Physics-informed multi-gated convolutional recurrent network for extreme wind speed prediction

Accurate wind speed prediction (WSP) is critical for maintaining grid stability and economical operation of wind-integrated power systems due to the cubic relationship between wind speed and power generation. However, current deep learning models often employ static architectures and loss functions that do not adapt to atmospheric volatility, which leads to large power and energy errors during extreme events. We present the Volatility-Regulated Temporal Evolutionary Network (VORTEX), a physics-informed deep learning framework for extreme WSP in operational settings. The method incorporates a physics-based volatility proxy derived from the pressure-gradient force discrepancy to regulate learning. The model uses a Membership-Gated Convolutional Recurrent Module that trains different sets of weights for different volatility conditions. Furthermore, a differentiable evolutionary mixture layer refines the learned features before the final prediction and optimizes feature representation. We evaluate the method on four geographically diverse locations, namely Dodge City (United States), Sarmiento (Argentina), Jiuquan (China) and Pincher Creek (Canada). Across the 6-hour forecasting horizon, VORTEX consistently achieves the best overall performance on all four datasets, reducing RMSE by 4.33%–9.23% and MAE by 4.59%–10.28% relative to the strongest baseline. Its advantage further widens under the most volatile conditions, with Q4 RMSE reductions of up to 13.30%, and the improvements remain statistically significant across all datasets. At the Jiuquan site, it reduces RMSE by 7.64% and MAE by 6.75% relative to the best baseline, indicating tangible benefits for short-term dispatch and reserve scheduling.