Machine learning based prediction of long-term energy consumption and overheating under climate change impacts using urban building energy modeling

In cities, well-informed decisions targeting improved building energy performance under climate change impacts require tools that can make long-term projections and devise effective strategies. Physics-based Urban Building Energy Models (UBEM) can calculate building performance for future years; however, this process is challenging as (i) future weather files are generated only for discrete years, and (ii) physics-based simulations are computationally demanding, hindering the evaluation of a high number of buildings for all future years. Alternatively, machine learning (ML) approaches can offer high-precision estimations at a lower computational cost. In this paper, a UBEM-assisted ML-based approach that predicts residential buildings' heating energy use and indoor overheating for the current and future years is proposed. A UBEM of a residential district is developed, and simulations are performed using weather files of the current year, 2050, and 2080 to develop training/testing datasets. Multi-layer perceptrons are trained to a very high predictive performance (with an R-2 score of 0.98 and 0.96 for the two output features), with a remarkable speed advantage (similar to 430 times faster than simulations). Finally, the results of the long-term analysis of three urban-scale retrofit scenarios are presented, which offers insights into the potential use of the proposed ML models.