Non-linear autoregressive neural network (NARNET) with SSA filtering for a university energy consumption forecast

Energy consumption forecast is essential for strategic planning in achieving a sustainable energy system. The hemispherical seasonal dependency of energy consumption requires intelligent forecast. This paper uses a non-linear autoregressive neural network (NARNET) for energy consumption forecast in a South African University with four campuses, using three-year daily energy consumption data. Singular Spectrum Analysis (SSA) technique was used for the data filtering. Three window lengths (L=54, 103 and 155) were obtained using periodogram analysis and R-values of network training at these window lengths were compared. Filtered data at L=103 gave the best R-values of 0.951, 0.983, 0.945 and 0.940 for campus A, B, C, and D respectively. The network validation and a short-term forecast were performed. Forecast accuracies of 85.87%, 75.62%, 85.02% and 76.83% were obtained for campus A, B, C and D respectively. The study demonstrates the significance of data filtering in forecasting univariate autoregressive series.