Photovoltaic power station day-ahead generation power prediction model integrated with corrected NWP data

Abstract

The invention relates to the technical field of photovoltaic power generation power prediction, in particular to a photovoltaic power station day-ahead power generation power prediction model integrated with corrected NWP data, which comprises the following steps: 1) performing correlation analysis on historical power generation power data and historical actual meteorological data, and determining input characteristics of the prediction model; 2) taking historical actual data of the meteorological factors determined as the input features as input, taking historical power data as output, and training to obtain a CNN-LSTM hybrid prediction model based on Bayesian optimization; and 3) decomposing the error sequence of the actual meteorological data and the NWP data by using variational mode decomposition to obtain a plurality of sub-components with different frequency characteristics, modeling and predicting each component by using a CNN-BiGRU-AM network, and reconstructing the prediction result of each component to realize overall correction and optimization of the NWP data. According to the method, the day-ahead power generation power prediction of the photovoltaic power station is more accurate, the NWP data is corrected while the hybrid prediction model is improved, and the power prediction precision is fully improved.

Description

Technical Field

[0001] This invention relates to the field of day-ahead power generation prediction technology for photovoltaic power plants, and in particular to a day-ahead power generation prediction model for photovoltaic power plants that incorporates corrected NWP data. Background Technology

[0002] With the depletion of fossil fuels and the resulting environmental pollution, new energy power generation has been vigorously developed globally. Photovoltaic power generation, due to its low cost and zero pollution, has gradually become one of the most promising renewable energy technologies. However, the uncertainty of meteorological factors such as cloud cover, rain, and smog above photovoltaic power plants makes it difficult to accurately predict photovoltaic power generation output. Therefore, improving the accuracy of photovoltaic power generation output prediction has become a key technical issue in the current engineering field.

[0003] To improve the accuracy of photovoltaic (PV) power generation forecasting, two approaches can be taken. First, the forecasting model can be improved by enhancing its performance to achieve high-precision PV power generation predictions. Second, day-ahead power forecasting often uses NWP data as model input, while actual meteorological information and actual power output are typically used as input and output during model training. Traditional weather forecasts have a large spatial resolution scale (usually at the kilometer level), while the coverage area of ​​a MW-level PV power plant may be smaller than the spatial scale of weather forecasts. Therefore, directly using low-spatial-resolution NWP data as input to the forecasting model may result in relatively low prediction accuracy. Thus, NWP spatial downscaling, by improving the accuracy of the input data, can also achieve high-precision PV power generation forecasting. Summary of the Invention

[0004] The purpose of this invention is to address the shortcomings of existing technologies by proposing a day-ahead power generation prediction model for photovoltaic power plants that incorporates corrected NWP data. The focus of this research is on optimizing the hybrid prediction model and spatial downscaling of the NWP data. The results obtained by this model show improved accuracy compared to those obtained using a single method.

[0005] To achieve the above objectives, the present invention adopts the following technical solution:

[0006] A photovoltaic power generation prediction model incorporating corrected NWP data includes the following steps:

[0007] S1. Collect data from the photovoltaic power station, including historical power data, historical actual data of various meteorological factors, and historical NWP (Numerical Weather Prediction) data, and perform correlation analysis based on this data. This mainly includes the following steps:

[0008] S1.1 Calculate the Pearson correlation coefficient and Spearman correlation coefficient between photovoltaic power generation and historical actual meteorological data to determine the input characteristics of the prediction model;

[0009] S1.2 Calculate the error between the historical actual data and NWP data of the input features, and analyze the error characteristics;

[0010] S2. Using historical actual data of meteorological factors as input and historical power data as output, a Bayesian-optimized CNN-LSTM hybrid prediction model is trained. This mainly includes the following steps:

[0011] S2.1 Given that meteorological variables usually have significant local spatiotemporal correlations, a convolutional neural network (CNN) module is introduced at the front end of the CNN-LSTM hybrid prediction model. Subsequently, a prediction model is established using a two-layer stacked long short-term memory (LSTM) network.

[0012] S2.2 To improve model performance, Bayesian optimization (BO) is introduced before training to jointly optimize some key hyperparameters with the goal of minimizing prediction error. This is achieved through efficient search using a Gaussian process surrogate model.

[0013] S3. To address the discrepancy between NWP data and actual meteorological data, a spatial downscaling model for NWP based on VMD-CNN-BiGRU-AM is proposed, which mainly includes the following steps:

[0014] S3.1 First, the error sequence between the actual meteorological data and the NWP data is decomposed using Variational Mode Decomposition (VMD) to obtain several sub-components with different frequency characteristics;

[0015] S3.2 Subsequently, a multi-channel neural network structure is designed, which includes a Convolutional Neural Network (CNN), a Bidirectional Gated Recurrent Unit (BiGRU), and an Attention Mechanism module.

[0016] S3.3 Next, the CNN-BiGRU-AM network is used to model and predict each component.

[0017] S3.4 Finally, the prediction results of each component are reconstructed to achieve overall correction and optimization of the NWP data;

[0018] S4. Combine the Bayesian optimization-based CNN-LSTM hybrid prediction model in S2 with the NWP spatial downscaling model in S3 to form a photovoltaic power generation prediction model that incorporates corrected NWP data.

[0019] S5. Substitute the original NWP data for the prediction day into the photovoltaic power generation prediction model for the day-ahead of the photovoltaic power plant that incorporates the corrected NWP data in S4, and evaluate the performance of the prediction model.

[0020] Preferably, in S2, the CNN-LSTM hybrid prediction model based on Bayesian optimization is implemented using the following steps:

[0021] 1) A convolutional neural network module is introduced at the front end of the CNN-LSTM hybrid prediction model. One-dimensional convolutional kernels are used to slide across the input sequence in the time dimension to extract local features. The ELU activation function is used to enhance the network's non-linear expressive power. Batch normalization and pooling operations are combined to enhance training stability and feature robustness. After processing by the CNN module, sequence unrolling and flattening operations are used to reconstruct its output features into a standard time series format, facilitating further prediction model building by the subsequent Long Short-Term Memory module.

[0022] 2) The LSTM part consists of two stacked layers. The first layer contains 50 hidden units for initial learning of temporal information, while the number of hidden units in the second layer is automatically searched and determined within a preset interval by a Bayesian optimization algorithm. The LSTM internally employs a standard gating structure, including forget gate, input gate, and output gate mechanisms, enabling the network to effectively capture long-term dependencies and suppress gradient vanishing. Subsequently, the CNN-LSTM hybrid prediction model outputs the photovoltaic power generation prediction result through a fully connected layer and a regression layer, using the Mean Squared Error (MSE) loss function, combined with an L2 regularization term to suppress overfitting and improve generalization ability.

[0023] 3) To ensure that the hyperparameter configuration of the CNN-LSTM hybrid prediction model reaches the optimal level, the Bayesian optimization method is used before network training to jointly optimize the three key hyperparameters: the number of LSTM units, the initial learning rate, and the L2 regularization coefficient. With the goal of minimizing the prediction error, the Gaussian process surrogate model is used to efficiently search the hyperparameter space.

[0024] Preferably, in S3, the NWP spatial downscaling model based on VMD-CNN-BiGRU-AM is implemented using the following steps:

[0025] 1) Data preprocessing: Calculate the error between the input NWP data and the actual weather data, and use it as the raw error sequence. The specific expression is as follows:

[0026] e(t) = y actual (t)-y NWP (t)

[0027] Among them, y NWP (t), y actual e(t) represents the NWP data input at time t and the actual weather data, respectively; e(t) represents the original error sequence at time t.

[0028] 2) Error Sequence Decomposition: Multimodal decomposition of the error sequence is performed using VMD, decomposing it into K Intrinsic Mode Functions (IMFs). The specific expressions are as follows:

[0029]

[0030] Where K is the number of modes; u k (t) represents the value of the k-th IMF component at time t.

[0031] 3) Supervision Sample Construction: For each IMF component, input and output supervision samples are constructed using a sliding window. Input feature X i By k im Composed of historical values, predicting target Y i For the future z im The value of the step. The specific expression is as follows:

[0032] X i =[u k (i),u k (i+1),…,u k (i+k im -1)]

[0033] Y i =u k (i+k im +z im -1)

[0034] Among them, X i Let k be the input feature vector of the i-th sample, consisting of a length of k. im Composed of continuous historical values; Y i Let k be the predicted target value for the i-th sample; im z is the historical time step (window length) of the input sequence; im To predict the step size.

[0035] 4) CNN-BiGRU-AM Model Design: For each IMF component, a multi-channel neural network structure is designed, comprising a convolutional neural network, bidirectional gated recurrent units, and an attention mechanism module. First, convolutional units are used to extract local spatial features; second, bidirectional gated recurrent units are used to model the forward and reverse information of the sequence, and an attention mechanism is used to enhance the response of salient features through channel weighting; finally, a regression layer outputs the predicted value, the main expression of which is as follows:

[0036] Y i * =f BiGRU-AM (f CNN (X i ))

[0037] Among them, f CNN f represents the convolutional feature extraction module; BiGRU-AM This represents a temporal prediction module that integrates a bidirectional gated recurrent unit and an attention mechanism; Y i * These are predicted values.

[0038] 5) Model training and prediction: Train a CNN-BiGRU-AM model independently for each IMF component and obtain its prediction results respectively;

[0039] 6) Component reconstruction

[0040] The prediction sequence of the original error is reconstructed by summing all the IMF prediction results, as shown in the following expression:

[0041]

[0042] Among them, e * (t) represents the prediction result at time t after reconstruction; This represents the prediction result of the k-th modal component at time t.

[0043] Finally, by combining the NWP data and the predicted error sequence, the corrected NWP data is obtained and used as input to the photovoltaic power generation prediction model to predict power output. The specific expression is as follows:

[0044] y * correct (t)=y NWP (t)+e * (t)

[0045] Among them, y * correct (t) represents the corrected NWP data at time t.

[0046] Compared with the prior art, the present invention has the following beneficial effects:

[0047] 1. To extract key features and characterize their temporal trends, this invention proposes a CNN-LSTM hybrid model based on Bayesian optimization. This model combines the spatial feature extraction capabilities of CNNs with the temporal modeling advantages of LSTMs, and adaptively adjusts hyperparameters through Bayesian optimization, thereby improving prediction accuracy and generalization ability.

[0048] 2. To address the systematic errors present in NWP data, this invention introduces the VMD algorithm to decompose the error sequence based on the existing CNN-BiGRU-AM error correction model. A separate CNN-BiGRU-AM model is constructed for each IMF component for modeling and prediction, and the prediction results of each component are reconstructed, further improving the accuracy of error correction.

[0049] 3. To verify the universality and stability of the NWP spatial downscaling model and the hybrid prediction model, this invention randomly selects one month's data in each quarter for testing, and uses 30 consecutive days of data as the training set to predict the photovoltaic power generation power of the next day. Attached Figure Description

[0050] Figure 1 This is a schematic diagram of the process of the present invention;

[0051] Figure 2 This is a comparison chart of actual meteorological data and NWP data for some sampled data in an embodiment of the present invention;

[0052] Figure 3 This is the NWP data correction result for sample 1 in this embodiment of the invention;

[0053] Figure 4 This is the NWP data correction result for sample 2 in this embodiment of the invention;

[0054] Figure 5 This is the NWP data correction result for sample 3 in this embodiment of the invention;

[0055] Figure 6 This is the NWP data correction result for sample 4 in this embodiment of the invention;

[0056] Figure 7 This is a comparison chart of the prediction results for each sample in an embodiment of the present invention. Detailed Implementation

[0057] The technical solutions in the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings, so that those skilled in the art can better understand the advantages and features of the present invention, thereby making a clearer definition of the scope of protection of the present invention. The embodiments described in this invention are only some embodiments of the present invention, not all embodiments. All other embodiments obtained by those skilled in the art based on the embodiments of the present invention without creative effort are within the scope of protection of the present invention.

[0058] Reference Figure 1 A photovoltaic power generation prediction model incorporating corrected NWP data includes the following steps:

[0059] S1. Collect data from the photovoltaic power station, including historical power data, historical actual data of various meteorological factors, and historical NWP (Numerical Weather Prediction) data, and perform correlation analysis based on this data. This mainly includes the following steps:

[0060] S1.1 Calculate the Pearson correlation coefficient and Spearman correlation coefficient between photovoltaic power generation and historical actual meteorological data to determine the input characteristics of the prediction model;

[0061] S1.2 Calculate the error between the historical actual data and NWP data of the input features, and analyze the error characteristics;

[0062] S2. Using historical actual data of meteorological factors as input and historical power data as output, a Bayesian-optimized CNN-LSTM hybrid prediction model is trained. This mainly includes the following steps:

[0063] S2.1 Given that meteorological variables usually have significant local spatiotemporal correlations, a convolutional neural network (CNN) module is introduced at the front end of the CNN-LSTM hybrid prediction model. Subsequently, a prediction model is established using a two-layer stacked long short-term memory (LSTM) network.

[0064] S2.2 To improve model performance, Bayesian optimization (BO) is introduced before training to jointly optimize some key hyperparameters with the goal of minimizing prediction error. A Gaussian process surrogate model is used to achieve efficient search.

[0065] S3. To address the discrepancy between NWP data and actual meteorological data, a spatial downscaling model for NWP based on VMD-CNN-BiGRU-AM is proposed, which mainly includes the following steps:

[0066] S3.1 First, the error sequence between the actual meteorological data and the NWP data is decomposed using Variational Mode Decomposition (VMD) to obtain several sub-components with different frequency characteristics;

[0067] S3.2 Subsequently, a multi-channel neural network structure is designed, which includes a Convolutional Neural Network (CNN), a Bidirectional Gated Recurrent Unit (BiGRU), and an Attention Mechanism module.

[0068] S3.3 Next, the CNN-BiGRU-AM network is used to model and predict each component.

[0069] S3.4 Finally, the prediction results of each component are reconstructed to achieve overall correction and optimization of the NWP data;

[0070] S4. Combine the Bayesian optimization-based CNN-LSTM hybrid prediction model in S2 with the NWP spatial downscaling model in S3 to form a photovoltaic power generation prediction model that incorporates corrected NWP data.

[0071] S5. Substitute the original NWP data for the prediction day into the photovoltaic power generation prediction model for the day-ahead of the photovoltaic power plant that incorporates the corrected NWP data in S4, and evaluate the performance of the prediction model.

[0072] Specifically, in S2, the CNN-LSTM hybrid prediction model based on Bayesian optimization follows these steps:

[0073] 1) A convolutional neural network module is introduced at the front end of the CNN-LSTM hybrid prediction model. One-dimensional convolutional kernels are used to slide across the input sequence in the time dimension to extract local features. The ELU activation function is used to enhance the network's non-linear expressive power. Batch normalization and pooling operations are combined to enhance training stability and feature robustness. After processing by the CNN module, sequence unrolling and flattening operations are used to reconstruct its output features into a standard time series format, facilitating further prediction model building by the subsequent Long Short-Term Memory module.

[0074] 2) The LSTM part consists of two stacked layers. The first layer contains 50 hidden units for initial learning of temporal information, while the number of hidden units in the second layer is automatically searched and determined within a preset interval by a Bayesian optimization algorithm. The LSTM internally employs a standard gating structure, including forget gate, input gate, and output gate mechanisms, enabling the network to effectively capture long-term dependencies and suppress gradient vanishing. Subsequently, the CNN-LSTM hybrid prediction model outputs the photovoltaic power generation prediction result through a fully connected layer and a regression layer, using the Mean Squared Error (MSE) loss function, combined with an L2 regularization term to suppress overfitting and improve generalization ability.

[0075] 3) To ensure that the hyperparameter configuration of the CNN-LSTM hybrid prediction model reaches the optimal level, the Bayesian optimization method is used before network training to jointly optimize the three key hyperparameters: the number of LSTM units, the initial learning rate, and the L2 regularization coefficient. With the goal of minimizing the prediction error, the Gaussian process surrogate model is used to efficiently search the hyperparameter space.

[0076] Specifically, in S3, the NWP spatial downscaling model based on VMD-CNN-BiGRU-AM follows these steps:

[0077] 1) Data preprocessing: Calculate the error between the input NWP data and the actual weather data, and use it as the raw error sequence. The specific expression is as follows:

[0078] e(t) = y actual (t)-y NWP (t)

[0079] Among them, y NWP (t), y actual e(t) represents the NWP data input at time t and the actual weather data, respectively; e(t) represents the original error sequence at time t.

[0080] 2) Error Sequence Decomposition: Multimodal decomposition of the error sequence is performed using VMD, decomposing it into K Intrinsic Mode Functions (IMFs). The specific expressions are as follows:

[0081]

[0082] Where K is the number of modes; u k (t) represents the value of the k-th IMF component at time t.

[0083] 3) Supervision Sample Construction: For each IMF component, input and output supervision samples are constructed using a sliding window. Input feature X i By k imComposed of historical values, predicting target Y i For the future z im The value of the step. The specific expression is as follows:

[0084] X i =[u k (i),u k (i+1),…,u k (i+k im -1)]

[0085] Y i =u k (i+k im +z im -1)

[0086] Among them, X i Let k be the input feature vector of the i-th sample, consisting of a length of k. im Composed of continuous historical values; Y i Let k be the predicted target value for the i-th sample; im z is the historical time step (window length) of the input sequence; im To predict the step size.

[0087] 4) CNN-BiGRU-AM Model Design: For each IMF component, a multi-channel neural network structure is designed, comprising a convolutional neural network, bidirectional gated recurrent units, and an attention mechanism module. First, convolutional units are used to extract local spatial features; second, bidirectional gated recurrent units are used to model the forward and reverse information of the sequence, and an attention mechanism is used to enhance the response of salient features through channel weighting; finally, a regression layer outputs the predicted value, the main expression of which is as follows:

[0088] Y i * =f BiGRU-AM (f CNN (X i ))

[0089] Among them, f CNN f represents the convolutional feature extraction module; BiGRU-AM This represents a temporal prediction module that integrates a bidirectional gated recurrent unit and an attention mechanism; Y i * These are predicted values.

[0090] 5) Model training and prediction: Train a CNN-BiGRU-AM model independently for each IMF component and obtain its prediction results respectively;

[0091] 6) Component reconstruction

[0092] The prediction sequence of the original error is reconstructed by summing all the IMF prediction results, as shown in the following expression:

[0093]

[0094] Among them, e * (t) represents the prediction result at time t after reconstruction; This represents the prediction result of the k-th modal component at time t.

[0095] Finally, by combining the NWP data and the predicted error sequence, the corrected NWP data is obtained and used as input to the photovoltaic power generation prediction model to predict power output. The specific expression is as follows:

[0096] y * correct (t)=y NWP (t)+e * (t)

[0097] Among them, y * correct (t) represents the corrected NWP data at time t.

[0098] Example:

[0099] A photovoltaic power generation prediction model incorporating corrected NWP data includes the following steps:

[0100] Step 1: Collect data from the photovoltaic power station, including historical power data, historical actual data of various meteorological factors, and historical NWP data. Based on this data, conduct correlation analysis to determine the input characteristics of the prediction model.

[0101] Referring to Table 1, the statistical table of the correlation between different meteorological factors and photovoltaic power output, the three types of meteorological factor data with the largest absolute values ​​of correlation coefficients (measured total solar radiation, measured temperature, and measured wind speed) were selected as the input features of the prediction model.

[0102] Table 1. Pearson and Spearman correlation coefficients between photovoltaic output and meteorological factors.

[0103]

[0104]

[0105] Step 2: Randomly select one month's sample each quarter for testing, and use 30 days of data as training samples to predict the photovoltaic power generation power in the next day.

[0106] Step 3: For the establishment of the hybrid prediction model, historical actual data of meteorological factors determined as input features are used as input, and historical power data are used as output. CNN-LSTM hybrid prediction models based on Bayesian optimization are trained separately. Table 2 shows the Bayesian optimization results for each sample.

[0107] Table 2. Results of Bayesian Optimization Parameters

[0108]

[0109] Step 4: For the establishment of the NWP spatial downscaling model, firstly, variational mode decomposition is used to decompose the error sequences of actual meteorological data and NWP data to obtain several sub-components with different frequency characteristics; then, CNN-BiGRU-AM networks are used to model and predict each component; finally, the prediction results of each component are reconstructed to achieve overall correction and optimization of the NWP data. Table 3 shows some of the parameter settings.

[0110] Table 3. Partial parameter settings in the calibration system

[0111]

[0112]

[0113] Step 5: For the NWP data in the test set of each sample, use the calibration system from Step 4 to calibrate the NWP data, obtaining high-precision NWP data at a smaller spatial scale, and evaluate the calibration results. See [link to specific calibration results] for details. Figure 3 .

[0114] refer to Figure 2 It can be seen that there is a significant discrepancy between actual meteorological data and NWP data. Using NWP data as input will obviously have a considerable impact on the prediction of photovoltaic power generation. Therefore, it is necessary to perform spatial downscaling on the existing NWP data to obtain high-precision NWP data at a smaller spatial scale.

[0115] refer to Figure 3 , Figure 4 , Figure 5 , Figure 6 As shown in Table 4, the high-resolution meteorological information obtained after correction by the system proposed in this invention is very close to the actual meteorological data, representing a significant improvement over the original NWP data. The regression coefficient of the corrected total solar radiation reaches over 0.99, and the mean absolute error and root mean square error are both less than 16.25 W / m². 2 35.43W / m 2The regression coefficient for corrected temperature reached over 0.99, with mean absolute error and root mean square error both less than 0.18℃ and 0.25℃, respectively. The regression coefficient for corrected wind speed reached 0.86, with mean absolute error and root mean square error reaching 0.22m / s and 0.28m / s, respectively.

[0116] Figure 3 , Figure 4 , Figure 5 , Figure 6 The errors of correction method 1 and correction method 2 are respectively the errors between meteorological data after CNN-BiGRU-AM correction and actual meteorological data, and the errors between meteorological data after VMD-CNN-BiGRU-AM correction and actual meteorological data.

[0117] Table 4 Comparison of NWP Data Correction Results

[0118]

[0119]

[0120]

[0121] Step 6: For each sample, input the corrected NWP data from Step 5 into the hybrid prediction model from Step 3 to obtain the final prediction result, and evaluate the performance of the model.

[0122] Refer to Table 5 and Figure 7 Regarding metrics such as root mean square error (RMSE), mean absolute error (MAE), and mean error, using the VMD-CNN-BiGRU-AM corrected NWP data as input to the BO-CNN-LSTM hybrid model yields a predicted power that is very close to the predicted power obtained using actual meteorological data as input. The RMSE, MAE, and mean error compared to the actual power are only 0.0693, 0.0461, and 0.0208, respectively. The correlation coefficient, accuracy, and pass rate reach 0.9546, 93.07%, and 98.31%, respectively. These evaluations fully demonstrate the feasibility and superiority of the NWP spatial downscaling system. By improving the accuracy of NWP data as input to the photovoltaic power prediction model, the accuracy of power prediction is ensured.

[0123] Figure 7 The prediction power 1, prediction power 2, and prediction power 3 are respectively the prediction power obtained by using actual meteorological data as input, the prediction power obtained by using meteorological data after CNN-BiGRU-AM correction as input, and the prediction power obtained by using meteorological data after VMD-CNN-BiGRU-AM correction as input.

[0124] Table 5 Comparison of Prediction Results for Different Input Types

[0125]

[0126] In summary, this invention enables more accurate prediction of day-ahead power generation (NWP) of photovoltaic (PV) power plants. By improving the hybrid prediction model and correcting the NWP data, the accuracy of power prediction is significantly enhanced, making it valuable for application and promotion in the field of PV power generation prediction.

[0127] The descriptions and practices disclosed in this invention are readily apparent and understandable to those skilled in the art, and various modifications and refinements can be made without departing from the principles of this invention. Therefore, any modifications or improvements made without departing from the spirit of this invention should also be considered within the scope of protection of this invention.

Claims

1. A photovoltaic power generation prediction model incorporating corrected NWP data, characterized in that, Includes the following steps: S1. Collect data from the photovoltaic power station, including historical power data, historical actual data of various meteorological factors, and historical NWP data, and perform correlation analysis based on this data, including the following steps: S1.1 Calculate the Pearson correlation coefficient and Spearman correlation coefficient between photovoltaic power generation and historical actual meteorological data to determine the input characteristics of the prediction model; S1.2 Calculate the error between the historical actual data and NWP data of the input features, and analyze the error characteristics; S2. Using historical actual data of meteorological factors as input and historical power data as output, train a Bayesian-optimized CNN-LSTM hybrid prediction model, including the following steps: S2.

1. A convolutional neural network (CNN) module is introduced at the front end of the CNN-LSTM hybrid prediction model. Subsequently, a prediction model is established using a two-layer stacked long short-term memory (LSTM) network. S2.2 Introduce Bayesian optimization method before training to jointly optimize some hyperparameters with the goal of minimizing prediction error, and achieve efficient search with the help of Gaussian process surrogate model; S3. To address the discrepancy between NWP data and actual meteorological data, a spatial downscaling model for NWP based on VMD-CNN-BiGRU-AM is proposed, including the following steps: S3.1 First, variational mode decomposition (VMD) is used to decompose the error sequence between actual meteorological data and NWP data to obtain several sub-components with different frequency characteristics. S3.2 Subsequently, a multi-channel neural network structure is designed, which includes a convolutional neural network (CNN), a bidirectional gated recurrent unit (BiGRU), and an attention mechanism module (AM). S3.3 Next, the CNN-BiGRU-AM network is used to model and predict each component. S3.4 Finally, the prediction results of each component are reconstructed to achieve overall correction and optimization of the NWP data; S4. Combine the Bayesian optimization-based CNN-LSTM hybrid prediction model in S2 with the NWP spatial downscaling model in S3 to form a photovoltaic power generation prediction model that incorporates corrected NWP data. S5. Substitute the original NWP data of the forecast date into the photovoltaic power generation forecast model that incorporates the corrected NWP data, and evaluate the performance of the forecast model.

2. The photovoltaic power generation prediction model incorporating corrected NWP data according to claim 1, characterized in that, In S2, the specific steps of the CNN-LSTM hybrid prediction model based on Bayesian optimization are as follows: 1) A convolutional neural network module is introduced at the front end of the CNN-LSTM hybrid prediction model. A one-dimensional convolutional kernel is used to slide the input sequence in the time dimension to extract local features. The non-linear expression ability of the network is improved by using the ELU activation function. Batch normalization and pooling operations are combined to enhance the stability of training and the robustness of features. After processing by the CNN module, the output features are reorganized into a standard time series format by sequence unrolling and flattening operations, so that the subsequent long short-term memory module can further build a prediction model. 2) The LSTM part consists of two stacked layers. The first layer contains 50 hidden units for initial learning of temporal information, while the number of hidden units in the second layer is automatically searched and determined by a Bayesian optimization algorithm within a preset interval. The LSTM uses a standard gating structure, including forget gate, input gate, and output gate mechanisms, which enables the network to effectively capture long-term dependencies and suppress gradient vanishing. Subsequently, the CNN-LSTM hybrid prediction model outputs the photovoltaic power generation prediction results through a fully connected layer and a regression layer, and uses the mean squared error (MSE) loss function, combined with an L2 regularization term to suppress overfitting and improve generalization ability. 3) To ensure that the hyperparameter configuration of the CNN-LSTM hybrid prediction model reaches the optimal level, the Bayesian optimization method is used before network training to jointly optimize the three hyperparameters: the number of LSTM units, the initial learning rate, and the L2 regularization coefficient. With the goal of minimizing the prediction error, the Gaussian process surrogate model is used to efficiently search the hyperparameter space.

3. The photovoltaic power generation prediction model incorporating corrected NWP data according to claim 1, characterized in that, In S3, the NWP spatial downscaling model based on VMD-CNN-BiGRU-AM follows these steps: 1) Data preprocessing: Calculate the error between the input NWP data and the actual weather data, and use it as the raw error sequence. The specific expression is as follows: e(t)=y actual (t)-y NWP (t) Among them, y NWP (t), y actual e(t) represents the NWP data input at time t and the actual weather data, respectively; e(t) represents the original error sequence at time t. 2) Error Sequence Decomposition: Multimode Decomposition (VMD) is used to decompose the error sequence into K Intrinsic Mode Functions (IMFs). The specific expressions are as follows: Where K is the number of modes; u k (t) represents the value of the k-th IMF component at time t; 3) Construction of supervised samples: For each IMF component, input and output supervised samples are constructed using a sliding window, with input feature X. i By k im Composed of historical values, predicting target Y i For the future z im The value of the step; the specific expression is as follows: X i =[u k (i),u k (i+1),…,u k (i+k im -1)] Y i =u k (i+k im +z im -1) Among them, X i Let k be the input feature vector of the i-th sample, consisting of a length of k. im Composed of continuous historical values; Y i Let k be the predicted target value for the i-th sample; im z is the historical time step of the input sequence; im To predict the step size; 4) CNN-BiGRU-AM Model Design: For each IMF component, a multi-channel neural network structure is designed, including a convolutional neural network, bidirectional gated recurrent units, and an attention mechanism module. First, convolutional units are used to extract local spatial features. Second, bidirectional gated recurrent units are used to model the forward and backward information of the sequence, and the attention mechanism is used to enhance the response of salient features through channel weighting. Finally, the predicted value is output through a regression layer, the expression of which is as follows: Y i * =f BiGRU-AM (f CNN (X i )) Among them, f CNN f represents the convolutional feature extraction module; BiGRU-AM This represents a temporal prediction module that integrates a bidirectional gated recurrent unit and an attention mechanism; Y i * This is a predicted value; 5) Model training and prediction: Train a CNN-BiGRU-AM model independently for each IMF component and obtain its prediction results respectively; 6) Component reconstruction The prediction sequence of the original error is reconstructed by summing all the IMF prediction results, as shown in the following expression: Among them, e * (t) represents the prediction result at time t after reconstruction; This represents the prediction result of the k-th modal component at time t; Finally, by combining the NWP data and the predicted error sequence, the corrected NWP data is obtained and used as the input to the photovoltaic power generation prediction model to predict the power output; the specific expression is as follows: y * correct (t)=y NWP (t)+e * (t) Among them, y * correct (t) represents the corrected NWP data at time t.

Images