Systematic evaluation of transformer-based time series forecasting models for post-processing WRF-simulated wind speed and predicting short-term power output

Accurate short-term (0–72 h) forecasting of wind speed and power is essential for wind energy integration, yet remains hindered by persistent biases in Numerical Weather Prediction (NWP) models. This study presents the first systematic evaluation of Transformer-based time series forecasting (TSF) models for post-processing Weather Research and Forecasting (WRF) simulations. Using three years (2020−2022) of data from four operational wind farms in northwestern China, we assessed Pyraformer, Autoformer, Reformer, and other TSF variants. Multi-resolution architectures, particularly Pyraformer, demonstrate clear superiority. Compared to raw WRF hub-height wind speeds, Pyraformer reduces root mean squared error (RMSE) by up to 27 %, and relative to other Transformers, improves power forecasting accuracy by 3–6 % and qualification rate by 3–9 %. Mechanistic analysis reveals that decomposition-based Transformers fail by spuriously fitting temporal trends in stationary WRF biases, while multi-resolution architectures effectively correct these regime-dependent errors through scale-aware attention. A unified evaluation framework shows that nonlinear wind-to-power conversion amplifies meteorological uncertainties, resulting in consistently higher normalized errors in power forecasts than in wind speed forecasts across lead times and wind regimes. All methods are implemented within an integrated Global Forecast System (GFS)-WRF-AI pipeline. These findings offer mechanistic insights for next-generation hybrid forecasting model development and demonstrate the potential of advanced TSF architectures for operational wind power forecasting.