A joint deterministic and probabilistic wind power forecasting method integrating wide-area meteorological representation learning

Local-area (LA) wind power variability is governed by the spatiotemporal evolution of wide-area (WA) meteorological fields. However, current forecasting approaches largely isolate wind farms from this broader context, primarily because incorporating high-dimensional WA Numerical Weather Prediction (NWP) data induces the “curse of dimensionality” and computational complexity. To bridge this gap, this study proposes a dual-branch framework designed to explicitly decouple the learning of WA meteorological patterns from LA temporal dependencies. Specifically, a novel discretized representation learning strategy is developed using Vector Quantized Weather Encoder (VQWE). By pre-training distinct WA meteorological ‘prototype patterns’ from a discrete codebook, this strategy achieves high-fidelity feature compression and extraction. Subsequently, a meteorological context perception layer based on attention mechanisms and adaptive residual gating is designed to effectively align and fuse multi-source features. To tackle the inherent stochasticity, a collaborative optimization objective is designed, which utilizes quantile regression theory to guarantee reliable and interpretable forecasts across various confidence levels. Experimental results demonstrate that the proposed method exhibits superior robustness across diverse regional and temporal scenarios, achieving a 3.64% reduction in RMSE compared to the best-performing SOTA baseline, while consistently maintaining interval prediction deviation within 4%. Moreover, comparative analysis confirms that the proposed pre-trained representation learning strategy outperforms the traditional WA-NWP processing strategies, reducing RMSE by 13.40%.