predict the power generation of the photovoltaic power station in the next week, and compared with the LSTM-based prediction model and the actual power. The results are as follows Figure 5 shown.

[0118] Table 1 Comparison of experimental results between the proposed method and LSTM

[0119] RMSE MAE <![CDATA[R 2 ]]> Method of the present invention 152 114 0.884 LSTM 185 132 0.811

[0120] It can be seen from the results that the model of the present invention is better than the traditional LSTM model in terms of prediction error RMSE, MAE and determination coefficient R 2 This shows that the model of the present invention that integrates multi-source data can extract richer features and model the changing trend of photovoltaic power generation more accurately.

[0121] It will be apparent to those skilled in the art that the invention is not limited to the details of the exemplary embodiments described above and that the invention can be implemented in other specific forms without departing from the spirit or essential features of the invention. Therefore, the embodiments should be considered in all respects as exemplary and non-limiting, and the scope of the invention is defined by the appended claims rather than the foregoing description, and it is intended that all variations within the meaning and scope of the elements of the claims be included in the invention. Any reference numeral in a claim should not be considered as limiting the claim to which it relates.

Claims

1. A short-term photovoltaic power generation prediction method based on multi-source data fusion, characterized in that: The following steps are involved: S1. Build a multimodal data input layer to receive data from different sources; S2. Use convolutional neural networks to process satellite image data to extract spatial features; S3. Use a bidirectional LSTM network to process time series data to extract temporal features; S4. Construct a multimodal fusion layer and concatenate the feature vectors of step 2 and step 3; S5. Build an attention layer to learn the weights of different modal features; S6. Perform weighted fusion on the features of each modality for the prediction of photovoltaic power generation.

2. A short-term photovoltaic power generation prediction method based on multi-source data fusion as claimed in claim 1, characterized in that: The convolutional neural network adopts a series structure of convolutional layer and pooling layer.

3. A short-term photovoltaic power generation prediction method based on multi-source data fusion as claimed in claim 1, characterized in that: The bidirectional LSTM network includes a forward LSTM and a backward LSTM, and concatenates the outputs of the two. Suppose the time series of the forward serialization is The output of BiLSTM is The calculation process of forward and backward LSTM is as follows: Forward LSTM calculation: Backward LSTM calculation: in, represents the forward input sequence, represents the reverse input sequence, is the output of the forward LSTM, is the output of the reverse LSTM; in order to fuse the outputs of the forward and backward LSTMs, the BiLSTM model connects the outputs of the two directions to obtain the final output H, as shown below: Among them, [] represents the connection operation, H is the final output feature sequence of BiLSTM; the hidden states of the forward and backward LSTM are concatenated together to form a feature vector with a size of 2H, where H is the hidden state dimension of the LSTM unit.

4. A short-term photovoltaic power generation prediction method based on multi-source data fusion as claimed in claim 1, characterized in that: The multimodal fusion layer adopts a splicing method; the feature vectors from CNN and BiLSTM are (F,) and (2H,) respectively; the two feature vectors are connected by splicing operation to form a comprehensive multimodal feature vector with a size of (F+2H,); and the ReLU activation function is used to activate the fused feature vector to increase nonlinearity.

5. The short-term photovoltaic power generation prediction method based on multi-source data fusion according to claim 1, characterized in that: The attention layer is implemented by a fully connected layer; consider two data sources, namely the feature vector of satellite image data (CNN output) denoted as F img , and the feature vector of the time series data (BiLSTM output) is denoted as F seq ; The calculation of attention weight α is achieved by the following formula: α=softmax(w T [F img ,F seq ]) Among them, [F img ,F seq ] indicates that F img and F seq The operation of splicing together; F seq is the learned weight vector, whose dimension is the same as [F img ,F seq ] has the same dimension; the softmax function will ensure that the sum of the weights α is 1, making it a normalized weight; based on the calculated attention weight α, the weighted feature vector F is calculated weighted as follows: F weighted =α·[F img ,F seq ] Weighted eigenvector F weighted Will contain information from different data sources.

6. A short-term photovoltaic power generation prediction method based on multi-source data fusion as claimed in claim 1, characterized in that: It also includes end-to-end training of the described method; ensuring that the different components of the model work together to maximize performance and perform unified parameter optimization; S6.1 Forward propagation: On the training data, forward propagate the input data through the entire multimodal deep learning framework; S6.2 Loss function: Define a loss function to measure the performance of the model; for time series prediction tasks, use mean square error; the specific calculation method is as follows: Where N is the number of samples; y i is the true target value; is the predicted value of the model; S6.3 Back propagation: According to the loss function, the gradient of each parameter is calculated by the back propagation algorithm; these gradients will be used for parameter update; S6.4 Parameter Update: Use gradient descent or its variants to update the parameters of the model; the updated parameters will be adjusted according to the gradient direction to minimize the loss function.

7. A short-term photovoltaic power generation prediction method based on multi-source data fusion as claimed in claim 6, characterized in that: The end-to-end training adopts the gradient descent method.

8. The short-term photovoltaic power generation prediction method based on multi-source data fusion according to claim 1, characterized in that: The multi-source data includes satellite images, meteorological data and historical power generation data.

9. A short-term photovoltaic power generation prediction system based on the prediction method according to any one of claims 1 to 8, characterized in that: It includes data acquisition module, data processing module, model training module and model prediction module.

Images