Digital twin model calibration of HVAC system using adaptive domain Nelder-Mead method

Digital twin control enhances HVAC systems by enabling real-time monitoring, energy optimization, and maintenance for improved performance and energy efficiency. To perform digital twin control effectively, it is necessary to have an accurate simulation model that can predict the performance of HVAC systems. However, simulation models have limitations in replicating real operating conditions, requiring calibration using actual measurement data. HVAC systems vary in operating environments, necessitating on-site embedded systems for customized operational optimization. However, existing research utilizing physics-based simulation models and metaheuristic algorithms incurs high computational costs, making practical implementation challenging. In this study, we developed the adaptive domain Nelder-Mead (ADNM) method, which optimizes calibration coefficients of the simulation model to minimize the error between predicted values and measured data in a short computation time. ADNM constructs an initial simplex using the Sobol sequence and previous optimal solutions, enabling robust minimization of the objective function even with changing operating conditions. The proposed method, validated with real HVAC data, reduced temperature RMSE by 60-70 % compared to the traditional NM algorithm and demonstrated up to 10-60 % lower temperature error than Bayesian optimization (BO) at similar computational times. Although genetic algorithm (GA) achieved lower errors, its computational time was over 3 to 6 times higher, highlighting the proposed method's superior efficiency and accuracy.