In large-scale air-conditioning heat supply systems, such as district heating and cooling (DHC), multiple heat source units are typically operated, with manual control based largely on operator experience. This study aims to develop an energy management system that optimizes operational planning for such facilities. Central to this initiative is the construction of a high-fidelity energy consumption estimation model for each heat source unit, leveraging actual operational data within a machine learning framework. To enhance model accuracy, the learning process systematically addresses operational biases, enabling precise characterization of unit-specific performance even when historical data are limited or unbalanced. Beyond predicting gas consumption based on cooling water temperature and part-load ratio, the proposed methodology explicitly quantifies the sensitivity of these variables and incorporates a simulation-driven data augmentation strategy. This approach expands the effective learning domain, allowing the machine learning model to capture operational characteristics across a broader spectrum of conditions, including those infrequently observed in practice. By integrating the developed performance model with forecasts of cooling water temperature and heat demand, this study introduces a method for estimating future energy consumption and generating optimized operation schedules that account for anticipated system dynamics. The application of the proposed framework to real plant data revealed that conventional manual operations often favored less efficient units, whereas the proposed approach achieved a potential reduction in energy consumption of approximately 22.7%. These results highlight the significant impact of accurate performance modeling on subsequent optimization.
Okuda et al. (Tue,) studied this question.