Building Energy Prediction and Design Optimisation Using EnergyPlus and Machine Learning

Window-to-wall ratio, wall insulation, shading and orientation are fixed early in architectural design, when a full energy simulation of every option is too slow. This project trains machine-learning surrogates on its own EnergyPlus simulations of a five-zone office in Chennai and New Delhi weather, so that the annual energy use intensity (EUI) of a design option is known in milliseconds with an uncertainty band, and searches the trade-off between EUI and an illustrative cost index with the NSGA-II genetic algorithm - and then re-simulates the Pareto designs in EnergyPlus to see whether the surrogate told the truth. On the public UCI Energy Efficiency data it also shows how a random split overstates accuracy for building forms a model has never seen. A web application evaluates design options, runs the optimisation, checks designs with EnergyPlus and keeps a review workflow, on five databases. An OpenStudio model of the same office, a GNU Octave cross-check and a complete MATLAB version are included.

Category: School of Architecture and Planning

Architecture & Planning Project AP01 from ProTechSC Academy, with graphical abstract, video demonstration and project guidance.