Experimental Investigation and Data-Driven Modeling for Variable Refrigerant Systems
by Po-Ching Hsu
Variable refrigerant flow (VRF) systems have gained widespread adoption in commercial and residential buildings due to their high part-load efficiency, zoning flexibility, and superior thermal comfort performance. As buildings account for a substantial portion of global energy consumption and carbon emissions, improving the modeling, control, and operational efficiency of VRF systems has become increasingly important. However, the growing structural and operational complexity of modern VRF systems, characterized by variable-speed compressors, multiple indoor units, electronic expansion valves, and heat recovery configurations, poses significant challenges for accurate system modeling and advanced control implementation. Traditional curve-based models, while computationally efficient and physically interpretable, often lack sufficient control-oriented inputs and struggle to capture transient dynamics. Conversely, purely data-driven models demonstrate strong predictive performance but may suffer from limited physical consistency, robustness, and generalizability. These limitations highlight the need for a hybrid modeling framework that combines the strengths of both approaches.
This dissertation presents a comprehensive experimental and modeling investigation of VRF systems based on long-term field test data. A multi-functional VRF (MFVRF) system installed in a campus office building was instrumented to capture detailed operational data across diverse seasonal and operating conditions. The resulting database enabled exploratory analysis of system behavior, part-load performance, and the impact of control strategies on energy consumption and thermal comfort.
Building on this dataset, multiple data-driven models were developed to capture VRF system behavior, including decision trees (DTs), artificial neural networks (ANNs), and long short-term memory (LSTMs). A systematic evaluation framework was established to assess predictive accuracy, physical consistency, computational efficiency, and model compactness. Results demonstrate that deep learning models, particularly LSTM-based architectures, effectively capture transient and nonlinear system behavior, achieving higher predictive accuracy than shallow machine learning (ML) models. To enhance robustness and physical interpretability, a hybrid modeling framework was proposed by integrating a modified VRF-SysCurve physics-based model with ML models such as ANN. The hybrid structure preserves key thermodynamic relationships while incorporating control-oriented variables, enabling improved predictive accuracy and greater flexibility for control applications. Transfer learning strategies were implemented to fine-tune sub-models for underrepresented indoor units using limited additional data, thereby improving generalizability and data efficiency. Compared with standalone curve-based and purely data-driven models, the proposed hybrid model demonstrates superior physical consistency and reduced performance degradation under extrapolated conditions.
The developed hybrid model was further embedded within a model predictive control (MPC) framework to evaluate advanced control strategies. The MPC implementation effectively regulates control inputs to reduce power consumption while maintaining thermal comfort. Pareto-front analyses reveal the trade-offs between energy savings and comfort objectives across cooling and heating seasons.
Overall, this research establishes an end-to-end digital-twin framework for VRF systems that integrates field experimentation, data-driven modeling, hybrid physics-informed modeling, and optimized control. The proposed framework advances the accuracy, robustness, and practical deployability of VRF modeling for real-time control and smart building applications.
Future work will enhance data representativeness for underutilized indoor units (IDUs) through targeted experiments to improve model generalization in rarely observed operating regions. The scalability of the proposed framework will be evaluated across VRF systems with different configurations. Multi-domain co-simulation integrating building thermal and distribution-network models will be explored to assess demand-side flexibility and grid-interactive performance. Finally, field implementation of the proposed MPC strategy will be conducted to validate energy savings, comfort impact, robustness, and operational reliability under practical conditions.
Doctoral Dissertation
Top