An adaptive distribution-matched recurrent network for wind power prediction using time-series distribution period division

Precise wind power prediction (WPP) can address the issue caused by large-scale wind power grid integration to the power system operation. Most WPP research focus on the randomness and high volatility problem of wind power but ignore the time-series distribution shift (TSDS) problem. To solve the TSDS issue, this study proposes a novel hybrid model that incorporates complementary ensemble empirical mode decomposition (CEEMD), time-series distribution period division (TSDPD) and adaptive distribution-matched GRU (ADMGRU). First, CEEMD is utilized to decompose the nonstationary data into in sub-sequence, reducing complexity and randomness. Second, TSDPD is employed to automatically identify the underlying temporal segments within wind power sequence by maximizing discrepancies in distribution information between two periods, determining the quantity and respective boundaries of periods. Finally, ADMGRU, comprising Pre-train and Boosting-based importance assessment components, learns prediction model accurately by dynamically matching distribution periods. The former component initializes predictive model parameters and the latter learns the importance of each hidden state to assigns corresponding weights to different distributions. Numerous comparative experiments demonstrate the CEEMD-TSDPD-ADMGRU hybrid model surpasses existing popular models. Especially in spring scenario, compared with other four advanced models, the maximum reduction in MAE and RMSE are 54.64 % and 73.33 %, respectively. •A hybrid model (CEEMD-TSDPD-ADMGRU) is proposed considering the time-series distribution shift issue in wind power.•Decompose nonstationary data into subsequences by CEEMD to reduce complexity and randomness.•TSDPD method is proposed to divide distribution segments of wind data.•ADMGRU is proposed to dynamically match distribution periods and achieve prediction.•CEEMD-TSDPD-ADMGRU outperforms other state-of-art methods in prediction accuracy and computational efficiency.