Enhancing Wind Power Forecasting at Local Peak Points: A Novel Seq2LPP Model

Abstract

Mining the potential of deep learning (DL)-based models for forecasting wind power during local peak points (LPPs) remains a crucial yet underexplored direction. Although existing DL-based models work exceptionally well in regular wind power forecasting (WPF), they primarily focus on optimizing the average accuracy of overall wind power predictions within an prediction horizon, thereby generating poor performance in the predictions of the LPPs. Due to the substantial fluctuations and nonstationarity of wind power specifically for the LPPs, it is more difficult for DL-based models to predict them. Considering a fact that there exists strong correlations between the LPPs and multisource numerical weather prediction (NWP) data, we propose a novel Seq2LPP model powered by the multisource NWP data to enhance the understandings of the LPPs. The proposed model specifically designs three key modules: an NWP-guided attention module to calculate weighted representations of LPPs using variables in NWP data, a patch-based feature learning module to capture trend-specific semantic information, and a mixture decoder module to output both the regular predictions and the predictions of the LPPs. We compare the proposed model with the state-of-the-art models on two real-word wind power dataset. The proposed model can obtain average enhancements of 14.8% and 11.6% for MAE and RMSE, respectively, for the predictions of the LPPs within an ultrashort-term prediction horizon ranging from 1 to 4 h. These findings underscore the proposed model's ability to attain greater accuracy and robustness in ultrashort-term WPF especially for the LPPs.