CityTFT: A temporal fusion transformer-based surrogate model for urban building energy modeling

Almost one-third of global greenhouse gas emissions are from buildings and nearly 70 % of energy is consumed in urban areas. To address this as a long-term issue, accurately predicting building energy use at the urban level is indispensable. Urban Building Energy Modeling (UBEM) is an emerging method to investigate urban design and energy systems at urban and neighborhood levels. UBEM methods are considered reasonably accurate in simulating the performance of almost any building combinations. However, customized urban projects, and optimization problems involving many UBEM scenarios, are time-consuming and labor-intensive, and based on the scalability, the simulation runtime can be exponentially high if a broad set of design variations is analyzed. Here, we propose a data-driven approach to generate a surrogate model for UBEM and accelerate the simulation process. Based on the extensively used forecasting model, Temporal Fusion Transformer (TFT), we extract the static covariate encoder and variable selection network from the TFT structure while adding a small neural network to model the probability of triggering heating and cooling needs. The major advantages compared to other surrogate models are: (i) A sequential input in the transformer-based model to improve the temporal accuracy. (ii) A physics-inspired training strategy that builds on weather dynamics and urban interactions for hourly energy demands concurrently with a customized loss function. (iii) Improved generalizability for the proposed surrogate model using open and public data. In our study of 114 buildings in different climate zones in the US, our model predicts heating and cooling triggers in unseen climate dynamics with an F1 score of 0.83 while the root mean squared error (RMSE) of hourly loads was 71.51 kWh. This model is available for interfacing with climate projections and city-scale analysis as part of atmospheric urban digital twins.