Time Series Prediction with Dynamic Neural Networks - MATLAB-Based NARX Implementation

Time series prediction carries substantial significance in both economic and engineering domains. In this paper, we present a dynamic neural network-based approach for time series forecasting, leveraging the inherent properties of dynamic neural networks for predictive modeling. The methodology employs a NARX network structure implemented in MATLAB, which incorporates tapped delay lines for handling temporal dependencies and feedback connections for capturing dynamic system behavior. We applied this approach to predict response time series of Duffing's equation, a nonlinear dynamical system, using MATLAB's Neural Network Toolbox. The implementation involves configuring input/output delays, selecting appropriate hidden layer sizes, and training the network with Bayesian regularization or Levenberg-Marquardt algorithms to prevent overfitting. Experimental results indicate that dynamic neural networks can effectively forecast response time series of dynamic systems, achieving accurate predictions through proper network architecture design and parameter optimization. This method enables improved understanding and prediction of time series data in economic and engineering applications, with MATLAB code implementations typically involving timeseries data preprocessing, network creation using 'narxnet' function, and closed-loop simulation for multi-step predictions.