DeepNWP4Wind: A Wind Forcasting Correction Solution in View of Multi-Criteria Evaluation

Wind speed and direction predictions bear profound significance in both societal benefits and human livelihoods. However, traditional numerical weather prediction (NWP) methods often yield unsatisfactory forecasting outcomes due to inherent limitations in initial conditions. To address this challenge, this study introduces an innovative NWP correction model, termed as DeepNWP4Wind, tailored for short-term wind speed and direction prediction within a 48-hour horizon. This approach pioneers an exploration into feature engineering deeply rooted in meteorological domain knowledge, encompassing transformative techniques like angle transformation, U and V wind component conversion, and spatiotemporal embedding. Beyond conventional regression metrics, the study introduces a novel evaluation perspective by introducing an interval level accuracy metric. Extensive experiments demonstrate the remarkable superiority of DeepNWP4Wind. The advancement in feature engineering techniques holds promise for enhancing deep learning methodologies in wind prediction. Furthermore, the broad applicability of the multi-criteria evaluation approach across various spatiotemporal forecasting tasks is anticipated.