Surrogate modelling methodology for predicting annual site energy for single-family wartime bungalow archetypes in Toronto

Current rates of retrofitting residential buildings are inadequate for meeting goals set by governments to reduce emissions. Plans, laws, policies, programs, and regulations must be tailored, yet adaptive to changes in technology, and have a way to communicate the evidence behind them. Surrogate models allow for accurate predictions much quicker than traditional brute-force methods with only small amounts of error. This research presents a methodology for developing a surrogate model that takes advantage of ANN’s ability to make accurate predictions. Modelling is supported through analysing the alignment of input and output interactions with building science knowledge through SHAP values. Surrogate models developed using this methodology were able to predict the total annual site energy of an archetype representing 21% of Toronto’s single-family housing stock with an average of 6% error, equivalent to 2.7 GJ per year. This research is another example of how the increased interpretability of black-box models support further development of surrogate models to solve optimisation problems such as developing effective retrofitting plans and policy strategies to accelerate rates of retrofitting existing housing.