An online adjusting RBF neural network for nonlinear system modeling

Aiming to improve the prediction accuracy and to obtain a compact structure, an online adjusting radial basis function neural network (OA-RBFNN) is proposed in this paper. The proposed OA-RBFNN realizes online modeling by combining the advantages of the sliding window strategy and clustering algorithm. First, with a small batch of samples in the first sliding window, the network is initialized using a density canopy-based k-means algorithm, and an optimal network structure and its initial parameters are determined automatically. Secondly, through the sliding window movement, the network parameters are adjusted by fine-tuning based on the changed samples, followed by the gradient-based online learning algorithm. Finally, the effectiveness of the proposed OA-RBFNN model is verified by four experiments: the function approximation, Mackey-Glass time series prediction, Lorenz time series prediction, and the effluent BOD prediction in wastewater treatment plant (WWTP), and the prediction accuracy obtained in these four experiments reached 97.11%, 99.25%, 99.69%, and 98.81%, respectively. The results demonstrate that the OA-RBFNN can achieve competitive prediction performance while having a more compact network structure than the existing online RBF neural networks.