Intraday Wind Power Forecasting by Ensemble of Overlapping Historical Numerical Weather Predictions

The numerical weather prediction (NWP) is crucial to improve intraday wind power forecasting (WPF) accuracy. However, conventional WPF methods relied solely on a latest reported single NWP, overlooking hidden information from sequentially reported multiple historical NWPs that are partially overlapped over time. Additionally, it's challenging to tackle intraday WPF as it involves both ultra-short-term and short-term horizons with different characteristics. Therefore, a novel spatio-temporal representation learning network is proposed for intraday WPF by ensemble of overlapping historical NWPs. Initially, an integrated mask-reconstruction representation learning pretraining strategy is employed to extract hidden representations of historical wind power measurements and overlapping historical NWPs, providing contextual information for the subsequent intraday WPF task. Then, the output layer is trained and end-to-end fine-tuning of the entire network is conducted to adapt to the specific forecasting task. Moreover, a multi-task learning strategy based on hard parameter sharing is adopted to ensure balanced predictive accuracy across each of forecasted wind farms. Case study and detailed ablation tests based on 5 real-world wind farms demonstrate that the proposed method enhances the forecasting accuracy of most wind farms by leveraging spatio-temporal correlation, achieving the best average performance across all time horizons compared to the baseline models.