A Surrogate Modelling Framework for Time Resolved Energy Prediction in High Performance Office Buildings

Abstract

High-performance buildings are an important component of the transition toward low-carbon and net-zero energy systems. However, evaluating building energy performance across a wide range of design and operational conditions typically requires large numbers of detailed simulations, which can be computationally expensive. This research investigates the energy behavior of high-performance office buildings and develops a surrogate modelling framework capable of predicting time-resolved energy consumption using machine learning techniques. The study follows a four-phase methodology. In the first phase, operational data from a high-performance office building located in Waterloo, Ontario are analyzed to examine daily energy-use behavior and identify the primary operational drivers of electricity consumption. This analysis provides empirical insight that informs the selection of key parameters used in subsequent modelling stages. In the second phase, a parametric physics-based building energy model is developed in EnergyPlus to represent office buildings operating under Waterloo’s climate conditions. The model incorporates variations in building geometry, envelope properties and configurations, and internal loads to represent a range of plausible design and operational scenarios. In the third phase, the parametric model is used to generate a synthetic dataset through systematic sampling of the input parameter space. The resulting dataset contains hourly energy simulation outputs across a wide range of building configurations and serves as the training and evaluation data for surrogate models. In the fourth phase, machine learning algorithms are developed, trained, and evaluated to predict building energy performance, including hourly electricity consumption and annual performance indicators. The trained models are also used to examine the relative influence of key building parameters on predicted energy outcomes. Three surrogate architectures were evaluated: a fully connected artificial neural network (ANN), a convolutional neural network (CNN), and a hybrid ANN–CNN model. The ANN showed stronger performance for annual energy prediction but limited accuracy in reproducing hourly temporal patterns, while the CNN captured hourly variations more effectively but produced less accurate annual energy estimates. The hybrid ANN–CNN model combined these complementary strengths, achieving comparable hourly prediction accuracy to the CNN and similar annual prediction performance to the ANN. The hybrid model achieved an hourly RMSE of approximately 12.89 kWh and an annual energy prediction RMSE of 1.08 kWh/m²·yr, providing the most balanced overall performance among the evaluated architectures. However, the surrogate models showed a tendency to underestimate peak energy demand, indicating limitations in capturing short-duration peak loads.