Predict the element instead of the sequence: ResNet surrogate method for very accurate predictions of hourly building energy

In building energy modelling, surrogate models (such as artificial neural networks) are usually trained to predict the annual or monthly energy consumption of physics-based building energy models (BEMs), and they usually show high accuracies. But predicting hourly consumptions is becoming increasingly important, as we tackle problems like predicting peak loads on electricity grids or calibrating BEMs with detailed measured energy data. Only a handful of studies address hourly predictions, but they show either low design space complexity or low accuracy. In this article, we develop an hourly surrogate modelling method that greatly outperforms other methods in the literature. Traditionally, surrogates of BEMs use one large sequence of data (such as the hourly weather over a year) to predict a large sequence of energy consumptions/loads. Instead, our large convolutional residual neural network (i.e., ResNet) uses several small sequences of data (the weather, building parameters, and schedules over the past 8 hours) to predict individual hours of heating and cooling loads. We apply our “predict the element instead of the sequence” method to a medium-sized office BEM case study under two weather scenarios: (1) training and testing on a single weather and (2) training on 112 climatically diverse weather files and testing on 30 unseen ones. Our model has R2 values above 0.9999 in the first scenario and R2 values around 0.99 in the second, thereby demonstrating that our model can generalise to any weather around the world. Our model’s largest errors occur during the hours of the weekday when the setpoint temperatures transition between their nighttime and daytime values. By focusing on these hours in future work, we can improve our surrogate model’s performance even further.