Elman Networks Based Time Series Prediction With Gpu Acceleration

Recently, accurately and rapidly predicting a time series is a hot research issue in current applied sciences field. Neural network with strong ability of memory, fault tolerance, robustness and self-learning can approximate nonlinear maps with arbitrary precision, which make it get more and more attentions. The technology of time series prediction based on neural network has been widely applied to various areas for some practical problems.Elman network is a typical recurrent neural network. Due to its internal feedback mechanism that adapts to time-varying systems, this network structure exhibits a better capability to model a dynamic system compared to the feedforward neural network. However, owing to its complicated structure, the main disadvantages of the existing gradient-based modeling methods lied in the slow convergence rate and the tendency falling into local optimum. The existing extended Kalman filtering (EKF)-based network modeling improved the convergence rate of training, but its computing for the Jacobian matrix were usually complicated and time-consuming. Therefore, in this study an Elman network modeling method based on EKF with graphics processing unit (GPU) acceleration is proposed. Considering the feature of Elman network and the parallel modeling demand, a new direct calculation of the Jacobian matrix for the networks is proposed and the corresponding matrix solution is clearly derived, which can greatly simplify the solving process and help to realize its parallelization. Given the industrial real-time demand, a parallelized method is then reported to model the Elman network, which shifts the computational intensive tasks of network training on GPU and has the logical transaction and the serial processing performed on CPU for the modeling efficiency.To demonstrate the reliability of the proposed method, a number of experimental instances are conducted, including the Mackey-Glass time series with additive Gaussian white noise and a real-world industrial data. To quantify its performance, several time series prediction methods are applied for the comparative experiments. The results indicate that the proposed method exhibits the remarkable merits of rapid modeling, strong generalization and good stability, especially for industrial data with strong dynamics and high-level noise.