This paper examines the applicability of neural network and DOE-2 models to predict hourly electrical demand and cooling load, as well as monthly electricity consumptions and cooling energy in a complex institutional building. Three-layered feed-forward neural networks trained with a fast non-iterative training algorithm were used in this study. Neural networks were optimized by removing redundant inputs. Possibility of improving the prediction accuracy by day typing was examined. Comparative results show that neural network models could predict hourly cooling load and electrical demand more accurately than DOE-2 simulations. However, DOE-2 simulations can give reasonably accurate estimations of monthly electricity consumption.