Systemic framework of time-series prediction via feed-forward neural networks

The evolution of neural networks has led to the development of various prediction applications such as the time series prediction models. Time series prediction techniques has as widely used in many real-life applications such character recognition, speech recognition, and bio-metric authentication as well person identification. In this work, we report on the comprehensive model for time-series prediction system using multi-Layer feedforward neural network (FFNN) model with the least possible training/prediction error. As a case study, an FFNN Model with minimum estimation error has been developed via MATLAB computing platform to model and predict the time series for yearly averaged sunspot activity during the period from 1719-2018 as a 10th order nonlinear one step ahead predictor. The developed model uses only 90% of the data set to train a two-layer feed forward neural network using generalized delta learning rule (gradient) algorithm. In order to achieve the minimum possible training error, we have trained the proposed FFNN with 4000 sweeps where it recorded a high accuracy levels with a mean square error (MSE) registered as 0.002227 after only 200 sweeps of the training process according to the plot that shows the number of training sweeps vs. the training error (i.e., MSE). Thus, the remaining unused 10% of the data set were used to test the proposed NN model. Therefore, all output signals (the desired output, the training output and testing output) were plotted together for comparison purposes and to gain more insights about the level of confidence achieved by applying the proposed FFNN.