Prediction of energy consumption in campus buildings using long short-term memory

Abstract

In this paper, Long Short-Term Memory (LSTM) was proposed to predict the energy consumption of an institutional building. A novel energy usage prediction method was demonstrated for daily day-ahead energy consumption by using forecasted weather data. It used weather forecasting data from a local meteorological organization, the Malaysian Meteorological Department (MET). The predictive model was trained by considering the dependencies between energy usage and weather data. The performance of the model was compared with Support Vector Regression (SVR) and Gaussian Process Regression (GPR). The experimental results with a dataset obtained from a building in Multimedia University, Malacca Campus from January 2018 to July 2021 outperformed the SVR and GPR. The proposed model achieved the best RMSE scores (561.692–592.319) when compared to SVR (3135.590–3472.765) and GPR (1243.307–1334.919). Through experimentation and research, the dropout method reduced overfitting significantly. Furthermore, feature analysis was done with SHapley Additive exPlanation to identify the most important weather variables. The results showed that temperature, wind speed, rainfall duration and the amount had a positive effect on the model. Thus, the proposed approach could aid in the implementation of energy policies because accurate predictions of energy consumption could serve as system fault detection and diagnosis for buildings.

Keywords

• Energy consumption
• Long short-term memory
• Support vector regression
• Gaussian process regression
• Weather forecasting