A photovoltaic power generation amount prediction method and system applied to a photovoltaic power station

Abstract

The application discloses a photovoltaic power generation amount prediction method and system applied to a photovoltaic power station, relates to the photovoltaic power generation prediction field, and comprises the following steps: performing Pearson correlation analysis and Spearman correlation analysis to determine key covariants; performing multivariate singular spectrum signal decomposition processing to obtain reconstructed sequence data; training a multi-scale covariant interaction model by using the reconstructed sequence data to obtain a local model; based on a multi-source domain collaborative training framework, uploading model parameters of the local model corresponding to all photovoltaic power stations to perform global training to obtain a trained global model; and adopting a transfer learning fine-tuning strategy to fine-tune the trained global model to obtain a photovoltaic power generation amount prediction model. The application can improve the generalization ability of the model used in the photovoltaic power generation amount prediction process, realizes high-precision and high-stability prediction of the photovoltaic power generation amount, and is especially suitable for application scenarios with poor data quality.

Description

Technical Field

[0001] This application relates to the field of photovoltaic power generation prediction, and in particular to a method and system for predicting photovoltaic power generation in photovoltaic power plants. Background Technology

[0002] Accurate photovoltaic (PV) power generation prediction algorithms can help predict whether a PV power generation system will experience redundancy or insufficiency in the future. This allows for the combined use of other power supply strategies to achieve rational power system scheduling, optimize energy utilization, and balance supply and demand. Therefore, research on PV power generation prediction algorithms is of great significance for improving the utilization efficiency of PV power generation, maintaining stable grid operation, and achieving energy conservation and emission reduction.

[0003] Currently, data-driven deep learning algorithms for photovoltaic (PV) power generation forecasting offer higher stability and reliability, eliminating the need for complex physical models and providing a more feasible alternative. Furthermore, most existing studies employ a covariate input strategy, integrating multiple relevant time-series data as auxiliary information to achieve superior performance compared to methods relying solely on a single target variable. However, existing multivariate input-based PV power generation forecasting methods still face two key challenges:

[0004] First, although multivariate input strategies can theoretically improve prediction accuracy, their actual effectiveness largely depends on data quality, the selection of covariates, and the model's ability to effectively utilize covariate information. Furthermore, improper use of covariates may reduce the overall performance of the model.

[0005] Second, photovoltaic power generation forecasting usually requires at least a full year of high-quality data. However, the data available for a certain period of time is of poor quality and has serious gaps, resulting in low accuracy in photovoltaic power generation forecasting. Summary of the Invention

[0006] The purpose of this application is to provide a method and system for predicting photovoltaic power generation in photovoltaic power plants, which can achieve high-precision and high-stability prediction of photovoltaic power generation, and is especially suitable for application scenarios with poor data quality.

[0007] To achieve the above objectives, this application provides the following solution:

[0008] Firstly, this application provides a method for predicting photovoltaic power generation in photovoltaic power plants, including:

[0009] Obtain operational datasets from multiple photovoltaic power plants; the operational datasets include target variables and multiple covariates of the photovoltaic power plants within a preset time period;

[0010] Pearson correlation analysis and Spearman correlation analysis were performed on multiple running datasets, and key covariates were determined by combining them with preset correlation thresholds;

[0011] For any photovoltaic power station, multivariate singular spectrum signal decomposition is performed based on the target variables and corresponding key covariates in the operation dataset to obtain reconstructed sequence data.

[0012] The reconstructed sequence data is used to train a multi-scale covariate interaction model to obtain a local model. Based on a multi-source domain collaborative training framework, the model parameters of the local models corresponding to all photovoltaic power stations are uploaded for training to obtain an initial global model. Then, the model parameters of the initial global model are distributed to each photovoltaic power station for the next round of local model training until the preset training termination condition is met, at which point the trained global model is obtained and then distributed to each photovoltaic power station.

[0013] For any photovoltaic power plant, based on the target variables and corresponding key covariates in the operational dataset, a transfer learning fine-tuning strategy is adopted to fine-tune the trained global model to obtain a photovoltaic power generation prediction model.

[0014] The target variables and corresponding key covariates of the photovoltaic power station are input into the photovoltaic power generation prediction model to obtain the future target variables.

[0015] Secondly, this application provides a photovoltaic power generation prediction system for photovoltaic power plants, comprising:

[0016] The operation data acquisition module is used to acquire operation datasets from multiple photovoltaic power plants; the operation datasets include target variables and multiple covariates of the photovoltaic power plants within a preset time period;

[0017] The key covariate extraction module is used to perform Pearson correlation analysis and Spearman correlation analysis on multiple running datasets, and determine key covariates by combining preset correlation thresholds.

[0018] The reconstructed sequence module is used to perform multivariate singular spectrum signal decomposition processing on any photovoltaic power station based on the target variables and corresponding key covariates in the operation dataset to obtain reconstructed sequence data.

[0019] The model training module is used to train the multi-scale covariate interaction model using the reconstructed sequence data to obtain a local model; based on the multi-source domain collaborative training framework, the model parameters of the local models corresponding to all the photovoltaic power stations are uploaded for training to obtain an initial global model, and then the model parameters of the initial global model are distributed to each of the photovoltaic power stations for the next round of local model training until the preset training end condition is met, at which point the trained global model is obtained and then distributed to each of the photovoltaic power stations;

[0020] The model fine-tuning module is used to fine-tune the trained global model for any photovoltaic power plant based on the target variables and corresponding key covariates in the running dataset, using a transfer learning fine-tuning strategy to obtain a photovoltaic power generation prediction model.

[0021] The prediction application module is used to input the target variables to be used by the photovoltaic power station and the corresponding key covariates into the photovoltaic power generation prediction model to obtain the future target variables.

[0022] According to the specific embodiments provided in this application, this application has the following technical effects: This application identifies key covariates through correlation analysis, which can fully utilize effective covariates and provide a data foundation for subsequent processing, which is of great significance for improving the predictive ability of the target variable. By introducing multivariate singular spectrum analysis, the feature representation ability of the input data is enhanced, and noise interference is effectively suppressed; through the application of a multi-scale covariate interaction model, deep interaction modeling between multivariate sequences in the time and feature dimensions is realized, which can fully explore the contribution of covariate information in the prediction task and significantly improve the prediction accuracy of the target variable; the reconstructed sequence data after multivariate singular spectrum signal decomposition is applied to the multi-source domain collaborative training framework to participate in model training, which can enhance the generalization ability of the model while protecting data privacy, and significantly improve the accuracy and stability of the prediction model under poor data quality conditions. Furthermore, based on model parameters from multiple photovoltaic sites, a global model that can be transferred across sites is constructed, realizing knowledge sharing and model collaboration. On this basis, combined with a transfer learning strategy based on a small number of high-quality samples, the models of each site are individually adjusted, avoiding dependence on large-scale data preprocessing. The aforementioned processing in this application improves the model's generalization ability under heterogeneous data conditions, effectively mitigates the negative impact of low-quality covariates on prediction results, and achieves information complementarity at the feature level, thereby enhancing the stability and adaptability of overall prediction performance. In summary, this application demonstrates good robustness in dealing with low-quality covariate information and is suitable for real-world complex data scenarios. Attached Figure Description

[0023] To more clearly illustrate the technical solutions in the embodiments of this application or the prior art, the drawings used in the embodiments will be briefly introduced below. Obviously, the drawings described below are only some embodiments of this application. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0024] Figure 1 This is a schematic diagram of a photovoltaic power generation prediction method applied to a photovoltaic power plant in one embodiment of this application.

[0025] Figure 2 This is a schematic diagram of a photovoltaic power generation prediction method applied to a photovoltaic power plant in another embodiment of this application.

[0026] Figure 3 This is a schematic diagram illustrating the implementation details of a multi-source domain collaborative training framework in another embodiment of this application. Detailed Implementation

[0027] The technical solutions of the embodiments of this application will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of this application, and not all embodiments. Based on the embodiments of this application, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of this application.

[0028] This application realizes multi-source domain collaborative photovoltaic power generation prediction based on covariate interaction and transfer learning enhancement. It can still achieve accurate and stable photovoltaic power generation prediction even when the data quality is poor, and has good practicality.

[0029] To make the above-mentioned objectives, features and advantages of this application more apparent and understandable, this application will be further described in detail below with reference to the accompanying drawings and specific embodiments.

[0030] In one exemplary embodiment, a method for predicting photovoltaic power generation in a photovoltaic power plant is provided. This method is executed by a computer device, specifically a terminal or server, or both. In this embodiment, the method implements multi-source domain collaborative photovoltaic power generation prediction based on covariate interaction and transfer learning enhancement, which can be referred to as MDCTL-MCI. Figure 1 As shown, the method includes steps 101 to 106.

[0031] Step 101: Obtain the operation dataset of multiple photovoltaic power plants; the operation dataset includes the target variable and multiple covariates of the photovoltaic power plants within a preset time period.

[0032] In practical applications, the selected photovoltaic power plants vary significantly in technical specifications and geographical location, with rated output capacities ranging from 30 MW to 130 MW. These power plants are located in northern, central, and northwestern China, covering various terrain types, including deserts, mountains, and plains. In the collected raw dataset, the target variable, namely photovoltaic power generation data, is relatively complete; however, the missing rates of covariate information, such as solar radiation and weather conditions, vary from 0% to 80% across different sites. The raw dataset itself provides both a version with missing data and a version with imputed data. This application uses the imputed version for experiments, i.e., as the running dataset.

[0033] In a specific application scenario, such as Figure 2 As shown, a total of four photovoltaic power stations were selected, corresponding to... Figure 2 The dataset includes photovoltaic (PV) sites 1, 2, 3, and 4. The preset duration is one year, but can be adjusted as needed. In short, the operational dataset comprises one year of continuous operational data from the PV power plants. This data includes PV power generation as the target variable, solar irradiance as a covariate, and weather data, among others. The sampling frequency is standardized, and timestamps are aligned for easier subsequent processing.

[0034] Step 102: Perform Pearson correlation analysis and Spearman correlation analysis on multiple running datasets, and determine key covariates by combining preset correlation thresholds. Specifically, Pearson correlation analysis is used to determine the linear correlation between each covariate and the target variable; Spearman correlation analysis is used to determine the non-linear correlation between each covariate and the target variable; finally, the covariates that show a strong correlation with the target variable in both of the above correlation analyses are identified as key covariates.

[0035] In a specific application, the key covariates determined based on preset correlation thresholds include: global horizontal irradiance (GHI), direct normal irradiance (DNI), and total solar irradiance (TSI).

[0036] In another specific application, the dataset can be divided into training, validation, and test sets in a 6:1:1 ratio for subsequent training processes.

[0037] Step 103: For any photovoltaic power station, multivariate singular spectrum signal decomposition processing is performed based on the target variable and corresponding key covariates in the operation dataset to obtain reconstructed sequence data. The multivariate singular spectrum signal decomposition processing, through four steps of trajectory matrix construction, singular spectrum decomposition, grouping, and recombination, can improve the expressive ability of time series information while denoising the data.

[0038] In a specific application, step 103 includes the following steps (31)-(35).

[0039] (31) For each variable in the running dataset corresponding to the running data subsequence, a trajectory matrix is ​​constructed by embedding; specifically, the target variable and the corresponding key covariate in the running dataset are defined as input to a multivariate time series X, with a length of N and variables p and x. ij Let be the value of the i-th variable at the j-th time point.

[0040] The running data subsequence X corresponding to the i-th variable i For example, X i By selecting appropriate embedding dimensions and delay parameters, the sliding window method is used to transform the running data subsequence into a trajectory matrix X. ′ i As shown in the following formula:

[0041]

[0042] Where L represents the size of the embedding window, and the dimension of the embedding matrix is ​​(N-L+1)×L.

[0043] (32) Perform singular spectral decomposition on the trajectory matrix to obtain a left singular vector matrix U, a right singular vector matrix V, and a diagonal matrix Σ containing singular values; wherein the left singular vector matrix and the right singular vector matrix represent X. ′ i The principal components of the row and column spaces, Σ, are diagonal matrices containing singular values, denoted as: Σ=diag(σ1,σ2,…,σ i ), σ i This represents singular values. Singular values ​​are arranged in descending order, with larger singular values ​​corresponding to the dominant signal component and smaller singular values ​​typically corresponding to noise.

[0044] (33) Based on the magnitude of the singular values ​​in the diagonal matrix, singular values ​​are filtered and the corresponding singular vectors are retained to obtain singular groups. Specifically, the top s singular values ​​are selected and their corresponding eigenvectors are retained to preserve the original information. To avoid redundancy and reduce computational complexity, in a specific experimental setting, s is fixed at 4.

[0045] (34) Based on the singular grouping, signal reconstruction is performed to obtain a reconstructed subsequence; specifically, based on the grouping results, signal reconstruction is performed on the selected singular value vectors to obtain the recombined sequence. Represented as:

[0046] (35) Concatenate and aggregate the reconstructed subsequences corresponding to all variables to obtain the reconstructed sequence data. Specifically, repeat the above process for each subsequence of running data corresponding to each variable, and then concatenate and aggregate the results to obtain the final reconstructed data X. r ,

[0047] Step 104: The reconstructed sequence data is used to train the multi-scale covariate interaction model to obtain a local model; based on the multi-source domain collaborative training framework, the model parameters of the local models corresponding to all photovoltaic power stations are uploaded for training to obtain an initial global model, and then the model parameters of the initial global model are distributed to each photovoltaic power station for the next round of local model training until the preset training end condition is met, at which point the trained global model is obtained and then distributed to each photovoltaic power station.

[0048] In this step, a multi-scale covariate interaction model is trained to achieve time-series feature modeling. This fully utilizes covariate information to improve the predictive ability of the target variable. Furthermore, it is extended to a multi-source domain collaborative training framework. By weighted aggregation of model parameters from multiple photovoltaic power plants (e.g., four), a global model that can be transferred across sites is constructed, realizing knowledge sharing and model collaboration.

[0049] Specifically, the Multi-Scale Covariate Interaction (MCI) model enables deep interaction modeling between multivariate sequences in both the time and feature dimensions. This allows for the full utilization of covariate information in the prediction task, significantly improving the prediction accuracy of the target variable. The MCI model includes a temporal mixing module, a feature fusion module, and a multi-scale convolution module. These modules are connected using residual connections to preserve information during transmission, thereby enhancing model stability and improving prediction accuracy.

[0050] The temporal mixing module is used to extract features from the reconstructed sequence data along the time dimension using an MLP with shared weights, in order to capture global dynamic dependencies across time steps, and inputs the obtained global dynamic features into the feature fusion module. Specifically, the input data... First it is transposed to obtain Then, the sequence is projected into a higher-dimensional feature space through a fully connected layer, and a non-linear transformation using the ReLU activation function is applied. Finally, the transformed features are mapped back to the original sequence length to obtain the output of the temporal mixing module. The data processing function corresponding to the time mixing module is:

[0051]

[0052] in, For global dynamic features, Dropout() is the regularization method, Linear2() is the function of the second fully connected layer, ReLU() is the activation function, and Linear1() is the function of the first fully connected layer. X in For the variables in the reconstructed sequence data that are used as input, For X in The transpose of .

[0053] The feature fusion module is used to extract features from the global dynamic features along the variable dimension to capture the complex interaction between the multidimensional covariates output from the multivariate singular spectral decomposition and the target variable, thereby extracting temporal interaction features that are beneficial to improving the accurate prediction of the target variable. The obtained temporal interaction features are then input into the multi-scale convolution module. The data processing function corresponding to the feature fusion module is:

[0054]

[0055] in, It is a temporal interaction feature.

[0056] The multi-scale convolution module is used to perform multi-scale cross-processing on the temporal interaction features to simultaneously extract local and global features, and obtain the final output features through fusion. The multi-scale convolution module can improve the model's ability to capture temporal information at multiple time scales, and it simultaneously extracts local features and global patterns from time series data.

[0057] Specifically, the input data is processed through two convolutional operations, Conv1 and Conv2, each with a different receptive field to generate feature maps X1 and X2 of different sizes. Conv1 uses a 1×1 convolutional kernel to capture short-term fluctuations in the time series, such as periodic components, while Conv2 uses a larger 3×3 convolutional kernel to extract long-term global trends, thereby enhancing its ability to model long-term dependencies. Subsequently, X1 and X2 are fused using a Hadamard product. Finally, temporal features are further extracted using a 1×1 convolutional kernel, Conv3, to obtain the final output features. The data processing function corresponding to the multi-scale convolution module is:

[0058]

[0059] in, For the final output features, Conv3 is the convolution kernel, X1 is the feature map of the local features obtained by performing convolution operation using the Conv1 convolution kernel, X2 is the feature map of the global features obtained by performing convolution operation using the Conv2 convolution kernel, and the Conv1 convolution kernel and the Conv2 convolution kernel are of different sizes, and ⊙ is the Hadamard product operation.

[0060] After determining the above model, a multi-source domain collaborative training framework is further used to train a global model. This process consists of five key stages: model initialization, local training, parameter uploading, global parameter aggregation, and parameter broadcasting. Figure 3 As shown.

[0061] (41) Model initialization: Based on the above multi-scale covariate interaction model, each local photovoltaic power station initializes its local model.

[0062] (42) Local training: In each round of communication, each local photovoltaic power station uses its own dataset D k Independently train its local model θ k The loss function is:

[0063]

[0064] Among them, D k For the reconstructed sequence data corresponding to the k-th photovoltaic power station, L k (θ) represents the loss function value of the multi-scale covariate interaction model θ for the k-th photovoltaic power station, x i To reconstruct the i-th sample variable in the sequence data, y i For x i The corresponding true value, f θ (x i ) for x i The corresponding predicted value, l(·), represents the loss function applied to each sample variable in the reconstructed sequence data.

[0065] (43) Parameter upload: The local model updates its parameters through gradient descent and then resends the updated weights to the central server.

[0066] (4) Global parameter aggregation: The central server uses a weighted average strategy to calculate the new global model parameters, which is implemented using the following function:

[0067]

[0068] Where, θ t+1 These are the model parameters for the updated global model. Let be the model parameters of the local model obtained by the k-th photovoltaic power station in the t-th round of training; K is the total number of photovoltaic power stations, |D k | represents the number of sample variables for the k-th photovoltaic power station.

[0069] (5) Parameter broadcasting: After aggregation, the central server will broadcast the updated global model θ t+1The data is broadcast to each local photovoltaic power station for the next round of training. Through multiple rounds of communication and iterative updates, the global model θ is finally obtained. global .

[0070] To improve computational efficiency, automatic mixed precision (AMS) is employed to accelerate training during the global model's training process. AMS aims to improve model deployment flexibility and training efficiency by accelerating computation and reducing GPU memory consumption while maintaining model accuracy. Specifically, during the forward propagation phase, AMS uses half-precision floating-point numbers for tensor calculations, significantly reducing memory usage and increasing computational throughput. During the backpropagation phase, a dynamic loss scaling mechanism is introduced to mitigate gradient underflow caused by half-precision training. This mechanism adaptively amplifies the loss value to ensure the numerical stability of gradients within the representable range of half-precision.

[0071] Step 105: For any photovoltaic power station, based on the target variables and corresponding key covariates in the operational dataset, a transfer learning fine-tuning strategy is used to fine-tune the trained global model to obtain a photovoltaic power generation prediction model. Specifically, transfer learning is used to locally fine-tune the global model, reducing the impact of data heterogeneity, improving prediction accuracy, and obtaining the final photovoltaic power generation prediction model.

[0072] In a specific application, step 105 includes the following steps (51)-(53).

[0073] (51) Based on preset conditions, data is extracted from the target variables and corresponding key covariates in the running dataset to obtain partial running sample data.

[0074] (52) Perform multivariate singular spectrum signal decomposition on the partial running sample data to obtain partially reconstructed sequence data. Thus, this application achieves denoising processing for a small amount of data to ensure the accuracy of subsequent fine-tuning.

[0075] (53) Based on the model parameters of the completed global model, the model is fine-tuned using the partially reconstructed sequence data. During the fine-tuning training, only the parameters of the output layer of the global model are adjusted until the preset training stop condition is met, thus obtaining the photovoltaic power generation prediction model.

[0076] Specifically, a global model is used as the pre-training weights for transfer learning. Before fine-tuning, the parameters of all layers except the output layer are fixed and denoted as θ. frozen No parameter updates are performed; only the output layer parameters are updated, denoted as θ. out , can be represented as: θ global ={θ frozen ,θ out}, θfrozen θ represents the frozen model parameters. out This represents the output layer model parameters that participate in the weight update.

[0077] The fine-tuning process aims to optimize the model's output layer parameters based on the local dataset, resulting in better model parameters. The following objective function is used during fine-tuning training:

[0078]

[0079] in, Let θ be the model parameters of the photovoltaic power generation prediction model obtained after fine-tuning training for the k-th photovoltaic power station. out The parameters of the output layer for weight updates, θ frozen For the model parameters in the global model other than the output layer, which serve as the frozen layer, |D1 k | represents the number of sample variables in the partially reconstructed sequence data of the k-th photovoltaic power station, x1 i For the i-th sample variable in the partially reconstructed sequence data, y1 i To be with x1 i The corresponding true value, f(x1) i ;θ frozen ,θ out ) is related to x1 i The corresponding predicted value, l(·), represents the loss function applied to each sample variable in the partially reconstructed sequence data.

[0080] To improve computational efficiency, an automatic mixed precision mode is also used to accelerate training during fine-tuning, which is similar to the global training process described above and will not be repeated here.

[0081] Thus, the photovoltaic power generation prediction models corresponding to each photovoltaic power station have been obtained. To verify the effectiveness of the technical solution of this application, two types of experiments can be carried out: Experiment 1 explores the ability of the proposed model to utilize covariates, and Experiment 2 explores the stability of the proposed model under different covariate missing rates.

[0082] Experiment 1: By considering two methods—using a single objective variable as input and adding covariates as input—the constructed multi-source domain collaborative photovoltaic power generation prediction method based on covariate interaction and transfer learning enhancement was validated. The impact of adding covariates on the prediction accuracy of different models was explored, and the ability of this application to utilize covariates was evaluated.

[0083] (1) Experimental parameter settings. The same hyperparameter settings were used for the four photovoltaic power stations. The training epoch was set to 10, the batch size was set to 32, the early stop mechanism was used, and the patience was set to 3. The learning rate was set to lr = 0.001, the number of hidden layer nodes was set to d_model = 1024, the historical observation step size was set to 48, the future prediction step size was set to 48, and a one-time prediction method was adopted instead of a step-by-step cyclic prediction method.

[0084] (2) Evaluation Indicators. Mean Absolute Error (MAE) and Mean Squared Error (MSE) are used to evaluate whether the addition of covariates improves the model's predictive performance. Mean Absolute Error measures the average level of the absolute error between predicted and true values; a smaller value indicates higher overall prediction accuracy. Mean Squared Error represents the average of the squares of the differences between predicted and true values, reflecting the overall bias of the model in numerical prediction. The formulas for calculating MAE and MSE are shown below:

[0085]

[0086] (3) Selection of comparison methods. In order to fully verify the performance of this application, we selected excellent time series prediction methods in recent years to carry out comparative experiments, including MLP-based models LightTS and TSMixer, local model MCI, and attention-based models Pyraformer, Transformer, Informer, iTransformer, PatchTST, TimeXer, etc.

[0087] (4) To avoid the randomness of the results, all experiments were repeated 5 times with different random number seeds, and the average of the 5 experiments was taken. The experimental results are shown in Table 1. SITE1 to SITE4 represent 4 different photovoltaic sites, Univariate represents the input result of a single objective variable, and Covariate represents the prediction result of covariate input.

[0088] Table 1. Schematic diagram of the impact of adding covariates on model prediction accuracy.

[0089]

[0090] As shown in Table 1, different models exhibit significant performance differences after incorporating covariate information, and the characteristics of the dataset itself also affect performance after adding covariates. Specifically, the MDCTL-MCI proposed in this application and its local model MCI demonstrate significant improvements in prediction performance across different photovoltaic sites after incorporating covariate information. This indicates that this application effectively utilizes covariate information, enhancing the model's predictive ability for the target variable. Among the baseline methods, the MLP-based method shows an overall improvement in prediction accuracy after incorporating covariates. In contrast, the attention mechanism-based method underutilizes covariates, showing little or no change in accuracy, and its overall performance is lower than that of the MLP-based method.

[0091] Experiment 2: Based on the actual data missing situation in the dataset, and while ensuring the integrity of the target variable, covariate missing rates of 0%, 10%, 30%, 50%, 70%, and 90% were set sequentially. The accuracy and robustness of this application were verified by incorporating covariate information with different missing rates.

[0092] (1) The experimental parameter settings and the selection of comparison file methods are consistent with those in Experiment 1 above.

[0093] (2) The evaluation metrics MAE and MSE remain consistent with those described in Experiment 1 above. To measure the dispersion of prediction error relative to the mean error, a new evaluation metric, the Coefficient of Variation (CV), is added. This metric can be used to quantify the sensitivity of prediction error under different levels of missing covariates, thereby assessing the robustness of the model under uncertainty. The calculation formula is shown below:

[0094]

[0095] Where σ represents the standard deviation of the prediction error, which reflects the volatility of the error; μ represents the mean of the prediction error; and CV is used to measure the overall degree of prediction deviation.

[0096] (3) To analyze the stability of the proposed method under different covariate missing rates, the experiment was conducted with a fixed random number seed. The experimental results are shown in Table 2. SITE1 to SITE4 represent four photovoltaic sites. Rate0 to Rate0.9 represent the six covariate missing rates set in this experiment. Rank represents the ranking of the average prediction accuracy of each method under different covariate missing rates across the four photovoltaic sites. CV shows the sensitivity of MAE and MSE to changes in covariate missing rate.

[0097] Table 2 Comparison of model prediction performance under different covariate missing rates

[0098]

[0099] As shown in Table 2, the MDCTL-MCI model of this application consistently achieved the highest average prediction accuracy across various photovoltaic power plants, maintaining stable prediction performance under various missing variable rates. Specifically, with a covariate missing rate of 0, compared to the baseline method, the MAE and MSE of the MDCTL-MCI model were reduced by 6.51% and 14.5%, respectively. Under conditions with covariate missing rates, the MDCTL-MCI model exhibited a more significant advantage, with MAE and MSE reduced by 9.7% and 20.9%, respectively. In the baseline method, attention-based models are generally less accurate than MLP-based models and are more sensitive to changes in the covariate missing rate.

[0100] Specifically, LightTS outperforms other baseline models in prediction performance, while TimeXer exhibits the highest accuracy among attention-based methods. However, compared to the MDCTL-MCI model proposed in this application, LightTS shows an increase of 4.16% and 7.97% in average MAE and MSE, respectively, while TimeXer shows increases of 4.07% and 19.7%, respectively. Compared to the local model proposed in this application, MCI's prediction accuracy is comparable to MDCTL-MCI in the absence of missing covariates, which can be attributed to its strong ability to utilize covariate data, thereby improving prediction accuracy. However, with the increase in the covariate missing rate, MCI cannot maintain stable prediction accuracy, leading to a significant increase in prediction error. In contrast, this application maintains high accuracy while having a smaller overall CV index, indicating that it maintains stable prediction performance under various site data conditions and various missing rate conditions.

[0101] In summary, this demonstrates the significant superiority of our application under conditions of low covariate quality.

[0102] Step 106: Input the target variable to be used and the corresponding key covariates of the photovoltaic power station into the photovoltaic power generation prediction model to obtain the future target variable.

[0103] This application provides an efficient method for modeling covariate information. By introducing multivariate singular spectral analysis (MSSA), it enhances the feature representation capability of input data and effectively suppresses noise interference, avoiding reliance on large-scale data preprocessing. The lightweight multi-scale covariate interaction model and automatic mixed-precision training strategy employed demonstrate competitive performance in terms of parameter counting, model complexity, iteration time, and memory consumption, thereby reducing computational burden while maintaining prediction accuracy and stability. Combined with a transfer learning strategy based on a small number of high-quality samples, the model is individually adjusted for each site. Furthermore, this application improves the model's generalization ability under heterogeneous data conditions, effectively mitigating the negative impact of low-quality covariates on prediction results and achieving feature-level information complementarity, thereby enhancing the stability and adaptability of overall prediction performance.

[0104] Based on the same inventive concept, this application also provides a photovoltaic power generation prediction system for photovoltaic power plants, which implements the methods described above. The solution provided by this system is similar to the implementation scheme described in the above methods; therefore, the specific limitations in one or more system embodiments provided below can be found in the limitations of the methods described above, and will not be repeated here.

[0105] The photovoltaic power generation prediction system for photovoltaic power plants provided in this application includes:

[0106] The operation data acquisition module is used to acquire operation datasets from multiple photovoltaic power plants; the operation datasets include target variables and multiple covariates of the photovoltaic power plants within a preset time period.

[0107] The key covariate extraction module is used to perform Pearson correlation analysis and Spearman correlation analysis on multiple running datasets, and determine key covariates by combining preset correlation thresholds.

[0108] The reconstructed sequence module is used to perform multivariate singular spectrum signal decomposition processing on any photovoltaic power station based on the target variables and corresponding key covariates in the operating dataset to obtain reconstructed sequence data.

[0109] The model training module is used to train a multi-scale covariate interaction model based on a multi-source domain collaborative training framework using the reconstructed sequence data to obtain a local model. The model parameters of the local models corresponding to all photovoltaic power stations are uploaded for global training to obtain a global model. Then, the model parameters of the global model are distributed to each of the photovoltaic power stations for the next round of local model training until the preset training termination conditions are met, at which point the trained global model is obtained and then distributed to each of the photovoltaic power stations.

[0110] The model fine-tuning module is used to fine-tune the trained global model for any photovoltaic power plant based on the target variables and corresponding key covariates in the running dataset, using a transfer learning fine-tuning strategy to obtain a photovoltaic power generation prediction model.

[0111] The prediction application module is used to input the target variables to be used by the photovoltaic power station and the corresponding key covariates into the photovoltaic power generation prediction model to obtain the future target variables.

[0112] In summary, this application employs multivariate singular spectral decomposition (MSD) to denoise the original time-series data of each photovoltaic power station, aiming to remove random interference while preserving the main trends, thereby enhancing the data's expressive power and modeling stability. Secondly, a lightweight multi-scale covariate interaction module is constructed. This module models the relationships between variables in both the variable and time dimensions, and uses multi-scale convolution operations to simultaneously extract long-term trends and local details, effectively capturing the dynamic correlations between variables. Subsequently, a multi-source domain collaborative training mechanism is proposed. Initial models are trained separately based on local data from each photovoltaic site. Without directly transmitting the original data, feature-level knowledge fusion is achieved through weighted averaging, thus constructing a global prediction model containing collaborative features from multiple sites. Finally, a transfer learning fine-tuning strategy is used. Based on the aforementioned global model, it is used as a pre-trained model, and a small amount of local data processed by multivariate SSD is introduced to fine-tune the model, making it better adaptable to the actual data characteristics of each site. This yields the final photovoltaic power generation prediction model for practical prediction applications.

[0113] Compared with the prior art, this application has the following advantages:

[0114] (1) The effectiveness of utilizing covariate features is enhanced by the multi-scale covariate interaction module, and the prediction accuracy is improved by using covariates to assist in the prediction of target variables.

[0115] (2) By employing a multivariate singular spectrum signal decomposition and multi-source domain collaborative training framework, the generalization ability of the model is enhanced while protecting data privacy, and the accuracy and stability of the prediction model are significantly improved even under conditions of poor data quality. Compared with traditional methods, this approach has strong practical value and promising prospects for wider application.

[0116] In one exemplary embodiment, a computer device is provided, including a memory and a processor, wherein the memory stores a computer program, and the processor executes the computer program to implement the steps in the above-described method embodiments.

[0117] In one exemplary embodiment, a computer-readable storage medium is provided storing a computer program that, when executed by a processor, implements the steps in the above-described method embodiments.

[0118] In one exemplary embodiment, a computer program product is provided, including a computer program that, when executed by a processor, implements the steps in the above-described method embodiments.

[0119] It should be noted that the user information (including but not limited to user device information, user personal information, etc.) and data (including but not limited to data used for analysis, data stored, data displayed, etc.) involved in this application are all information and data authorized by the user or fully authorized by all parties, and the collection, use and processing of the relevant data must comply with relevant regulations.

[0120] Those skilled in the art will understand that all or part of the processes in the methods of the above embodiments can be implemented by a computer program instructing related hardware. The computer program can be stored in a non-volatile computer-readable storage medium, and when executed, it can include the processes of the embodiments of the above methods. Any references to memory, databases, or other media used in the embodiments provided in this application can include at least one of non-volatile and volatile memory. Non-volatile memory can include read-only memory (ROM), magnetic tape, floppy disk, flash memory, optical memory, high-density embedded non-volatile memory, resistive random access memory (ReRAM), magnetic random access memory (MRAM), ferroelectric random access memory (FRAM), phase change memory (PCM), graphene memory, etc. Volatile memory can include random access memory (RAM) or external cache memory, etc. By way of illustration and not limitation, RAM can take many forms, such as Static Random Access Memory (SRAM) or Dynamic Random Access Memory (DRAM).

[0121] The databases involved in the embodiments provided in this application may include at least one type of relational database and non-relational database. Non-relational databases may include, but are not limited to, blockchain-based distributed databases. The processors involved in the embodiments provided in this application may be general-purpose processors, central processing units, graphics processing units, digital signal processors, programmable logic devices, quantum computing-based data processing logic devices, etc., and are not limited to these.

[0122] The technical features of the above embodiments can be combined in any way. For the sake of brevity, not all possible combinations of the technical features in the above embodiments are described. However, as long as there is no contradiction in the combination of these technical features, they should be considered to be within the scope of this specification.

[0123] This document uses specific examples to illustrate the principles and implementation methods of this application. The descriptions of the above embodiments are only for the purpose of helping to understand the methods and core ideas of this application. Furthermore, those skilled in the art will recognize that, based on the ideas of this application, there will be changes in the specific implementation methods and application scope. Therefore, the content of this specification should not be construed as a limitation of this application.

Claims

1. A photovoltaic power generation amount prediction method applied to a photovoltaic power plant, characterized by, The method comprises: obtaining a plurality of operating data sets of photovoltaic power stations; the operating data set comprises a target variable and a plurality of covariates of the photovoltaic power station within a preset time length; performing Pearson correlation analysis and Spearman correlation analysis on a plurality of the operating data sets, and determining key covariates in combination with a preset correlation threshold; for any photovoltaic power station, performing multivariate singular spectrum signal decomposition processing on the target variable and the corresponding key covariates in the operating data set to obtain reconstructed sequence data; performing multivariate singular spectrum signal decomposition processing on the target variable and the corresponding key covariates in the operating data set to obtain reconstructed sequence data, comprising: for each variable corresponding to the operating data subsequence in the operating data set, constructing a trajectory matrix through embedding; performing singular spectrum decomposition on the trajectory matrix to obtain a left singular vector matrix, a right singular vector matrix, and a diagonal matrix containing singular values; performing singular value screening and corresponding singular vector reservation based on the size of the singular values in the diagonal matrix to obtain a singular group; performing signal reconstruction based on the singular group to obtain a reconstructed subsequence; and splicing and aggregating the reconstructed subsequences corresponding to all variables to obtain reconstructed sequence data; training a multi-scale covariate interaction model using the reconstructed sequence data to obtain a local model; based on a multi-source domain collaborative training framework, uploading model parameters of the local model corresponding to all the photovoltaic power stations for training to obtain an initial global model, and then downloading the model parameters of the initial global model to each of the photovoltaic power stations for training of the next round of local models, until a preset training end condition is met, to obtain a trained global model, which is then downloaded to each of the photovoltaic power stations; the multi-scale covariate interaction model comprises a time mixing module, a feature fusion module, and a multi-scale convolution module; the time mixing module is used to extract features of the reconstructed sequence data along the time dimension using an MLP with shared weights, and input the obtained global dynamic features into the feature fusion module; the feature fusion module is used to extract features of the global dynamic features along the variable dimension, and input the obtained time sequence interaction features into the multi-scale convolution module; the multi-scale convolution module is used to perform multi-scale cross processing on the time sequence interaction features to simultaneously extract local features and global features, and obtain final output features through fusion; for any photovoltaic power station, based on the target variable and the corresponding key covariates in the operating data set, a transfer learning fine-tuning strategy is used to fine-tune the trained global model to obtain a photovoltaic power generation capacity prediction model; inputting the target variable to be used of the photovoltaic power station and the corresponding key covariates into the photovoltaic power generation capacity prediction model to obtain a future target variable.

2. The photovoltaic power generation amount prediction method for a photovoltaic power plant according to claim 1, characterized by, The data processing function corresponding to the time mixing module is: ； wherein, is a global dynamic feature, () is a regularization method, () is a function of the second fully connected layer, () is an activation function, () is a function of the first fully connected layer, is a variable in the reconstructed sequence data as input, is a transpose of . The data processing function corresponding to the feature fusion module is: ； wherein, is a timing interaction feature; The data processing function corresponding to the multi-scale convolution module is: ； wherein, is the final output feature, is a convolution kernel, is a Hadamard product operation, is a feature map of local features obtained by performing a convolution operation, is a Hadamard product operation, is a feature map of global features obtained by performing a convolution operation with the convolution kernel, and is different from a size of the convolution kernel, is a Hadamard product operation.

3. The photovoltaic power generation amount prediction method for a photovoltaic power plant according to claim 1, characterized by, In the process of training the multi-scale covariate interaction model using the reconstructed sequence data to obtain a local model, the loss function is: ； wherein, is the reconstruction sequence data corresponding to the kth photovoltaic power station, is the multi-scale covariant interaction model of the kth photovoltaic power station is the loss function value of the multi-scale covariant interaction model, is the ith sample variable in the reconstruction sequence data, is the true value corresponding to is the predicted value corresponding to is the predicted value corresponding to is the predicted value corresponding to represents the loss function applied to each sample variable in the reconstruction sequence data.

4. The photovoltaic power generation amount prediction method for a photovoltaic power plant according to claim 1, characterized by, The step of uploading model parameters of all local models corresponding to the photovoltaic power stations to perform global training to obtain a global model is implemented by using the following function: ； wherein, are model parameters of the updated global model; are model parameters of the local model of the kth photovoltaic power station obtained in the tth round of training; is the total number of photovoltaic power stations, is the number of sample variables of the kth photovoltaic power station. 5.The photovoltaic power generation amount prediction method for a photovoltaic power plant according to claim 1, characterized by, In the training process of the global model, an automatic mixed precision mode is used for training acceleration.

6. The photovoltaic power generation amount prediction method for a photovoltaic power plant according to claim 1, characterized by, Based on the target variables and corresponding key covariates in the operation data set, a transfer learning fine-tuning strategy is used to fine-tune the trained global model to obtain a photovoltaic power generation prediction model, including: Based on a preset condition, the target variables and corresponding key covariates in the operation data set are intercepted to obtain partial operation sample data; The partial operation sample data is subjected to multivariate singular spectrum signal decomposition processing to obtain partial reconstructed sequence data; Based on the model parameters of the trained global model, the partial reconstructed sequence data is used for model fine-tuning training, and only the parameters of the output layer of the global model are adjusted during the fine-tuning training process until the preset training stopping condition is met, to obtain a photovoltaic power generation prediction model.

7. The photovoltaic power generation amount prediction method for a photovoltaic power plant according to claim 6, characterized by, In the fine-tuning training process, the following objective function is used: ； wherein, is the model parameter of the photovoltaic power generation amount prediction model obtained after fine-tuning training for the kth photovoltaic power station, is the parameter of the output layer of the weight update, is the model parameter of the global model except the output layer as a frozen layer, is the number of sample variables in the partial reconstruction sequence data of the kth photovoltaic power station, is the ith sample variable in the partial reconstruction sequence data, is the corresponding true value, is the corresponding predicted value, is the corresponding predicted value, is the corresponding predicted value, represents the loss function applied to each sample variable in the partial reconstruction sequence data.

8. A photovoltaic power generation amount prediction system for a photovoltaic power plant, in which the photovoltaic power generation amount prediction method for a photovoltaic power plant according to any one of claims 1 to 7 is applied, characterized by The system comprises: An operation data acquisition module configured to acquire operation data sets of a plurality of photovoltaic power stations; the operation data set comprises target variables and a plurality of covariates of a photovoltaic power station within a preset time length; A key covariate extraction module configured to perform Pearson correlation analysis and Spearman correlation analysis on a plurality of the operation data sets, and determine key covariates in combination with a preset correlation threshold; A reconstructed sequence module configured to, for any photovoltaic power station, perform multivariate singular spectrum signal decomposition processing on the target variables and corresponding key covariates in the operation data set to obtain reconstructed sequence data; A model training module configured to train a multi-scale covariate interaction model using the reconstructed sequence data to obtain a local model; based on a multi-source domain collaborative training framework, upload model parameters of all local models corresponding to the photovoltaic power stations to perform training to obtain an initial global model, then download the model parameters of the initial global model to each of the photovoltaic power stations for training of the next round of local models, until a preset training end condition is met, to obtain a trained global model, and then download to each of the photovoltaic power stations; A model fine-tuning module configured to, for any photovoltaic power station, based on the target variables and corresponding key covariates in the operation data set, use a transfer learning fine-tuning strategy to fine-tune the trained global model to obtain a photovoltaic power generation prediction model; A prediction application module configured to input target variables and corresponding key covariates to be used of the photovoltaic power station into the photovoltaic power generation prediction model to obtain future target variables.

Images