An incremental neuronal-activity-based RBF neural network for nonlinear system modeling

Abstract

Abstract In this paper, a novel incremental radial basis function (RBF) neural network is proposed for nonlinear systems modeling. The hidden layer is constructed dynamically on the basis of the neuronal activity (NA), which is measured by the local field potential (LFP) and the average firing rate (AFR), with the goal of enhancing the structural compactness. Simultaneously, a modified second-order algorithm is utilized to train the neuronal activity-based RBF (NARBF) neural network, which can decrease the convergence time and improve the generalization performance. Then, three benchmark nonlinear system modeling simulations are employed to evaluate the proposed NARBF neural network, indicating that the proposed neural network can obtain good generalization performance with a compact structure after fast training. Finally, the NARBF neural network is applied to wastewater treatment process modeling, which demonstrates that the proposed algorithm can predict the key water quality variable precisely.