Development of a simulation-based ANN framework for predicting energy consumption metrics: a case study of an office building

Abstract

Building energy performance assessments are essential in High-Performance Building Design (HPBD) in order to reduce energy consumption and carbon emissions. With advancement in data analytics, rapid and accurate machine learning-based building energy consumption prediction models have emerged. These models can be used by non-professionals as an alternative to time-consuming energy simulation software, offering benefits in HPBD. Therefore, the main objective of the present study is to develop a prediction model using data generated by physics-based simulations of a typical open-plan office space. The model predicts annual energy consumption, CO₂ emissions, and percentage of comfort hours during the design phase. Various configurations of Artificial Neural Network (ANN), Support Vector Machine (SVM), Random Forrest (RF), and K-Nearest Neighbors (KNN) algorithms were trained and tested with the generated data via the DesignBuilder software. The technical parameters considered as inputs include U-values of envelope constructions, Window to Wall Ratio (WWR), orientation, and Heating, Ventilation, and Air-Conditioning (HVAC) systems. The results indicate that there is no clear linear relationship between individual inputs and the target indicators. However, ANN, with its ability to handle non-linear relationships, performed the best, achieving a maximum Coefficient of Determination (R²) value of 0.997 for predicting percentage of comfort hours and outperforms the other algorithms. Furthermore, the results show that RF is the next best algorithm, with 0.96 ≤ R²_Test ≤ 0.98 for predicting the various target variables. SVM with Radial Basis Function (SVM-RBF) follows, with 0.89 ≤ R²_Test ≤ 0.95. Contrary to ANN, SVM, and RF algorithms with high abilities to learn complex pattern between various independent parameters and the target variable, KNN exhibits the poorest performance, with 0.88 ≤ R²_Test ≤ 0.91. Additionally, it is observed that with a maximum time cost of 619 s, ANN with three layers is able to learn the relationships between the inputs and target indicators at a convenient speed. Since knowledge-based decision making in the early design stages is crucial for achieving the optimum solutions to reduce energy consumption and related CO₂ emissions while ensuring occupants’ comfort and minimizing future modifications and costs, high-speed and accurate prediction methods for design stage evaluation are essential.