Comparison between Adaptive and Non-adaptive HRBF Neural Network in Multiple Steps Time Series Forecasting

This paper compares the performance of adaptive and non-adaptive learning approaches of the Hybrid Radial Basis Function (HRBF) neural network in multiple steps time series forecasting. The HRBF was trained by using the Adaptive Fuzzy C-Means Clustering (AFCMC) and Exponential Weighted Recursive Least Square (e-WRLS) algorithms. Both approaches were set to produce up to 25 steps ahead forecasting on two time series data: Mackey Glass and Set A Data from Santa Fe Competition. The performance of both approaches in multiple steps ahead forecasting was measured using Mean Square Error Test and Coefficient of Determination Test between the actual and forecasted data for the 25 steps ahead forecasting. Results show that both approaches perform comparatively equal for shorter forecasting distance. However for longer forecasting distance (10 steps ahead onwards), the adaptive approach performs significantly better to compare with non-adaptive approach.