Buildingenergy.ninja: A web-based surrogate model for instant building energy time series for any climate.
Abstract Machine learning-based surrogate models are trained on building energy simulation input and output data. Their key advantage is their computational speed allowing them to produce building performance estimates in fractions of a second. In this work we showcase the use of deep convolutional neural network surrogate models embedded into a web application, allowing users to rapidly explore building performance at high spatio-temporal resolution. Users can pick any climate on an interactive map, customize a building design with thirteen decisive design parameters, and the surrogate model allows them to retrieve hourly heating and cooling load time series data in fractions of a second. In this work, we further show that the surrogate model reaches an accuracy of R 2 > 0.93 (MAE < 0.27 kWh) for unseen design specifications and climates. These results motivate the use of computationally cheap surrogate models to replace building energy simulation for a wide variety of tasks in the future.