Short-Term Residential Electric Power Consumption Forecast Using Machine and Deep Learning Algorithms

Abstract

The objective of this research is to create prediction models for electric power consumption, in the short term, using the multiple linear regression algorithm and an artificial neural network. It is a descriptive-explanatory type of research, and the methodology of data science is used to achieve the stated objective. The models are trained and evaluated using monthly invoiced electricity data in Uruguay, during the period 2000-2022, in addition to measurements of ambient temperature and relative humidity, as well as the percentage variation of the consumer protection index. Through correlation analysis and principal component analysis, the importance of the selected explanatory variables is verified, prior to obtaining the models. The linear regression model had an R2 of 0.782 and the neural network model of 0.797, the residuals were normally distributed for both models, and the RMSE, MAE, and MAPE metrics had similar values in both models as well. The neural network model had a better performance, compared to the linear regression model, regarding the prediction of electrical energy for the months of July and August of the year 2022.