Mitigating systematic underestimation in short-term wind speed forecasting via time-dependent weighted quantile loss (TDWQLoss)

Accurate short-term wind speed forecasting is pivotal for wind power integration. Deep learning models have shown promise but often suffer from distributional bias. Since wind speed follows a right-skewed Weibull distribution, the scarcity of high-wind-speed samples creates a data imbalance that predisposes deep learning models to systematic underestimation. To address this, we propose a lightweight strategy incorporating quantile loss to guide model training. By imposing asymmetric penalties, this method forces the model to better capture the tail of the distribution. We evaluate this approach using three representative spatiotemporal models (Conv3D, SimVP, and MFWPN). The results show that the proposed method not only effectively mitigates systematic negative bias but also significantly enhances the detection of extreme wind events. Specifically, the models exhibit improved Probability of Detection (POD) and Critical Success Index (CSI) scores, demonstrating that quantile loss is a robust and generalizable solution for high-impact wind speed forecasting.