Multi-Scale Convolutional Echo State Network With an Effective Pre-Training Strategy for Solar Irradiance Forecasting

Abstract

In this article, a new kind of neural network model named multi-scale convolutional echo state network (MCESN) is proposed for solar irradiance prediction, which integrates the strong feature extraction capability of convolutional neural network (CNN) and the fast yet efficient prediction ability of echo state network (ESN). Firstly, the feature information at different time scales of solar irradiance (one dimensional series) data are extracted and selected by multi-scale CNN (MCNN) in the pre-training stage. Then, the trained features extracted above are concatenated and passed to ESN module as the input signal, which can be further encoded into high-dimensional state space; Meanwhile, the target solar irradiance value is fitted and predicted by ESN in the prediction phase. Finally, the effectiveness of MCESN is evaluated by hourly solar irradiance prediction. In experiment, RMSE, MAE, MAPE and R are chosen as four metrics to evaluate the performance of the proposed model. Simulation results demonstrate that the proposed MCESN perform better than classical ESN, MCNN, backpropagation (BP) random forest (RF), long short time memory (LSTM) and deep ESN (DESN) algorithms.

Keywords

• Echo state network
• multi-scale convolution
• pre-processing
• solar irradiance prediction