딥러닝 기반 HVAC 시스템 최적제어 알고리즘에 대한 효과 검증 및 확장성 향상에 관한 연구

Abstract

When developing energy prediction models or verifying the effects of optimal control algorithms in a simulation environment, it is difficult to accurately identify interactions between devices or changes in the indoor environment due to changes in control setpoints. In order to develop a machine learning model with high prediction accuracy for optimal control, it is necessary to secure operation data for various outdoor conditions, load conditions, and control settings. However, due to the unique characteristics of buildings and air conditioning systems, data acquisition for individual buildings is essential, and this process causes time and cost issues. Furthermore, developing a prediction model for optimal control is challenging as it is difficult to obtain actual operation data for all boundary conditions while varying the control setpoints. In this study, a physics-based heat source system simulation model was developed using Matlab Simulink and calibrated using actual operation data to improve accuracy. After generating sufficient learning data for all control settings using the calibrated simulation model, a DNN-based heat source system power consumption prediction model was developed using Python, and a highly accurate prediction model was secured through Bayesian optimization among hyperparameter optimization methods. Next, a cooling water temperature and flow rate control setting value optimization algorithm based on a DNN prediction model (hereinafter referred to as the cooling water algorithm) and a rule-based chilled water temperature control algorithm (hereinafter referred to as the chilled water algorithm) were developed. To verify the developed algorithm, it was applied to a target building located in Hanoi, Vietnam for a total of 244 days from October 27, 2023 to June 26, 2024, and the energy saving effect of the algorithm and its impact on indoor temperature and humidity were evaluated through comparative analysis of operation data before and after the application of the algorithm. The verified DNN prediction model was used as a source model to apply transfer learning to evaluate the possibility of expansion to other systems. At this time, the degree of improvement in prediction accuracy was derived and compared by focusing on fine tuning, a technique for relearning the weights of the source model among transfer learning methods.