A novel dynamic ensemble of numerical weather prediction for multi-step wind speed forecasting with deep reinforcement learning and error sequence modeling

Abstract

Accurate wind forecasts for one day ahead or longer periods have significant impacts on the safe and efficient dispatch of power grids, where Numerical Weather Prediction (NWP) serves as the essential tool, such as ensemble NWP integrating multiple single simulations. Typically, ensembles include all single members with fixed weights; however, the relative accuracy of each member may change over time. This study introduces an attractive idea: improving ensemble performance by dynamically recognizing and avoiding low-performing members. It proposes a dynamic ensemble strategy based on NWP, reinforcement learning and error sequence correction. The process begins with Weather Research and Forecasting ensemble simulations. A dynamic framework is then constructed by mapping the multi-step ensemble problem into a Markov decision process, which is further solved using deep deterministic policy gradient. Subsequently, a hybrid deep learning model, comprising temporal convolutional network and bidirectional long short-term memory, is constructed for error sequence estimation of dynamic ensemble, using the high-frequency information of NWP as input. Conducting experiments at two wind farms, and focusing on the 24-h wind speed forecast with a 15-min time resolution, the proposed system demonstrates a reliable and stable ensemble throughout the entire forecasting horizon, significantly reducing the probability of large forecasting errors.