Ramp-Feature-Aware Short-Term Power Forecasting for Wind Farm Clusters Based on Multi-Graph Fusion

Abstract

To address the insufficient modeling of multi-source spatial relationships, inadequate coupling of ramp propagation features, and large prediction errors during severe fluctuation periods in short-term power forecasting for wind farm clusters, this paper proposes an MGAT-CGAN-LSTM forecasting method based on ramp features and adaptive multi-graph fusion. First, event features including ramp magnitude, duration, ramp rate, and ramp direction are extracted, and a conditional generative adversarial network is used to directionally augment scarce ramp samples in the training set. Then, multi-source graph structures are constructed from four perspectives: station distance, temporal pattern similarity, feature-variable coupling, and ramp propagation relationships, so as to characterize the multi-source spatial correlations and ramp propagation features within the wind farm cluster. In the forecasting model, a multi-graph attention network is employed to extract spatial features under different graph structures, and an adaptive graph-weight learning mechanism is used to dynamically fuse multi-source graph information. Furthermore, a long short-term memory network is introduced to capture temporal dependencies, while a ramp-weighted loss and a trend constraint mechanism are incorporated to enhance the model’s learning capability for rapid power variations. Case-study results show that MGAT-CGAN-LSTM achieves the lowest prediction errors among the comparison models, with MAPE, MAE, and RMSE values of 2.21%, 9.60 MW, and 14.16 MW, respectively, demonstrating good trend-tracking and error-control performance. Single-farm forecasting, multi-horizon forecasting, ablation experiments, and ramp-period-specific analysis further indicate that conditional ramp sample augmentation, multi-source graph fusion, adaptive graph-weight learning, and ramp-aware loss can jointly improve forecasting performance, especially for short-term power forecasting under ramping and severe fluctuation conditions.