Abstract
The study presents a proposed approach towards developing the core engine for a simplified Rapid Overheating ASSessment Tool (ROASST), which is intended to help assist early-stage analysis of the risks of indoor overheating for apartments located in Greater London. Using a discrete number of plan forms selected from case studies, a virtual risk database was populated with the results of a large number of parametric dynamic thermal simulations based on the EnergyPlus calculation engine and including aspects such as location within Greater London, orientation, fenestration size and natural ventilation, which are associated with known overheating risk factors. Alternative statistical meta-models were developed with both explanatory and predictive purposes, correlating the simulation input with the overheating risk predictions expressed according to multiple metrics. Results from multiple linear regression analysis show that while all factors considered are relevant towards determining the propensity to overheating, window opening and natural ventilation capacity are by far the strongest predictors among those considered. The implementation of machine learning algorithms is shown to improve the accuracy of the meta-model, producing very high coefficients of determination (R2) and lower prediction errors (RMSE). The development of a meta-model demonstrates the ability of returning accurate predictions with limited input, albeit with significant limitations. Possibilities of further improvements to the tool are briefly outlined, including the coupling with a User Interface for applicability in a design environment for early-stage design advice.