Transfer learning of surrogate models for building energy prediction

Researchers are increasingly using machine learning models known as “surrogate models” to emulate complex physics-based building energy models (BEMs). Surrogate models are usually trained to predict the energy outputs of a single base BEM.

However, this paradigm requires training a new surrogate for every new base model, weather, etc. In this article, we study the effectivenss of “transfer learning”, where we pretrain a surrogate model to predict the annual heating and cooling loads of a source BEM, then retrain it using a small number of samples to represent a new target BEM. On average, our models are able to predict the outputs from the target BEM within 1 CVRMSE percentage point of standalone models, while using 125% fewer samples.