An Effective and Efficient Renewable Energy Generation Forecasting via Meteorological Assistance

Accurate signal pattern mining of renewable energy generation forecasting (REGF) is important to the daysahead power scheduling of renewable energy power systems. Despite achieving excellent performance with current methods, two issues still persist. (1) They solely utilize historical meteorological signal data to assist in power signal forecasting and neglect valuable information in future information of meteorological signals, consequently limiting their performance. (2) They pursue predictive performance by designing complex architectures and mechanisms, which may lead to insufficient model generalization. To this end, an effective and efficient MLP architecture is proposed to mine REGF signal patterns in renewable energy power systems (SPM-REPS), which contains power signal forecast architecture and meteorological signal forecast architecture. Two architectures seamlessly collaborate in forecasting power generation patterns, which achieves better performance. Meanwhile, time-correlation and feature-correlation strategies are devised within MLP networks to capture both intra-sequence and inter-sequence correlations of signal variables like transformer- and RNNbased methods. Furthermore, a theoretical analysis of linear architecture is given to prove the progressiveness of SPM-REPS. Finally, numerous experiments, conducted on common datasets (CSG-PV and CSG-wind) from Chinese State Grid, demonstrate SPM-REPS sets a new benchmark in mining REGF signal patterns of REPS.