Bias prediction vs. parameter Estimation: Calibration Framework comparison for urban Building energy modeling considering application scenarios

Abstract

Accurate Urban Building Energy Modeling (UBEM) is always a prerequisite during the design of a city's climate actions, plans and renewals, but is often subject to limit information of buildings and usage in large scale. Thus, researchers often simplify the UBEM models and improve the accuracy through calibration. Result-based bias prediction (BP) and parameter-based estimation (PE) are the two of the most-commonly-used calibration frameworks, but the data requirements, capability boundaries, and applicable scenarios are still unclear in previous studies due to the insufficient sample space. Therefore, the study carried out a systematic comparison between the two calibration frameworks through a case study on 3442 public buildings with 256,968 monthly electricity consumption records. Nine types of buildings' energy consumption were modelled and simulated, in which the PE provided calibrated archetype models, and the BP provided the prediction on primary simulation bias. As a conclusion, the BP showed better accuracy, generalization, and robustness crossing different tasks than the PE did, and so would be more feasible for most of the analytical simulations focusing on an accurate energy assessment, especially those in building maintenance or Urban Sustainability Assessment scenarios. The PE was suggested if the persistence on model interpretability to explain the composition or sensitivity of energy consumption overperform the accuracy, which is widely required in design stage. In addition, the implementation of Transfer Learning could improve BP targeting at mix-use building simulations. This study contributes to the selection of calibration methods cross the main UBEM applications.