Explainable temporal dependence in multi-step wind power forecast via decomposition based chain echo state networks.

Wind power is one of the most promising renewable energy for its abundant resources, economically competitive, and environmentally friendly. Nevertheless, the wind power is challenging used in the power generation system due to its intermittency. Therefore, to improve the utilization ratio of wind power, the common method is adopting a prediction model for scheduling the generation industries. However, the information offered by single-step models hardly assists managers control their generators, and existing multi-step prediction models ignore the temporal dependence among predicted steps. Thus, a hybrid method based on a deep-chain echo state network (DCESN) and variational mode decomposition (VMD) is proposed to enhance the mapping capability for wind power multi-step prediction. The multiple reservoirs of deep-chain echo state network are concatenated as a chain structure, which could congregate the temporal relations among future steps shown in visualized graphs. Three comparative experiments demonstrate that the proposed hybrid method has promising performance on wind power multi-step prediction. • A novel hybrid model explains the temporal dependence in multi-step prediction. • Deep-chain echo state network maps the correlation in prediction. • Variational mode decomposition further improves the hybrid model's performance. • A temporal dependence coefficient visualizes the correlation among the steps. [ABSTRACT FROM AUTHOR]