Abstract
•Real-time energy use prediction for complex residential precincts aids quick design decisions.•ML and DL surrogate models eliminate the need for time-intensive energy simulations.•Hourly energy predictive accuracy improves at monthly and annual resolutions.•MLP model offers the highest accuracy; HGBoost achieves similar results with faster training.•Study presents an end-to-end workflow for surrogate modeling of residential precincts.
Machine learning and deep learning have increasingly replaced computationally intensive simulations in building energy performance assessments, expediting early-stage decision-making processes. However, predicting urban-scale energy performance at an hourly resolution throughout the year for heterogeneous thermal zones remains complex. This study proposed a surrogate modeling framework to predict annual site energy consumption at an hourly resolution for high-rise, high-density residential precincts in a tropical city, supporting progress toward global net-zero energy (NZE) goals.
Using Latin Hypercube Sampling (LHS), 2000 unique Singaporean public housing precinct designs were generated through a novel, in-house developed parametric model incorporates urban-scale planning features such as density, block spacing, and inter-building shading. Simulations produced a dataset of 17.5 million hourly energy consumption records and 23 independent variables—comprising climate data, spatial logic and model parameters—with energy use (EUse) as the target variable. Data preprocessing, including exploratory data analysis and feature importance evaluation, helped refine the dataset.
Seven surrogate models—Linear Regression, LASSO, Ridge, Elastic-Net, Histogram Gradient Boosting (HGBoost), Multilayer Perceptron (MLP), and Recurrent Neural Networks (RNN)—were developed. Among them, MLP achieved the highest accuracy (MAPE: 18.1% hourly, 3.3% monthly, and 3.5% annually), while HGBoost ranked second with one-fourth the training time of MLP. This study demonstrated that rapid, early-stage NZE analysis is possible without the need for an urban-scale thermal model. The novelty lies in combining urban-scale spatial complexity, climatic variables, and multi-zone thermal logic into a scalable ML workflow. A well-performing surrogate model enables quick, informed decision-making during early design stages by supporting computationally intensive tasks—such as iterative design optimization, sensitivity analysis, and uncertainty analysis—at a significantly reduced computational cost.