A Novel Multivariable Nonlinear Time Series Prediction Method for APSO-Elman Network

Abstract

Elman neural network is a local dynamic neural network with good approximation fitting ability, which is suitable for the prediction of complex nonlinear time series models. When multi-variable, multi-step nonlinear industrial process prediction is involved. However, the traditional Elman neural network learning parameters are too slow, difficult to find an optimal parameter. To overcome the weakness of the traditional Elman neural network, this paper proposed a new adaptive particle swarm optimization Elman (APSO-Elman) neural network, which can achieve better results for nonlinear, multi-step, multivariate time series prediction. Find the optimal weight parameters for the neural network. Firstly, the data mining method is applied to select the appropriate correlation variables, then the APSO-Elman algorithm is used to find the optimal neural network weight parameters. To verify the effectiveness of the method, the non-isothermal continuous tank stirred reactor discharge concentration is used to predict. Compared with the traditional Elman neural network and PSO-Elman network, the results show that the method can accelerate the learning rate of parameters and find the optimal parameters, thus improving the accuracy of discharge concentration prediction.