Forecasting the Total Non-coincidental Monthly System Peak Demand in the Philippines: A Comparison of Seasonal Autoregressive Integrated Moving Average Models and Artificial Neural Networks

Abstract

This paper aims to determine suitable seasonal autoregressive integrated moving average (SARIMA) and feed-forward neural network (FFNN) models to forecast the total non-coincidental monthly system peak demand in the Philippines. To satisfy the stationary requirement of the SARIMA model, seasonal differencing, and first-differencing were applied. The findings reveal that SARIMA (0,1,1)(0,1,1)12 is the appropriate SARIMA model. All the model parameters were statistically significant. Also, the residuals were normally distributed. For the feed-forward neural networks, the NNAR (10,1,6)12 was found to be the appropriate model. The evaluation statistics indicate that the models developed are suitable for forecasting. A comparison of the models has been performed by examining their respective root mean square error (RMSE), mean absolute error (MAE), and mean absolute percent error (MAPE) values. It was found that the FFNN performs better and is the most suitable model to forecast peak demand.