A novel approach for training neural networks for long-term prediction

Neural networks have been widely used in performing time series prediction. Long-term prediction is generally far more difficult than short-term prediction, because of the difficulty in modeling the system dynamics far ahead. In this paper, we present a novel approach for training neural networks to perform long-term prediction. Our approach relies on the utilization of traditional time series analysis, based on Box-Jenkins methodology (1976), to: (1) determine the appropriate neural network architecture, (2) select the inputs to the neural network, and (3) determine the appropriate lead time for updating the connection-weights of the neural network during training. We demonstrate the effectiveness of this approach in producing accurate multistep ahead prediction on some real-world problems as well as on simulated time series data