Short-term wind speed forecasting using deep reinforcement learning with improved multiple error correction approach

Abstract The safe and stable operation of wind power systems requires the support of wind speed prediction. To ensure the controllability and stability of smart grid dispatching, a novel hybrid model consisting of data-adaptive decomposition, reinforcement learning ensemble, and improved error correction is established for short-term wind speed forecasting. In decomposition module, empirical wavelet transform algorithm is used to adaptively disassemble and reconstruct the wind speed series. In ensemble module, Q-learning is utilized to integrate gated recurrent unit, bidirectional long short-term memory, and deep belief network. In error correction module, wavelet packet decomposition and outlier-robust extreme learning machine are combined to developing predictable components. An appropriate correction shrinkage rate is used to obtain the best correction effect. Ljung-Box Q-Test is utilized to judge the termination of the error correction iteration. Four real data are utilized to validate model performance in the case study. Experimental results show that: (a) The proposed hybrid model can accurately capture the changes of wind data. Taking 1-step prediction results as an example, the mean absolute errors for site #1, #2, #3, and #4 are 0.0829 m/s, 0.0661 m/s, 0.0906 m/s, and 0.0803 m/s, respectively; (b) Compared with several state-of-the-art models, the proposed model has the best prediction performance.