Wind power plant cluster power prediction method

Abstract

The invention discloses a wind power plant cluster power prediction method. The method comprises the following steps: step 1, embedding priori knowledge into a learning module; step 2, constructing a memory enhanced adaptive graph structure; step 3, an endogenous feature learning module; step 4, an exogenous feature learning module; 5, performing memory enhancement gating fusion on the MAGF; and step 6, outputting and predicting. According to the method, special learning modules are respectively constructed by decoupling endogenous power signals and exogenous meteorological factors, and a dynamic graph structure adaptive to a sample is retrieved by utilizing a meta-memory library, so that context sensing modeling of a complex space-time dependency relationship is realized. Meanwhile, a memory enhanced gating fusion mechanism (MAGF) is introduced, adaptive semantic alignment and fusion of double-flow features are realized, the precision and robustness of the model in multiple prediction time domains are remarkably improved, and more reliable technical support is provided for power grid dispatching and wind field operation.

Description

Technical Field

[0001] This invention relates to the fields of artificial intelligence and time series forecasting technology, and in particular to a method for predicting the power of a wind farm cluster. Background Technology

[0002] As wind power continues to account for an increasing proportion of the global energy mix, wind farm cluster power prediction (WFCPF) has become a key technology for ensuring stable grid operation, improving dispatch efficiency, and promoting the integration of new energy sources. However, wind power output is highly intermittent and uncertain, with complex spatial dynamic dependencies. It is strongly influenced by aerodynamic coupling effects such as wake effects and flow field shielding. At the same time, there are significant semantic heterogeneities and temporal inconsistencies between power series and exogenous variables such as meteorology, posing a great challenge to accurate modeling.

[0003] Currently, deep learning methods based on graph neural networks (GNN) and Transformers have been widely used in wind power prediction, which can capture spatiotemporal dependencies to a certain extent.

[0004] Existing wind farm cluster power prediction methods are mainly based on graph neural networks (GNNs) and Transformer architectures, aiming to capture the spatiotemporal dependencies between wind turbines. One typical approach relies on static prior graph structures, such as constructing spatial adjacency matrices through geographical distance or dynamic time warping (DTW) of historical power sequences to explicitly encode the spatial relationships between wind turbines, and then combining this with a time series modeling module for prediction. Another approach attempts to fuse multimodal exogenous variables (such as meteorological data), but its fusion strategies (such as direct splicing or simple attention mechanisms) fail to fully consider the semantic differences between endogenous power and exogenous environmental factors.

[0005] For example, some studies construct a geographic adjacency matrix between wind turbines using weighted geographic distance and extract spatial features using graph convolutional networks; others introduce dynamic clustering strategies, grouping samples based on temporal pattern similarity or delayed correlation, and constructing a specific graph structure for each group to improve the flexibility of spatial representation. However, existing prediction methods still have significant limitations in dynamic spatial dependency modeling and heterogeneous feature fusion.

[0006] Although these methods have improved the ability to represent the complex dynamics of wind farms to some extent, their spatial structure modeling still lacks sample-level adaptability, and the feature fusion mechanism has failed to achieve semantic alignment, which limits the accuracy and robustness of the model in longer prediction time domains and under complex meteorological conditions.

[0007] Existing methods have the following limitations:

[0008] 1. Static graph structures cannot capture dynamic spatial dependencies:

[0009] Most graph neural network-based methods rely on predefined static graph structures (such as geographic distance or historical sequence similarity), which cannot adapt to changes in the spatiotemporal context specific to the samples. The interactions between wind turbines are constantly evolving under the influence of meteorological conditions, wake effects, and turbulence changes, while existing methods lack the ability to model dynamic topologies, resulting in insufficient representation of spatial relationships in variable wind field environments.

[0010] 2. Heterogeneous feature fusion strategies suffer from semantic confusion:

[0011] Existing models typically concatenate endogenous power sequences with exogenous meteorological variables directly or fuse them using a unified attention mechanism, ignoring their fundamental differences in physical mechanisms, temporal patterns, and semantic levels. This homogenization can easily lead to information interference, weakening the model's ability to identify key driving factors, especially resulting in a significant decline in generalization performance under extreme weather or multimodal input scenarios.

[0012] 3. Dynamic graph learning methods are inefficient and lack generalization ability:

[0013] While some studies have attempted to construct dynamic graph structures through clustering or similarity grouping, these methods often rely on non-differentiable operations or complex grouping strategies. The training process frequently requires rebuilding the graph topology, resulting in high engineering complexity and difficulty in scaling. Furthermore, these methods cannot achieve end-to-end adaptive inference for samples, limiting their feasibility in large-scale applications in real-world wind fields.

[0014] These shortcomings make it difficult for existing prediction models to maintain robust performance in complex wind field layouts, variable weather conditions, and long-term prediction tasks. Therefore, there is an urgent need for a new prediction framework that can achieve dynamic structure awareness, semantic alignment fusion, and efficient context modeling. Summary of the Invention

[0015] In view of this, the present invention provides a wind farm cluster power prediction method to solve the problems of existing methods relying on static graph structures or simple feature splicing strategies, which make it difficult to accurately capture the dynamic interaction between wind turbines as weather and flow field change, and also cannot effectively handle the semantic heterogeneity and temporal mismatch between endogenous power and exogenous environmental variables, resulting in limited prediction accuracy and insufficient generalization ability.

[0016] To solve the above-mentioned technical problems, the present invention adopts the following technical solution:

[0017] A method for predicting the power of a wind farm cluster includes the following steps:

[0018] Step 1: Embedding prior knowledge into the learning module

[0019] The static topology and time-periodic information of the wind farm are encoded into a unified prior representation. A Gaussian kernel spatial adjacency matrix is ​​constructed based on the geographical coordinates of the wind turbines, and spatial embedding is generated through Node2Vec. At the same time, the time attributes are extracted and one-hot encoded and linearly projected to form a time embedding. The two are then added together after dimensional alignment by MLP to generate an embedding representation that integrates spatiotemporal priors.

[0020] Step 2, Construction of Memory-Enhanced Adaptive Graph Structure

[0021] Construct a shared memory matrix Where m represents the number of memory units, d m This represents the dimension of each memory vector; for each input sample, context-related node embeddings are retrieved from the memory bank through a hypernetwork projection mechanism, and two types of directed graph structures are dynamically generated accordingly;

[0022] Step 3, Endogenous Feature Learning Module

[0023] Variational mode decomposition (VMD) is performed on the wind power sequence to decompose the non-stationary signal into multiple intrinsic mode functions (IMFs) in order to extract multi-scale frequency domain features.

[0024] Step 4, Exogenous Feature Learning Module

[0025] Given the environment input and previously embedded Feature alignment and projection are performed, and an adaptive graph structure under exogenous conditions is constructed. And perform structure-aware query enhancement;

[0026] Step 5, Memory-enhanced gating fusion MAGF

[0027] We design a bidirectional cross-attention mechanism to realize semantic interaction between endogenous and exogenous feature streams, enhance cross-modal representation consistency, further introduce a learnable memory matrix for soft route matching, generate context-aware fusion weights, and finally achieve adaptive feature fusion through a gating mechanism.

[0028] Step 6: Output and Prediction

[0029] The fused features are processed by projection, residual connection and feedforward network, and the future multi-step wind power prediction value is generated through linear output layer.

[0030] Preferably, step 1 specifically includes:

[0031] (1) Spatial prior embedding: Based on the coordinates of the wind turbine generator, a spatial adjacency matrix is ​​constructed using a Gaussian kernel on the pairwise Euclidean distances.

[0032]

[0033] In the formula, d ij Let be the Euclidean distance between wind turbines i and j, σ be the bandwidth parameter, and ∈ = 0.8 be the sparsity threshold.

[0034] Then Node2Vec is applied to perform a biased random walk from A. geo Extract the structural context and generate spatial embeddings.

[0035]

[0036] Where, d s The embedding dimension is used to encode the topological semantics of the wind farm layout for downstream modeling.

[0037] (2) Time Prior Embedding: Time embedding is constructed based on calendar attributes, denoted as follows: and These categorical features are encoded using one-hot vectors and then concatenated:

[0038]

[0039] Where σ is the ReLU activation, C(·) is the dimensionality connection sign, and W is the trainable weight matrix. For the generated temporal embedding;

[0040] (3) Alignment and fusion of prior embeddings: MLPs integrate spatial embeddings and time embedding They are projected onto the same feature space, and then merged together by element-wise addition, as shown in the formula:

[0041]

[0042] in, D is the dimension of the unified embedding space used for downstream modeling, which incorporates priors. The spatial topology and temporal periodicity are encoded to provide global structural guidance for downstream modules.

[0043] Preferably, step 2 specifically includes:

[0044] Memory is projected through two independent paths:

[0045]

[0046] Among them, W e1 , These are learnable hypernetwork projection weights;

[0047] Perform L2 normalization on node embeddings n1 and n2, and calculate the similarity matrix:

[0048]

[0049] Here, ∈ is added for numerical stability;

[0050] To capture directionality and construct a directed graph structure of asymmetric wind turbine dependencies, a bidirectional adaptive adjacency matrix is ​​computed:

[0051]

[0052] Among them, τ controls the attention temperature, ReLU removes negative interactions, and softmax is calculated along the row dimension to ensure that the sum of each row is 1;

[0053] To ensure numerical stability and preserve node information, this invention addresses G... m1 and G m2 A self-loop was added, and it was re-normalized as follows:

[0054]

[0055] Where α = 0.1 is the self-loop connection weight parameter;

[0056] The final adaptive atlas is denoted as:

[0057]

[0059] Preferably, step 3 specifically includes:

[0060] Variational mode decomposition (VMD) is performed on the wind power sequence to decompose the non-stationary signal into multiple intrinsic mode functions (IMFs) to extract multi-scale frequency domain features.

[0061]

[0062] in, For generator n, the u-th IMF;

[0063] VMD formulates this decomposition as a constrained variational problem that jointly estimates the mode function {ρ} by minimizing the total bandwidth of the modulated analytic signal in the frequency domain. u} and its associated center frequency {ω u};

[0064] Specifically, the optimization objective is given by the following formula:

[0065]

[0066] Where δ(t) is the Dirac function, This represents the Hilbert kernel that converts a signal into its analytical form. It is ω u The frequency shifter centered on the frequency shifter, The time derivative of the bandwidth estimate;

[0067] The optimization is solved using the augmented Lagrange method, employing the alternating direction method of the multiplier.

[0068] Let the decomposition result be

[0069] To fully integrate information and strengthen knowledge connections, MLP is used for X. VMD Perform feature mapping to associate it with pre-built spatiotemporal prior embedding knowledge. Organic combination, thereby effectively enhancing X VMD Its own static spatiotemporal prior characteristics;

[0070] Final fusion input

[0071]

[0072] Where Φ(·) is the dimension expansion symbol.

[0073] This study utilizes a frequency-aware self-attention mechanism and adaptive graph convolution (AGCN) to model the dynamic correlations between different wind turbine nodes under different frequency modes, enhancing the discriminative power of spatiotemporal representation. The objective of the l-th encoder layer is to model inter-frequency correlations using a graph-guided multi-head frequency attention mechanism. Given the input... First, the query, key, and value matrix is ​​obtained through non-linear projection:

[0074]

[0075] Where φ(·) is a nonlinear projection function, its expression is:

[0076] φ(x)=Gelu(Wx+b)

[0077] Where W and b are learnable parameter matrices, and Gelu is the activation function;

[0078] Using graph generator f adp (·) Construct a frequency-adaptive adjacency matrix And enhance Q using an Adaptive Graph Convolutional Network (AGCN):

[0079]

[0080] AGCN uses multi-order graph convolutions to capture frequency propagation across different neighborhoods and bidirectional node dependencies;

[0081] Then, the attention weight between nodes i and v of head k is calculated using the dot product as follows:

[0082]

[0083] in, and K i (k,l) Let V represent the query vector and key vector, and let V represent the values ​​of nodes v and i at level l. Let j∈V be a normalized index for all candidate nodes, and <·, ·> are inner product operators.

[0084] Note the coefficients. The effect of the kth attention head node i on node v is measured at the frequency level.

[0085] Next, the value vectors are weighted and summed to obtain the output for each attention head:

[0086]

[0087] After aggregating all attention heads, the final layer output of the intrinsic feature learning module is... The results were obtained through splicing and fusion calculations:

[0088]

[0089] Preferably, step 4 specifically includes:

[0090] Given the environment input and previously embedded Feature alignment and projection are performed using the following formula:

[0091]

[0092] And construct an adaptive graph structure under exogenous conditions. via f adp (·), and perform structure-aware query enhancement:

[0093]

[0094] Employing a multi-head graph attention mechanism symmetric to the endogenous module, scaled dot product attention is computed:

[0095]

[0096] The attention output of each head is:

[0097]

[0098] Then all the heads are aggregated and projected through a fusion layer:

[0099]

[0100] This encoder enables the model to display the spatial coordination relationships between captured meteorological factors and their transvariable impact on wind turbine status.

[0101] Preferably, the specific formula for step 5 is as follows:

[0102]

[0103] Where, d c The hidden dimension of the query vector used for attention scaling;

[0104] Conversely, exogenous features act as queries, while the focus is on endogenous features:

[0105]

[0106] The resulting interactive participation representation and An enhanced interaction between endogenous and exogenous features was captured;

[0107] They are integrated into their respective streams through residual connections and layer normalization:

[0108]

[0109]

[0110] We then introduce a learnable external memory matrix. Will and Mapping to route query vector and

[0111] Then, use the memory matrix. The attention score is calculated as follows:

[0112]

[0113] in, and They represent and The j-th memory unit, Represents the global memory matrix The j-th memory item in;

[0114] Then based on respectively and Weighted retrieval of memories yields memory routing response representations for both intrinsic and extrinsic channels. and

[0115]

[0116] The two sets of routing responses are concatenated and fed into a fusion gate network to generate adaptive gate weights, controlling the fusion ratio of endogenous and exogenous traffic.

[0117] z=σ(FC(C(r en ,r ex )))

[0118] Where σ is the sigmoid activation function. FC(·) is a fully connected layer;

[0119] The final fusion is represented as:

[0120]

[0121] Here, ⊙ represents element-wise multiplication.

[0122] Preferably, the specific formula for step 6 is as follows:

[0123]

[0124]

[0125] Among them, W pj and b pj It is a learnable projection parameter matrix and bias, σ is the GELU activation function, and X is the original embedding of wind power.

[0126] The present invention achieves the following technical effects compared to the prior art:

[0127] (1) Overall improvement in prediction accuracy: On multiple real datasets (SDWPF, WSTD2) and different prediction time domains (1h to 18h), MAMNET outperforms various existing benchmark models (including Transformer, GNN and special exogenous variable models) under the experimental conditions shown. Its advantages are more significant in long-term prediction, and it significantly reduces RMSE and MAE errors.

[0128] (2) Effective modeling of dynamic spatial dependence: Through the memory-enhanced graph structure construction module, it is possible to adaptively generate context-aware dynamic graph topology for different samples, which overcomes the problem that traditional static graph methods cannot accurately capture the dynamic evolution of spatiotemporal relationships caused by wakes, turbulence, etc.

[0129] (3) Heterogeneous feature semantic alignment fusion: The modular dual-stream design explicitly decouples endogenous power and exogenous environmental variables, and combines a memory-guided gating fusion mechanism to achieve cross-modal semantic alignment, effectively avoiding information confusion and improving the discriminative ability of feature expression;

[0130] (4) Good generalization and robustness: The framework does not depend on fixed wind field layout or specific meteorological models. It can adapt to the changing external environment and complex wind field structure. It shows stable performance on datasets with different spatiotemporal resolutions and has strong engineering application potential. Attached Figure Description

[0131] Figure 1 This is a flowchart of the present invention. Detailed Implementation

[0132] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0133] This invention discloses a method for predicting the power of a wind farm cluster, comprising the following steps:

[0134] Step 1: Embedding prior knowledge into the learning module

[0135] A Gaussian kernel spatial adjacency matrix is ​​constructed based on the geographical coordinates of the wind turbines, and spatial embeddings are generated using Node2Vec. Simultaneously, time attributes (hour, day, year) are extracted and one-hot encoded and linearly projected to form temporal embeddings. The two embeddings are then added together after dimension alignment using MLP to generate an embedding representation that integrates spatiotemporal priors, providing structural guidance for subsequent modules.

[0136] (1) Spatial prior embedding: Based on the coordinates of the wind turbine generator, a spatial adjacency matrix is ​​constructed using a Gaussian kernel on the pairwise Euclidean distances.

[0137]

[0138] In the formula, d ij Let be the Euclidean distance between wind turbines i and j, σ be the bandwidth parameter, and ∈ = 0.8 be the sparsity threshold.

[0139] Then Node2Vec is applied to perform a biased random walk from A. geo Extract the structural context and generate spatial embeddings.

[0140]

[0141] Where, d s The embedding dimension is used to encode the topological semantics of the wind farm layout for downstream modeling.

[0142] (2) Time Prior Embedding: Time embedding is constructed based on calendar attributes (hour / day, weekday / day, year / day), denoted as _____. and These categorical features are encoded using one-hot vectors and then concatenated:

[0143]

[0144] Where σ is the ReLU activation, C(·) is the dimensionality connection sign, and W is the trainable weight matrix.

[0145] For the generated time embedding.

[0146] (3) Alignment and fusion of prior embeddings: In order to form a unified spatiotemporal prior, this invention uses separate MLPs to embed the spatial priors. and time embedding They are projected onto the same feature space and then merged together by element-wise addition; the formula is:

[0147]

[0148] in, D is the dimension of the unified embedding space used for downstream modeling, which incorporates priors. The spatial topology and temporal periodicity are encoded to provide global structural guidance for downstream modules.

[0149] Step 2, Construction of Memory-Enhanced Adaptive Graph Structure

[0150] First, construct a shared memory matrix. Where m represents the number of memory units, d m This represents the dimension of each memory vector. This memory is used to store typical dependency structures between wind turbines, frequency components, or input features, thereby enabling the reuse of structural knowledge across samples.

[0151] For each input sample, context-related node embeddings are retrieved from the memory using a hypernetwork projection mechanism, and two types of directed graph structures are dynamically generated accordingly.

[0152] Memory is projected through two independent paths:

[0153]

[0154] in, These are learnable hypernetwork projection weights. This dual-view projection allows nodes to access memories from complementary semantic perspectives.

[0155] Perform L2 normalization on node embeddings n1 and n2, and calculate the similarity matrix:

[0156]

[0157] Here, ∈ is added for numerical stability. To capture directionality and construct a directed graph structure of asymmetric wind turbine dependencies, a bidirectional adaptive adjacency matrix is ​​computed:

[0158]

[0159] In this method, τ controls the attention temperature, ReLU removes negative interactions, and softmax is calculated along the row dimension to ensure that the sum of each row is 1. Furthermore, to ensure numerical stability and preserve node information, this invention uses G... m1 and G m2 A self-loop was added, and it was re-normalized as follows:

[0160]

[0161] Where α = 0.1 is the self-loop connection weight parameter. The final adaptive graph atlas is denoted as:

[0162]

[0163] This mechanism enables flexible and fine-grained graph modeling under specific input conditions of the samples, and it has been applied to both endogenous and exogenous feature learning modules.

[0164] Step 3, Endogenous Feature Learning Module

[0165] Variational mode decomposition (VMD) is performed on the wind power sequence to decompose the non-stationary signal into multiple intrinsic mode functions (IMFs) to extract multi-scale frequency domain features.

[0166]

[0167] in, Let u be the IMF of generator n.

[0168] VMD formulates this decomposition as a constrained variational problem that jointly estimates the mode function {ρ} by minimizing the total bandwidth of the modulated analytic signal in the frequency domain. u} and its associated center frequency {ω u Specifically, the optimization objective is given by the following formula:

[0169]

[0170] Where δ(t) is the Dirac function, This represents the Hilbert kernel that converts a signal into its analytical form. It is ω u The frequency shifter centered on the frequency shifter, This represents the time derivative of the bandwidth estimate.

[0171] The optimization is solved using the augmented Lagrange method, employing an alternating direction method with multipliers to ensure effective convergence and spectral separation. Let the decomposition result be...

[0172] To fully integrate information and strengthen knowledge connections, MLP is used for X. VMD Perform feature mapping to associate it with pre-built spatiotemporal prior embedding knowledge. Organic combination, thereby effectively enhancing X VMD Its own static spatiotemporal prior features. The final fusion input. for:

[0173]

[0174] Where Φ(·) is the dimension expansion symbol.

[0175] This study utilizes a frequency-aware self-attention mechanism and adaptive graph convolution (AGCN) to model the dynamic correlations between different wind turbine nodes under different frequency modes, enhancing the discriminative power of the spatiotemporal representation. The goal of the l-th encoder layer is to model inter-frequency correlations using a graph-guided multi-head frequency attention mechanism. Given input... First, the query, key, and value matrix is ​​obtained through non-linear projection:

[0176]

[0177] Where φ(·) is a nonlinear projection function, its expression is:

[0178] φ(x)=Gelu(Wx+b)

[0179] Where W and b are learnable parameter matrices, and Gelu is the activation function.

[0180] Using graph generator f adp (·) Construct a frequency-adaptive adjacency matrix And enhance Q using an Adaptive Graph Convolutional Network (AGCN):

[0181]

[0182] AGCN uses multi-order graph convolutions to capture frequency propagation across different neighborhoods and bidirectional node dependencies.

[0183] Then, the attention weight between nodes i and v of head k is calculated using the dot product as follows:

[0184]

[0185] in, and Let V represent the query vector and key vector, and let V represent the values ​​of nodes v and i at level l. Let j∈V be a normalized index for all candidate nodes, and <·, ·> are inner product operators.

[0186] Note the coefficients. The influence of node i on node v under the k-th attention head at the frequency level is measured. Next, the value vectors are weighted and summed to obtain the output of each attention head:

[0187]

[0188] After aggregating all attention heads, the final layer output of the intrinsic feature learning module is... The results were obtained through splicing and fusion calculations:

[0189]

[0190] Step 4, Exogenous Feature Learning Module

[0191] Given the environment input and previously embedded Feature alignment and projection are performed using the following formula:

[0192]

[0193] And construct an adaptive graph structure under exogenous conditions. via f adp (·), and perform structure-aware query enhancement:

[0194]

[0195] Employing a multi-head graph attention mechanism symmetric to the endogenous module, scaled dot product attention is computed:

[0196]

[0197] The attention output of each head is:

[0198]

[0199] Then all the heads are aggregated and projected through a fusion layer:

[0200]

[0201] This encoder enables the model to explicitly capture the spatial coordination relationships between meteorological factors and their transvariable impact on wind turbine status.

[0202] Step 5, Memory-Enhanced Gated Fusion (MAGF)

[0203] A bidirectional cross-attention mechanism is designed to achieve semantic interaction between endogenous and exogenous feature streams, enhancing cross-modal representation consistency. Furthermore, a learnable memory matrix is ​​introduced for soft route matching, generating context-aware fusion weights. Finally, an adaptive feature fusion is achieved through a gating mechanism. The specific formula is as follows:

[0204]

[0205] Where, d c This is the hidden dimension of the query vector used for attention scaling. Conversely, exogenous features act as the query, focusing on endogenous features:

[0206]

[0207] The resulting interactive participation representation and Enhanced interactions between endogenous and exogenous features were captured. They are integrated into their respective streams via residual connections and layer normalization:

[0208]

[0209] We then introduce a learnable external memory matrix. Will and Mapping to route query vector and Then, use the memory matrix. The attention score is calculated as follows:

[0210]

[0211]

[0212] in, and They represent and The j-th memory unit. Represents the global memory matrix The j-th memory item in the memory.

[0213] Then based on respectively and Weighted retrieval of memories yields memory routing response representations for both intrinsic and extrinsic channels. and

[0214]

[0215] The two sets of routing responses are concatenated and fed into a fusion gate network to generate adaptive gate weights, controlling the fusion ratio of endogenous and exogenous traffic.

[0216] z=σ(FC(C(r en ,r ex )))

[0217] Where σ is the sigmoid activation function. FC(·) is a fully connected layer. The final fusion is represented as:

[0218]

[0219] Here, ⊙ represents element-wise multiplication.

[0220] Step 6: Output and Prediction

[0221] The fused features are processed by projection, residual connection, and feedforward network, and future multi-step wind power prediction values ​​are generated through a linear output layer. The specific formula is as follows:

[0222]

[0223] Among them, W pj and b pj It represents the learnable projection parameter matrix and bias. σ is the GELU activation function. X is the original embedding of wind power.

[0224] Example 1:

[0225] 1. Experimental conditions

[0226] This invention was conducted on a hardware platform equipped with an NVIDIA RTX 3090 GPU (24GB VRAM), using the PyTorch deep learning framework for model training and inference. The experimental environment included CUDA 11.7 and Python 3.9, utilizing mixed precision training (AMP) and gradient pruning techniques to improve training efficiency and stability.

[0227] The methods compared in the experiment are as follows:

[0228] ITransformer (M1): Based on the feature-inverted Transformer architecture, focusing on variable-level dependency modeling;

[0229] Informer (M2): Employs the ProbSparse self-attention mechanism, suitable for long sequence prediction tasks;

[0230] FEDformer (M3): A Transformer variant that combines frequency enhancement and decomposition structures;

[0231] PatchTST(M4): A channel-independent time-series patch prediction model;

[0232] TIMEXER(M5): A time-series Transformer model that supports exogenous variables;

[0233] TESTAM (M6): A hybrid architecture that integrates time experts and static / dynamic graph models;

[0234] WeaGAN(M7): A graph attention network for weather perception.

[0235] MAGNN (M8): Multi-attention map neural network, enhancing spatial relationship representation;

[0236] 2. Experimental Dataset

[0237] The experiment uses two publicly available wind farm cluster datasets:

[0238] The SDWPF dataset contains sampled data from 134 wind turbines over 245 days, with a time resolution of 10 minutes. Features include power, wind speed, wind direction, and temperature.

[0239] WSTD2 dataset: Contains hourly power and wind speed data for 200 wind turbines from 2010 to 2012.

[0240] During preprocessing, missing values ​​were linearly imputed, negative power values ​​were set to zero, and all features were Z-score standardized. The training, validation, and test sets were divided in a 7:1:2 ratio, and performance was evaluated using multi-step prediction tasks (1h / 6h / 12h / 18h).

[0241] 3. Experimental Evaluation Indicators

[0242] The normalized root mean square error (RMSE) and mean absolute error (MAE) were used as evaluation metrics. The model hyperparameters included: number of memory units M = 20, number of VMD decomposition modes U = 6, number of graph attention heads K = 4, self-connection weights in the adaptive graph α = 0.1, the Adam optimizer was used, the initial learning rate was 0.001, and an exponential decay strategy was adopted.

[0243] 4. Comparative Experiment

[0244] To verify the effectiveness of the MAMNET framework, this invention conducted a comprehensive comparative experiment with eight representative baseline models on two real-world wind farm cluster datasets, SDWPF and WSTD2, with prediction ranges covering power output for the next 1, 6, 12, and 18 hours. The combined results of multiple evaluation metrics (RMSE, MAE) are shown in Table 1.

[0245] Table 1: Prediction performance of different models on two datasets at different prediction step sizes.

[0246]

[0247]

[0248] Experimental results show that MAMNET consistently outperforms all comparison models across all prediction time steps and on both datasets. On the SDWPF dataset, in the 18-hour prediction task, MAMNET reduces MAE by approximately 30%-35% and RMSE by over 25% compared to TimeXer and WeaGAN models, which focus on exogenous variable modeling; compared to the graph neural network baselines TESTAM and MAGNN, it reduces RMSE by over 45%. In short- to medium-term predictions (6h and 12h), MAMNET also significantly outperforms, with its MAE improving by an average of over 35% compared to WeaGAN. A similar consistent trend is observed on the WSTD2 dataset, where MAMNET reduces MAE and RMSE by approximately 20%-30% and over 15% compared to TimeXer and WeaGAN, respectively, in 1-hour predictions, maintaining a stable advantage in long-term predictions, with an overall error reduction of approximately 8%.

[0249] Analysis of the performance of different model categories: Transformer-type models (such as ITransformer, Informer) perform reasonably well in short-term predictions, but in long-term predictions, due to the lack of explicit spatial modeling capabilities, their average RMSE is more than 40% higher than MAMNET; Graph neural network models (such as TESTAM, MAGNN) can capture some spatial relationships, but are limited by static graph structures and have difficulty adapting to dynamically changing field topologies, resulting in a significant amplification of their prediction errors at 12h and 18h; Exogenous variable-specific models (such as TimeXer, WeaGAN) show improved performance after introducing meteorological factors, but their prediction accuracy and robustness are still limited due to the lack of explicit modeling of feature semantic heterogeneity and dynamic structures.

[0250] MAMNET’s superior performance is attributed to its modular feature decoupling structure, memory-enhanced dynamic graph construction, and semantically aware gated fusion mechanism, which enables it to effectively capture complex spatiotemporal dependencies and maintain high accuracy and strong generalization ability in variable wind power scenarios.

[0251] The above description is merely a preferred embodiment of the present invention and does not constitute any limitation on the technical scope of the present invention. Therefore, any minor modifications, equivalent changes, and alterations made to the above embodiments based on the technical essence of the present invention shall still fall within the scope of the technical solution of the present invention.

Claims

1. A method for predicting the power of a wind farm cluster, characterized in that, Includes the following steps: Step 1: Embedding prior knowledge into the learning module The static topology and time-periodic information of the wind farm are encoded into a unified prior representation. A Gaussian kernel spatial adjacency matrix is ​​constructed based on the geographical coordinates of the wind turbines, and spatial embedding is generated through Node2Vec. At the same time, the time attributes are extracted and one-hot encoded and linearly projected to form a time embedding. The two are then added together after dimensional alignment by MLP to generate an embedding representation that integrates spatiotemporal priors. Step 2, Construction of Memory-Enhanced Adaptive Graph Structure Construct a shared memory matrix Where m represents the number of memory units, d m This represents the dimension of each memory vector; for each input sample, context-related node embeddings are retrieved from the memory bank through a hypernetwork projection mechanism, and two types of directed graph structures are dynamically generated accordingly; Step 3, Endogenous Feature Learning Module Variational mode decomposition (VMD) is performed on the wind power sequence to decompose the non-stationary signal into multiple intrinsic mode functions (IMFs) in order to extract multi-scale frequency domain features. Step 4, Exogenous Feature Learning Module Given the environment input and previously embedded Feature alignment and projection are performed, and an adaptive graph structure under exogenous conditions is constructed. via f adp (·), and perform structure-aware query enhancement; Step 5, Memory-enhanced gating fusion MAGF We design a bidirectional cross-attention mechanism to realize semantic interaction between endogenous and exogenous feature streams, enhance cross-modal representation consistency, further introduce a learnable memory matrix for soft route matching, generate context-aware fusion weights, and finally achieve adaptive feature fusion through a gating mechanism. Step 6: Output and Prediction The fused features are processed by projection, residual connection and feedforward network, and the future multi-step wind power prediction value is generated through linear output layer.

2. The method for predicting the power of a wind farm cluster according to claim 1, characterized in that, Step 1 specifically includes: (1) Spatial prior embedding: Based on the coordinates of the wind turbine generator, a spatial adjacency matrix is ​​constructed using a Gaussian kernel on the pairwise Euclidean distances. In the formula, d ij Let be the Euclidean distance between wind turbines i and j, σ be the bandwidth parameter, and ∈ = 0.8 be the sparsity threshold. Then Node2Vec is applied to perform a biased random walk from A. geo Extract the structural context and generate spatial embeddings. Where, d s To increase the embedding dimension, these embeddings encode the topological semantics of the wind farm layout for downstream modeling. (2) Time Prior Embedding: Time embedding is constructed based on calendar attributes, denoted as follows: and These categorical features are encoded using one-hot vectors and then concatenated: Where σ is the ReLU activation, C(·) is the dimensionality connection sign, and W is the trainable weight matrix. For the generated temporal embedding; (3) Alignment and fusion of prior embeddings: MLPs integrate spatial embeddings and time embedding They are projected onto the same feature space, and then merged together by element-wise addition, as shown in the formula: in, D is the dimension of the unified embedding space used for downstream modeling, which incorporates priors. The spatial topology and temporal periodicity are encoded to provide global structural guidance for downstream modules.

3. The method for predicting the power of a wind farm cluster according to claim 1, characterized in that, Step 2 specifically includes: Memory is projected through two independent paths: in, These are learnable hypernetwork projection weights; Perform L2 normalization on node embeddings n1 and n2, and calculate the similarity matrix: Here, ∈ is added for numerical stability; To capture directionality and construct a directed graph structure of asymmetric wind turbine dependencies, a bidirectional adaptive adjacency matrix is ​​computed: Among them, τ controls the attention temperature, ReLU removes negative interactions, and softmax is calculated along the row dimension to ensure that the sum of each row is 1; To ensure numerical stability and preserve node information, this invention addresses G... m1 and G m2 A self-loop was added, and it was re-normalized as follows: G′ m1 =G m1 +αI,G′ m2 =G m2 +αI Where α = 0.1 is the self-loop connection weight parameter; The final adaptive atlas is denoted as:

4. The method for predicting the power of a wind farm cluster according to claim 1, characterized in that, Step 3 specifically includes: Variational mode decomposition (VMD) is performed on the wind power sequence to decompose the non-stationary signal into multiple intrinsic mode functions (IMFs) to extract multi-scale frequency domain features. in, For generator n, the u-th IMF; VMD formulates this decomposition as a constrained variational problem that jointly estimates the mode function {ρ} by minimizing the total bandwidth of the modulated analytic signal in the frequency domain. u } and its associated center frequency {ω u }; Specifically, the optimization objective is given by the following formula: Where δ(t) is the Dirac function, This represents the Hilbert kernel that converts a signal into its analytical form. It is ω u The frequency shifter centered on the frequency shifter, The time derivative of the bandwidth estimate; The optimization is solved using the augmented Lagrange method, employing the alternating direction method of the multiplier. Let the decomposition result be To fully integrate information and strengthen knowledge connections, MLP is used for X. VMD Perform feature mapping to associate it with pre-built spatiotemporal prior embedding knowledge. Organic combination, thereby effectively enhancing X VMD Its own static spatiotemporal prior characteristics; Final fusion input for: Where Φ(·) is the dimension expansion symbol; This study utilizes a frequency-aware self-attention mechanism and adaptive graph convolution (AGCN) to model the dynamic correlations between different wind turbine nodes under different frequency modes, enhancing the discriminative power of the spatiotemporal representation. The objective of the l-th encoder layer is to model inter-frequency correlations using a graph-guided multi-head frequency attention mechanism. Given the input... First, the query, key, and value matrix is ​​obtained through non-linear projection: Where φ(·) is a nonlinear projection function, its expression is: φ(x)=Gelu(Wx+b) Where W and b are learnable parameter matrices, and Gelu is the activation function; Using graph generator f adp (·) Construct a frequency-adaptive adjacency matrix And enhance Q using an Adaptive Graph Convolutional Network (AGCN): AGCN uses multi-order graph convolutions to capture frequency propagation across different neighborhoods and bidirectional node dependencies; Then, the attention weight between nodes i and v of head k is calculated using the dot product as follows: in, and Let V represent the query vector and key vector, and let V represent the values ​​of nodes v and i at level l. Let j∈V be a normalized index for all candidate nodes, and <·, ·> are inner product operators. Note the coefficients. The effect of the kth attention head node i on node v is measured at the frequency level. Next, the value vectors are weighted and summed to obtain the output for each attention head: After aggregating all attention heads, the final layer output of the intrinsic feature learning module is... The results were obtained through splicing and fusion calculations:

5. The method for predicting the power of a wind farm cluster according to claim 1, characterized in that, Step 4 specifically includes: Given the environment input and previously embedded Feature alignment and projection are performed using the following formula: And construct an adaptive graph structure under exogenous conditions. via f adp (·), and perform structure-aware query enhancement: Employing a multi-head graph attention mechanism symmetric to the endogenous module, scaled dot product attention is computed: The attention output of each head is: Then all the heads are aggregated and projected through a fusion layer: This encoder enables the model to display the spatial coordination relationships between captured meteorological factors and their transvariable impact on wind turbine status.

6. The method for predicting the power of a wind farm cluster according to claim 1, characterized in that, The specific formula for step 5 is as follows: Where, d c The hidden dimension of the query vector used for attention scaling; Conversely, exogenous features act as queries, while the focus is on endogenous features: The resulting interactive participation representation and An enhanced interaction between endogenous and exogenous features was captured; They are integrated into their respective streams through residual connections and layer normalization: We then introduce a learnable external memory matrix. Will and Mapping to route query vector and Then, use the memory matrix. The attention score is calculated as follows: in, and They represent and The j-th memory unit, Represents the global memory matrix The j-th memory item in; Then based on respectively and Weighted retrieval of memories yields memory routing response representations for both intrinsic and extrinsic channels. and The two sets of routing responses are concatenated and fed into a fusion gate network to generate adaptive gate weights, controlling the fusion ratio of endogenous and exogenous traffic. z=σ(FC(C(r en ,r ex ))) Where σ is the sigmoid activation function, and FC(·) is a fully connected layer; The final fusion is represented as: Here, ⊙ represents element-wise multiplication.

7. The method for predicting the power of a wind farm cluster according to claim 1, characterized in that, The specific formula for step 6 is as follows: Among them, W pj and b pj It is a learnable projection parameter matrix and bias, σ is the GELU activation function, and X is the original embedding of wind power.

Images