Experiment on Electricity Consumption Prediction using Long Short-Term Memory Architecture on Residential Electrical Consumer

Abstract

Renewable energy is an alternative for carbon-intensive energy sources that reduce global warming emissions. The electricity demand prediction helps to predict the consumption patterns on the demand side. The historical dataset of electricity usage is an essential source required to perform electricity prediction. This paper proposed the addition of independent variables that includes special days or holidays, weekend, seasons, and daylight duration into the basic electricity usage dataset that helps to increase the prediction accuracy. There were two datasets used in this study, basic electricity usage dataset that consists of date, time, and usage features, and extended electricity usage dataset that consists of the basic and independent variables features. Each dataset produced one model, basic model and extended model, respectively, from the training sessions conducted. The basic electricity usage dataset model was used as a benchmark to evaluate the quality of the model with extended features, extended model. Long-Short Term Memory (LSTM) was the selected machine learning architecture due to its ability to solve the regression problem in time series. All models produced were evaluated using two evaluation metrics, mean squared error (MSE) and mean absolute error (MAE). The application of the proposed methodology, LSTM with the proposed extended features had the lowest error rate with an MSE value of 0.1238 and an MAE value of 0.0388. These results showed that adding independent variables into the dataset improved the model generated from the training session.