Multi-reservoir echo state computing for solar irradiance prediction: A fast yet efficient deep learning approach

Abstract

Accurate solar irradiance prediction plays an important role in renewable energy systems. Based on time series analysis, a serially connected multi-reservoir echo state network (MR-ESN) is developed to predict solar irradiance. MR-ESN is a fast yet efficient approach, which makes use of the high efficiency of ESN and the advantages of deep learning. MR-ESN consists of multiple reservoirs in series, which are responsible for encoding the input signals into a richer state representation. The time series analysis is adopted to provide more appropriate input and output for MR-ESN. Various prediction horizons including one-hour-ahead and multi-hour-ahead prediction are conducted, respectively. The effect of reservoir layer number on the MR-ESN performance is explored in detail. Three internal qualitative indicators are adopted to investigate the performance differences of MR-ESN, i.e., probability distribution, correlation analysis, and principal component analysis (PCA) of network states. Simulation results demonstrate that MR-ESN outperforms than traditional ESN, backpropagation (BP) and Elman neural networks.