Method and system for multi-wind farm power interval prediction based on adaptive cooperative VMD and dynamic spatiotemporal graph convolution

Abstract

The application discloses a kind of based on adaptive collaborative VMD and dynamic space-time graph convolution multi-wind farm power interval prediction method and system, specifically: to the historical power data of multiple wind farms is preprocessed, constructs normalized power data matrix;Through the introduction of spatial collaborative constraint and frequency alignment constraint adaptive multi-wind farm collaborative variation modal decomposition, obtain spatially aligned multi-modal feature sequence;Geospatial adjacency matrix and power correlation adjacency matrix are constructed, and the node embedding of current time step is learned using dynamic graph generation module to generate dynamic adjacency matrix, then layered weighted fusion obtains static fusion adjacency matrix and the layered fusion adjacency matrix of the combination of dynamic adjacency matrix;Based on enhanced dynamic space-time graph convolution network, design progressive constraint loss function to train, finally output each wind farm at multiple confidence levels under the predicted interval.The application significantly improves the accuracy and robustness of power interval prediction.

Description

Technical Field

[0001] This invention relates to the field of new energy power generation prediction or smart grid dispatching technology, and in particular to a method and system for predicting the power range of multiple wind farms based on adaptive collaborative VMD and dynamic spatiotemporal graph convolution. Background Technology

[0002] Against the backdrop of rapid growth in installed capacity of new energy sources and the continuous advancement of electricity market transactions, new energy power forecasting has become an important technology for ensuring the safe and stable operation of the power system. Accurate wind power forecasting helps grid dispatching departments to rationally arrange power generation plans, reduce reserve capacity requirements, reduce wind and solar curtailment, and support new energy power plants to participate in electricity market transactions.

[0003] Existing wind power forecasting methods mainly include physical methods, statistical methods, machine learning methods, and deep learning methods. Physical methods rely on numerical weather prediction and wind farm physical models for forecasting, and are highly dependent on the accuracy of weather forecasts. Statistical methods, such as Auto-Regressive Integrated Moving Average (ARIMA) and Support Vector Machine (SVM), are suitable for short-term forecasting but struggle to capture complex nonlinear relationships. Machine learning methods, such as Random Forest and Gradient Boosting Decision Tree (GBDT), can handle high-dimensional features and nonlinear relationships, but have limited ability to model time-series dependencies. Deep learning methods, such as Long Short-Term Memory (LSTM) and Convolutional Neural Network (CNN), excel at extracting time-series features, but still fall short in characterizing the complex spatial relationships and feature interactions between multiple wind farms.

[0004] In recent years, Graph Neural Networks (GNNs) have made significant progress in the field of spatiotemporal data modeling. Spatial-Temporal Graph Convolutional Networks (STGCNs) capture spatial dependencies through graph convolutional layers and extract temporal features through temporal convolutional layers, demonstrating excellent performance in areas such as traffic flow prediction and air quality prediction. However, existing STGCN methods mostly employ fixed graph structures built based on geographical distance or historical correlations, making it difficult to adapt to the dynamically changing relationships between wind farms and unable to adaptively adjust node connection strength according to the current operating status, thus limiting the model's ability to model spatial dependencies.

[0005] Interval forecasting is an important method for uncertainty quantification. The quality of a forecast interval is usually evaluated by the Prediction Interval Coverage Probability (PICP) and the Prediction Interval Normalized Average Width (PINAW), which reflect the reliability and sharpness of the forecast interval, respectively. Existing interval forecasting methods mainly include quantile regression, Bayesian methods, and neural network methods. However, most methods generate forecast intervals independently for a single wind farm, ignoring the collaborative relationships between adjacent or related wind farms under similar meteorological conditions. This makes it difficult to guarantee the spatial consistency and coverage coordination of forecast intervals for multiple wind farms.

[0006] In signal decomposition, Variational Mode Decomposition (VMD) is an adaptive signal decomposition method that can decompose complex non-stationary signals into several Intrinsic Mode Functions (IMFs), each IMF representing the signal components at different frequency scales. Compared to methods such as Empirical Mode Decomposition (EMD), VMD has a better mathematical foundation and noise resistance, and has been widely used in wind power prediction. However, existing methods typically treat VMD as an independent preprocessing step, decomposing each wind farm before modeling it, without fully considering the consistency of frequency characteristics between adjacent wind farms, and lacking deep integration of multimodal features and graph neural networks.

[0007] In summary, existing multi-wind farm power prediction methods still have the following shortcomings: First, the graph structure is usually fixed, making it impossible to adaptively adjust the node connection strength according to meteorological conditions, wind direction changes, and current power status, thus failing to characterize dynamic spatial relationships. Second, prediction intervals are mostly generated independently for individual wind farms, lacking spatial coordination constraints, making it difficult to ensure the smoothness of the boundaries and the consistency of coverage of adjacent or highly correlated wind farm intervals. Third, traditional models output multiple point prediction results, lacking effective quantification of uncertainties. Fourth, existing signal decomposition methods are mostly independent decompositions, lacking cross-site frequency component alignment mechanisms, affecting subsequent spatial dependency learning. Fifth, existing interval prediction loss functions mainly focus on individual prediction objects, making it difficult to simultaneously constrain interval width, coverage, spatial consistency, and the nesting relationship of multiple confidence level intervals. Sixth, in multi-objective training, each loss term usually uses manually fixed weights, making it difficult to adapt to differences in the scale of different loss terms and changes in the training stage, resulting in high parameter tuning costs. Therefore, it is necessary to propose a wind power interval prediction method that integrates the dynamic spatial relationships of multiple wind farms, multi-modal frequency characteristics, and prediction uncertainty constraints to improve the accuracy and reliability of joint prediction of multiple wind farms. Summary of the Invention

[0008] The purpose of this invention is to provide a method and system for predicting the power range of multiple wind farms based on adaptive collaborative VMD and dynamic spatiotemporal graph convolution. Through collaborative VMD spatial alignment denoising, dynamic graph hierarchical fusion, and progressive loss automatic weighting, adaptive high-reliability range prediction of multiple wind farms is achieved.

[0009] The technical solution to achieve the purpose of this invention is: a multi-wind farm power interval prediction method based on adaptive collaborative VMD and dynamic spatiotemporal graph convolution, comprising the following steps:

[0010] Step 1, Data Acquisition and Preprocessing: Historical power data is collected from the data acquisition systems of multiple wind farms, and the historical power data is preprocessed to obtain a normalized power data matrix with each wind farm as a node;

[0011] Step 2, Adaptive Multi-Wind Farm Cooperative VMD Signal Decomposition: Input the normalized power data matrix obtained in Step 1 into the adaptive multi-wind farm cooperative variational mode decomposition model. Through the joint optimization objective including VMD bandwidth constraints, spatial cooperative constraints and frequency alignment constraints, the power sequences of multiple wind farms are cooperatively decomposed. The decomposition parameters are adaptively determined according to energy entropy, spatial adjacency weight and spectral entropy to obtain multiple intrinsic mode components of each wind farm spatially aligned under the same mode index.

[0012] Step 3: Construction of geospatial adjacency matrix and power correlation adjacency matrix: Calculate the geographical distance between wind farms based on the geographical coordinates of each wind farm, and convert the geographical distance into a geospatial adjacency matrix; at the same time, calculate the correlation between the power sequences of different wind farms based on the normalized power data matrix obtained in Step 1, and obtain the power correlation adjacency matrix.

[0013] Step 4: Generation of adaptive dynamic adjacency matrix: Extract and organize the multimodal feature sequence obtained in Step 2 according to time steps to obtain the node feature matrix of the current time step, and input it into the dynamic graph generation module. Learn the node embedding of each wind farm node in the current time step through graph convolutional network, normalize and calculate the similarity of the node embedding, and generate the dynamic adjacency matrix of the corresponding time step after normalization.

[0014] Step 5: Construction of the hierarchical fusion graph structure: Perform a first-layer weighted fusion on the geospatial adjacency matrix and the power correlation adjacency matrix obtained in Step 3 to obtain a static fusion adjacency matrix; then perform a second-layer weighted fusion on the static fusion adjacency matrix and the dynamic adjacency matrix obtained in Step 4 to obtain a hierarchical fusion adjacency matrix;

[0015] Step 6, Progressive Constraint Loss Training: Based on the prediction interval results and true power values ​​output by the enhanced dynamic spatiotemporal graph convolutional network during the training phase, a progressive constraint loss function based on conditional probability chain decomposition is constructed, and the progressive constraint loss function is used to train the dynamic graph generation module, hierarchical fusion parameters, enhanced dynamic spatiotemporal graph convolutional network, and output layer parameters.

[0016] Step 7: Generation of prediction intervals at multiple confidence levels: The multimodal feature sequences obtained in Step 2 and the hierarchical fusion adjacency matrix obtained in Step 5 are input into the enhanced dynamic spatiotemporal graph convolutional network trained in Step 6. Initial feature extraction is performed through two-dimensional convolution. The time dependence features of the power sequences of each wind farm and the spatial correlation features between different wind farms are extracted through the spatiotemporal graph convolutional layer to obtain the spatiotemporal feature representation. Finally, the spatiotemporal feature representation is mapped to the upper and lower bounds of the prediction intervals for each wind farm at multiple confidence levels through the output layer to obtain the multi-confidence level interval prediction results for the power of multiple wind farms.

[0017] A multi-wind farm power interval prediction system based on adaptive cooperative VMD and dynamic spatiotemporal graph convolution is disclosed. This system implements the aforementioned multi-wind farm power interval prediction method based on adaptive cooperative VMD and dynamic spatiotemporal graph convolution. The system includes a data preprocessing module, an adaptive multi-wind farm cooperative VMD signal decomposition module, a hybrid adjacency matrix construction module, a dynamic graph generation module, a spatiotemporal graph convolution modeling module, a progressive constraint loss function training module, and a multi-confidence level interval output module.

[0018] The data preprocessing module is responsible for data acquisition and normalization.

[0019] The adaptive multi-wind farm collaborative VMD signal decomposition module decomposes the power sequence of each wind farm into spatially aligned multi-frequency components through a joint optimization objective.

[0020] The hybrid adjacency matrix construction module integrates geographical distance and power correlation to construct a static graph;

[0021] The dynamic graph generation module learns an adaptive adjacency matrix through a graph neural network;

[0022] The spatiotemporal graph convolutional modeling module extracts spatiotemporal features through a layered and fused graph structure;

[0023] The progressive constraint loss function training module achieves automatic weight adjustment through conditional probability chain decomposition.

[0024] The multi-confidence level interval output module generates prediction intervals for multiple confidence levels.

[0025] A computer-readable storage medium having a computer program stored thereon, which, when executed by a processor, implements the steps in the multi-wind farm power range prediction method based on adaptive cooperative VMD and dynamic spatiotemporal graph convolution.

[0026] Compared with the prior art, the significant advantages of this invention are:

[0027] (1) Collaborative VMD decomposition: Through spatial and frequency dual regularization, adaptive decomposition of multiple wind farms with the same mode alignment is achieved, effectively denoising and extracting consistent multimodal features;

[0028] (2) Dynamic graph and hierarchical fusion: Combining static prior graphs with dynamic graphs generated step by step, the time-varying characteristics of spatial correlation are adaptively modeled in a parameter-efficient hierarchical manner, thereby improving interpretability and prediction accuracy;

[0029] (3) Progressive constraint loss: Based on the conditional probability chain decomposition, a four-layer progressive loss is constructed. Only one temperature parameter is needed to automatically balance multi-objective optimization, reduce the difficulty of parameter tuning, and enhance spatial consistency.

[0030] (4) Business adaptability: The end-to-end framework supports any number of wind farms and forecast windows, directly outputs multiple confidence level intervals, flexibly adapts to different scheduling and trading scenarios, and has wide application value. Attached Figure Description

[0031] Figure 1 This is a flowchart of the multi-wind farm power range prediction method based on adaptive collaborative VMD and dynamic spatiotemporal graph convolution of the present invention.

[0032] Figure 2 This is a schematic diagram of the enhanced dynamic graph neural network structure in this invention.

[0033] Figure 3 This is a multi-interval prediction result curve of wind farm 1 to wind farm 10 in an embodiment of the present invention.

[0034] Figure 4 This is a multi-interval prediction result curve of wind farm 11 to wind farm 20 in an embodiment of the present invention.

[0035] Figure 5 This is a multi-interval prediction result curve of wind farm 21 to wind farm 25 in an embodiment of the present invention. Detailed Implementation

[0036] This invention proposes a multi-wind farm power interval prediction method based on adaptive multi-wind farm cooperative VMD-dynamic graph convolutional network. This method employs adaptive multi-wind farm cooperative variational mode decomposition (VMD) to jointly decompose the power signals of multiple wind farms, generating spatially aligned intrinsic mode functions (IMFs), and using these IMFs as multi-channel inputs to graph nodes. A hybrid adjacency matrix is ​​constructed to fuse geospatial and power correlation information. A dynamic graph generation module is designed to learn the adaptive adjacency matrix. A hierarchical fusion strategy is used to integrate static and dynamic graphs. Spatiotemporal features are extracted through an enhanced spatiotemporal graph convolutional network. A progressive constraint loss function based on conditional probability chain decomposition is designed to dynamically adjust the strength of subsequent constraints using the pre-loss quality factor. Finally, a prediction interval with multiple confidence levels is output.

[0037] Combination Figure 1 This invention relates to a multi-wind farm power range prediction method based on adaptive collaborative VMD and dynamic spatiotemporal graph convolution, comprising the following steps:

[0038] Step 1, Data Acquisition and Preprocessing: Historical power data is collected from the data acquisition systems of multiple wind farms, and the historical power data is preprocessed to obtain a normalized power data matrix with each wind farm as a node;

[0039] Step 2, Adaptive Multi-Wind Farm Cooperative VMD Signal Decomposition: Input the normalized power data matrix obtained in Step 1 into the adaptive multi-wind farm cooperative variational mode decomposition model. Through the joint optimization objective including VMD bandwidth constraints, spatial cooperative constraints and frequency alignment constraints, the power sequences of multiple wind farms are cooperatively decomposed. The decomposition parameters are adaptively determined according to energy entropy, spatial adjacency weight and spectral entropy to obtain multiple intrinsic mode components of each wind farm spatially aligned under the same mode index.

[0040] Step 3: Construction of geospatial adjacency matrix and power correlation adjacency matrix: Calculate the geographical distance between wind farms based on the geographical coordinates of each wind farm, and convert the geographical distance into a geospatial adjacency matrix; at the same time, calculate the correlation between the power sequences of different wind farms based on the normalized power data matrix obtained in Step 1, and obtain the power correlation adjacency matrix.

[0041] Step 4: Generation of adaptive dynamic adjacency matrix: Extract and organize the multimodal feature sequence obtained in Step 2 according to time steps to obtain the node feature matrix of the current time step, and input it into the dynamic graph generation module. Learn the node embedding of each wind farm node in the current time step through graph convolutional network, normalize and calculate the similarity of the node embedding, and generate the dynamic adjacency matrix of the corresponding time step after normalization.

[0042] Step 5: Construction of the hierarchical fusion graph structure: Perform a first-layer weighted fusion on the geospatial adjacency matrix and the power correlation adjacency matrix obtained in Step 3 to obtain a static fusion adjacency matrix; then perform a second-layer weighted fusion on the static fusion adjacency matrix and the dynamic adjacency matrix obtained in Step 4 to obtain a hierarchical fusion adjacency matrix;

[0043] Step 6, Progressive Constraint Loss Training: Based on the prediction interval results and true power values ​​output by the enhanced dynamic spatiotemporal graph convolutional network during the training phase, a progressive constraint loss function based on conditional probability chain decomposition is constructed, and the progressive constraint loss function is used to train the dynamic graph generation module, hierarchical fusion parameters, enhanced dynamic spatiotemporal graph convolutional network, and output layer parameters.

[0044] Step 7: Generation of prediction intervals at multiple confidence levels: The multimodal feature sequences obtained in Step 2 and the hierarchical fusion adjacency matrix obtained in Step 5 are input into the enhanced dynamic spatiotemporal graph convolutional network trained in Step 6. Initial feature extraction is performed through two-dimensional convolution. The time dependence features of the power sequences of each wind farm and the spatial correlation features between different wind farms are extracted through the spatiotemporal graph convolutional layer to obtain the spatiotemporal feature representation. Finally, the spatiotemporal feature representation is mapped to the upper and lower bounds of the prediction intervals for each wind farm at multiple confidence levels through the output layer to obtain the multi-confidence level interval prediction results for the power of multiple wind farms.

[0045] As a specific example, step 1 includes:

[0046] 1.1) Input Data Acquisition

[0047] Historical power data was collected from data acquisition systems of multiple wind farms, including timestamps, wind farm identifiers, and actual power generation values ​​(unit: MW). The data was in the form of a long table, with each row containing three fields: timestamp, wind farm name, and power value.

[0048] 1.2) Preprocess historical power data, including format conversion, missing value imputation, and normalization:

[0049] ① Format conversion: Convert the data format to a wide table, with row indexes as timestamps, column indexes as wind farm names, and cell values ​​as power;

[0050] ② Missing value handling: Check for missing values ​​in the data and fill them with linear interpolation. For missing values ​​at the beginning and end, use backfilling and forwardfilling methods respectively.

[0051] ③ Data Normalization: Normalize the power data for each wind farm, scaling the data to... The normalization formula for the interval is:

[0052]

[0053] in, This is the normalized power value. This is the original power value. , These represent the minimum and maximum historical power outputs of the wind farm, respectively.

[0054] As a specific example, step 2 of this invention proposes an adaptive multi-wind farm collaborative VMD decomposition method considering spatial collaborative constraints, overcoming the limitations of traditional VMD's independent decomposition of individual wind farms. By introducing a spatial Laplace regularization term and a frequency alignment mechanism, collaborative decomposition of power signals from multiple wind farms and spatial consistency optimization of frequency components are achieved. This method ensures that wind farms with similar geographical locations can extract IMFs with similar frequency characteristics under the same modal index, providing high-quality multimodal features for subsequent spatial collaborative modeling of graph convolutional layers. The specific implementation steps are as follows:

[0055] 2.1) Optimization objective of collaborative VMD decomposition

[0056] Traditional VMD decomposition optimizes each wind farm independently. This invention extends traditional VMD decomposition to a joint optimization problem involving multiple wind farms. Power signals of a wind farm , The optimization objective of collaborative VMD is:

[0057]

[0058]

[0059] in, For wind farm indexing, For the number of wind farms, For modal indexing, For modal number, For time variables, For the first The raw power signal of a wind farm. For the first The first wind farm One modal component, For the first The first wind farm The center frequency of each mode For time The partial derivative operator, For the Dirac function, The square of the L2 norm, The imaginary unit;

[0060] The first term in the formula is the bandwidth constraint term of traditional VMD (the sum of all wind farms). For spatial co-regularization, For the weight coefficients of the spatial cooperative regularization term, For frequency alignment regularization, These are the weighting coefficients for the frequency alignment regularization term;

[0061] 2.2) Spatial Co-regularization Term Design

[0062] The spatial cooperative regularization term, based on the graph Laplace regularization theory, constrains the same-frequency modes of spatially adjacent wind farms to have similar waveforms and frequency characteristics:

[0063]

[0064] in, For wind farm indexing, For the first The first wind farm One modal component, For the first The first wind farm The center frequency of each mode; Geospatial adjacency matrix The Middle The wind farm and the first The adjacency weights between wind farms, calculated based on the Gaussian kernel function, are expressed as follows:

[0065]

[0066] in, The bandwidth parameter is set according to the typical spatial distribution scale of the wind farm; For the first The wind farm and the first Haversine distance between wind farms;

[0067] First item The second term constrains the similarity of the same-frequency modal waveforms of adjacent wind farms. The center frequencies of the same-frequency modes of adjacent wind farms are close, and wind farms in similar geographical locations are affected by similar meteorological conditions. Therefore, the frequency characteristics of their power fluctuations should have spatial continuity.

[0068] The spatial co-regularization term is weighted by the spatial adjacency matrix, which makes the constraint strength between wind farm pairs that are closer together greater.

[0069] 2.3) Design of Frequency Alignment Regularization Terms

[0070] The frequency alignment regularization term constrains the IMFs (Indexed Mode Factors) of all wind farms to have similar frequency ranges, thus avoiding frequency aliasing.

[0071]

[0072] in, For all wind farms The variance of the center frequencies of each mode. For the first The average center frequency of each mode To correct the linear unit.

[0073] The frequency alignment regularization term ensures that the IMFs with the same modal index represent the same frequency range in all wind farms, and that there is a clear frequency separation between different modes, arranged in an orderly manner from high frequency to low frequency.

[0074] 2.4) Solving Cooperative VMD using the Alternating Direction Multiplier Method (ADMM)

[0075] The cooperative VMD optimization problem is solved iteratively using the extended ADMM algorithm:

[0076] ① Initialization: Perform standard VMD decomposition independently on each wind farm to obtain the initial modes. and frequency Initialize Lagrange multipliers and penalty parameters ;

[0077] ② Modal update: Fixed center frequency and Lagrange multipliers Update modes in the frequency domain:

[0078]

[0079] in, This is a modal summation index used to iterate through all modal summations except the current index. The remaining modes other than the one mode; This is the index for the number of iterations in the ADMM algorithm; Indicates the first The result of the iteration is the first The first wind farm Modal components Fourier transform; Indicates the first The first wind farm Modal components Fourier transform; Indicates the first Power signal of a wind farm Fourier transform; Indicates the first Lagrange multipliers corresponding to each wind farm Fourier transform; For frequency domain variables; For VMD bandwidth constraint parameters; spatial coordination term in the denominator Used to represent the coupling constraints between adjacent wind farms;

[0080] ③ Frequency update: Fixed mode Update # The iteration of the ... The first wind farm The center frequency of each mode :

[0081]

[0082] in, For the first The first wind farm Modal components Fourier transform.

[0083] ④ Lagrange multiplier update:

[0084]

[0085] in, and They represent the first Second and third Lagrange multipliers in the next iteration;

[0086] ⑤ Convergence criterion: When the residual Spatial consistency index All less than the tolerance Stop iteration when the time is right;

[0087] 2.5) Parameter Adaptive Selection Strategy

[0088] ① Modal number K: Adaptively determined using the energy entropy criterion; the decomposition results under different K values ​​are calculated, and the K value that minimizes the average energy entropy is selected; energy entropy is defined as:

[0089]

[0090] in, For energy entropy, the first The proportion of modal energy to the total modal energy , For the first Energy of each mode.

[0091] ② Spatial Cooperative Regularization Term Weight Coefficient The weight coefficients of the spatial coordination regularization term are adaptively set based on the density of the wind farm's spatial distribution.

[0092]

[0093] in, This is the scaling factor for the weight coefficients of the spatial collaborative regularization term, adjusted according to the actual scenario. The denser the spatial distribution, the better. The larger the value, the stronger the collaborative constraints.

[0094] ③ Frequency alignment regularization term weight coefficient : Adaptively set according to signal frequency complexity; for the first Power signals of a wind farm Perform a Fourier transform to obtain the frequency domain signal. And calculate the spectral entropy based on the power spectrum, the first The wind farm in the first The power spectral energy of a frequency sampling point is defined as:

[0095]

[0096] No. The wind farm in the first The normalized spectral energy percentage of each frequency sampling point is defined as:

[0097]

[0098] No. The spectral entropy of a wind farm is defined as:

[0099]

[0100] The average spectral entropy of all wind farms is defined as:

[0101]

[0102] The frequency alignment regularization term weights are adaptively set based on the average spectral entropy:

[0103]

[0104] in, For the first The frequency corresponding to each frequency sampling point For frequency sampling point index, This represents the total number of frequency sampling points. For the first The wind farm in the first Power spectral energy at each frequency sampling point For the first The wind farm in the first Normalized spectral energy percentage of each frequency sampling point This serves as the index for summing frequency sampling points in normalization calculations. To prevent positive constants with a denominator of zero, For the first The spectral entropy of a wind farm The average spectral entropy of all wind farms, This is the scaling factor for the weights of the frequency-aligned regularization term; the higher the frequency complexity, the better. The larger, The larger the value, the stronger the frequency alignment constraint.

[0105] As a specific example, step 3 constructs a hybrid adjacency matrix that integrates geospatial information and power correlation information, serving as the static graph structure of the graph convolutional network. The hybrid adjacency matrix can simultaneously characterize the geographical proximity relationships between wind farms and the correlation of power changes, providing prior knowledge for subsequent graph convolution operations. The specific implementation steps are as follows:

[0106] 3.1) Construction of geospatial adjacency matrix

[0107] Based on the geographical coordinates of wind farms, namely longitude and latitude, the actual distance between wind farms is calculated, and the distance is converted into adjacency weights using a Gaussian kernel function.

[0108] ① Wind farm coordinate definition: Define latitude and longitude geographic coordinates for each wind farm and store them in a dictionary structure.

[0109] ②Haversine distance calculation: For any two wind farms and Calculate the distance between them (in kilometers) using the Haversine formula:

[0110]

[0111]

[0112]

[0113] in, , Wind farm , Latitude; , Wind farm , Longitude; This is an intermediate variable in the Haversine formula. For the first The wind farm and the first The central angle between the wind farms The average radius of the Earth;

[0114] ③ Gaussian kernel transformation: Convert distance to adjacency weights using the Gaussian kernel function:

[0115]

[0116] ④ Diagonal processing: This indicates that the diagonal elements of the adjacency matrix representing a self-join are set to 0, i.e. ;

[0117] 3.2) Construction of Power Correlation Adjacency Matrix

[0118] The correlation coefficient between wind farms is calculated based on historical power data, and a correlation adjacency matrix is ​​constructed.

[0119] ① Correlation coefficient calculation: Based on the normalized power data matrix, calculate the Pearson correlation coefficient matrix between each wind farm:

[0120]

[0121] in, Power-related adjacency matrix The Middle Line number The element of the column represents the first element. The wind farm and the first Pearson correlation coefficient between power series of wind farms; , Wind farm , The power sequence, For covariance, The standard deviation is denoted as .

[0122] ② Physical meaning of the correlation coefficient matrix: The correlation coefficient reflects the synchronicity of power changes between wind farms. A high positive correlation indicates that the power change trends of the two wind farms are consistent and are affected by similar meteorological conditions; a low or negative correlation indicates that the power change trends are different and are affected by different meteorological systems.

[0123] As a specific example, step 4 learns node embeddings through a graph neural network to generate an adaptive dynamic adjacency matrix. The dynamic graph can adaptively adjust the connection strength between nodes according to the current power state, capturing the dynamically changing spatial relationships between wind farms and learning the correlation strength between wind farms at a given moment, thereby improving the model's ability to capture complex spatiotemporal patterns. Combined with... Figure 2 The specific implementation steps are as follows:

[0124] 4.1) Construction of Node Feature Matrix

[0125] For any time step in the input historical time window The multimodal feature sequence obtained in step 2 is extracted and organized according to time steps to obtain the node feature matrix of the current time step. ,in, , Represents the composition of real numbers OK Column matrix space, The The line represents the first A wind farm in time step The following is from The modal feature vector is composed of 1 eigenmodes, and the node feature matrix is... As input to the dynamic graph generation module;

[0126] 4.2) First-layer graph convolution embedding calculation

[0127] The node feature matrix at the current time step Input the first layer of the graph convolutional network in the dynamic graph generation module, and use the geospatial adjacency matrix obtained in step 3. Perform graph convolution to obtain the embedding matrix of intermediate nodes at the current time step. The calculation method is as follows:

[0128]

[0129] in, The weight matrix to be learned for the first layer of the graph convolutional network. For time steps The embedding matrix of intermediate nodes output by the first layer of the graph convolutional network;

[0130] 4.3) Second-layer graph convolution embedding calculation

[0131] Embed the intermediate node into the matrix Input the second-layer graph convolutional network, and again base it on the geospatial adjacency matrix. Perform graph convolution to obtain the node embedding matrix at the current time step. The calculation method is as follows:

[0132]

[0133] in, The weight matrix to be learned for the second layer of the graph convolutional network. For time steps The node embedding matrix output by the second-layer graph convolutional network; the second-layer graph convolutional network does not have an activation function set to preserve the linear features of the node embedding representation;

[0134] 4.4) Node embedding normalization processing

[0135] Node embedding matrix The embedding vector corresponding to each wind farm node is L2 normalized to obtain the normalized node embedding matrix. The calculation method is as follows:

[0136]

[0137] in, Embedding matrix for nodes The Middle Embedding vectors corresponding to each wind farm node normalized node embedding matrix The Middle The normalized embedding vector corresponding to each wind farm node Represents the L2 norm;

[0138] 4.5) Dynamic Adjacency Matrix Generation

[0139] Calculate the normalized node embedding matrix The inner product between the embedding vectors of any two wind farm nodes yields the dynamic adjacency matrix for the current time step. The calculation method is as follows:

[0140]

[0141] in, normalized node embedding matrix The transpose of the matrix, For time steps The dynamically generated adjacency matrix; Indicates the first The wind farm node and the first Each wind farm node at time step The embedding similarity is used to characterize the correlation strength between two corresponding wind farms under the current power state.

[0142] 4.6) Normalization of dynamic adjacency matrix

[0143] For dynamic adjacency matrix Each row is Softmax normalized to obtain the normalized dynamic adjacency matrix. ,in, The Middle The wind farm node points to the first Dynamic connection weights of individual wind farm nodes The calculation method is as follows:

[0144]

[0145] in, For the summation index in Softmax normalization, is a natural constant; the Softmax normalization makes the sum of the outgoing edge weights corresponding to each wind farm node equal to 1;

[0146] 4.7) Dynamic generation and processing of time steps

[0147] For each time step in the input historical time window The process of constructing the node feature matrix at the current time step, calculating the first-layer graph convolutional embedding, calculating the second-layer graph convolutional embedding, normalizing the node embedding, generating the dynamic adjacency matrix, and normalizing the dynamic adjacency matrix are repeated to obtain the normalized dynamic adjacency matrix for each time step. The normalized dynamic adjacency matrix It is used to characterize the real-time correlation strength between wind farms as the current power state changes, and serves as the dynamic graph input for hierarchical fusion with the static fusion adjacency matrix in step 5.

[0148] As a specific example, step 5 uses learnable weight parameters to hierarchically fuse the static and dynamic adjacency matrices to form the final adjacency matrix used for graph convolution operations. This hierarchical fusion strategy adaptively balances prior knowledge and data-driven learning results. The specific implementation steps are as follows:

[0149] 5.1) First layer fusion: Fusion of static graphs

[0150] First, merge the two static adjacency matrices, namely the geospatial adjacency matrix. and correlation matrix Through the weight parameters that need to be learned Perform a weighted average:

[0151]

[0152] in, This is a statically merged adjacency matrix; These are the scalar parameters that need to be learned in the static fusion adjacency matrix fusion process, and their value range is constrained within... Within the interval, the initial value is set to 0.5, indicating that the two static graphs initially have the same weights; during training, Automatic learning and adjustment are performed to achieve the optimal static image fusion ratio;

[0153] 5.2) Second layer fusion: Fusion of static and dynamic graphs

[0154] The static adjacency matrix obtained from the first layer fusion With normalized dynamic adjacency matrix To perform the fusion, another weight parameter needs to be learned. Perform a weighted average:

[0155]

[0156] in, For time steps The hierarchical fusion adjacency matrix below, The scalar parameters that need to be learned in the fusion of static and dynamic graphs have their value range constrained within... Within the interval, the initial value is set to 0.5; during training, It automatically learns and adjusts to achieve the optimal static-dynamic image fusion ratio.

[0157] As a specific example, step 6 includes:

[0158] The multi-objective optimization problem of multi-wind farm interval prediction is modeled as a joint probability maximization problem, based on the chain decomposition theory of conditional probability:

[0159]

[0160] The total loss is expressed as a negative logarithmic joint probability, comprising four progressive loss terms, where This represents the WinklerScore basic loss. Indicates spatial smoothing loss, This indicates the correlation coverage loss. Indicates nested interval loss;

[0161] 6.1) Winkler Score Basic Loss

[0162] The classic Winkler Score is used to comprehensively evaluate interval width and coverage violations for the first... The wind farm in the first The prediction interval at each confidence level is defined as:

[0163]

[0164] Winkler Score Basic Loss for:

[0165]

[0166] in, For confidence level indexing, The total number of confidence levels. For the first One confidence level , The first individual wind farms at confidence level The upper and lower bounds of the prediction interval are below. For the first The actual power value of a wind farm at the predicted time; Temperature parameter for the Winkler Score base loss (controls the sharpness of the probability distribution, default value 1.0);

[0167] Log probability converted to Winkler Score basic loss

[0168]

[0169] 6.2) Spatial smoothing loss

[0170] Based on graph Laplace regularization, the smoothness of the boundary between adjacent wind farms is constrained, resulting in spatial smoothing loss. Represented as:

[0171]

[0172] in, , The first The wind farm in the first The upper and lower bounds of the prediction interval at each confidence level are constrained by the dynamic adjustment of the temperature parameter in the Winkler Score base loss.

[0173]

[0174] in, The temperature parameter is used to represent the spatial smoothing loss, which is then converted into the log probability of the spatial smoothing loss. :

[0175]

[0176] 6.3) Correlation Coverage Loss

[0177] The constraints require that the relative positions of the true values ​​of highly correlated wind farms within the interval be similar. The actual power value of the wind farm at the first The relative position within the prediction interval corresponding to each confidence level is defined as follows: Correlation coverage loss Defined as:

[0178]

[0179] in, , The first The actual power value of each wind farm , No. The actual power value of each wind farm In the Each confidence level corresponds to a relative position within the prediction interval; This is a correlation threshold mask; the correlation coefficient is set to 1 if it exceeds a preset threshold, and 0 otherwise. Constraints are only applied to strongly correlated wind farm pairs. The constraint strength is jointly adjusted by the first two temperature parameters.

[0180]

[0181] in, The temperature parameter is used to represent the correlation coverage loss, which is then converted into the log probability of the correlation coverage loss. :

[0182]

[0183] 6.4) Nested Interval Loss

[0184] An interval with higher confidence should include an interval with lower confidence; therefore, the loss is constrained by this rule. The confidence levels are arranged from highest to lowest confidence level, i.e. The nested interval loss is defined as:

[0185]

[0186] in, , The first The wind farm in the first The upper and lower bounds of the prediction interval at each confidence level; the constraint strength is adjusted by all prerequisite temperature parameters.

[0187]

[0188] in, The temperature parameter represents the nested interval loss.

[0189] 5) Total Loss Function for:

[0190]

[0191] As a specific example, step 7 extracts the spatiotemporal features of the power sequence by combining a temporal convolutional network and a graph convolutional network. Temporal convolution captures temporal dependencies, while graph convolution captures spatial dependencies. The combination of the two can comprehensively model the spatiotemporal evolution of power across multiple wind farms. The specific implementation steps are as follows:

[0192] 7.1) Input Data

[0193] The input is a multimodal feature sequence after adaptive multi-wind farm VMD decomposition. The features include sample batch size, historical time window length, number of wind farms, and number of decomposition modes.

[0194] 7.2) Initial Feature Extraction

[0195] Initial feature extraction is performed on the input data using 2D convolutional layers. Zero padding is used to maintain the output size. After Leaky ReLU activation, the features are obtained. Each modal feature is mapped to a higher-dimensional feature space to capture local patterns in the temporal and spatial dimensions;

[0196] 7.3) Spatiotemporal graph convolutional layer

[0197] Construct multiple spatiotemporal graph convolutional layers, each containing two sub-modules: temporal convolution and graph convolution.

[0198] ① Temporal Convolution Submodule: Performs a one-dimensional dilated convolution operation independently on the time series data of each wind farm, with a convolution kernel size of [missing value]. expansion rate Zero padding is used to maintain the sequence length, and the output is denoted as... ;

[0199] ② Graph Convolution Submodule: Performs graph convolution operation on the spatial data at each time step to capture spatial dependencies. The operation formula is:

[0200]

[0201] in, The weight matrix that the graph convolution module needs to learn;

[0202] Graph convolution operations aggregate the features of adjacent nodes through an adjacency matrix, thereby achieving the propagation and fusion of spatial information.

[0203] ③ Residual connection: The input and output of the spatiotemporal graph convolutional layer are added together to form a residual connection, which alleviates the gradient vanishing problem and accelerates model training;

[0204] 7.4) Time step traversal

[0205] Since the dynamic adjacency matrix is ​​generated independently for each time step, the graph convolution operation needs to be performed separately for each time step. The specific process is as follows:

[0206] ① Regarding time steps , extract the node features of this time step ,in The length of the historical time window;

[0207] ② Input the dynamic graph generation module to generate the dynamic adjacency matrix for that time step. ;

[0208] ③ Use a layered fusion strategy to... With static adjacency matrix Fusion, to obtain ;

[0209] ④ Use Perform graph convolution to obtain the output at that time step. ;

[0210] ⑤ Stack the outputs of all time steps to form a complete timing output;

[0211] 7.5) Output Layer

[0212] After multiple spatiotemporal graph convolutional layers, a fully connected layer is used to map the features to the output space of the prediction interval; each wind farm output Each value corresponds to The upper and lower bounds of the prediction intervals for each confidence level are denoted as follows: Since we only need to predict the next point in time, we take... Output of the last time step .

[0213] This invention also provides a multi-wind farm power interval prediction system based on adaptive cooperative VMD and dynamic spatiotemporal graph convolution. This system is used to implement the aforementioned multi-wind farm power interval prediction method based on adaptive cooperative VMD and dynamic spatiotemporal graph convolution. The system includes a data preprocessing module, an adaptive multi-wind farm cooperative VMD signal decomposition module, a hybrid adjacency matrix construction module, a dynamic graph generation module, a spatiotemporal graph convolution modeling module, a progressive constraint loss function training module, and a multi-confidence level interval output module, wherein:

[0214] The data preprocessing module is responsible for data acquisition and normalization.

[0215] The adaptive multi-wind farm collaborative VMD signal decomposition module decomposes the power sequence of each wind farm into spatially aligned multi-frequency components through a joint optimization objective.

[0216] The hybrid adjacency matrix construction module integrates geographical distance and power correlation to construct a static graph;

[0217] The dynamic graph generation module learns an adaptive adjacency matrix through a graph neural network;

[0218] The spatiotemporal graph convolutional modeling module extracts spatiotemporal features through a layered and fused graph structure;

[0219] The progressive constraint loss function training module achieves automatic weight adjustment through conditional probability chain decomposition.

[0220] The multi-confidence level interval output module generates prediction intervals for multiple confidence levels.

[0221] The present invention also provides a computer-readable storage medium having a computer program stored thereon, wherein the program, when executed by a processor, implements the steps in the multi-wind farm power range prediction method based on adaptive cooperative VMD and dynamic spatiotemporal graph convolution.

[0222] This invention addresses the practical needs of power range prediction for multiple wind farms. Through innovative technologies such as adaptive multi-wind farm collaborative VMD decomposition, dynamic graph generation and hierarchical fusion strategies, and a progressive constraint loss function based on conditional probability chain decomposition, it constructs a complete spatiotemporal range prediction methodology. This solves problems in existing prediction technologies, including limitations of static graph structures, lack of spatial coordination constraints in multi-wind farm range prediction, inability to quantify uncertainty in point prediction, and fixed VMD parameters without considering spatial coordination. The core technical points are as follows:

[0223] 1. Adaptive Multi-Wind Farm Cooperative VMD Decomposition Method

[0224] An adaptive multi-wind farm collaborative VMD decomposition method considering spatial coordination constraints is proposed, overcoming the limitations of traditional VMD independent decomposition of each wind farm:

[0225] Collaborative optimization objective: Based on the traditional VMD bandwidth constraint term, add spatial collaborative regularization term and frequency alignment regularization term to unify the decomposition problem of each wind farm into a multi-wind farm joint optimization problem, so as to achieve global optimum rather than local optimum.

[0226] Spatial Cooperative Regularization: Based on the graph Laplace regularization theory, a spatial adjacency matrix is ​​constructed using the geographical distance between wind farms to constrain the waveform similarity and center frequency proximity of the same frequency modes of adjacent wind farms, thus ensuring the consistency of spatial cooperative decomposition.

[0227] Frequency alignment regularization: Constrains all wind farms to maintain the same frequency range as the mode index IMF, and forces different modes to maintain an orderly arrangement from high frequency to low frequency to avoid frequency aliasing.

[0228] Extended ADMM solution: Introduce the coupling constraint term of adjacent wind farms into the frequency domain formula of the modal update subproblem, add the weighted average of the frequencies of adjacent wind farms into the frequency update subproblem, and achieve an effective solution of the cooperative constraints through the extended alternating direction multiplier method.

[0229] Adaptive parameter selection: The number of modes is adaptively determined by the energy entropy criterion, the spatial coordination weight is adaptively set according to the density of the spatial distribution of the wind farm, and the frequency alignment weight is adaptively set according to the signal spectrum entropy, so that the method can adapt to different wind farm layouts and meteorological conditions.

[0230] 2. Dynamic Graph Generation Module and Layered Fusion Strategy

[0231] A three-layer hybrid graph structure integrating geospatial, power correlation, and data-driven dynamic graphs is constructed, and dynamic graph generation and hierarchical fusion are achieved through a two-layer graph convolutional network:

[0232] Geospatial adjacency matrix: The spherical distance between wind farms is calculated using the Haversine formula and converted into adjacency weights through a Gaussian kernel function to characterize the spatial propagation characteristics of geographical proximity and meteorological conditions.

[0233] Power correlation adjacency matrix: The Pearson correlation coefficient matrix is ​​calculated based on historical power data to characterize the statistical synchronicity of power changes among wind farms.

[0234] GCN dynamic graph generation: Taking the multimodal features of collaborative VMD decomposition as input, it learns node embeddings through a two-layer graph convolutional network, performs L2 normalization on the embedding vectors, and then performs Softmax normalization to generate a dynamic adjacency matrix, which adaptively characterizes the real-time correlation strength between wind farms according to the power state at the current moment.

[0235] Time-dependent dynamic graph evolution: A dynamic adjacency matrix is ​​generated independently for each time step to capture the dynamic evolution of the relationship between wind farms as meteorological conditions and wind direction change.

[0236] Layered fusion strategy: A static map is obtained by weighted fusion of the geospatial matrix and the power correlation matrix through learnable parameters. Then, the static map is fused with the dynamic map through another learnable parameter. Only two parameters are needed to automatically balance prior knowledge and data-driven learning results. The parameters are efficient and interpretable.

[0237] 3. Progressive Constraint Loss Function Based on Conditional Probability Chain Decomposition

[0238] Using the chain decomposition theory of conditional probability in probability theory, a progressively constrained loss function is designed to achieve automatic weight adjustment of multiple loss terms:

[0239] Theoretical framework: Each loss term is modeled as the negative logarithm of the conditional probability. Based on the chain decomposition of the conditional probability of the joint probability, the total loss is expressed as the sum of four progressive loss terms. The dependencies between the loss terms are automatically adjusted by the pre-parameters.

[0240] The design of the fusion loss function is proposed: a progressive loss function is proposed that integrates Winkler Score basic loss, spatial smoothing conditional loss, correlation coverage conditional loss and interval nesting conditional loss. This enables progressive optimization from the basic interval quality of a single wind farm to the spatial coordination consistency of multiple wind farms, and then to the logical constraints of multiple confidence levels. The constraint strength of each loss term is automatically adjusted by the pre-set quality factor, without the need for manual weight setting.

[0241] The quality factor dynamic adjustment mechanism is as follows: the better the quality of the preceding loss, the smaller the relative weight of the subsequent constraints will be, forming a progressive optimization process from the basic quality of a single wind farm to the spatial coordination of multiple wind farms. The entire loss function only requires a temperature parameter.

[0242] 4. Enhanced Spatiotemporal Graph Convolutional Network Architecture

[0243] By combining collaborative VMD multimodal features, dynamic graph generation, and hierarchical fusion strategies, an end-to-end spatiotemporal prediction network is constructed:

[0244] Initial feature extraction: Two-dimensional convolutional layers are used to extract initial features from the multimodal input of co-VMD decomposition, mapping multiple IMF modes to a higher-dimensional feature space and capturing local patterns in the temporal and spatial dimensions.

[0245] Spatiotemporal graph convolutional layer: Construct multiple spatiotemporal graph convolutional layers, each containing a temporal convolutional submodule and a graph convolutional submodule. The temporal convolutional submodule uses dilated convolution with an increasing dilation rate to increase the temporal receptive field, while the graph convolutional submodule uses a hierarchical fusion adjacency matrix to aggregate spatial features. The two are connected in series to achieve joint extraction of spatiotemporal features.

[0246] Residual connections: Residual connections are added between spatiotemporal graph convolutional layers. When the input and output dimensions are different, pointwise convolution is used for dimension projection, which alleviates the gradient vanishing problem in deep networks.

[0247] Multi-confidence level output: The output layer uses a fully connected layer to map features to the upper and lower bounds of the prediction intervals for multiple confidence levels, thus completing the construction of prediction intervals for multiple confidence levels in a single prediction process.

[0248] The effectiveness and superiority of the enhanced dynamic graph convolutional network proposed in this invention in predicting power ranges across multiple wind farms are verified through the following embodiments.

[0249] Example

[0250] This embodiment designs detailed real-world application cases and conducts comprehensive ablation experiment analysis.

[0251] 1) Experimental Dataset and Preprocessing

[0252] This experiment uses data from a real wind farm power output dataset in a region of Australia, spanning one year with a data sampling resolution of 15 minutes. This dataset contains complete seasonal variations and wind power fluctuation characteristics under complex meteorological conditions throughout the year, effectively testing the model's ability to model the spatial coordination of multiple wind farms.

[0253] The dataset was divided into training, validation, and test sets in an 8:1:1 ratio according to chronological order. Missing values ​​were processed using linear interpolation, and the power data from all wind farms were Min-Max normalized and scaled to [specific scale value]. The interval is used to eliminate the influence of dimensions and accelerate model convergence.

[0254] 2) Experimental Design and Results Analysis

[0255] To verify the actual effectiveness of each core innovative module in this invention, this embodiment sets the prediction confidence levels to 85%, 90%, and 95%, respectively, and designs a comparative experiment between the model of this invention and three ablation models. The adaptive multi-wind farm collaborative VMD module, the enhanced dynamic graph neural network module, and the proposed progressive constraint loss function based on conditional probability chain decomposition are verified one by one.

[0256] The multi-interval prediction results obtained by the model proposed in this invention on the real wind power dataset are as follows: Figures 3-5 As shown, the red curve represents the actual power value of the wind farm during this period, and the different shades of blue represent three sets of prediction intervals at confidence levels of 85%, 90%, and 95%, respectively.

[0257] Comparative experimental results show that, compared with standard variational mode decomposition, the model prediction results using the adaptive multi-wind farm collaborative VMD module are significantly closer to the confidence level in terms of interval coverage. Its accuracy is improved by an average of more than 8%, and the interval width is reduced by more than 10%. Using the Winkler Score, which is widely recognized in the field, as a comprehensive evaluation index, the index is reduced by an average of more than 8%. This indicates that the proposed improved VMD decomposition method can align the spatial frequencies of IMF features of adjacent wind farms, providing cleaner features for graph convolution and effectively controlling the overall prediction accuracy.

[0258] Compared with the standard Winkler Score as the loss function, the prediction model using the proposed progressive constraint loss function based on conditional probability chain decomposition does not significantly improve the error in interval coverage and confidence level, but the interval width is reduced by an average of more than 16%. Analysis of the comparison curves shows that in time periods with similar power changes in adjacent wind farms, the model proposed in this invention exhibits better interval shrinkage capability, and its Winkler Score comprehensive evaluation is reduced by more than 13%.

[0259] Compared with ordinary static graph networks, graph neural networks using dynamic matrices as input show a significant improvement in interval coverage, with interval coverage error reduced by more than 12% and interval width reduced by an average of more than 7%. Due to the use of dynamic graph fusion mechanism, the model proposed in this invention shows better capture ability when facing dynamic changes such as wind direction and airflow, especially during periods of drastic power changes, with its Winkler Score overall evaluation reduced by more than 9%.

[0260] In summary, the present invention, based on adaptive collaborative VMD and dynamic spatiotemporal graph convolution, achieves the following technical effects in its method and system for predicting power ranges in multiple wind farms:

[0261] (1) Denoising and feature extraction of adaptive multi-wind farm collaborative VMD decomposition;

[0262] This invention proposes an adaptive multi-wind farm collaborative variational mode decomposition method, overcoming the limitations of traditional VMD's independent decomposition of individual wind farms. By introducing a spatial Laplace regularization term into the optimization objective, it achieves collaborative constraints on the waveforms and frequencies of the same-frequency modes of wind farms with similar geographical locations. Through a frequency alignment regularization term, it ensures that the IMFs with the same mode index represent similar frequency ranges across all wind farms, avoiding frequency aliasing. The number of modes is adaptively determined using the energy entropy criterion, and the spatial collaborative weights and frequency alignment weights are adaptively set according to the density of the wind farm spatial distribution and the signal frequency complexity, respectively, enabling the method to adapt to different wind farm layouts and meteorological conditions. The multi-modal features obtained from the collaborative decomposition exhibit good spatial alignment consistency, providing high-quality multi-channel input features for subsequent spatial collaborative modeling of graph convolutional layers, significantly improving the model's ability to capture the spatiotemporal fluctuation patterns of power from multiple wind farms.

[0263] (2) The complementary advantages and spatial modeling enhancement of dynamic graph generation and hierarchical fusion strategies;

[0264] This invention, through a dynamic graph generation module and a hierarchical fusion strategy, achieves a three-layer hybrid graph structure integrating geospatial prior knowledge, historical power correlation statistics, and data-driven real-time association, effectively overcoming the limitations of static graph structures in adapting to dynamic spatial associations. The dynamic graph generation module employs a two-layer graph convolutional network to learn node embeddings, independently generating an adaptive adjacency matrix for each time step to capture the dynamic evolution of relationships between wind farms as meteorological conditions and wind direction change. The hierarchical fusion strategy requires only two learnable parameters, first fusing two types of static graphs, then fusing the static and dynamic graphs, aligning with the cognitive hierarchy from prior knowledge to data-driven learning. Compared to a three-graph weighted average, this approach is more parameter-efficient and interpretable. It significantly improves the model's ability to model complex spatiotemporal patterns, achieving a comprehensive improvement in prediction accuracy and model interpretability.

[0265] (3) Progressive constraint loss function and automatic weight adjustment based on conditional probability chain decomposition;

[0266] This invention employs conditional probability chain decomposition theory to design a progressive constraint loss function, achieving automatic weight adjustment in multi-objective optimization for multi-wind farm interval prediction. This avoids the problem of imbalanced optimization objectives caused by manually setting weight coefficients. Four loss terms form a progressive constraint hierarchy: basic interval quality assurance, boundary spatial smoothing constraint between adjacent wind farm intervals, coverage consistency constraint for highly correlated wind farms, and nested relationship constraint between multiple confidence level intervals. The better the pre-constraints are satisfied, the stricter the subsequent constraints become automatically, forming a progressive optimization process from single wind farm to multi-wind farm collaboration. The entire loss function requires only one temperature parameter, significantly reducing the difficulty of parameter tuning. It fully utilizes the spatial correlation information between multiple wind farms, significantly improving the spatial consistency and overall reliability of multi-wind farm joint prediction, while reducing reliance on professional parameter tuning experience, thus improving generalizability and practicality.

[0267] (4) Extensive business adaptability and application value;

[0268] This invention addresses the practical needs of power range prediction for multiple wind farms, providing a complete end-to-end spatiotemporal range prediction framework with high business scalability and flexibility. It supports any number of wind farms and prediction time windows of arbitrary length. Hyperparameters such as the number of modes in the co-VMD decomposition, time window length, and number of graph convolution layers can be flexibly configured according to specific application scenarios, adapting to power fluctuation characteristics under different wind farm layouts, seasons, and meteorological conditions. The model directly outputs prediction ranges with multiple confidence levels, meeting the power dispatching decision-making needs of different risk preferences, providing reliable uncertainty quantification information for grid dispatching, and supporting the safe and stable operation of the grid and the consumption of new energy. It can provide high-precision range prediction results for new energy power plants participating in electricity market transactions, reducing assessment costs, and has broad application prospects and significant socio-economic value.

[0269] The above description is only a preferred embodiment of the present invention, but the scope of protection of the present invention is not limited thereto. Any changes or substitutions that can be easily conceived by those skilled in the art within the scope of the technology disclosed in the present invention should be included within the scope of protection of the present invention.

Claims

1. A method for predicting the power range of multiple wind farms based on adaptive collaborative VMD and dynamic spatiotemporal graph convolution, characterized in that, Includes the following steps: Step 1, Data Acquisition and Preprocessing: Historical power data is collected from the data acquisition systems of multiple wind farms, and the historical power data is preprocessed to obtain a normalized power data matrix with each wind farm as a node; Step 2, Adaptive Multi-Wind Farm Cooperative VMD Signal Decomposition: Input the normalized power data matrix obtained in Step 1 into the adaptive multi-wind farm cooperative variational mode decomposition model. Through the joint optimization objective including VMD bandwidth constraints, spatial cooperative constraints and frequency alignment constraints, the power sequences of multiple wind farms are cooperatively decomposed. The decomposition parameters are adaptively determined according to energy entropy, spatial adjacency weight and spectral entropy to obtain multiple intrinsic mode components of each wind farm spatially aligned under the same mode index. Step 3: Construction of geospatial adjacency matrix and power correlation adjacency matrix: Calculate the geographical distance between wind farms based on the geographical coordinates of each wind farm, and convert the geographical distance into a geospatial adjacency matrix; at the same time, calculate the correlation between the power sequences of different wind farms based on the normalized power data matrix obtained in Step 1, and obtain the power correlation adjacency matrix. Step 4: Generation of adaptive dynamic adjacency matrix: Extract and organize the multimodal feature sequence obtained in Step 2 according to time steps to obtain the node feature matrix of the current time step, and input it into the dynamic graph generation module. Learn the node embedding of each wind farm node in the current time step through graph convolutional network, normalize and calculate the similarity of the node embedding, and generate the dynamic adjacency matrix of the corresponding time step after normalization. Step 5: Construction of the hierarchical fusion graph structure: Perform a first-layer weighted fusion on the geospatial adjacency matrix and the power correlation adjacency matrix obtained in Step 3 to obtain the static fusion adjacency matrix; The static fused adjacency matrix is ​​then fused with the dynamic adjacency matrix obtained in step 4 using a second-level weighted fusion to obtain a hierarchical fused adjacency matrix. Step 6, Progressive Constraint Loss Training: Based on the prediction interval results and true power values ​​output by the enhanced dynamic spatiotemporal graph convolutional network during the training phase, a progressive constraint loss function based on conditional probability chain decomposition is constructed, and the progressive constraint loss function is used to train the dynamic graph generation module, hierarchical fusion parameters, enhanced dynamic spatiotemporal graph convolutional network, and output layer parameters. Step 7: Generation of prediction intervals at multiple confidence levels: The multimodal feature sequences obtained in Step 2 and the hierarchical fusion adjacency matrix obtained in Step 5 are input into the enhanced dynamic spatiotemporal graph convolutional network trained in Step 6. Initial feature extraction is performed through two-dimensional convolution. The time dependence features of the power sequences of each wind farm and the spatial correlation features between different wind farms are extracted through the spatiotemporal graph convolutional layer to obtain the spatiotemporal feature representation. Finally, the spatiotemporal feature representation is mapped to the upper and lower bounds of the prediction intervals for each wind farm at multiple confidence levels through the output layer to obtain the multi-confidence level interval prediction results for the power of multiple wind farms.

2. The multi-wind farm power interval prediction method based on adaptive cooperative VMD and dynamic spatiotemporal graph convolution as described in claim 1, characterized in that, Step 1 includes: 1.1) Input Data Acquisition Historical power data was collected from multiple wind farm data acquisition systems, including timestamps, wind farm identifiers, and actual power generation values. The data was in the form of a long table, with each row containing three fields: timestamp, wind farm name, and power value. 1.2) Preprocess historical power data, including format conversion, missing value imputation, and normalization: ① Format conversion: Convert the data format to a wide table, with row indexes as timestamps, column indexes as wind farm names, and cell values ​​as power; ② Missing value handling: Check for missing values ​​in the data and fill them with linear interpolation. For missing values ​​at the beginning and end, use backfilling and forwardfilling methods respectively. ③ Data Normalization: Normalize the power data for each wind farm, scaling the data to... The normalization formula for the interval is: in, This is the normalized power value. This is the original power value. , These represent the minimum and maximum historical power outputs of the wind farm, respectively.

3. The multi-wind farm power interval prediction method based on adaptive cooperative VMD and dynamic spatiotemporal graph convolution according to claim 2, characterized in that, Step 2 includes: 2.1) Optimization objective of collaborative VMD decomposition Traditional VMD decomposition optimizes each wind farm independently. Extending traditional VMD decomposition to a joint optimization problem involving multiple wind farms is more effective. Power signals of a wind farm The optimization objective of collaborative VMD is: in, For wind farm indexing, For the number of wind farms, For modal indexing, For modal number, For time variables, For the first The raw power signal of a wind farm. For the first The first wind farm One modal component, For the first The first wind farm The center frequency of each mode For time The partial derivative operator, For the Dirac function, The square of the L2 norm. The imaginary unit; The first term in the formula is the bandwidth constraint term of VMD, which is the sum of all wind farms. For spatial co-regularization, For the weight coefficients of the spatial cooperative regularization term, For frequency alignment regularization, These are the weighting coefficients for the frequency alignment regularization term; 2.2) Spatial Co-regularization Term Design The spatial cooperative regularization term is based on graph Laplace regularization theory, specifically as follows: in, For wind farm indexing, For the first The first wind farm One modal component, For the first The first wind farm The center frequency of each mode; Geospatial adjacency matrix The Middle The wind farm and the first The adjacency weights between wind farms, calculated based on the Gaussian kernel function, are expressed as follows: in, The bandwidth parameter is set according to the typical spatial distribution scale of the wind farm; For the first The wind farm and the first Haversine distance between wind farms; The spatial cooperative regularization term is weighted by the spatial adjacency matrix, which makes the constraint strength between wind farm pairs that are closer together greater; 2.3) Design of Frequency Alignment Regularization Terms The frequency alignment regularization term is as follows: in, For all wind farms The variance of the center frequencies of each mode. For the first The average center frequency of each mode To correct the linear unit; The frequency alignment regularization term ensures that the IMFs with the same mode index represent the same frequency range in all wind farms, and that there is a clear frequency separation between different modes, arranged in an orderly manner from high frequency to low frequency; 2.4) Solving with the Alternating Direction Multipliers (ADMM) of Cooperative VMD The cooperative VMD optimization problem is solved iteratively using the extended ADMM algorithm: ① Initialization: Perform standard VMD decomposition independently on each wind farm to obtain the initial modes. and frequency Initialize Lagrange multipliers and penalty parameters ; ② Modal update: fixed center frequency and Lagrange multipliers Update modes in the frequency domain: in, This is a modal summation index used to iterate through all modal summations except the current index. The remaining modes other than the one mode; This is the index for the number of iterations in the ADMM algorithm; Indicates the first The result of the iteration is the first The first wind farm Modal components Fourier transform; Indicates the first The first wind farm Modal components Fourier transform; Indicates the first Power signal of a wind farm Fourier transform; Indicates the first Lagrange multipliers corresponding to each wind farm Fourier transform; For frequency domain variables; For VMD bandwidth constraint parameters; spatial coordination term in the denominator Used to represent the coupling constraints between adjacent wind farms; ③ Frequency update: Fixed mode Update # The iteration of the ... The first wind farm The center frequency of each mode : in, For the first The first wind farm Modal components Fourier transform; ④ Lagrange multiplier update: in, and They represent the first Second and third Lagrange multipliers in the next iteration; ⑤ Convergence criterion: When the residual Spatial consistency index All less than the tolerance Stop iteration when the time is right; 2.5) Parameter Adaptive Selection Strategy ① Modal number K: Adaptively determined using the energy entropy criterion; the decomposition results under different K values ​​are calculated, and the K value that minimizes the average energy entropy is selected; energy entropy is defined as: in, For energy entropy, the first The proportion of modal energy to the total modal energy , For the first Energy of each mode; ② Spatial Cooperative Regularization Term Weight Coefficient The weight coefficients of the spatial coordination regularization term are adaptively set based on the density of the wind farm's spatial distribution. in, This is the scaling factor for the weight coefficients of the spatial co-regularization term; ③ Frequency alignment regularization term weight coefficient : Adaptively set according to signal frequency complexity; for the first Power signals of a wind farm Perform a Fourier transform to obtain the frequency domain signal. And calculate the spectral entropy based on the power spectrum, the first The wind farm in the first The power spectral energy of a frequency sampling point is defined as: No. The wind farm in the first The normalized spectral energy percentage of each frequency sampling point is defined as: No. The spectral entropy of a wind farm is defined as: The average spectral entropy of all wind farms is defined as: The frequency alignment regularization term weights are adaptively set based on the average spectral entropy: in, For the first The frequency corresponding to each frequency sampling point For frequency sampling point index, This represents the total number of frequency sampling points. For the first The wind farm in the first Power spectral energy at each frequency sampling point For the first The wind farm in the first Normalized spectral energy percentage of each frequency sampling point This serves as the index for summing frequency sampling points in normalization calculations. To prevent positive constants with a denominator of zero, For the first The spectral entropy of a wind farm The average spectral entropy of all wind farms, This is the scaling factor for the weights of the frequency-aligned regularization term; the higher the frequency complexity, the better. The larger, The larger the value, the stronger the frequency alignment constraint.

4. The multi-wind farm power interval prediction method based on adaptive cooperative VMD and dynamic spatiotemporal graph convolution according to claim 3, characterized in that, Step 3 includes: 3.1) Construction of geospatial adjacency matrix Based on the geographical coordinates of wind farms, namely longitude and latitude, the actual distance between wind farms is calculated, and the distance is converted into adjacency weights using a Gaussian kernel function. ① Wind farm coordinate definition: Define latitude and longitude geographic coordinates for each wind farm and store them in a dictionary structure. ②Haversine distance calculation: For any two wind farms and The distance between them can be calculated using the Haversine formula: in, , Wind farm , Latitude; , Wind farm , longitude; This is an intermediate variable in the Haversine formula. For the first The wind farm and the first The central angle between the wind farms The average radius of the Earth; ③ Gaussian kernel transformation: Convert distance to adjacency weights using the Gaussian kernel function: ④ Diagonal processing: This indicates that the diagonal elements of the adjacency matrix representing a self-join are set to 0, i.e. ; 3.2) Construction of Power Correlation Adjacency Matrix The correlation coefficient between wind farms is calculated based on historical power data, and a correlation adjacency matrix is ​​constructed. ① Correlation coefficient calculation: Based on the normalized power data matrix, calculate the Pearson correlation coefficient matrix between each wind farm: in, Power-related adjacency matrix The Middle Line number The element of the column represents the first element. The wind farm and the first Pearson correlation coefficient between power series of wind farms; , Wind farm , The power sequence, For covariance, The standard deviation is denoted as .

5. The multi-wind farm power interval prediction method based on adaptive cooperative VMD and dynamic spatiotemporal graph convolution according to claim 4, characterized in that, Step 4 includes: 4.1) Construction of Node Feature Matrix For any time step in the input historical time window The multimodal feature sequence obtained in step 2 is extracted and organized according to time steps to obtain the node feature matrix of the current time step. ,in, , Represents the composition of real numbers OK Column matrix space, The The line represents the first A wind farm in time step The following is from The modal feature vector is composed of 1 eigenmodes, and the node feature matrix is... As input to the dynamic graph generation module; 4.2) First-layer graph convolution embedding calculation The node feature matrix at the current time step Input the first layer of the graph convolutional network in the dynamic graph generation module, and use the geospatial adjacency matrix obtained in step 3. Perform graph convolution to obtain the embedding matrix of intermediate nodes at the current time step. The calculation method is as follows: in, The weight matrix to be learned for the first layer of the graph convolutional network. For time step The embedding matrix of intermediate nodes output by the first layer of the graph convolutional network; 4.3) Second-layer graph convolution embedding calculation Embed the intermediate node into the matrix Input the second-layer graph convolutional network, and again base it on the geospatial adjacency matrix. Perform graph convolution to obtain the node embedding matrix at the current time step. The calculation method is as follows: in, The weight matrix to be learned for the second layer of the graph convolutional network. For time step The node embedding matrix output by the second layer graph convolutional network; the second layer graph convolutional network does not have an activation function set to preserve the linear features of the node embedding representation; 4.4) Node embedding normalization processing Node embedding matrix The embedding vector corresponding to each wind farm node is L2 normalized to obtain the normalized node embedding matrix. The calculation method is as follows: in, Embedding matrix for nodes The Middle Embedding vectors corresponding to each wind farm node normalized node embedding matrix The Middle The normalized embedding vector corresponding to each wind farm node Represents the L2 norm; 4.5) Dynamic Adjacency Matrix Generation Calculate the normalized node embedding matrix The inner product between the embedding vectors of any two wind farm nodes yields the dynamic adjacency matrix for the current time step. The calculation method is as follows: in, normalized node embedding matrix The transpose of the matrix, For time step The dynamically generated adjacency matrix; Indicates the first The wind farm node and the first Each wind farm node at time step The embedding similarity is used to characterize the correlation strength between two corresponding wind farms under the current power state. 4.6) Normalization of dynamic adjacency matrix For dynamic adjacency matrix Each row is Softmax normalized to obtain the normalized dynamic adjacency matrix. ,in, The Middle The wind farm node points to the first Dynamic connection weights of individual wind farm nodes The calculation method is as follows: in, For the summation index in Softmax normalization, is a natural constant; the Softmax normalization makes the sum of the outgoing edge weights corresponding to each wind farm node equal to 1; 4.7) Dynamic generation and processing of time steps For each time step in the input historical time window The process of constructing the node feature matrix at the current time step, calculating the first-layer graph convolutional embedding, calculating the second-layer graph convolutional embedding, normalizing the node embedding, generating the dynamic adjacency matrix, and normalizing the dynamic adjacency matrix are repeated to obtain the normalized dynamic adjacency matrix for each time step. The normalized dynamic adjacency matrix It is used to characterize the real-time correlation strength between wind farms as the current power state changes, and serves as the input for the dynamic graph in step 5 to perform hierarchical fusion with the static fusion adjacency matrix.

6. The multi-wind farm power interval prediction method based on adaptive cooperative VMD and dynamic spatiotemporal graph convolution according to claim 5, characterized in that, Step 5 includes: 5.1) First layer fusion: Fusion of static graphs First, merge the two static adjacency matrices, namely the geospatial adjacency matrix. and correlation matrix Through the weight parameters that need to be learned Perform a weighted average: in, This is a statically merged adjacency matrix; These are the scalar parameters that need to be learned in the static fusion adjacency matrix fusion process, and their value range is constrained within... Within the interval, the initial value is set to 0.5, indicating that the two static graphs initially have the same weights; during training, Automatic learning and adjustment are performed to achieve the optimal static image fusion ratio; 5.2) Second layer fusion: Fusion of static and dynamic graphs The static adjacency matrix obtained from the first layer fusion With normalized dynamic adjacency matrix To perform the fusion, another weight parameter needs to be learned. Perform a weighted average: in, For time step The hierarchical fusion adjacency matrix below, The scalar parameters to be learned in the fusion of static and dynamic graphs have their value range constrained within... Within the interval, the initial value is set to 0.5; during training, It automatically learns and adjusts to achieve the optimal static-dynamic image fusion ratio.

7. The multi-wind farm power interval prediction method based on adaptive collaborative VMD and dynamic spatiotemporal graph convolution according to claim 6, characterized in that, Step 6 includes: The multi-objective optimization problem of multi-wind farm interval prediction is modeled as a joint probability maximization problem, based on the chain decomposition theory of conditional probability: The total loss is expressed as a negative logarithmic joint probability, comprising four progressive loss terms, where This represents the Winkler Score basic loss. Indicates spatial smoothing loss, This indicates the correlation coverage loss. Indicates nested interval loss; 6.1) Winkler Score Basic Loss The classic Winkler Score is used to comprehensively evaluate interval width and coverage violations for the first... The wind farm in the first The prediction interval at each confidence level is defined as: Winkler Score Basic Loss for: in, For confidence level indexing, The total number of confidence levels. For the first One confidence level , The first individual wind farms at confidence level The upper and lower bounds of the prediction interval are below. For the first The actual power value of a wind farm at the predicted time; Temperature parameters for the Winkler Score baseline loss; Log probability converted to Winkler Score basic loss 6.2) Spatial smoothing loss Based on graph Laplace regularization, the smoothness of the boundary between adjacent wind farms is constrained, resulting in spatial smoothing loss. Represented as: in, , The first The wind farm in the first The upper and lower bounds of the prediction interval at each confidence level are constrained by the dynamic adjustment of the temperature parameter in the Winkler Score base loss. in, The temperature parameter for spatial smoothing loss is used to convert it into the log probability of spatial smoothing loss. : 6.3) Correlation Coverage Loss No. The actual power value of the wind farm at the first The relative position within the prediction interval corresponding to each confidence level is defined as follows: Correlation coverage loss Defined as: in, , The first The actual power value of each wind farm , No. The actual power value of each wind farm In the Each confidence level corresponds to a relative position within the prediction interval; This is a correlation threshold mask; the correlation coefficient is set to 1 if it exceeds a preset threshold, and 0 otherwise. The constraint strength is jointly adjusted by the first two temperature parameters. in, The temperature parameter is used to represent the correlation coverage loss, which is then converted into the log probability of the correlation coverage loss. : 6.4) Nested Interval Loss right The confidence levels are arranged from highest to lowest confidence level. The nested interval loss is defined as: in, , The first The wind farm in the first The upper and lower bounds of the prediction interval at each confidence level; the constraint strength is adjusted by all prerequisite temperature parameters. in, The temperature parameter represents the nested interval loss. 5) Total Loss Function for: 。 8. The multi-wind farm power interval prediction method based on adaptive cooperative VMD and dynamic spatiotemporal graph convolution according to claim 1, characterized in that, Step 7 includes: 7.1) Input Data The input is a multimodal feature sequence after adaptive multi-wind farm VMD decomposition. The features include sample batch size, historical time window length, number of wind farms, and number of decomposition modes. 7.2) Initial Feature Extraction Initial feature extraction is performed on the input data using 2D convolutional layers. Zero padding is used to maintain the output size. After Leaky ReLU activation, the features are obtained. Each modal feature is mapped to a higher-dimensional feature space to capture local patterns in the temporal and spatial dimensions; 7.3) Spatiotemporal graph convolutional layer Construct multiple spatiotemporal graph convolutional layers, each containing two sub-modules: temporal convolution and graph convolution. ① Temporal Convolution Submodule: Performs a one-dimensional dilated convolution operation independently on the time series data of each wind farm, with a convolution kernel size of [missing value]. expansion rate Zero padding is used to maintain the sequence length, and the output is denoted as... ; ② Graph Convolution Submodule: Performs graph convolution operation on the spatial data at each time step to capture spatial dependencies. The operation formula is: in, The weight matrix that the graph convolution module needs to learn; Graph convolution operations aggregate the features of adjacent nodes through an adjacency matrix, thereby achieving the propagation and fusion of spatial information. ③ Residual connection: The input and output of the spatiotemporal graph convolutional layer are added together to form a residual connection; 7.4) Time step traversal Since the dynamic adjacency matrix is ​​generated independently for each time step, the graph convolution operation needs to be performed separately for each time step. The specific process is as follows: ① Regarding time steps , extract the node features of this time step ,in The length of the historical time window; ② Input the dynamic graph generation module to generate the dynamic adjacency matrix for that time step. ; ③ Use a layered fusion strategy to... With static adjacency matrix Fusion, to obtain ; ④ Use Perform graph convolution to obtain the output at that time step. ; ⑤ Stack the outputs of all time steps to form a complete timing output; 7.5) Output Layer After multiple spatiotemporal graph convolutional layers, a fully connected layer is used to map the features to the output space of the prediction interval; each wind farm output Each value corresponds to The upper and lower bounds of the prediction intervals for each confidence level are denoted as follows: Since we only need to predict the next point in time, we take... Output of the last time step .

9. A multi-wind farm power range prediction system based on adaptive cooperative VMD and dynamic spatiotemporal graph convolution, characterized in that, This system is used to implement the multi-wind farm power interval prediction method based on adaptive cooperative VMD and dynamic spatiotemporal graph convolution as described in any one of claims 1 to 8. The system comprises a data preprocessing module, an adaptive multi-wind farm cooperative VMD signal decomposition module, a hybrid adjacency matrix construction module, a dynamic graph generation module, a spatiotemporal graph convolution modeling module, a progressive constraint loss function training module, and a multi-confidence level interval output module, wherein: The data preprocessing module is responsible for data acquisition and normalization. The adaptive multi-wind farm collaborative VMD signal decomposition module decomposes the power sequence of each wind farm into spatially aligned multi-frequency components through a joint optimization objective. The hybrid adjacency matrix construction module integrates geographical distance and power correlation to construct a static graph; The dynamic graph generation module learns an adaptive adjacency matrix through a graph neural network; The spatiotemporal graph convolutional modeling module extracts spatiotemporal features through a layered and fused graph structure; The progressive constraint loss function training module achieves automatic weight adjustment through conditional probability chain decomposition. The multi-confidence level interval output module generates prediction intervals for multiple confidence levels.

10. A computer-readable storage medium having a computer program stored thereon, characterized in that, When the program is executed by the processor, it implements the steps in the multi-wind farm power range prediction method based on adaptive cooperative VMD and dynamic spatiotemporal graph convolution as described in any one of claims 1 to 8.

Images