Time-varying Echo State Networks: Harnessing Dynamic Parameters for Robust Time-Series Analysis

This work presents a novel variant of Echo State Network (ESN) known as Time-Varying Echo State Network (TV-ESN) and conducts a comprehensive comparative analysis with the standard ESN model. TV-ESN introduces dynamic variations in the leaking rate (a) and spectral radius (rho) over time, offering adaptability to diverse temporal patterns. The evaluation is performed on benchmark datasets, including the Mackey Glass system and actual electricity demand in the Spanish Electricity Market. Results indicate that TV-ESN achieves superior forecasting performance, as evidenced by lower Normalized Root Mean Square Error (NRMSE) and Mean Absolute Error (MAE) compared to the conventional ESN. Moreover, TV-ESN exhibits enhanced stability in forecasting. The findings suggest the potential of TV-ESN in capturing rich temporal characteristics.