Pangu-Weather Driven Dual-Path Network with Dynamic Gated Attention for Ultra-Short-Term Wind Power Forecasting

Abstract

Conventional numerical weather prediction is limited by high computational cost and delayed availability, while single-path models cannot effectively integrate multi-scale historical power information with future meteorological drivers. To address these issues, this work proposes a dual-path ultra-short-term wind power forecasting method based on Pangu-Weather and dynamic gated attention fusion. The raw wind power series is decomposed using complete ensemble empirical mode decomposition with adaptive noise, and meteorological forecasts are generated by Pangu-Weather. In the proposed network, the power path uses extreme gradient boosting to model short-window historical power features. The meteorological path comprises a temporal convolutional network branch and a meteorological feature branch, which respectively extract long-range temporal features and future meteorological driving features. A dynamic gated attention fusion module is then designed to fuse the outputs of the two paths through step-wise adaptive weighting, inter-step dependency modeling, and residual correction. Experiments on a real-world 387.45 MW wind farm demonstrate that the proposed model achieves an root mean square error of 21.28 MW, outperforming the best baseline by 13.97% and surpassing several baseline models overall. Ablation studies further validate the necessity of the dual-path design and the dynamic fusion mechanism. These results demonstrate that AI-weather-driven forecasting and dual-path dynamic fusion can provide a promising low-latency meteorological-input strategy for ultra-short-term wind power forecasting.