Multi-timescale Forecast of Solar Irradiance Based on Multi-task Learning and Echo State Network Approaches
Zhou Wu, Qian Li, Xiaohua Xia
Abstract
Solar irradiance forecast is closely related with efficiency and reliability of renewable energy systems. Multi-timescale irradiance forecast is a new and efficient way to simultaneously predict solar energy generation on different timescales for hierarchical decision making. This article newly adopts the multi-task learning mechanism to study the multi-timescale forecast for improving accuracy and computational efficiency. A novel multi-timescale (MTS) prediction framework is presented to fulfill the multi-task application, and echo state network (ESN) is studied in the proposed MTS framework. The multi-timescale ESN (MTS-ESN) is proposed to enhance the information sharing among correlated tasks. Simulation results of hourly solar data demonstrate that the proposed MTS-ESN could achieve promising performance at both hourly and daily level in parallel. The MTS-ESN outperforms the single-timescale ESN (STS-ESN), which indicates the information sharing in the multi-task learning is effective in this application.