Multi-scale tensor diagram space-time network wind power prediction method fusing space-time correlation and environmental factors

Abstract

The invention provides a multi-scale tensor diagram space-time network wind power prediction method fusing space-time correlation and environmental factors, which comprises the following steps of: firstly, realizing high-quality completion of incomplete wind power data on the basis of tensor ring decomposition and in combination with diagram structure constraint, time smooth constraint and non-negative constraint; the importance of each constraint item is dynamically balanced through an adaptive weight mechanism; then, extracting local space-time characteristics under different time granularities by adopting multi-scale three-dimensional convolution generated by a dynamic convolution kernel and combining frequency domain transformation; and finally, constructing a double-layer dynamic graph structure consisting of a long-term dependency graph and a short-term dependency graph, dynamically adjusting an edge weight in combination with a physical model, and realizing multi-time-span time sequence modeling in multi-scale expansion time convolution. According to the method, missing data can be effectively reconstructed, the complex space-time dependency relationship between fans can be accurately described, minute-level wind speed mutation and hour-level environment change characteristics are considered, and higher precision and stability are shown in a long-time prediction task.

Description

Technical Field

[0001] This invention belongs to the field of wind power prediction technology, specifically involving a multi-scale tensor graph spatiotemporal network wind power prediction method that integrates spatiotemporal correlation and environmental factors, applicable to smart grid dispatching and new energy management. Background Technology

[0002] Wind power forecasting is a key technology for achieving stable grid operation and efficient utilization of new energy sources. With the rapid growth of wind power installed capacity, traditional forecasting methods are no longer sufficient to meet the accuracy requirements of modern power systems. Statistical learning methods such as ARIMA and support vector regression, while computationally efficient, are essentially univariate time series models and cannot capture the complex spatiotemporal relationships between multiple wind turbines in a wind farm. Deep learning methods such as Long Short-Term Memory networks and gated recurrent units improve time series modeling capabilities to some extent, but the data processing requires forcibly flattening the data, which has a three-dimensional structure of turbines, time, and features, into a two-dimensional sequence, inevitably losing important high-order feature interaction information. More advanced spatiotemporal graph neural networks, while modeling spatial correlations by introducing graph structures, still have significant limitations. These models have high requirements for the completeness of input data, but actual wind farm monitoring data often contains a large number of missing and outlier values, and conventional interpolation methods can destroy the inherent spatiotemporal characteristics of the data. Furthermore, these models typically use static adjacency matrices based on geographical distance to define the connections between wind turbines, which cannot adapt to the dynamic characteristics of wind speed propagation paths changing with meteorological conditions.

[0003] The challenges of wind power forecasting also lie in multi-scale feature extraction. Instantaneous changes in wind speed often occur on a minute-by-minute scale, while changes in environmental parameters such as temperature and humidity exhibit hourly or even longer-term periodicity. Most existing prediction models employ single-scale feature extraction methods, making it difficult to simultaneously capture the changing patterns across these different time scales. Furthermore, current mainstream tensor completion methods, such as CP decomposition, lack explicit constraints on the spatial distribution characteristics of wind turbines during data reconstruction, resulting in insufficient physical consistency in the completion results. These technical bottlenecks severely restrict further improvements in prediction accuracy, especially in large-scale wind farm scenarios with high data loss rates and strong spatiotemporal coupling; the prediction performance of existing methods often falls short of practical application requirements. Summary of the Invention

[0004] To address the aforementioned issues, this invention provides a multi-scale tensor graph spatiotemporal network wind power prediction method that integrates spatiotemporal correlations and environmental factors. This method innovatively combines multi-constraint tensor completion technology with graph neural networks, achieving high-quality reconstruction of incomplete wind power data through a triple regularization mechanism that incorporates graph structure constraints, temporal smoothing constraints, and non-negativity constraints. Simultaneously, the designed multi-scale tensor convolution module can extract local features at different temporal granularities in parallel, while the dynamic graph learning mechanism can adaptively capture the spatiotemporal correlations between wind turbines as meteorological conditions change. This innovative framework effectively solves the technical bottlenecks of existing methods in areas such as data missingness handling, multi-scale feature extraction, and dynamic spatiotemporal modeling, providing a novel technical path for improving the accuracy of wind power prediction.

[0005] To achieve the above objectives, the present invention provides the following technical solution: a wind power prediction method based on multi-scale tensor graph spatiotemporal networks, comprising the following steps:

[0006] Step 1: Data preprocessing and high-order tensor construction. For multi-source data from wind farms, construct a third-order tensor X∈R for wind turbine-time-feature. N×T×F Where N is the number of wind turbines, T is the time step, and F is the feature dimension, outlier data is cleaned through spatial correlation analysis and temporal interpolation strategies, specifically including the following sub-steps:

[0007] Step 1-1: Cluster the wind turbines based on their geographical location and fill in missing values ​​using the within-group mean.

[0008] Steps 1-2 involve using time series interpolation to fill in the remaining missing values, ensuring data integrity.

[0009] Step 2, multi-constraint tensor completion, using tensor ring (TR) decomposition to complete the tensor. A low-rank approximation is performed, and graph structure constraints, time smoothing constraints, and non-negativity constraints are introduced to achieve high-quality reconstruction of incomplete wind power data. The specific sub-steps include:

[0010] Step 2-1: Through tensor ring decomposition, the original tensor X is represented as a series of third-order tensor factors {G}. (1) G (2) G (3)}, where G (k) Let R be the k-th third factor tensor of the tensor ring. k represents the connection dimension between adjacent factors. The tensor ring decomposition form is: Where i∈[1,N], j∈[1,T], k∈[1,F];

[0011] Step 2-2, Structural constraints in the design drawings: in Let d be the spatial factor matrix. s As a potential dimension of space, The graph Laplace matrix reflects the spatial topology of the wind turbine, and Tr(·) is the matrix trace operation.

[0012] Steps 2-3 introduce time smoothing constraints: Ensure time continuity, among which Let d be the time factor matrix. t As a potential dimension of time, This represents the t-th row of the time factor matrix, where ||·||2 is the L2 norm, used to measure the continuity between adjacent time steps.

[0013] Steps 2-4: Add nonnegativity constraint: L nonneg =||ReLU(-U (f) )||1, Enhance the physical rationality of features, among which Let d be the eigenfactor matrix. f is the latent dimension of the feature, ReLU(·) is the linear rectified function that sets negative values ​​to zero, and ||·||1 is the L1 norm, used to measure the strength of nonnegativity constraints.

[0014] Steps 2-5 involve determining the weights of each constraint term using an adaptive weighting mechanism. An adaptive weighting network f is then constructed. θ (·), the input is the missing pattern matrix M, where missing values ​​are 1 and non-missing values ​​are 0) and the time-varying statistical features, and the output is the weights (λ). g ,λ t ,λ n The final optimization objective is: L tensor =L rec +λ g L graph +λ t L time +λ n L nonneg ,in, For tensor reconstruction error, ||·|| F It is the Frobenius norm. This is the result of tensor ring decomposition and reconstruction.

[0015] Compared to existing technologies, the advantages of this step are:

[0016] (1) Dynamic weight allocation to adapt to different missing patterns: The weights of each constraint can be automatically adjusted according to the missing pattern matrix to adapt to different missing rates and missing distribution conditions, significantly improving the generalization ability of tensor completion.

[0017] (2) Combine time-varying characteristics to achieve time-series adaptiveness: By using the statistical characteristics of time-varying input, the weights are dynamically adjusted over time, and the optimal completion performance can be maintained during both the wind speed change period and the stable period.

[0018] (3) End-to-end learnable, reducing manual parameter tuning: The weight allocation process is jointly optimized with tensor decomposition and prediction tasks in a unified framework, without relying on repeated trial and error based on human experience, thus reducing the cost of model migration and deployment.

[0019] (4) Improve the consistency between completion and prediction: Achieve constraint balance between the completion process and the prediction task to ensure that the completion result has both low reconstruction error and improved downstream prediction performance, avoiding the problem of "high completion accuracy but decreased prediction performance" in existing methods.

[0020] Step 3, Multi-scale tensor feature extraction: Perform multi-scale convolution operations on the completed tensor to extract local spatiotemporal features, specifically including the following sub-steps:

[0021] Step 3-1: Adaptively generate three-dimensional convolutional kernels of different sizes and weights through the Dynamic Kernel Generator (DKG) module. The kernel size can be dynamically selected or weighted and fused in {3×3×3, 5×5×5, 7×7×7} to adapt to different meteorological models and spatiotemporal dependence characteristics.

[0022] Step 3-2: Before convolution, perform a frequency domain transformation (Short Time Fourier Transform, STFT, or Wavelet Transform) on the padded tensor to obtain the frequency domain feature tensor X. f Combine it with the original time-domain tensor X t Concatenate along the channel dimension and input multi-scale convolutional layers;

[0023] Step 3-3: Concatenate the features extracted by multi-scale convolution and reduce the dimensionality through 1×1×1 convolution to obtain the fused feature F. fusion , where F fusion =Conv 1×1×1 ([F 3×3×3 ;F 5×5×5 ;F 7×7×7 The symbol [;] indicates channel splicing;

[0024] Steps 3-4 involve concatenating the fused features with the original input and then feeding the concatenated features into a multilayer perceptron (MLP) to obtain the final feature representation F. final The MLP consists of several fully connected layers and activation functions (such as ReLU) to achieve nonlinear transformations and multi-feature interactions in a high-dimensional feature space.

[0025] Compared to existing technologies, the advantages of this step are:

[0026] (1) Fusion preserves original information: The fusion features extracted by multi-scale convolution are directly spliced ​​with the original input, which can preserve the fine-grained information in the original signal and avoid the loss of details that may occur during the convolution process.

[0027] (2) Enhanced feature representation capability: By performing nonlinear mapping and feature combination on the spliced ​​high-dimensional features through MLP, more complex spatiotemporal relationships and environmental factor interaction patterns can be modeled, thereby improving downstream prediction performance.

[0028] (3) Balancing local and global features: The fused features include local pattern information extracted by multi-scale convolution, while the original input retains global background information. The concatenated input to the MLP can utilize both types of information simultaneously to achieve local-global collaborative modeling.

[0029] (4) Improve model generalization: Fusion and nonlinear combination in the feature layer helps to alleviate the risk of overfitting and maintain good generalization ability under different wind farms and meteorological conditions.

[0030] Step 4, Spatial graph modeling and prediction: A spatiotemporal network is constructed based on graph convolution (GCN) and dilated temporal convolution (DTCN) to achieve dynamic prediction of wind power. This includes the following sub-steps:

[0031] Step 4-1: Construct a two-layer dynamic graph adjacency matrix:

[0032] (1) Long-term dependency graph Based on long-term historical data statistics, it reflects the stable spatial topological relationship between wind turbines;

[0033] (2) Short-term dependency graph The dynamic graph learning module updates in real time, capturing instantaneous spatial dependencies caused by weather changes;

[0034] (3) Adjacency matrix update combined with meteorological physics model, based on wind direction θ w Wind speed v w And dynamically adjust edge weights based on terrain features, for example: A S (i,j)←A S (i,j)+α·max(0,cos(θ ij -θ w ))·v w , where θ ij Let α be the azimuth angle of fan i pointing to j, and α be the adjustment coefficient.

[0035] Step 4-2, respectively at A L With A S Perform graph convolution on top: in, W is the adjacency matrix with added self-loops and normalized. L W S Let σ be the graph convolution weight matrix, and σ(·) be the activation function (such as ReLU).

[0036] Step 4-3: Fuse features of long-term and short-term graphs using the graph fusion module: H = βH L +(1-β)H S , where β∈[0,1] are learnable fusion coefficients;

[0037] Step 4-4: Generate wind power prediction results for the future time step T′ through the output layer:

[0038] The advantages of this scheme are as follows: (1) Dynamically capturing spatial dependencies, taking into account both long-term and short-term features: By constructing a two-layer dynamic graph adjacency matrix, it distinguishes between long-term stable spatial topology (long-term dependency graph) and instantaneous spatial dependencies caused by meteorological changes (short-term dependency graph). Traditional graph models usually only use static adjacency matrices, ignoring the dynamic changes of environment and meteorology. This scheme effectively combines static and dynamic aspects, improving the precision and adaptability of spatial dependency modeling.

[0039] (2) Dynamic adjustment of adjacency matrix using meteorological physical models: The edge weights of the adjacency matrix are dynamically adjusted using physical factors such as wind direction, wind speed, and terrain, which enhances the physical rationality of the graph structure. Compared with simple data-driven adjacency matrix learning, this physical guidance makes the model more interpretable and generalizable, especially in predicting more accurately under complex meteorological conditions.

[0040] (3) Graph convolution fusion of multi-scale spatial information: After performing graph convolution on the long-term and short-term adjacency matrices respectively, the two features are fused through learnable fusion coefficients to flexibly capture multi-scale spatial dependencies. This graph fusion mechanism effectively makes up for the limitations of single graph convolution in extracting spatial information and improves the ability to express spatiotemporal features.

[0041] (4) Capturing long-term series dependencies by combining dilated temporal convolution: Dilated temporal convolution (DTCN) used in conjunction with GCN effectively expands the temporal receptive field and can capture the dynamic changes of wind power over long time scales. Compared with traditional time series models (such as ordinary convolution or LSTM), DTCN is more efficient in modeling time series dependencies and has stronger predictive performance.

[0042] (5) Strong real-time dynamic prediction capability: Through real-time updates of the short-term dependency graph and dynamic adjustment of the adjacency matrix, the model has the ability to respond quickly to changes in the wind farm environment, making it suitable for dynamic change scenarios in actual wind power prediction.

[0043] Step 5: Model training and optimization, using a composite loss function. End-to-end training is performed, and parameters are tuned using the Adam optimizer, where L... total This is the total loss function, used to guide the updating of all network parameters. The model predicts the future T ′ The wind power value at the time step, where N is the number of wind turbines. Represents the future T as actually observed ′ Wind power output at time step This is the prediction error term, which is the squared error between the predicted result and the true value (L2 norm squared), used to measure the prediction accuracy. tensor It is the loss function of the tensor completion process, used to constrain the accuracy and rationality of tensor decomposition and completion.

[0044] As a further technical solution of the present invention, the graph structure adjacency matrix in step 2-2 is composed of a two-layer structure, including a long-term dependency graph and a short-term dependency graph. The long-term dependency graph is obtained based on long-term historical data statistics and reflects the stable spatial topological relationship between wind turbines. The short-term dependency graph is updated in real time by a dynamic graph learning module and the edge weights are dynamically adjusted in combination with meteorological physical models (wind direction, wind speed propagation path and terrain features) to more accurately reflect the physical relationship between wind turbines.

[0045] As a further technical solution of the present invention, the multi-scale convolution kernel in step 3-1 is adaptively generated by the dynamic convolution kernel generation module. The size of the convolution kernel can be dynamically selected or weighted and fused in 3×3×3, 5×5×5, and 7×7×7. The convolution kernel parameters are adaptively adjusted according to the spatiotemporal characteristics of the input tensor through an attention mechanism.

[0046] As a further technical solution of the present invention, when the spatiotemporal network in step 4-2 performs graph convolution calculation, it performs feature propagation on long-term dependency graphs and short-term dependency graphs respectively, and fuses the two types of graph convolution results by a graph fusion module according to learnable weight coefficients, and then inputs an expanded temporal convolutional layer with an expansion coefficient of 2, 4 or 8 to capture temporal dependencies at different time scales.

[0047] Compared with the prior art, the present invention has the following advantages and beneficial effects:

[0048] This invention innovatively integrates multi-constraint tensor completion technology with spatiotemporal graph neural networks. It achieves high-quality data reconstruction through a triple regularization mechanism, designs a multi-scale feature extraction module to capture the changing patterns at different time granularities, and employs a dynamic graph learning mechanism to adaptively represent spatiotemporal correlations, thus constructing an end-to-end joint modeling framework. This method significantly improves prediction accuracy, reduces computational complexity, and effectively addresses the shortcomings of existing technologies in areas such as data missingness handling, multi-scale feature extraction, and dynamic spatiotemporal modeling, providing a more reliable technical solution for wind power prediction.

[0049] Furthermore, this invention has optimized the design of several key technical aspects, further improving the performance and adaptability of the model:

[0050] (1) The dual-layer dynamic graph structure is combined with physical driving. The dual-layer structure of long-term dependency graph and short-term dependency graph is introduced in the construction of dynamic graph. At the same time, the edge weights of the short-term graph are dynamically adjusted by combining meteorological physical models such as wind direction, wind speed propagation path and terrain features. It not only relies on data-driven learning, but also incorporates the physical prior knowledge of wind field, thereby improving the robustness and prediction stability of the model when meteorological conditions change rapidly.

[0051] (2) Multi-scale dynamic convolution kernel generation mechanism: In multi-scale feature extraction, a dynamic convolution kernel generation module is adopted to adaptively generate three-dimensional convolution kernels of different sizes and weights (including 3×3×3, 5×5×5, 7×7×7, etc.) according to the spatiotemporal pattern of the input tensor. This overcomes the problem of insufficient adaptability of fixed convolution kernels under complex weather patterns and improves the extraction accuracy of features at different time granularities.

[0052] (3) Fusion modeling of frequency domain and time domain features: Before extracting convolution features, frequency domain transformation (short-time Fourier transform or wavelet transform) is introduced to extract periodic features, and then fused with the time domain multi-scale convolution results to realize simultaneous modeling of wind speed mutation and periodic changes, thereby improving the model's ability to express multiple types of wind power time series patterns.

[0053] (4) Multi-graph fusion and multi-scale temporal convolution collaboration: In spatiotemporal modeling, graph convolution feature extraction is performed on long-term dependency graphs and short-term dependency graphs respectively. The outputs of the two types of graphs are fused by learnable weight coefficients and then input into multi-scale dilated temporal convolutional layers with dilation coefficients of 2, 4, 8, etc. This takes into account the temporal dependency modeling capabilities of minute-level, hour-level and longer time spans, thereby significantly improving the accuracy and stability of long-term prediction.

[0054] (5) Adaptive constraint weight optimization: An adaptive weight network is introduced in the tensor completion process to dynamically allocate the weights of graph structure constraints, time smoothing constraints and non-negativity constraints according to the missing pattern and time change characteristics, so that the completion process can maintain the optimal reconstruction effect under different missing rates and data patterns. Attached Figure Description

[0055] Figure 1 This is a flowchart of the method of the present invention.

[0056] Figure 2 This is a schematic diagram of the model of the present invention. Detailed Implementation

[0057] The technical solutions provided by the present invention will be described in detail below with reference to specific implementations. It should be understood that the following specific implementations are only used to illustrate the present invention and are not intended to limit the scope of the present invention.

[0058] Example: This invention provides a multi-scale tensor graph spatiotemporal network wind power prediction method that integrates spatiotemporal correlations and environmental factors, employing a technical approach of multi-constraint tensor modeling and graph neural network co-optimization. This method mainly includes three key steps: First, by introducing Tucker tensor decomposition with graph structure constraints, time smoothing constraints, and non-negativity constraints, high-quality reconstruction of incomplete wind power data is achieved; second, a multi-scale 3D convolution module is used, employing parallel convolutional structures to extract features at different temporal granularities, such as minute-level wind speed mutations and hour-level environmental changes; finally, a dynamic graph spatiotemporal network is constructed, using a learnable adjacency matrix to adaptively represent the spatiotemporal correlations between wind turbines, and multi-scale temporal modeling is achieved by fusing dilated temporal convolutions. This method, through end-to-end joint optimization of tensor completion, feature extraction, and spatiotemporal prediction, effectively solves key technical challenges in wind power prediction such as data loss, multi-scale feature fusion, and dynamic spatiotemporal modeling, providing a novel technical solution for new energy power prediction.

[0059] Specifically, the flowchart and model diagram of this invention are as follows: Figure 1 , Figure 2 As shown, it includes the following steps:

[0060] Step 1: Data preprocessing and construction of higher-order tensors.

[0061] For multi-source data from wind farms, construct a third-order tensor X∈R: wind turbine-time-feature. N×T×F Where N is the number of wind turbines, T is the time step, and F is the feature dimension, outlier data is cleaned through spatial correlation analysis and temporal interpolation strategies, specifically including the following sub-steps:

[0062] Step 1-1: Cluster the wind turbines based on their geographical location and fill in missing values ​​using the within-group mean.

[0063] Steps 1-2 involve using time series interpolation to fill in the remaining missing values, ensuring data integrity.

[0064] Step 2, multi-constraint tensor completion.

[0065] Tensor ring (TR) decomposition of tensors A low-rank approximation is performed, and graph structure constraints, time smoothing constraints, and non-negativity constraints are introduced to achieve high-quality reconstruction of incomplete wind power data. The specific sub-steps include:

[0066] Step 2-1: Through tensor ring decomposition, the original tensor X is represented as a series of third-order tensor factors {G}. (1) G (2) G (3)}, where G (k) Let R be the k-th third factor tensor of the tensor ring. k represents the connection dimension between adjacent factors. The tensor ring decomposition form is: Where i∈[1,N], j∈[1,T], k∈[1,F];

[0067] Step 2-2, Structural constraints in the design drawings: in Let d be the spatial factor matrix. s As a potential dimension of space, The graph Laplace matrix reflects the spatial topology of the wind turbine, and Tr(·) is the matrix trace operation.

[0068] Steps 2-3 introduce time smoothing constraints: Ensure time continuity, among which Let d be the time factor matrix. t As a potential dimension of time, This represents the t-th row of the time factor matrix, where ||·||2 is the L2 norm, used to measure the continuity between adjacent time steps.

[0069] Steps 2-4: Add nonnegativity constraint: L nonneg =||ReLU(-U (f) )||1, Enhance the physical rationality of features, among which Let d be the eigenfactor matrix. f is the latent dimension of the feature, ReLU(·) is the linear rectified function that sets negative values ​​to zero, and ||·||1 is the L1 norm, used to measure the strength of nonnegativity constraints.

[0070] Steps 2-5 involve determining the weights of each constraint term using an adaptive weighting mechanism. An adaptive weighting network f is then constructed. θ (·), the input is the missing pattern matrix M, where missing values ​​are 1 and non-missing values ​​are 0) and the time-varying statistical features, and the output is the weights (λ). g ,λ t ,λ n The final optimization objective is: Ltensor =L rec +λ g L graph +λ t L time +λ n L nonneg ,in, For tensor reconstruction error, ||·|| F It is the Frobenius norm. This is the result of tensor ring decomposition and reconstruction.

[0071] Step 3, multi-scale tensor feature extraction.

[0072] Perform multi-scale convolution operations on the padded tensor to extract local spatiotemporal features, specifically including the following sub-steps:

[0073] Step 3-1: Adaptively generate three-dimensional convolutional kernels of different sizes and weights through the Dynamic Kernel Generator (DKG) module. The kernel size can be dynamically selected or weighted and fused in {3×3×3, 5×5×5, 7×7×7} to adapt to different meteorological models and spatiotemporal dependence characteristics.

[0074] Step 3-2: Before convolution, perform a frequency domain transformation (Short Time Fourier Transform, STFT, or Wavelet Transform) on the padded tensor to obtain the frequency domain feature tensor X. f Combine it with the original time-domain tensor X t Concatenate along the channel dimension and input multi-scale convolutional layers;

[0075] Step 3-3: Concatenate the features extracted by multi-scale convolution and reduce the dimensionality through 1×1×1 convolution to obtain the fused feature F. fusion , where F fusion =Conv 1×1×1 ([F 3×3×3 ;F 5×5×5 ;F 7×7×7 The symbol [;] indicates channel splicing;

[0076] Steps 3-4 involve concatenating the fused features with the original input and then feeding the concatenated features into a multilayer perceptron (MLP) to obtain the final feature representation F. final .

[0077] Step 4, Empty graph modeling and prediction.

[0078] A spatiotemporal network is constructed based on Graph Convolution (GCN) and Dilated Temporal Convolution (DTCN) to achieve dynamic prediction of wind power. The specific sub-steps include:

[0079] Step 4-1: Construct a two-layer dynamic graph adjacency matrix:

[0080] (1) Long-term dependency graph Based on long-term historical data statistics, it reflects the stable spatial topological relationship between wind turbines;

[0081] (2) Short-term dependency graph The dynamic graph learning module updates in real time, capturing instantaneous spatial dependencies caused by weather changes;

[0082] (3) Adjacency matrix update combined with meteorological physics model, based on wind direction θ w Wind speed v w And dynamically adjust edge weights based on terrain features, for example: A S (i,j)←A S (i,j)+α·max(0,cos(θ ij -θ w ))·v w , where θ ij Let α be the azimuth angle of fan i pointing to j, and α be the adjustment coefficient.

[0083] Step 4-2, respectively at A L With A S Perform graph convolution on top: in, W is the adjacency matrix with added self-loops and normalized. L W S Let σ be the graph convolution weight matrix, and σ(·) be the activation function (such as ReLU).

[0084] Step 4-3: Fuse features of long-term and short-term graphs using the graph fusion module: H = βH L +(1-β)H S , where β∈[0,1] are learnable fusion coefficients;

[0085] Step 4-4: Generate wind power prediction results for the future time step T′ through the output layer:

[0086] Step 5: Model training and optimization.

[0087] Using a composite loss function End-to-end training is performed, and parameters are tuned using the Adam optimizer, where L... total This is the total loss function, used to guide the updating of all network parameters. The model predicts the future T ′ The wind power value at the time step, where N is the number of wind turbines. Represents the future T as actually observed′ Wind power output at time step This is the prediction error term, which is the squared error between the predicted result and the true value (L2 norm squared), used to measure the prediction accuracy. tensor It is the loss function of the tensor completion process, used to constrain the accuracy and rationality of tensor decomposition and completion.

[0088] Example:

[0089] Experimental conditions: All experiments were implemented using the PyTorch framework on an NVIDIA GeForce RTX 2080 GPU. Experimental parameters were set as follows: historical observation window length was set to 288 (corresponding to 48 hours of data), batch size was set to 32, the Adam optimizer was used for model training, the initial learning rate was 0.001, and a cosine annealing learning rate scheduling strategy was adopted. The model training epoch was set to 30, and an early stopping mechanism was applied (training was terminated when the validation set loss did not decrease for 5 consecutive epochs).

[0090] This experiment was validated on a real wind farm dataset provided by Longyuan Power Group, which contains 245 consecutive days of SCADA monitoring data from 134 wind turbines (sampling interval of 10 minutes). To comprehensively evaluate model performance, the first 196 days of data were used as the training set, the last 49 days as the validation set, and 17 independent days of data were reserved as the test set. The prediction task was set as wind power prediction for the next 24 and 48 hours, and compared with five benchmark models: Historical Average (HA), Support Vector Regression (SVR), LightGBM, GRU, and STGCN.

[0091] The evaluation metrics used were root mean square error (RMSE), mean absolute error (MAE), and a comprehensive score (Score), where the Score was the average of RMSE and MAE. Table 1 shows a performance comparison between the present invention and five baseline models on the SDWPF real wind power dataset. The MTGTRNet model of the present invention significantly outperformed the comparison models in all metrics across all prediction periods, particularly demonstrating stronger stability in long-term prediction (48-hour) tasks. Experimental results validated the model's advantages in handling missing data, multi-scale feature extraction, and dynamic spatiotemporal modeling.

[0092] Table 1 Comparison of Model Prediction Performance

[0093]

[0094] The technical means disclosed in this invention are not limited to those disclosed in the above embodiments, but also include technical solutions composed of any combination of the above technical features. It should be noted that those skilled in the art can make various improvements and modifications without departing from the principles of this invention, and these improvements and modifications are also considered within the scope of protection of this invention.

Claims

1. A multi-scale tensor graph spatiotemporal network wind power prediction method integrating spatiotemporal correlation and environmental factors, characterized in that, Includes the following steps: Step 1: Data preprocessing and high-order tensor construction. For multi-source data from wind farms, construct a third-order tensor X∈R for wind turbine-time-feature. N×T×F Where N is the number of wind turbines, T is the time step, and F is the feature dimension, outlier data is cleaned through spatial correlation analysis and time interpolation strategy; Step 2, multi-constraint tensor completion, using tensor ring (TR) decomposition to complete the tensor. We perform low-rank approximation and introduce graph structure constraints, time smoothing constraints, and non-negativity constraints to achieve high-quality reconstruction of incomplete wind power data. Step 3, multi-scale tensor feature extraction: perform multi-scale convolution operation on the completed tensor to extract local spatiotemporal features; Step 4, Spatial graph modeling and prediction: A spatiotemporal network is constructed based on graph convolution (GCN) and dilated temporal convolution (DTCN) to achieve dynamic prediction of wind power. Step 5: Model training and optimization.

2. The multi-scale tensor graph spatiotemporal network wind power prediction method integrating spatiotemporal correlation and environmental factors as described in claim 1, characterized in that, Step 1 specifically includes the following sub-steps: Step 1-1: Cluster the wind turbines based on their geographical location and fill in missing values ​​using the within-group mean. Steps 1-2 involve using time series interpolation to fill in the remaining missing values, ensuring data integrity.

3. The multi-scale tensor graph spatiotemporal network wind power prediction method integrating spatiotemporal correlation and environmental factors as described in claim 1, characterized in that, Step 2 specifically includes the following sub-steps: Step 2-1: Through tensor ring decomposition, the original tensor X is represented as a series of third-order tensor factors {G}. (1) G (2) G (3) }, where G (k) Let R be the k-th third factor tensor of the tensor ring. k , representing the connection dimension between adjacent factors, the tensor ring decomposition form is: Where i∈[1,N], j∈[1,T], k∈[1,F]; Step 2-2, Structural constraints in the design drawings: in Let d be the spatial factor matrix. s As a potential dimension of space, The graph Laplace matrix reflects the spatial topology of the wind turbine, and Tr(·) is the matrix trace operation; Steps 2-3 introduce time smoothing constraints: Ensure time continuity, among which Let d be the time factor matrix. t As a potential dimension of time, This represents the t-th row of the time factor matrix, where ||·||2 is the L2 norm, used to measure the continuity between adjacent time steps; Steps 2-4: Add nonnegativity constraint: L nonneg =||ReLU(-U (f) )||1, Enhance the physical rationality of features, among which Let d be the eigenfactor matrix. f The feature latent dimension is ReLU(·), which is a linear rectified function that sets negative values ​​to zero, and ||·||1 is the L1 norm, used to measure the strength of nonnegativity constraints. Steps 2-5 involve using an adaptive weighting mechanism to determine the weights of each constraint term and constructing an adaptive weighting network f. θ (·), the input is the missing pattern matrix M, where missing values ​​are 1 and non-missing values ​​are 0) and the time-varying statistical features, and the output is the weights (λ). g ,λ t ,λ n The final optimization objective is: L tensor =L rec +λ g L graph +λ t L time +λ n L nonneg ,in, For tensor reconstruction error, ||·|| F It is the Frobenius norm. This is the result of tensor ring decomposition and reconstruction.

4. The multi-scale tensor graph spatiotemporal network wind power prediction method integrating spatiotemporal correlation and environmental factors according to claim 1, characterized in that, Step 3 specifically includes the following sub-steps: Step 3-1: Adaptively generate three-dimensional convolutional kernels of different sizes and weights through the Dynamic Kernel Generator (DKG) module. The kernel size can be dynamically selected or weighted and fused in {3×3×3, 5×5×5, 7×7×7} to adapt to different meteorological models and spatiotemporal dependence characteristics. Step 3-2: Before convolution, perform a frequency domain transformation (Short Time Fourier Transform, STFT, or Wavelet Transform) on the padded tensor to obtain the frequency domain feature tensor X. f Combine it with the original time-domain tensor X t Concatenate along the channel dimension and input multi-scale convolutional layers; Step 3-3: Concatenate the features extracted by multi-scale convolution and reduce the dimensionality through 1×1×1 convolution to obtain the fused feature F. fusion F fusion =Conv 1×1×1 ([F 3×3×3 ;F 5×5×5 ;F 7×7×7 The symbol [;] indicates channel splicing; Steps 3-4 involve concatenating the fused features with the original input and then feeding the concatenated features into a multilayer perceptron (MLP) to obtain the final feature representation F. final .

5. The multi-scale tensor graph spatiotemporal network wind power prediction method integrating spatiotemporal correlation and environmental factors according to claim 1, characterized in that, Step 4 specifically includes the following sub-steps: Step 4-1: Construct a two-layer dynamic graph adjacency matrix: (1) Long-term dependency graph Based on long-term historical data statistics, it reflects the stable spatial topological relationship between wind turbines; (2) Short-term dependency graph The dynamic graph learning module updates in real time, capturing instantaneous spatial dependencies caused by weather changes; (3) Adjacency matrix update combined with meteorological physics model, based on wind direction θ w Wind speed v w And dynamically adjust edge weights based on terrain features, for example: A S (i,j)←A S (i,j)+α·max(0,cos(θ ij -θ w ))·v w , where θ ij Let α be the azimuth angle of fan i pointing to j, and α be the adjustment coefficient. Step 4-2, respectively at A L With A S Perform graph convolution on top: in, W is the adjacency matrix with added self-loops and normalized. L W S Let be the graph convolution weight matrix, and σ(·) be the activation function. Step 4-3: Fuse features of long-term and short-term graphs using the graph fusion module: H = βH L +(1-β)H S , where β∈[0,1] are learnable fusion coefficients; Step 4-4: Generate wind power prediction results for the future time step T′ through the output layer:

6. The multi-scale tensor graph spatiotemporal network wind power prediction method integrating spatiotemporal correlation and environmental factors according to claim 1, characterized in that, Step 5 specifically includes the following sub-steps: using a composite loss function. End-to-end training is performed, and parameters are tuned using the Adam optimizer, where L... total This is the total loss function, used to guide the updating of all network parameters. This represents the wind power value predicted by the model at the future time step T′, where N is the number of wind turbines. This represents the wind power value at the actual observed future time step T′. This is the prediction error term, which is the squared error between the predicted result and the true value (L2 norm squared), used to measure the prediction accuracy. tensor It is the loss function of the tensor completion process, used to constrain the accuracy and rationality of tensor decomposition and completion.

7. The multi-scale tensor graph spatiotemporal network wind power prediction method integrating spatiotemporal correlation and environmental factors according to claim 1, characterized in that: The graph adjacency matrix in step 2-2 consists of a two-layer structure, including a long-term dependency graph and a short-term dependency graph. The long-term dependency graph is obtained based on long-term historical data statistics and reflects the stable spatial topological relationship between wind turbines. The short-term dependency graph is updated in real time by the dynamic graph learning module and the edge weights are dynamically adjusted in combination with the meteorological physics model to more accurately reflect the physical relationship between wind turbines.

8. The multi-scale tensor graph spatiotemporal network wind power prediction method integrating spatiotemporal correlation and environmental factors according to claim 1, characterized in that: In step 4-2, the spatiotemporal network performs feature propagation on long-term dependency graphs and short-term dependency graphs respectively when performing graph convolution calculations. The two types of graph convolution results are then fused by the graph fusion module according to the learnable weight coefficients. The result is then input into an expanded temporal convolutional layer with an expansion coefficient of 2, 4 or 8 to capture temporal dependencies at different time scales.

9. A computer-readable storage medium storing computer instructions thereon, characterized in that: When the computer instructions are executed by the processor, they implement the steps of the multi-scale tensor graph spatiotemporal network wind power prediction method that integrates spatiotemporal correlations and environmental factors as described in any one of claims 1-8.

10. An electronic device, characterized in that, include: processor; A memory for storing processor-executable instructions; wherein the processor is configured to execute the instructions to implement the multi-scale tensor graph spatiotemporal network wind power prediction method that integrates spatiotemporal correlations and environmental factors as described in any one of claims 1 to 8.

Images