Probabilistic Wind Power Forecasting With Limited Data Based on Efficient Parameter Updating Rules

Abstract

In this paper, we propose a meta-optimizer-based approach for probabilistic wind power forecasting (WPF) with limited historical data, including offline training and online adaptation procedures. In the offline training part, a WPF meta-optimizer is constructed based on the long short-term memory network (LSTM) via meta-training first, and subsequently used to effectively train probabilistic forecast models under limited historical data scenarios. This meta-training-based process achieves learning to learn probabilistic wind power forecast algorithms directly from wind power data. In the online adaptation part, the performance of the forecast model trained offline is further improved by continuously adapting it to newly collected wind power data online with online updating strategies. Therein, the WPF meta-optimizer is also updated based on these online data to provide more adaptive updating rules for the parameters of the forecast model. Numerical tests were conducted on real-world wind power data sets. Simulation results validate the superiority of the proposed method under limited historical data situations compared with existing alternatives considering the accordance with reality and prediction interval width.