An improved Pi-Sigma neural network with error feedback for physical time series prediction

Abstract

Forecasting natural occurring phenomena has been addressed and analyzed in many domains of science and gets more attention because of its vast range of applications. Traditional time series forecasting tools have some limitations like slow training process, less efficient training methods that effect on performance. Higher Order Neural Network (HONN) using recurrent feedback appear as a powerful technique and it has the ability to expand the input space, make them more efficient for solving complex problems and perform high learning abilities in the time series forecasting. Recurrent networks commonly used the network output as the feedback terms. This study proposed a model called improved Pi-Sigma Neural Network with Error Feedback (PSNN-EF) which combines the properties of Pi-Sigma Neural Network (PSNN), recurrence and error feedback. PSNN-EF uses back propagation gradient descent algorithm for training purpose and tested with time series signals namely; humidity, evaporation and wind direction datasets. The benefit of using recurrence and error feedback is that it generates more accurate results of prediction and provide more promising results. Based on the obtained results, our proposed method shows better performance and can be an alternative solution to Jordan Pi-Sigma Neural Network (JPSN) and PSNN for one step ahead prediction of those three datasets.