A method for ultra-short-term wind power prediction based on a multi-graph diffusion Mamba model

Abstract

This invention provides an ultra-short-term wind power prediction method based on a multi-graph diffusion Mamba model, belonging to the field of wind power prediction technology. The method includes: constructing a multi-graph architecture to decouple the spatiotemporal dependencies between wind turbines into a meteorological causal graph and a system adaptive graph; introducing a dynamic gating fusion mechanism to adaptively fuse the two graph structures through real-time confidence factors; inputting the fused graph structure into a topology-aware Mamba module, embedding graph diffusion convolution into a selective scanning process, synchronously updating node states along the graph topology, and jointly extracting spatial dependencies and temporal dynamic features to achieve end-to-end spatiotemporal joint modeling. Experimental results show that, compared to the baseline model, this invention improves prediction accuracy in ultra-short-term power prediction tasks ranging from 15 minutes to 4 hours, and possesses good model interpretability and adaptability to different operating scenarios.

Description

Technical Field

[0001] This invention relates to the field of wind power prediction technology, specifically to an ultra-short-term wind power prediction method based on a multi-graph diffusion Mamba model. Background Technology

[0002] Wind power exhibits significant intermittency, volatility, and uncertainty, making accurate forecasting crucial for ensuring the safe and economical operation of the power system. With the increasing scale and clustering of wind farms, the wake effect between adjacent turbines intensifies, and the spatiotemporal coupling relationship of wind energy resources becomes increasingly complex, significantly increasing the difficulty of ultra-short-term power forecasting. As a critical front-end component of the energy management system (EMS), the accuracy of ultra-short-term wind power forecasting directly determines the reliability of dispatch decisions and the clearing efficiency of the electricity market. Currently, although data-driven forecasting models have made significant progress, existing ultra-short-term wind power forecasting methods still have the following shortcomings when facing complex and variable meteorological processes: 1. Limitations in Spatial Topology Construction. Traditional graph neural network methods typically construct a single topology to describe the relationships between wind turbines. On the one hand, purely data-driven topologies (such as graphs built based on correlation or attention mechanisms) lack physical boundaries, easily leading to overfitting and mislearning artificial power rationing or sensor noise as spatial dependencies. On the other hand, topologies based on physical information (such as graphs built based on wind turbine geographical location or wind direction), while encoding prior information about wind energy resources, have a fixed structure and cannot adapt to time-varying wind directions or operational constraints that are difficult to predict in advance. Existing technologies lack an algorithmic mechanism that can dynamically balance aerodynamic mechanisms and system constraints; 2. The spatiotemporal modeling architecture suffers from coupling defects. Most mainstream methods adopt a cascaded architecture, performing time series processing after completing spatial graph modeling. This separate processing approach is not conducive to the co-evolution of spatiotemporal features, and is prone to error accumulation and memory decay in multi-step prediction, making it difficult to effectively capture the inherent spatiotemporal coupling dynamic trends of wind farms.

[0003] Based on the current state of research, engineering practice requires a wind power prediction technology that can integrate physical mechanisms and data-driven advantages to achieve spatiotemporal joint modeling, in order to overcome the aforementioned shortcomings. Summary of the Invention

[0004] This invention provides an ultra-short-term wind power prediction method based on a multi-graph diffusion Mamba model. By constructing a dual-graph parallel architecture, a dynamic gating fusion mechanism, and a topology-aware Mamba module, it achieves adaptive fusion and end-to-end joint modeling of physical mechanisms and data-driven approaches, effectively improving prediction accuracy and model interpretability. The technical solution includes the following steps: 1. Obtain historical operating data and meteorological data of multiple wind turbine generators in the target wind farm, including active power, wind speed, wind direction, temperature, air pressure and humidity; preprocess the historical operating data and meteorological data, including outlier removal and normalization; use empirical wavelet transform to adaptively decompose the original power sequence and extract multiple time-frequency domain feature components. 2. Construct a multi-graph spatial topology, wherein the multi-graph spatial topology includes: Physics-driven meteorological causal graph. Based on the transfer entropy, the causal correlation strength between different wind turbine generators is calculated, and a directed weighted adjacency matrix is ​​constructed. Then, based on real-time wind direction information, the corresponding adjacency matrix is ​​dynamically selected from multiple pre-built wind direction state space causal matrices. A data-driven adaptive graph for the system. A learnable node embedding matrix is ​​initialized, and an adaptive adjacency matrix is ​​generated by performing an inner product operation on this matrix and passing it through an activation function to capture implicit operational constraints. 3. Input the multi-scale modal features and historical power data into the meteorological causal graph and the system adaptive graph, respectively, and perform graph diffusion convolution operation to extract spatial dependency features; the graph diffusion convolution adopts a bidirectional random walk diffusion process to aggregate the information flow of upstream and downstream nodes; 4. Through a dynamic gating fusion mechanism, gating factors are generated based on real-time operating status (including wind speed, wind direction, and power level). The first spatial features extracted from the meteorological causal graph and the second spatial features extracted from the system adaptive graph are adaptively weighted and fused to obtain fused spatial features. 5. The fused spatial features are input into the topology-aware Mamba module, which is built based on a structured state space model. Its control matrix is ​​dynamically generated by graph diffusion convolution operation based on the fused spatial features, thereby realizing the temporal state update and neighborhood spatial information aggregation at each time step and outputting spatiotemporal joint features. 6. Based on the aforementioned spatiotemporal joint features, an encoder-decoder architecture is adopted, and the wind power prediction results for multiple future time points are generated recursively through the decoder.

[0005] The technical effects of this invention are as follows: 1. This invention decouples aerodynamic mechanisms from data-driven implicit patterns by constructing a physics-driven meteorological causal graph and a learnable system adaptive graph. Through adaptive fusion using a dynamic gating mechanism, it effectively avoids the overfitting and structural rigidity problems of a single topology. 2. This invention embeds graph-diffused convolution into Mamba's selective state-space model, enabling the model to simultaneously aggregate spatial neighborhood information during state updates at each time step, achieving collaborative modeling of spatiotemporal information. This topology-aware selective scanning mechanism effectively suppresses error accumulation and memory decay in sequence prediction, making it suitable for wind power prediction scenarios with highly non-stationary and multi-scale fluctuations. Simultaneously, the linear time complexity of the Mamba module ensures model efficiency during the inference phase, making it suitable for real-time deployment. 3. The dynamic gating factor designed in this invention has physical interpretability. Under stable weather conditions, its value approaches 1, indicating that the model preferentially relies on the physical causal graph; under extreme weather or active power grid curtailment scenarios, the value of the gating factor decreases, indicating that the model relies on a data-driven adaptive graph to avoid the failure of the physical model, thus enhancing the reliability of the prediction system. 4. By introducing a robust task loss based on Huber loss and a structure regularization based on the Frobenius norm, this invention effectively suppresses the interference of outliers and noise during training. Attached Figure Description

[0006] Figure 1 This is an overall framework diagram of the present invention; Figure 2 This is a schematic diagram of the multi-graph architecture of the present invention; Figure 3 This is a conceptual schematic diagram of the graph diffusion convolution operation of the present invention; Figure 4 This is a schematic diagram of the gating fusion mechanism of the present invention; Figure 5 This is a schematic diagram of the topology-aware Mamba module of the present invention; Figure 6 This is a schematic diagram of the encoder-decoder (Seq2Seq) prediction architecture of the present invention; Figure 7 This is a curve comparison chart of the embodiment of the present invention and the existing benchmark model in the 1-hour prediction time domain; Figure 8 This is a comparison chart of ablation experiment results in the embodiments of the present invention; Figure 9 The response curve of the dynamic gating factor under special operating conditions; Figure 10 This is a graph showing the sensitivity analysis of parameters. Detailed Implementation

[0008] Example 1

[0009] This embodiment verifies the effectiveness of the multi-graph diffusion Mamba model on the ultra-short-term wind power prediction task through comparative experiments. The experimental data is derived from four wind turbines (FD1-FD4) of an onshore wind farm, with a time resolution of 15 minutes and a time span from January 1, 2019 to December 31, 2020. Model inputs include historical power and meteorological variables such as wind speed, wind direction, temperature, air pressure, and humidity. The data is divided into eight subsets by season, and each subset is split into a training set, a validation set, and a test set at a ratio of 70%:15%:15%. The lookback window is set to 24 (corresponding to 6 hours), and the prediction step size is set to {1,4,8,12,16}, corresponding to prediction durations from 15 minutes to 4 hours.

[0010] The comparative models include Temporal Convolutional Network (TCN), Fully Connected Long Short-Term Memory Network (FC-LSTM), Spatio-Temporal Graph Convolutional Network (STGCN), Attention-based Spatio-Temporal Graph Convolutional Network (ASTGCN), and Graph Wavelet Network (GWNet), covering three paradigms of pure time series, static spatio-temporal graph, and adaptive graph network. All models are trained on the same training set and evaluated on the same test set, with evaluation metrics being Normalized Mean Absolute Error (NMAE), Normalized Root Mean Square Error (NRMSE), and Coefficient of Determination (R²).

[0011] The experimental results show that the model achieves optimal performance at all prediction step sizes. As shown in Table 1, in the 15-minute prediction task, the NRMSE of the model is 0.0253, the NMAE is 0.0183, and the R² is 0.941; under the 4-hour prediction horizon, the NRMSE is 0.1527, the NMAE is 0.1213, and the R² is 0.773. Compared with the optimal benchmark model GWNet (whose NRMSE at 15 minutes and 4 hours is 0.0591 and 0.1916 respectively, and R² is 0.912 and 0.747 respectively), the model has significant advantages. The pure time series models TCN and FC-LSTM lack spatial constraints, so their errors increase significantly under long prediction horizons; although STGCN and ASTGCN based on fixed graph topology have improved in short-term prediction, they cannot capture time-varying wind direction propagation.

[0012] Appen Figure 7 shows the trajectories, scatter plots and residual distributions of the 1-hour-ahead power prediction for FD1 and FD2 by each model within three days in late October 2019. TCN and FC-LSTM have obvious phase lag and are difficult to capture the sharp rise and fall of power; ASTGCN and STGCN have overshoot and undershoot when wind conditions mutate; GWNet occasionally misses sudden rising events. In contrast, the prediction trajectory of the model is more consistent with the true value, the scatter points are more concentrated along the ideal diagonal, and the residuals are symmetrically distributed around zero, showing stronger sensitivity to high-frequency transients and extreme turning points.

[0013]

[0014] Example 2

[0015] This embodiment explores the effectiveness and working mechanism of each component of the MGDM framework through ablation experiments, gating factor response analysis, and hyperparameter sensitivity analysis.

[0016] ablation experiment To quantify the contribution of each core component, five ablation experiments were designed, all of which evaluated 15-minute prediction performance (NRMSE) on the test set. The settings for each group are as follows:

[0017] G1 (Complete Model): Includes Empirical Wavelet Transform (EWT) preprocessing, Meteorological Causal Graph (MCG), System Adaptive Graph (SAG), and Mamba Temporal Network.

[0018] G2 (w / o EWT): Removes EWT preprocessing and directly inputs the original non-stationary sequence into the network.

[0019] G3 (w / o MCG): Removes the MCG branch; spatial diffusion depends only on the system adaptive graph.

[0020] G4 (w / o SAG): Remove SAG and its learnable node embeddings; spatial aggregation relies solely on meteorological causal graphs. G5 (w / o Mamba): Replaces the Mamba temporal network with a recurrent neural network (RNN).

[0021] Table 2 shows the NRMSE results for each wind turbine. The complete model G1 achieved the lowest errors of 0.0253, 0.0386, 0.0241, and 0.0141 on FD1-FD4, respectively. Removing EWT (G2) caused the NRMSE of FD3 to increase from 0.0241 to 0.0317, confirming its role in suppressing high-frequency noise. Removing MCG (G3) and SAG (G4) both led to a significant increase in error, with the removal of SAG having a particularly significant impact on the downstream wind turbine FD2 (NRMSE reaching 0.0967), indicating that both physical priors and data-driven learning are indispensable. Replacing Mamba with RNN (G5) caused severe performance degradation, with the NRMSE of FD3 increasing to 0.1124, demonstrating the crucial role of the selective scanning mechanism in preventing long-range memory decay and achieving robust spatiotemporal modeling.

[0022]

[0023] Appendix Figure 8The prediction errors and tracking performance of various variants during the 48-hour fluctuation period at the end of December 2019 are presented. The complete model G1 has the smallest and most concentrated error, while G2-G4 exhibit an asymmetric error distribution and obvious outliers, especially showing tracking lag during sharp power changes. G5 has the highest error, producing a high-amplitude residual spike under transient shocks.

[0024] Gated factor response investigation

[0025] To investigate the internal working principle of the dynamic gating fusion mechanism, this embodiment examines the changing behavior of the gating factor under two typical non-stationary scenarios, using the FD1 wind turbine as an example. The gating factor is a real-time confidence coefficient between 0 and 1, used to adaptively balance the contributions of the physical driving branch and the data driving branch.

[0026] Appendix Figure 9 (a) illustrates the response under extreme wind conditions. When wind speed is normal, g remains stable in the high range of 0.6 to 0.8, indicating that the network mainly relies on meteorological causal priors. However, when the wind speed suddenly increases to above 15 m / s (approaching or exceeding the rated wind speed), the wind turbine initiates pitch control, and the power output tends to saturate. The model captures this physical saturation phenomenon, resulting in a significant decrease in the gating factor value. This adaptive fusion mechanism effectively prevents the network from incorrectly predicting unrealistic power values ​​under saturation events.

[0027] Appendix Figure 9 (b) illustrates a case of active power curtailment caused by grid dispatch. During this period, wind speed remained in the range of 2-12 m / s, but due to external dispatch instructions, actual power output was limited to near 0 MW. In this scenario, the dynamic gating mechanism frequently lowered the gating factor, switching the model to a data-driven mode and effectively maintaining prediction accuracy.

[0028] The above results show that the gating factor, as an interpretable adaptive index, can flexibly switch between meteorological-dominated and data-dominated modes according to changes in operational status, and the model has good adaptive capabilities.

[0029] Sensitivity analysis To investigate the impact of the hyperparameter λ on prediction performance, this embodiment performs a sensitivity analysis on λ. λ, in the joint loss function, is used to balance physical constraints (meteorological priors) with data-driven flexibility, and its expression is:

[0030] Taking the FD1 fan as an example, with Figure 10 Showing when lThe distribution of NRMSE varies from 0.15 to 0.65. The results show that when λ < 0.3 (insufficient constraints), the physical constraints are too weak, the model overfits local noise, leading to a median NRMSE exceeding 0.20 and a significant increase in the interquartile range (IQR). When λ approaches 0.35, the NRMSE reaches its minimum, and the IQR is narrowest with the fewest outliers. Experimental results indicate that λ ≈ 0.352 (around 0.35) achieves the optimal balance between physical constraint enforcement and adaptive flexibility. When λ > 0.4 (over-constraints), the NRMSE increases monotonically, the model is over-anchored to deterministic physical airflow, limiting the adaptive branch's ability to adjust the graph topology when grid scheduling conflicts with natural aerodynamic characteristics.

[0031] Based on the above results, this embodiment verifies the effectiveness of each core component of the model (EWT preprocessing, dual-graph topology, dynamic gating fusion, and Mamba temporal network) through ablation experiments, gating factor response analysis, and hyperparameter sensitivity analysis. It also reveals the mechanism by which the gating factor, as an interpretability indicator, reflects the switching of the model's operating state and determines the optimal value of the tradeoff coefficient λ.

Claims

1. A method for ultra-short-term wind power prediction based on a multi-graph diffusion Mamba model, characterized in that, The method includes the following steps: Data preprocessing steps. Acquire historical operating data and meteorological data of multiple wind turbine generators in the target wind farm. The data should include output active power, wind speed, and wind direction. Preprocess the historical operating data and meteorological data, and use empirical wavelet transform to adaptively decompose the original power sequence to extract multi-scale modal features. The steps of multi-graph spatiotemporal feature topology modeling are to construct a physics-driven meteorological causal graph and a data-driven system adaptive graph. The diffusion convolution step is used to input multi-scale modal features and historical operational data into the meteorological causal graph and the system adaptive graph, respectively, and perform graph diffusion convolution operation to extract spatiotemporal dependency features. The multi-map fusion step involves generating a gating factor based on the real-time operating status through a dynamic gating fusion mechanism. This factor is then used to adaptively fuse the first spatiotemporal features extracted from the meteorological causal graph and the second spatiotemporal features extracted from the system adaptive graph to obtain fused spatial features. The spatiotemporal joint modeling step involves inputting the fused spatial features into a topology-aware Mamba module. The Mamba module embeds graph-diffusion convolutions at each time step of the sequence to collaboratively update the temporal state of each node and aggregate neighborhood spatial information, outputting spatiotemporal joint features. Based on these spatiotemporal joint features, a decoder recursively generates wind power prediction results for multiple future time points.

2. The method according to claim 1, characterized in that, The steps for constructing the physical-driven meteorological causal graph include: calculating the causal correlation strength of physical information between different wind turbine generators based on transfer entropy, and constructing a directed weighted adjacency matrix; dividing the global wind direction into a finite number of discrete state spaces according to real-time wind direction information, and establishing a causal matrix corresponding to each wind direction state; and dynamically selecting the corresponding causal matrix according to the wind direction vector at the current moment during the forward propagation of the model to form a time-varying meteorological causal graph adjacency matrix.

3. The method according to claim 1, characterized in that, The steps for constructing a data-driven adaptive system graph include: initializing a learnable node embedding matrix, where row vectors correspond to the embedding representations of each wind turbine node; calculating an adaptive adjacency matrix by performing an inner product operation on the node embedding matrix and passing it through an activation function; and capturing implicit operational constraints not encoded by the meteorological causal graph caused by grid scheduling, measurement errors, or extreme operating conditions through the adaptive adjacency matrix.

4. The method according to claim 1, characterized in that, The graph diffusion convolution operation includes: defining a bidirectional diffusion process on the graph based on the in-degree matrix and the out-degree matrix; for meteorological causal graphs, truncating the diffusion process into a finite number of steps, and applying bidirectional diffusion convolution filtering to each convolutional layer to simultaneously aggregate the information flow of upstream and downstream nodes; for undirected system adaptive graphs, using unidirectional diffusion convolution for simplified calculation.

5. The method according to claim 1, characterized in that, The dynamic gating fusion mechanism includes: concatenating the first spatial feature with the second spatial feature; passing the concatenated feature through a nonlinear mapping layer containing a trainable weight matrix and a bias vector, and outputting a gating factor with a value range of (0,1) via an activation function; multiplying the first spatial feature with the gating factor element-wise to obtain the physical information feature component; multiplying the difference between the second spatial feature and the identity matrix minus the gating factor element-wise to obtain the system adaptive feature component; and summing the physical information feature component and the system adaptive feature component to obtain the fused spatial feature.

6. The method according to claim 1, characterized in that, The topology-aware Mamba module is constructed based on a structured state-space model; wherein the time step parameters, input control matrix, and output control matrix used to control state updates and outputs are dynamically generated by graph diffusion convolution operations based on the fusion space features.

7. The method according to claim 1, characterized in that, The model is trained end-to-end using a joint loss function based on physical information.

Images