Forecasting the impact of climate change on thermal comfort using a weighted ensemble of supervised learning models

Abstract This work presents a weighted ensemble of supervised learning models for estimating the days of the year where compliance with the Adaptive Model of thermal comfort is exceeded. The ensemble combines several gradient boosting on decision tree algorithms and Bayesian logistic regression. In a presented case study, the model is trained on three summers of hourly weather data and indoor air temperature data of a south-facing, naturally-ventilated office in Vancouver, Canada. The model is then used to predict thermal comfort exceedance under possible climate change scenarios. It is found that the ensemble outperforms its individual models in terms of accuracy, precision, and similar metrics. In eight of nine trials using the ensemble to re-assess known history, the ensemble predicts total comfort-exceeding days within a margin of one day. Under the RCP 8.5 global climate change scenario, the model predicts annual comfort-exceeding days will double by the 2050s, by that point exceeding current local thermal comfort compliance guidelines. Future applications of the presented methodology may assist other areas of data-driven forecasting, such as peak energy demand prediction. It may also assist analysis of emerging space cooling approaches such as radiant cooling of free-running buildings.