Photovoltaic load prediction method and system based on space-time mode cooperation and uncertainty perception
By employing multimodal spatiotemporal feature preprocessing and a three-dimensional collaborative attention mechanism, the problem of data complementarity and dynamic capture of spatiotemporal correlation in photovoltaic power generation load forecasting is solved, achieving high-precision photovoltaic power generation load forecasting and uncertainty quantification, supporting grid dispatching and energy storage configuration.
Patent Information
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2025-12-04
- Publication Date
- 2026-03-13
Smart Images
Figure CN121663476A_ABST
Abstract
Description
Technical Field
[0001] This invention belongs to the field of photovoltaic load forecasting technology, specifically relating to a photovoltaic load forecasting method and system based on spatiotemporal modal coordination and uncertainty perception. Background Technology
[0002] The statements in this section are merely background information related to the present invention and do not necessarily constitute prior art.
[0003] Photovoltaic power generation load is influenced by meteorological factors such as irradiance, temperature, and cloud movement, as well as spatial factors such as geographical topography and power plant layout. It exhibits three main characteristics: strong intermittency, strong volatility, and spatiotemporal correlation, posing severe challenges to grid frequency regulation, energy storage configuration, and renewable energy consumption. Accurate photovoltaic power generation load forecasting is a key technological support for reducing curtailment rates and ensuring the safe and economical operation of the power grid.
[0004] In power systems, photovoltaic (PV) power generation data typically exists in time-series format and is influenced by various heterogeneous data sources, including meteorological data, satellite remote sensing data, and equipment operation data. These data differ significantly in temporal resolution, spatial scale, and data format, creating a complex multimodal heterogeneous data environment. Traditional forecasting methods often rely on a single data source or homogeneous data, failing to fully utilize the complementary value of multimodal heterogeneous data and thus unable to comprehensively reflect the complexity of PV power generation systems.
[0005] Existing prediction methods have the following limitations: First, the adaptability of multimodal features is insufficient, failing to consider the dynamic differences in the contribution of different modes in different spatiotemporal scenarios, resulting in low feature utilization. Second, the spatiotemporal correlation modeling is fixed, making it impossible to dynamically capture the evolution of spatiotemporal correlation. Third, redundant information interference is significant, and there is a lack of targeted redundancy removal mechanisms, leading to redundant model parameters, reduced learning efficiency, and limited prediction accuracy. Fourth, there is a lack of uncertainty quantification, making it impossible to provide uncertainty information such as prediction confidence intervals, which makes it difficult to support risk decision-making in power grid dispatching. Summary of the Invention
[0006] To address the aforementioned issues, this invention proposes a photovoltaic load forecasting method and system based on spatiotemporal modal collaboration and uncertainty awareness. The invention uses multi-source data (including historical power load data, real-time meteorological data, and power plant archive data) collected by a photovoltaic power generation monitoring system as initial input. Multimodal spatiotemporal feature preprocessing is completed through spatiotemporal collaborative cleaning, spatiotemporal feature enhancement, and spatiotemporal block partitioning. Then, an adaptive graph topology learning module is constructed using spatiotemporal joint feature embedding and a sliding window mechanism to generate a dynamic spatiotemporal topology matrix that evolves over time to capture the spatiotemporal correlation of photovoltaic power load. Subsequently, a cross-modal attention mechanism based on spatiotemporal-modal three-dimensional collaboration dynamically allocates weights from modality, time, and space dimensions to achieve deep fusion of the two types of features and obtain a multimodal spatiotemporal fusion representation vector. A spatiotemporal redundancy removal module combining gating units and residual networks is used to remove repetitive semantic information from the fused features, obtaining a comprehensive representation vector of photovoltaic power load. Finally, point prediction is achieved through spatiotemporal ConvLSTM, while spatiotemporally aware Monte Carlo Dropout is combined to achieve photovoltaic power load prediction values, improving prediction accuracy.
[0007] According to some embodiments, the present invention adopts the following technical solution: A photovoltaic load forecasting method based on spatiotemporal modal coordination and uncertainty perception includes the following steps: Acquire multi-source data, perform multimodal spatiotemporal feature preprocessing, and obtain standardized multimodal spatiotemporal feature data through spatiotemporal collaborative cleaning, spatiotemporal feature enhancement, and spatiotemporal block partitioning; Based on multimodal spatiotemporal feature data, a spatiotemporal joint feature embedding and sliding window mechanism are used to generate a dynamic spatiotemporal topology matrix that evolves over time, capture the spatiotemporal correlation of photovoltaic power generation load, and obtain spatiotemporal correlation feature vectors. Based on multimodal spatiotemporal feature data and spatiotemporal associated feature vectors, a cross-modal attention mechanism with spatiotemporal-modal three-dimensional collaboration is adopted to dynamically allocate weights from three dimensions: modality, time, and space, to achieve deep fusion of multimodal features and spatiotemporal features and obtain a multimodal spatiotemporal fusion representation vector; By removing redundant semantic information from the fusion features of the multimodal spatiotemporal fusion representation vector, a comprehensive representation vector of photovoltaic power generation load is obtained; Based on the comprehensive representation vector of photovoltaic power generation load, a spatiotemporal collaborative probabilistic prediction model is constructed and trained to achieve point prediction. Combined with spatiotemporally aware Monte Carlo Dropout technology, the point prediction value and confidence interval of photovoltaic power generation load are obtained. The trained spatiotemporal collaborative probabilistic prediction model is used to process the acquired target multi-source data to obtain the final photovoltaic power generation load prediction value.
[0008] As an alternative implementation, the multi-source data includes historical power generation load data, real-time meteorological data, and power plant archive data.
[0009] As an alternative implementation method, the process of obtaining standardized multimodal spatiotemporal feature data through spatiotemporal collaborative cleaning, spatiotemporal feature enhancement, and spatiotemporal block partitioning includes: the multimodal spatiotemporal feature data is... Each component represents a standardized feature tensor of the meteorological mode, equipment mode, and load mode, respectively. For each modal data Standardized methods are used for processing:
[0010] in, They represent the first The mean and standard deviation of each modal data point. Indicates the first The modal data is standardized through spatiotemporal alignment and block partitioning to form a standard multimodal spatiotemporal dataset. .
[0011] As an alternative implementation method, a spatiotemporal joint feature embedding and sliding window mechanism is used to generate a dynamic spatiotemporal topology matrix that evolves over time. The process of capturing the spatiotemporal correlation of photovoltaic power generation load includes: defining the graph structure. ,in Indicates at time step The set of power station nodes, Let be the set of edges. This is a dynamic adjacency matrix, its elements are... Represents a node and In time The association strength is obtained through adaptive learning using the inner product of node features and a sliding window mechanism:
[0012] in, For nodes In time eigenvectors, , For a learnable parameter matrix, For feature dimension, For activation functions; A spatiotemporal convolution module combining graph convolutional networks and temporal convolutional networks is used to extract spatiotemporal correlation feature vectors. :
[0013]
[0014] in, To add self-connected adjacency matrices, Its degree matrix, For GCN The weight matrix of the layer, For activation function, and They represent the first Layer and first The feature representation vector of the layer.
[0015] As an alternative implementation, a spatiotemporal-modal three-dimensional collaborative cross-modal attention mechanism is adopted. The process of dynamically allocating weights from the three dimensions of modality, time, and space includes: concatenating features from different modalities with spatiotemporal features to form joint features. Attention weights are calculated in the three dimensions of modality, time, and space.
[0016] As a further step, the query vector for each dimension Key vector Sum value vector By joint features Obtained through linear transformation, based on the query vector. Key vector Sum value vector Calculate the corresponding attention score and the multimodal spatiotemporal fusion representation vector. The result is obtained by weighting the attention across three dimensions:
[0017] in, , , These are learnable weight parameters used to balance the importance of different dimensions.
[0018] As an alternative implementation, the process of removing repetitive semantic information from the fusion features of the multimodal spatiotemporal fusion representation vector to obtain the comprehensive representation vector of photovoltaic power generation load includes: defining a gating vector. and residual vector as follows:
[0019]
[0020] in, It is the Sigmoid activation function. , For learnable parameters, This represents a vector concatenation operation; Photovoltaic power generation load comprehensive representation vector We obtain the following through gating weighting and normalization:
[0021] in, This represents element-wise multiplication. Presentation layer normalization operation.
[0022] As an alternative implementation method, the process of constructing and training a spatiotemporal collaborative probabilistic prediction model based on the photovoltaic power generation load comprehensive representation vector to achieve point prediction includes: performing point prediction through spatiotemporal ConvLSTM:
[0023]
[0024] in, This is the predicted value for the photovoltaic power generation load point; , These are the output layer parameters; , These are the hidden state vectors at time t+1 and t in the ConvLSTM operation, respectively. and These are the intermediate state vectors at time t+1 and t respectively in the ConvLSTM operation.
[0025] As an alternative implementation, the process of combining spatiotemporally aware Monte Carlo Dropout technology includes: introducing spatiotemporally aware Monte Carlo Dropout, and conducting testing during the testing phase. Each forward propagation randomly discards a portion of the spatiotemporal connections, resulting in a set of prediction samples. The predicted confidence interval is obtained by calculating the sample statistic:
[0026]
[0027] in, This represents the mean. This represents the standard deviation; the final output confidence interval is... ,in The standard score corresponds to the required confidence level.
[0028] As an alternative implementation method, the training process of the spatiotemporal collaborative probability prediction model uses the mean squared error loss function for model training, with the loss... Defined as:
[0029] in, Indicates the number of samples. This is an L2 regularization term to prevent overfitting; This is a hyperparameter used to balance the loss.
[0030] As an alternative implementation, the following steps are also included: after the spatiotemporal collaborative probability prediction model is trained, it is tested using a test sample set. Based on the obtained prediction results, it is compared with the actual photovoltaic power generation load data, and the underlying data information of the spatiotemporal collaborative probability prediction model is updated to optimize the data weight values in the spatiotemporal collaborative probability prediction model.
[0031] A photovoltaic load forecasting system based on spatiotemporal modal coordination and uncertainty perception includes: The multimodal spatiotemporal feature preprocessing module is configured to acquire multi-source data, perform multimodal spatiotemporal feature preprocessing, and obtain standardized multimodal spatiotemporal feature data through spatiotemporal collaborative cleaning, spatiotemporal feature enhancement, and spatiotemporal block partitioning. The spatiotemporal adaptive graph topology learning module is configured to generate a dynamic spatiotemporal topology matrix that evolves over time based on multimodal spatiotemporal feature data and employs a spatiotemporal joint feature embedding and sliding window mechanism. This captures the spatiotemporal correlation of photovoltaic power generation loads and obtains spatiotemporal correlation feature vectors. The spatiotemporal cross-modal attention fusion module is configured to use a spatiotemporal-modal three-dimensional collaborative cross-modal attention mechanism based on multimodal spatiotemporal feature data and spatiotemporal associated feature vectors. It dynamically allocates weights from three dimensions: modality, time, and space, to achieve deep fusion of multimodal features and spatiotemporal features and obtain a multimodal spatiotemporal fusion representation vector. The spatiotemporal redundancy removal module is configured to remove redundant semantic information from the fusion features of the multimodal spatiotemporal fusion representation vector to obtain the comprehensive representation vector of photovoltaic power generation load; The prediction model construction and training module is configured to build and train a spatiotemporal collaborative probabilistic prediction model based on the comprehensive representation vector of photovoltaic power generation load to achieve point prediction, and combine spatiotemporally aware Monte Carlo Dropout technology to obtain the point prediction value and confidence interval of photovoltaic power generation load. The photovoltaic power generation load prediction module is configured to use a trained spatiotemporal collaborative probabilistic prediction model to process the acquired target multi-source data and obtain the final photovoltaic power generation load prediction value.
[0032] Compared with the prior art, the beneficial effects of the present invention are as follows: (1) Based on the multi-source data collected by the photovoltaic power generation monitoring system, this invention introduces a spatiotemporal collaborative cleaning and spatiotemporal feature enhancement mechanism through multimodal spatiotemporal feature preprocessing, effectively preserves the spatiotemporal local consistency of the data, mines local spatiotemporal correlation information, and provides high-quality feature input for subsequent modeling.
[0033] (2) This invention designs a spatiotemporal-modal three-dimensional collaborative attention fusion mechanism to dynamically balance the contributions of the three dimensions of modality, time, and space, realize deep collaboration between multimodal heterogeneous features and spatiotemporal features, and improve the pertinence and effectiveness of feature fusion; and proposes a spatiotemporal redundancy removal module, which combines gating units and residual networks to effectively suppress redundant information in the fusion process, alleviate the gradient vanishing problem, and improve model learning efficiency and prediction accuracy. (3) The present invention constructs a spatiotemporal collaborative probability prediction framework, and outputs point prediction values and confidence intervals to realize the quantification of prediction uncertainty, providing comprehensive decision support for applications such as power grid dispatch and energy storage configuration, and enhancing the practicality of prediction results.
[0034] To make the above-mentioned objects, features and advantages of the present invention more apparent and understandable, preferred embodiments are described below in detail with reference to the accompanying drawings. Attached Figure Description
[0035] The accompanying drawings, which form part of this invention, are used to provide a further understanding of the invention. The illustrative embodiments of the invention and their descriptions are used to explain the invention and do not constitute an improper limitation of the invention.
[0036] Figure 1 A flowchart of a photovoltaic power generation load forecasting method provided in one embodiment; Figure 2 A flowchart of a photovoltaic power generation load prediction embodiment is provided as one example; Figure 3 This is a schematic diagram of a photovoltaic power generation load forecasting system provided in one embodiment. Detailed Implementation
[0037] The present invention will be further described below with reference to the accompanying drawings and embodiments.
[0038] It should be noted that the following detailed description is illustrative and intended to provide further explanation of the invention. Unless otherwise specified, all technical and scientific terms used herein have the same meaning as commonly understood by one of ordinary skill in the art to which this invention pertains.
[0039] It should be noted that the terminology used herein is for the purpose of describing particular embodiments only and is not intended to limit the scope of exemplary embodiments according to the invention. As used herein, the singular form is intended to include the plural form as well, unless the context clearly indicates otherwise. Furthermore, it should be understood that when the terms "comprising" and / or "including" are used in this specification, they indicate the presence of features, steps, operations, devices, components, and / or combinations thereof.
[0040] Where there is no conflict, the embodiments and features described in this application may be combined with each other.
[0041] Example 1 like Figure 1 As shown in this embodiment, a photovoltaic load forecasting method based on spatiotemporal modal coordination and uncertainty perception includes: A. Collect and preprocess multi-source heterogeneous data from the photovoltaic power generation monitoring system. Specifically, the monitoring data of a photovoltaic power station cluster in a certain region is used as the dataset. This dataset includes power generation load data, meteorological monitoring data, and power station equipment records of 58 photovoltaic power stations in the region from January 2025 to June 2025.
[0042] Define the preprocessed multimodal spatiotemporal feature data as follows: Each component represents a standardized feature tensor of the meteorological mode, equipment mode, and load mode, respectively.
[0043] For each modal data The Z-Score normalization method is used for processing:
[0044] in, They represent the first The mean and standard deviation of each modality. Indicates the first Each modal data point is standardized, and then spatiotemporal alignment and block partitioning are performed to ultimately form a standard multimodal spatiotemporal dataset. .
[0045] In some embodiments, such as Figure 2 As shown, it also includes the process of filling in missing values and normalizing the data in the sample set.
[0046] B. Based on the preprocessed multimodal spatiotemporal feature data, this embodiment constructs an adaptive graph topology learning module to capture the dynamic spatiotemporal correlation between photovoltaic power plant groups.
[0047] Define graph structure ,in Indicates at time step The set of power station nodes, Let be the set of edges. This is a dynamic adjacency matrix, its elements are... Represents a node and In time The association strength is obtained through adaptive learning using the inner product of node features and a sliding window mechanism:
[0048] in, For nodes In time eigenvectors, , For a learnable parameter matrix, For feature dimension, This is the activation function.
[0049] Subsequently, a spatiotemporal convolution module combining a graph convolution network (GCN) and a temporal convolutional network (TCN) is used to extract spatiotemporal correlation feature vectors. :
[0050]
[0051] in, To add self-connected adjacency matrices, Its degree matrix, For GCN The weight matrix of the layer, For activation function, and They represent the first Layer and first The feature representation vector of the layer.
[0052] C. Obtain multimodal spatiotemporal feature data based on step A. Spatiotemporal correlation feature vector with step B This embodiment designs a spatiotemporal-modal three-dimensional collaborative cross-modal attention mechanism to achieve deep fusion of multimodal features.
[0053] First, features from different modalities are concatenated with spatiotemporal features to form joint features. Subsequently, attention weights were calculated in the modality, time, and space dimensions.
[0054] Taking modal dimension as an example, query vector Key vector Sum value vector By joint features Through linear transformation, we obtain:
[0055]
[0056]
[0057] in , , Both represent weight matrices.
[0058] Modal attention score The calculation is as follows:
[0059] Similarly, computation of time-dimension attention Spatial Dimension Attention The final multimodal spatiotemporal fusion representation vector The result is obtained by weighting the attention across three dimensions:
[0060] in, , , These are learnable weight parameters used to balance the importance of different dimensions.
[0061] D. The fusion representation vector obtained in step C In this embodiment, the gating unit is merged into the residual structure, and a spatiotemporal redundancy removal module is designed to remove duplicate semantic information that may be generated during the fusion process.
[0062] Define the gate vector and residual vector as follows:
[0063]
[0064] in, It is the Sigmoid activation function. , For learnable parameters, This indicates a vector concatenation operation.
[0065] Final photovoltaic power generation load comprehensive representation vector We obtain the following through gating weighting and normalization:
[0066] in, This represents element-wise multiplication. Presentation layer normalization operation.
[0067] E. The comprehensive representation vector obtained in step D Input the spatiotemporal collaborative probability prediction module. This module first performs point prediction using spatiotemporal ConvLSTM:
[0068]
[0069] in, This is the predicted value for the photovoltaic power generation load point; , These are the output layer parameters; , These are the hidden state vectors at time t+1 and t in the ConvLSTM operation, respectively. and These are the intermediate state vectors at time t+1 and t respectively in the ConvLSTM operation.
[0070] To quantify prediction uncertainty, this embodiment introduces spatiotemporally aware Monte Carlo Dropout. This is implemented during the testing phase. Each forward propagation randomly discards a portion of the spatiotemporal connections, resulting in a set of prediction samples. The predicted confidence interval is obtained by calculating the sample statistic:
[0071]
[0072] in, This represents the mean. It represents the standard deviation.
[0073] The final output confidence interval is ,in The standard score corresponds to the required confidence level.
[0074] F. Based on this embodiment, the mean squared error loss function is used for model training. Loss Defined as:
[0075] in, Indicates the number of samples. This is an L2 regularization term to prevent overfitting; This is a hyperparameter used to balance the loss.
[0076] After the model is trained, the predicted results of the test sample set are output and compared with the actual photovoltaic power generation load data. The underlying data information is updated in response to the feedback, thereby continuously optimizing the data weight values in the model and continuously improving the photovoltaic power generation load prediction.
[0077] To verify the performance of the proposed method in this embodiment, traditional LSTM, ConvLSTM, and Spatial-Temporal Graph Convolutional Network (ST-GCN) were selected as comparison models, and Mean Absolute Error (MAE), Root Mean Square Error (RMSE), and Prediction Interval Coverage Probability (PICP) were selected as evaluation metrics. Based on the results in Table 1, the photovoltaic load forecasting method proposed in this embodiment outperforms other methods.
[0078] Table 1. Performance comparison of different prediction methods
[0079] Example 2 A photovoltaic load forecasting system based on spatiotemporal modal coordination and uncertainty perception, such as Figure 3 As shown, it includes: The multimodal spatiotemporal feature preprocessing module is configured to acquire multi-source data, perform multimodal spatiotemporal feature preprocessing, and obtain standardized multimodal spatiotemporal feature data through spatiotemporal collaborative cleaning, spatiotemporal feature enhancement, and spatiotemporal block partitioning. The spatiotemporal adaptive graph topology learning module is configured to generate a dynamic spatiotemporal topology matrix that evolves over time based on multimodal spatiotemporal feature data and employs a spatiotemporal joint feature embedding and sliding window mechanism. This captures the spatiotemporal correlation of photovoltaic power generation loads and obtains spatiotemporal correlation feature vectors. The spatiotemporal cross-modal attention fusion module is configured to use a spatiotemporal-modal three-dimensional collaborative cross-modal attention mechanism based on multimodal spatiotemporal feature data and spatiotemporal associated feature vectors. It dynamically allocates weights from three dimensions: modality, time, and space, to achieve deep fusion of multimodal features and spatiotemporal features and obtain a multimodal spatiotemporal fusion representation vector. The spatiotemporal redundancy removal module is configured to remove redundant semantic information from the fusion features of the multimodal spatiotemporal fusion representation vector to obtain the comprehensive representation vector of photovoltaic power generation load; The prediction model construction and training module is configured to build and train a spatiotemporal collaborative probabilistic prediction model based on the comprehensive representation vector of photovoltaic power generation load to achieve point prediction, and combine spatiotemporally aware Monte Carlo Dropout technology to obtain the point prediction value and confidence interval of photovoltaic power generation load. The photovoltaic power generation load prediction module is configured to use a trained spatiotemporal collaborative probabilistic prediction model to process the acquired target multi-source data and obtain the final photovoltaic power generation load prediction value.
[0080] Those skilled in the art will understand that embodiments of the present invention can be provided as methods, systems, or computer program products. Therefore, the present invention can take the form of a completely hardware embodiment, a completely software embodiment, or an embodiment combining software and hardware aspects. Furthermore, the present invention can take the form of one or more computer-usable storage media (including, but not limited to, disk storage, etc.) containing computer-usable program code. CD - ROM It takes the form of a computer program product implemented on (such as optical memory, etc.).
[0081] This invention is described with reference to flowchart illustrations and / or block diagrams of methods, apparatus (systems), and computer program products according to embodiments of the invention. It will be understood that each block of the flowchart illustrations and / or block diagrams, and combinations of blocks in the flowchart illustrations and / or block diagrams, can be implemented by computer program instructions. These computer program instructions can be provided to a processor of a general-purpose computer, special-purpose computer, embedded processor, or other programmable data processing apparatus to produce a machine, such that the instructions, which execute via the processor of the computer or other programmable data processing apparatus, generate instructions for implementing the flowchart illustrations and / or block diagrams. Figure 1 One or more processes and / or boxes Figure 1 A device that provides the functions specified in one or more boxes.
[0082] These computer program instructions may also be stored in a computer-readable storage medium that can direct a computer or other programmable data processing device to function in a particular manner, such that the instructions stored in the computer-readable storage medium produce an article of manufacture including instruction means, which are implemented in a process Figure 1 One or more processes and / or boxes Figure 1 The function specified in one or more boxes.
[0083] These computer program instructions may also be loaded onto a computer or other programmable data processing equipment to cause a series of operational steps to be performed on the computer or other programmable equipment to produce a computer-implemented process, thereby providing instructions that execute on the computer or other programmable equipment for implementing the process. Figure 1 One or more processes and / or boxes Figure 1 The steps of the function specified in one or more boxes.
[0084] The above description is merely a preferred embodiment of the present invention and is not intended to limit the invention. Various modifications and variations can be made to the present invention by those skilled in the art. Any modifications, equivalent substitutions, improvements, etc., made by those skilled in the art without creative effort within the spirit and principles of the present invention should be included within the scope of protection of the present invention.
Claims
1. A photovoltaic load forecasting method based on spatiotemporal modal coordination and uncertainty perception, comprising the following steps: Acquire multi-source data, perform multimodal spatiotemporal feature preprocessing, and obtain standardized multimodal spatiotemporal feature data through spatiotemporal collaborative cleaning, spatiotemporal feature enhancement, and spatiotemporal block partitioning; Based on multimodal spatiotemporal feature data, a spatiotemporal joint feature embedding and sliding window mechanism are used to generate a dynamic spatiotemporal topology matrix that evolves over time, capture the spatiotemporal correlation of photovoltaic power generation load, and obtain spatiotemporal correlation feature vectors. Based on multimodal spatiotemporal feature data and spatiotemporal associated feature vectors, a cross-modal attention mechanism with spatiotemporal-modal three-dimensional collaboration is adopted to dynamically allocate weights from three dimensions: modality, time, and space, to achieve deep fusion of multimodal features and spatiotemporal features and obtain a multimodal spatiotemporal fusion representation vector; By removing redundant semantic information from the fusion features of the multimodal spatiotemporal fusion representation vector, a comprehensive representation vector of photovoltaic power generation load is obtained; Based on the comprehensive representation vector of photovoltaic power generation load, a spatiotemporal collaborative probabilistic prediction model is constructed and trained to achieve point prediction. Combined with spatiotemporally aware Monte Carlo Dropout technology, the point prediction value and confidence interval of photovoltaic power generation load are obtained. The trained spatiotemporal collaborative probabilistic prediction model is used to process the acquired target multi-source data to obtain the final photovoltaic power generation load prediction value.
2. The photovoltaic load forecasting method based on spatiotemporal modal coordination and uncertainty perception as described in claim 1, characterized in that, The process of obtaining standardized multimodal spatiotemporal feature data through spatiotemporal collaborative cleaning, spatiotemporal feature enhancement, and spatiotemporal block partitioning includes: the multimodal spatiotemporal feature data is... Each component represents a standardized feature tensor of the meteorological mode, equipment mode, and load mode, respectively. For each modal data Standardized methods are used for processing: in, They represent the first The mean and standard deviation of each modal data point. Indicates the first The modal data is standardized through spatiotemporal alignment and block partitioning to form a standard multimodal spatiotemporal dataset. .
3. The photovoltaic load forecasting method based on spatiotemporal modal coordination and uncertainty perception as described in claim 1, characterized in that, The process of generating a dynamic spatiotemporal topology matrix that evolves over time, using a spatiotemporal joint feature embedding and sliding window mechanism, to capture the spatiotemporal correlation of photovoltaic power generation loads includes: defining the graph structure. ,in Indicates at time step The set of power station nodes, Let be the set of edges. This is a dynamic adjacency matrix, its elements are... Represents a node and In time The association strength is obtained through adaptive learning using the inner product of node features and a sliding window mechanism: in, For nodes In time eigenvectors, , For a learnable parameter matrix, For feature dimension, For activation functions; A spatiotemporal convolution module combining graph convolutional networks and temporal convolutional networks is used to extract spatiotemporal correlation feature vectors. : in, To add self-connected adjacency matrices, Its degree matrix, For GCN The weight matrix of the layer, For activation function, and They represent the first Layer and first The feature representation vector of the layer.
4. The photovoltaic load forecasting method based on spatiotemporal modal coordination and uncertainty perception as described in claim 1, characterized in that, The cross-modal attention mechanism employing spatiotemporal-modal three-dimensional collaboration dynamically allocates weights across modality, time, and space. This process includes concatenating features from different modalities with spatiotemporal features to form joint features. Attention weights are calculated in three dimensions: modality, time, and space; the query vector for each dimension... Key vector Sum value vector By joint features Obtained through linear transformation, based on the query vector. Key vector Sum value vector Calculate the corresponding attention score and the multimodal spatiotemporal fusion representation vector. The result is obtained by weighting the attention across three dimensions: in, , , These are learnable weight parameters used to balance the importance of different dimensions.
5. The photovoltaic load forecasting method based on spatiotemporal modal coordination and uncertainty perception as described in claim 1, characterized in that, The process of removing redundant semantic information from the fusion features of the multimodal spatiotemporal fusion representation vector to obtain the comprehensive representation vector of photovoltaic power generation load includes: defining a gating vector. and residual vector as follows: in, It is the Sigmoid activation function. , For learnable parameters, This represents a vector concatenation operation; Photovoltaic power generation load comprehensive representation vector We obtain the following through gating weighting and normalization: in, This represents element-wise multiplication. Presentation layer normalization operation.
6. The photovoltaic load forecasting method based on spatiotemporal modal coordination and uncertainty perception as described in claim 1, characterized in that, The process of constructing and training a spatiotemporal collaborative probabilistic prediction model based on the photovoltaic power generation load comprehensive representation vector to achieve point prediction includes: point prediction through spatiotemporal ConvLSTM: in, This is the predicted value for the photovoltaic power generation load point; , These are the output layer parameters; , These are the hidden state vectors at time t+1 and t in the ConvLSTM operation, respectively. and These are the intermediate state vectors at time t+1 and t respectively in the ConvLSTM operation.
7. The photovoltaic load forecasting method based on spatiotemporal modal coordination and uncertainty perception as described in claim 1, characterized in that, The process of combining spatiotemporally aware Monte Carlo Dropout technology includes: introducing spatiotemporally aware Monte Carlo Dropout, and conducting testing during the testing phase. Each forward propagation randomly discards a portion of the spatiotemporal connections, resulting in a set of prediction samples. The predicted confidence interval is obtained by calculating the sample statistic: in, This represents the mean. This represents the standard deviation; the final output confidence interval is... ,in The standard score corresponds to the required confidence level.
8. The photovoltaic load forecasting method based on spatiotemporal modal coordination and uncertainty perception as described in claim 1, characterized in that, The training process of the spatiotemporal collaborative probabilistic prediction model uses the mean squared error loss function for model training. Defined as: in, Indicates the number of samples. This is an L2 regularization term to prevent overfitting; This is a hyperparameter used to balance the loss.
9. The photovoltaic load forecasting method based on spatiotemporal modal coordination and uncertainty perception as described in claim 1, characterized in that, It also includes the following steps: After the spatiotemporal collaborative probabilistic prediction model is trained, it is tested using a test sample set. Based on the obtained prediction results, it is compared with the actual photovoltaic power generation load data to update the underlying data information of the spatiotemporal collaborative probabilistic prediction model and optimize the data weight values in the spatiotemporal collaborative probabilistic prediction model.
10. A photovoltaic load forecasting system based on spatiotemporal modal coordination and uncertainty perception, characterized in that it includes: The multimodal spatiotemporal feature preprocessing module is configured to acquire multi-source data, perform multimodal spatiotemporal feature preprocessing, and obtain standardized multimodal spatiotemporal feature data through spatiotemporal collaborative cleaning, spatiotemporal feature enhancement, and spatiotemporal block partitioning. The spatiotemporal adaptive graph topology learning module is configured to generate a dynamic spatiotemporal topology matrix that evolves over time based on multimodal spatiotemporal feature data and employs a spatiotemporal joint feature embedding and sliding window mechanism. This captures the spatiotemporal correlation of photovoltaic power generation loads and obtains spatiotemporal correlation feature vectors. The spatiotemporal cross-modal attention fusion module is configured to use a spatiotemporal-modal three-dimensional collaborative cross-modal attention mechanism based on multimodal spatiotemporal feature data and spatiotemporal associated feature vectors. It dynamically allocates weights from three dimensions: modality, time, and space, to achieve deep fusion of multimodal features and spatiotemporal features and obtain a multimodal spatiotemporal fusion representation vector. The spatiotemporal redundancy removal module is configured to remove redundant semantic information from the fusion features of the multimodal spatiotemporal fusion representation vector to obtain the comprehensive representation vector of photovoltaic power generation load; The prediction model construction and training module is configured to build and train a spatiotemporal collaborative probabilistic prediction model based on the comprehensive representation vector of photovoltaic power generation load to achieve point prediction, and combine spatiotemporally aware Monte Carlo Dropout technology to obtain the point prediction value and confidence interval of photovoltaic power generation load. The photovoltaic power generation load prediction module is configured to use a trained spatiotemporal collaborative probabilistic prediction model to process the acquired target multi-source data and obtain the final photovoltaic power generation load prediction value.