전이학습 기반 제어 설정값 최적화 알고리즘의 확대 적용 및 실증 연구 = Extended Application and Empirical Study of a Transfer Learning-Based Control Setpoint Optimization Algorithm

다국어 초록 (Multilingual Abstract)

Expanding artificial intelligence–based HVAC control algorithms from one building to another is challenging due to differences in system configurations, operating conditions, and the limited availability of training data in newly applied buildings. In control applications, prediction models must maintain not only high accuracy but also stable prediction tendencies, as their outputs directly determine control setpoints and influence actual system operation and energy consumption.
This study investigates the scalability of a deep learning–based cooling water temperature optimization algorithm using a transfer learning approach. A prediction model trained with operational data from a source building was adopted as a source model, and its learned weights were fine-tuned using limited data from a target building to construct a target prediction model. The reconstructed model was integrated into the control algorithm and applied to an actual heat source system for performance verification.
The algorithm was implemented during a cooling season over approximately five months, with algorithm-based operation and conventional operation alternated on a biweekly basis to reduce the influence of varying boundary conditions. Cooling water temperature behavior, steam consumption, electricity consumption, and total operating cost were analyzed using measured data. The energy-saving performance was quantitatively evaluated based on the International Performance Measurement and Verification Protocol Option C with a regression-based baseline model.
The results show that the proposed algorithm enabled dynamic cooling water temperature control in response to outdoor conditions, unlike the fixed setpoint operation of the conventional method. Although electricity consumption increased due to enhanced cooling tower operation, steam consumption decreased significantly as a result of improved heat source system efficiency. Consequently, the total operating cost was reduced, achieving an overall cost reduction of approximately 6.4 percent. These findings demonstrate that transfer learning–based prediction models can effectively support the expansion of AI-based HVAC control algorithms to new buildings with limited data availability.

목차 (Table of Contents)

• 제 1장 서 론 1
• 1.1 연구 배경 1
• 1.2 연구 동향 6
• 1.3 연구 목적 8
• 제 2장 연구 방법 11
• 2.1 전이학습(Transfer Learning) 11
• 2.1.1 특성 추출(Feature Extraction) 14
• 2.1.2 미세 조정(Fine-Tuning) 15
• 2.1.3 하이퍼파라미터 최적화 18
• 2.2 대상 건물 및 HVAC 시스템 20
• 2.3 HVAC 시스템 설정값 최적화 알고리즘 26
• 2.3.1 운전데이터 구성 및 입력 처리 27
• 2.3.2 알고리즘 주기 설정 28
• 2.3.3 최적 설정값 산출 과정 28
• 2.3.4 제어시스템 연동 30
• 제 3장 전이학습 기반 열원시스템 운전비용 예측모델 개발 31
• 3.1 Source 모델 31
• 3.1.1 Source 모델 개발 32
• 3.1.2 Source 모델 구성 35
• 3.1.3 Source 모델 성능 37
• 3.2 Target 모델 39
• 3.2.1 Target 모델 개발 40
• 3.2.2 Target 모델 구성 43
• 3.2.3 Target 모델 성능 46
• 제 4장 알고리즘 실증 결과 50
• 4.1 효과 검증 방법 50
• 4.2 알고리즘 실증 결과 52
• 4.2.1 냉각수 온도 비교 53
• 4.2.2 운전비용 절감효과 비교 56
• 제 5장 결 론 62