Short term load forecasting (STLF) is important, since it is used to maintain optimal performance in the day-to-day operation of electric utility systems. The autoregressive integrated moving average (ARIMA) model is a linear prediction method that has been used for STLF. However, it has a weakness. It assumes a linear relationship between current and future values of load and a linear relationship between weather variables and load consumption. Neural networks have the ability to model complex and nonlinear relationships. Therefore, they can be used as a robust method for nonlinear prediction, and they can be trained with historical hourly load data. The purpose of this work is to describe how neural networks can transform linear ARIMA models to create short term load prediction tools. This thesis introduces a new neural network architecture - the periodic nonlinear ARIMA (PNARIMA) model. In this work, first, we make linear predictions of the daily load using ARIMA models. Then we test the PNARIMA predictor. The predictors are tested using load data (from May 2009 - April 2011) from Batam, Indonesia. The results show that the PNARIMA predictor is better than the ARIMA predictor for all testing periods. This demonstrates that there are nonlinear characteristics of the load that cannot be captured by ARIMA models. In addition, we demonstrate that a single model can provide accurate predictions throughout the year, demonstrating that load characteristics do not change substantially between the wet and dry seasons of the tropical climate of Batam, Indonesia.