A Behavioral-Based Machine Learning Approach for Predicting Building Energy Consumption

In recent years, artificial intelligence (AI) techniques, and in particular machine learning (ML), have been adopted for forecasting building energy consumption and performance. This data-driven approach relies heavily on either: (1) real data collected through energy meters and sensors or (2) simulated data modeled via building energy simulation tools such as EnergyPlus. However, both types of data suffer from several deficiencies that hinder the full potential of the learning algorithm. On one hand, real-data include noise, missing values, and outliers which affect the performance of the prediction models significantly. On the other hand, simulated data is affected by predefined conditions making its forecasting often inaccurate. To address these shortcomings, this paper presents an amalgamation of a behavioral-based simulation and a machine learning algorithm that can be ideally used during the early design stages, or for an existing building where real data is limited due to technical or economic difficulties. A parametric and behavioral analysis is first performed using agent-based modeling (ABM) to predict the hourly energy consumption of an office space under design. An artificial neural network (ANN) model is then trained with the simulated data and tested against the total energy consumption for an existing office having the same parametric features. Results confirm the potential of the proposed hybrid model in accurately predicting information about the patterns governing energy demand.