Time-variant post-processing method for long-term numerical wind speed forecasts based on multi-region recurrent graph network

Abstract

Weather Research and Forecasting (WRF) is widely used for long-term wind speed prediction. To reduce the inherent systematic error of the WRF, a graph-based wind speed prediction model is proposed by post-processing the WRF. The proposed model contains four stages. In stage 1, the WRF model is applied to generate raw wind speed prediction results. In stage 2, Planar Maximally Filtered Graph (PMFG) is used to construct an informative graph of the wind field. In stage 3, Deep Autoencoder-like Nonnegative Matrix Factorization (DANMF) and Subgraph Alignment and Region Organization (SARO) are utilized to divide the graph into several regions. In stage 4, Multi-region Recurrent Graph Network (MRGN) model is proposed to build multiple regional models, aggregate them via time-variant ensemble weights, and generate improved long-term wind speed prediction results. The wind speed prediction results on 25 real meteorological monitoring nodes show that: (1) the proposed model outperforms WRF and 3 state-of-the-art models with a 95% confidence level; (2) the proposed model tends to produce better performance in strongly dynamic wind speed; (3) the proposed model has enough computational efficiency for practice. In conclusion, the proposed model is effective to improve WRF performance and has application potential.