Explainable multi-objective optimization for building retrofit considering uncertainty

Retrofitting existing buildings is crucial for promoting sustainable urban development. Nevertheless, uncertainty, speed, and explainability remain challenging in the performance simulations for retrofit decisions. Uncertainty originates from the randomness inherent in buildings and gaps in knowledge, which includes thermal performance parameters, occupant behaviour schedules, and others. This study proposes an explainable multi-objective optimization framework for building retrofit considering uncertainty, which utilises measured data, optimization algorithms, and simulation software. The RBEOpt algorithm was employed to rapidly calibrate 13 uncertainty parameters. The retrofit objectives encompassed the energy use intensity (EUI), thermal comfort duration (TCD), and net present value (NPV). Multi-objective optimization was performed utilising the NSGA-II algorithm, with seven widely implemented building retrofit measures (BRM) being selected. The data generated during the multi-objective optimization was applied to train six rapid prediction models and the best-performing XGBoost model was subsequently elucidated by SHAP analysis. Experimental results demonstrated that model calibration has the potential to substantially reduce the uncertainty of simulation models. A 42.75 to 110.74 kWh/m2 reduction in the EUI of each building is possible, whilst a 2.90–13.01 % increase in the TCD can be expected. Furthermore, a substantial number of Pareto solutions generate an NPV that exceeds zero. Simultaneously, the complex relationship between BRMs and retrofit objectives is elucidated. The proposed method calibrated uncertainty parameters and trained rapid predictive models, significantly enhancing the accuracy and explainability of multi-objective optimization results for building retrofits. The research findings can provide valuable data and technical support for urban regeneration and building retrofit policies.