Deep learning combined wind speed forecasting with hybrid time series decomposition and multi-objective parameter optimization

This study proposes an effective combined model system for wind speed forecasting tasks. In this model, (a) improved hybrid time series decomposition strategy (HTD) is developed to concurrently extract the linear patterns and frequency-domain features from raw wind speed; (b) novel multi-objective binary backtracking search algorithm (MOBBSA) is exploited to optimize the decomposition parameters; (c) advanced Sequence-to-Sequence (Seq2Seq) predictor is utilized to uniformly process the component series, and predictions of multiple different Seq2Seq models are averaged to construct the final results. Real-world experiments from the National Wind Power Technology Center are implemented. The step-average mean absolute percentage errors of the proposed model in four datasets are 1.58%, 1.98%, 2.62%, and 2.95% respectively, which are much lower than those of eighteen benchmarks. Compared with state-of-the-art techniques, the average improvement percentage of proposed model reaches 59.92%. The non-parametric Kruskal-Wallis test is further implemented to explore the effectiveness of three designed modules (HTD, MOBBSA, and Seq2Seq), and test results demonstrate remarkable contributions of proposed modules compared with existing decomposition strategies, optimization techniques, and deep learning predictors, which indicates that the proposed model is a promising alternative for complex wind speed forecasting applications.

Automated summary

This study proposes an effective combined model system for wind speed forecasting tasks. The proposed model utilizes an improved time series decomposition strategy, a novel binary backtracking search algorithm, and an advanced sequence-to-sequence predictor, achieving significant improvements in wind speed prediction accuracy compared to existing techniques.