Echo state neural network based ensemble deep learning for short-term load forecasting

Abstract

Precise electricity load forecasts assist in planning, maintaining, and developing power systems. However, the electricity load's un-stationary and non-linear characteristics impose substantial challenges in anticipating future demand. Recently, a deep echo state network (DESN) with multi-scale features has been proposed for sequential tasks. Inspired by its structure, this paper offers a novel ensemble deep learning algorithm, the ensemble deep ESN (edESN), for load forecasting. First, hierarchical reservoirs are stacked to enforce the deep representation similar to the DESN. Then, instead of computing the readout weights based on the global states, the edESN trains a different readout layer for each scale. Finally, the network combines the outputs from each scale as the final prediction. The edESN is evaluated on twenty publicly available load datasets. This paper compares the edESN with eleven forecasting methods, and the comparative results demonstrate the proposed model's superiority in load forecasting.