Applying neural networks for short and long-term hourly electricity consumption forecasting in universities: A simultaneous approach for energy management

Abstract

Predictive models for energy consumption are increasingly relevant for buildings with mediumhigh demands, such as public buildings. These models enable effective energy management, especially in contexts of budget constraints, maximizing resource efficiency. Thus, university buildings are interesting cases because they meet the above-mentioned conditions, they can be managed by the public or private sector and have high societal impact. This work aims to simultaneously develop two predictive models for hourly electricity demand in a university building: one for the short-term (next hour) and one for the long-term (all hours of the year). The methodology consists in: analyzing the data using descriptive statistics to identify relevant variables; building the neural network models in Python; and evaluating them with performance indicators. The short-term model, which includes the previous hour’s energy as independent variable, achieved a relative error of 16.0 %, an R2 = 91.17, RMSE = 0.512 kWh, and MAE = 0.325 kWh. The long-term model, which omits this variable, showed a relative error of 31.7 %, R2 = 74.69, RMSE = 0.864 kWh, and MAE = 0.566 kWh. The study highlights the effectiveness of neural networks in energy forecasting. It also emphasizes the importance of using the previous hour’s energy as a variable to obtain short-term accuracy; but also the possibility of building models without it, allowing for long-term approaches. The novelty lies in the simultaneous approach of two predictive horizons in a university building. This approach can contribute by providing accurate information to energy control algorithms that integrate renewable energy, storage, grid power, etc., to optimize the use of resources and improving environmental sustainability.