RBF-BL time series model and its application in modeling and prediction

An improved nonlinear model was proposed based on the traditional linear and nonlinear models, that is radial basis function neural network based state dependent bilinear(RBF-BL) model. The sum of squares of the model residuals was taken as the objective function and the parameter estimation algorithm was introduced. The sunspot data, the Mackey-Glass series data and the creeping displacement data of machine table were taken as numerical examples. General expression for nonlinear autoregressive(GNAR), back propagation(BP), RBF and RBF-BL models were used for data modeling and prediction. Mean squared error of modeling(MSEM), mean squared error of prediction(MSEP), mean relative error of modeling(MREM), and mean relative error of prediction(MREP) were taken as the error indicators. The results show that RBF-BL model exhibits better modeling and prediction performance compared with the traditional models. For the sunspot data, the error indicators of RBF-BL model are 0.009 6, 0.026 6, 0.002 7, and 0.003 9. For Mackey-Glass series data, the error indicators of RBF-BL model are 7.982×10-6, 6.400×10-4, 0.002 5, and 0.025 0. For the creeping displacement data of machine table, the error indicators of RBF-BL model are 7.590×10-4, 0.010 1, 0.038 8, and 0.023 8. © 2020, Editorial Department of Journal of Southeast University. All right reserved.