Predictive artificial neural network models to forecast the seasonal hourly electricity consumption for a University Campus

Abstract

This paper proposes artificial neural network (ANN) models to forecast the seasonal hourly electricity consumption for three areas of a university campus, Japan. A total of six parameters including day of week, hour of day, hourly dry-bulb temperature, hourly relative humidity, hourly global irradiance, and previous hourly electricity consumption are used as input variables. The ANN models are developed to predict the future seasonal hourly electricity consumption for the three areas, considering the Feed-forward ANN trained with the Levenberg-Marquardt (LM) back-propagation algorithms. The correlation coefficient (R2) and root mean square error (RMSE) metrics are adopted to evaluate the accuracy of proposed ANN models. It showed that the R2 between actual measurement and ANN models prediction ranges between 0.96 and 0.99 at training stage, and between 0.95 and 0.99 at testing stage. It showed that RMSE in Science and Technology area of the university campus is the largest among three areas, followed by Humanities College area. The Old Liberal Arts area has the smallest RMSE and its difference. As future work, more input variables such as class schedule and indoor human activities etc. are suggested to be added to the input layer of ANN models for improving their forecast accuracy.