운전 데이터와 물리 모델을 활용한 디지털트윈 기반 건물 에너지시스템 최적 제어 = Digital twin based optimal control of building energy systems using physics-informed models with common operational data

다국어 초록 (Multilingual Abstract)

A In order to predict the energy consumed by the HVAC system during building operation and develop optimal conservation scenarios, it is necessary to develop learning-based models that can accurately describe the actual measurement environment. However, most buildings manage only partial data to regulate the indoor thermal environment, which makes it difficult to develop models with high predictive performance. In addition, heat sources are the most energy-consuming equipment in buildings, so it is important to manage and control them efficiently. However, a lot of data is required to calibrate the entire model, from load to heat source, to a level that can describe the actual physical behavior. Therefore, in the past, models were developed by limiting the scope from the indoor space to the air conditioner or installing additional sensors.
In this paper, an integrated system model that is organically connected from the load to the heat source is developed by using simulation results without installing additional sensors for a commercial building, and an optimal control scenario for energy saving is derived by applying reinforcement learning.
First, a physics-informed model was developed. The main components of the building system were partially calibrated using only air conditioning data, and the rest of the model was developed based on the rated performance of the blueprints and coupled in the TRNSYS platform.
As a result of evaluating the performance of the physics-informed model based on hourly room temperature and supply air temperature and daily fuel consumption, it was found that the prediction accuracy was greatly improved by direct calibration of the load side, but the prediction accuracy of fuel consumption was still insufficient. Therefore, a hybrid model development method was proposed to improve the prediction performance by using simulation results. The hybrid model was used to predict the next day's gas consumption of the boiler, and the CVRMSE was 12.3% and 10.5%, respectively, showing high prediction accuracy.
Because the physics-informed model directly calibrated the load side, it could better describe the physical behavior of the load, and the fuel consumption was more accurately predicted by the hybrid model. Therefore, connecting these two models effectively can develop a digital twin from the load to the heat source that can accurately predict the fuel consumption corresponding to the load. Digital twin can predict the gas consumption corresponding to the load, which is useful for simulating operational scenarios in advance and identifying energy saving measures before actual operation. In this process, a surrogate model was developed that allows the simulation results to be matched according to the control variables and used as an organic coupling between the two models.
Finally, reinforcement learning was applied to the digital twin with the surrogate model to derive an optimal control scenario for energy saving, with the goal of similarly managing the operating hours of the two boilers while maintaining the indoor comfort level and minimizing gas consumption by driving the boilers to operate optimally and efficiently. The analysis showed that the optimal control scenario could reduce gas consumption by about 23%.

목차 (Table of Contents)

• Overview 1
• 1. 서론 9
• 1.1. 연구 배경 및 목표 9
• 1.2. 연구방법 및 범위 14
• 2. 이론적 고찰 21
• 2.1. 기존 모델 개발 방법 22
• 2.2. Hybrid 모델 개발 방법 26
• 2.3. 디지털트윈 기반 건물에너지 관리 방법 30
• 2.4. 소결 34
• 3. 물리 모델 37
• 3.1. 개발 개요 39
• 3.2. 분석대상 건물 41
• 3.2.1. 건물 물성과 시스템 상세 41
• 3.2.2. 측정 데이터 상세 44
• 3.3. 모델 calibration 및 결과 46
• 3.3.1. 부하모델 calibration 46
• 3.3.2. 난방코일 모델 calibration 51
• 3.3.3. 전체 시스템 개발 55
• 3.3.4. 분석결과 59
• 3.3.4.1. 급기온도 분석결과 59
• 3.3.4.2. 실내온도 분석결과 63
• 3.3.4.3. 가스소비량 분석결과 67
• 3.4. 소결 70
• 4. Hybrid 모델 73
• 4.1. 개발 개요 76
• 4.2. 모델 학습 및 결과 80
• 4.2.1. 모델 학습 80
• 4.2.2. 분석결과 86
• 4.3. 소결 89
• 5. 디지털트윈 통합모델 91
• 5.1. 개발 개요 93
• 5.2. 통합 모델 개발 95
• 5.2.1. Surrogate 모델 95
• 5.2.2. 모델 결합 98
• 5.3. 분석결과 100
• 5.3.1. Surrogate 모델 성능 평가 100
• 5.3.2. 통합모델 성능 평가 101
• 5.4. 디지털트윈 특성 및 활용성 고찰 105
• 5.5. 소결 109
• 6. 디지털트윈 기반 최적제어 사례 113
• 6.1. 최적 제어 114
• 6.2. 분석결과 118
• 6.2.1. 최적제어 성능 평가 118
• 6.3. 소결 127
• 7. 결론 129
• Bibliography 133
• Appendix A 실내온도 hybrid 모델 개발 143
• A.1. 데이터 수집 및 전처리 147
• A.2. Hybrid 모델링 149
• A.3. Hybrid 모델 성능평가 152
• A.4. Black-box 모델과 성능비교 156
• A.5. 학습 범위 밖에서의 성능평가 157
• Appendix B 학습 영역 외에서 가스소비량 예측성능 평가 161