ABSTRACT
Electricity demand forecasting is a central and integral process for planning periodical operations and facility expansion in the electricity sector. Demand pattern is very complex due to the highly unpredictable behavior of consumers load consumption. Therefore, finding an appropriate forecasting model for a specific electricity network at peak demand is not an easy task for the utilities and policymakers. Many load forecasting methods developed in the past decades were characterized by poor precision, and large forecast error because of their inability to adapt to changes in dynamics of load demand. To fill this gap, this research has developed an improved short-term daily peak load forecasting model based on Seasonal Autoregressive Integrated Moving Average (SARIMA) and Nonlinear Autoregressive Neural Network (NARX). The developed model used SARIMA to captures the linear pattern (trend) and seasonality of the load time series but due to seasonal and cyclical nature of the load behavior which cannot accurately describe by linear regression model, NARX neural network was combined with SARIMA in order to improve and captures the non-linear patterns of the data series to minimize it forecast error. The structures of NARX was optimized by the tenets of chaos theory to avoid trial by error approach during training. A daily peak load data of Nigeria power system grid and daily average weather data for ten years, from January 1st, 2006 to December 31st, 2015 were used in this study to complete the short-term load forecasting using MATLAB 2015a environment for simulation and mean absolute percentage error (MAPE) as a measure of accuracy. The model forecast result was validated and compared with real peak load demand data of Nigeria grid in 2015 to measure the performance of the method. The evaluation results showed that the developed model trained with Levenberg-Marquardt training algorithm (LM) is more effective and performs better than classical SARIMA model with MAPE of 2.41%, correlation coefficient of 96.59% which is equivalent to an improvement of 63.70% in error reduction. Performance of different training methods also compare on the developed method and results shows that developed model training with LMshows more superiority and high precision over Bayesian regularization training algorithm (Br) with 1.6318% in error reduction equivalent to an improvement of 40.37%. Finally, the proposed model was further used to forecasts the daily peak load demand of year 2017 and 2018 successfully for planning and operations of the grid.