Machine Learning Baseline Energy Model (MLBEM) to Evaluate Prediction Performances in Building Energy Consumption

Abstract

Electric Energy Consumption (EEC) prediction for building operations can be performed using a Baseline Energy Model (BEM), which is vital to ensure the efficiency of the EEC estimates with its respective independent variables. However, developing the BEM to represent the relationship between independent variables can be a complex task due to the EEC variability in an educational building that differs during its operation period. The best-suited BEM must be continuously improvised to achieve good modeling with accurate and reliable predictions that capture the building operations’ current dynamics. This study aims to conduct a comparative performance assessment between deep learning, machine learning, and statistical models to develop the BEM and, therefore, predict the EEC of the building for 24, 48, 72, and 96 hours, while considering the operation of the lecture weeks and the associated number of students and staff. The hours and temperature are considered as independent variables to be tested with residual error evaluations, whilst the correlation coefficient, coefficient of determination, and training time are also taken into account. Three models with different categories involving Long Short-Term Memory (LSTM), Support Vector Regression (SVR), and AutoRegressive Integrated Moving Average with Exogenous inputs (ARIMAX) were compared, concluding that SVR was the best and can be used as a universal model in the Machine Learning Baseline Energy Model (MLBEM) studies. Accurate EEC prediction will offer a huge advantage for building operators to properly monitor, plan, and manage the EEC, hence avoiding excessive utility bills.