PREDICTING ELECTRIC POWER DEMAND ON A UFMG UNIVERSITY CAMPUS: A STUDY USING RECURRENT NEURAL NETWORKS

Abstract

Energy forecasting is a widely used tool for energy management and planning, particularly given the scenarios of rising consumption and energy tariffs in Brazil. Fundamentally, this study aimed to forecast the electric energy demand of a UFMG university campus while evaluating the performance of three main recurrent neural network architectures: Classic RNN, LSTM, and GRU. A time series with daily granularity spanning 48 months was used, supplied by CEMIG. The models were implemented in Python, and performance was assessed using MAPE, MAE, and RMSE metrics. The proposed recurrent neural network models yielded similar results for daily energy demand forecasting; the GRU model stood out, achieving a MAPE of 10.51%. In the monthly demand analysis, the GRU model performed slightly better, with a MAPE of 5.44%, demonstrating greater accuracy in identifying monthly demand peaks. For this dataset—a time series spanning a relatively short period—simpler architectures proved competitive, with the GRU model showing slightly superior performance.