Summary
This webinar presents a comprehensive exploration of Mitsubishi Electric’s innovative approach to refrigerant mass estimation in HVAC systems, particularly air conditioners and heat pumps. Traditional measurement methods in this domain tend to be invasive or inaccurate, but Mitsubishi’s use of the Julia programming language, JuliaSim software platform, and machine learning techniques enables highly precise, non-invasive predictions with less than 2% error.
The session opens with introductions of the speakers—Chris Laughman from Mitsubishi Electric Research Laboratories and Avinash Subramanian from JuliaHub—who bring expertise in multiphysical system modeling and energy process engineering, respectively. The presentation outlines the motivation for digital twins in HVAC systems and the challenges they face, such as dynamic behavior, nonlinearities, and multiscale time dynamics.
JuliaSim is introduced as a unified modeling environment built on Julia that integrates multiple engineering workflows—modeling, control, machine learning, and optimization—into a single, interoperable platform. Unlike traditional HVAC modeling tools that are fragmented or proprietary, JuliaSim combines symbolic and numeric computation with domain-specific libraries tailored for vapor compression cycles.
Chris elaborates on the concept of digital twins as real-time, data-informed virtual replicas of physical HVAC assets, emphasizing that these models must continuously update with operational data to track system performance and detect faults. The digital twin approach goes beyond static simulation by assimilating live data to maintain model fidelity and offer insights into unmeasurable system states such as refrigerant mass.
The session highlights key challenges in HVAC modeling: the highly nonlinear and coupled nature of vapor compression cycles, wide-ranging time scales from milliseconds to days, and data richness but information sparsity—where the available sensor data often lacks sufficient excitation for robust model calibration. JuliaSim addresses these issues through pre-built, high-fidelity components like compressors, valves, and heat exchangers with accurate refrigerant thermodynamics.
A live demonstration showcases JuliaSim’s textual interface and component-based modeling capabilities, emphasizing flexibility for both industrial users (who prefer drag-and-drop modeling) and power users (who require access to advanced features like automatic differentiation and solver customization). The platform supports efficient dynamic simulation, model calibration, and control design.
Chris further discusses the innovative achievement of refrigerant mass estimation using minimal sensor inputs—primarily temperature and pressure measurements—without the need for intrusive or costly mass flow measurements. This is enabled by robust state estimators built on JuliaSim’s solver transparency and automatic differentiation, which allow the system to maintain physical constraints while assimilating data in real time.
The presentation concludes by reflecting on the broader implications of this technology, particularly its potential to reduce environmental impact by early leak detection and improve energy efficiency through enhanced digital twins. The integration of machine learning with physics-based modeling is seen as a promising direction for advancing HVAC system design, operation, and grid interaction.
Highlights
Mitsubishi Electric leverages JuliaSim and machine learning for non-invasive refrigerant mass estimation with less than 2% error.
JuliaSim offers a unified platform integrating modeling, control, machine learning, and optimization workflows for HVAC systems.
Digital twins dynamically update models with real-time data, enabling performance tracking and fault detection in HVAC assets.
High-fidelity vapor compression cycle components with accurate refrigerant thermodynamics enable realistic simulations.
JuliaSim supports both drag-and-drop industrial users and power users requiring advanced numerical and symbolic tools.
Minimal sensor input (temperature and pressure) is sufficient for robust refrigerant mass estimation via advanced state estimators.
Accurate leak detection and system monitoring can mitigate environmental impacts of refrigerants with high global warming potential.
Key Insights
Digital Twins as Dynamic, Data-Driven Models
Unlike traditional static simulation models, digital twins in HVAC are continuously updated with field data, allowing them to reflect real-time system states and behaviors. This dynamic data assimilation is critical for accurate performance monitoring, fault detection, and predictive maintenance, providing value beyond design-time simulations.
JuliaSim’s Integrated Modeling Environment
JuliaSim’s combination of a domain-specific language (similar to Modelica), pre-built component libraries, and native interoperability across workflows sets it apart from legacy tools. This integration eliminates the need to juggle multiple software environments, thereby enhancing productivity and enabling more complex, realistic HVAC system models.
Managing Multiscale Dynamics and Nonlinearities
HVAC systems operate across vastly different time scales—from millisecond pressure changes in refrigerants to day-long building temperature variations. JuliaSim’s robust numerical solvers and symbolic manipulation capabilities enable the handling of stiff, nonlinear systems with multiple time scales, a challenge that traditional tools struggle to address effectively.
Balancing Data Richness with Information Scarcity
Despite abundant sensor data, HVAC systems often suffer from low excitation scenarios, limiting the observability of certain system states. The approach demonstrated leverages physics-based constraints embedded in the models alongside sparse sensor data to reliably estimate otherwise unmeasurable parameters such as refrigerant mass.
State Estimation Using Automatic Differentiation and Solver Transparency
The ability to extract solver internals and perform automatic differentiation enables the construction of sophisticated nonlinear state and parameter estimators. This is vital for real-time digital twin updates and is a technical capability largely missing from traditional equation-based modeling tools like Modelica.
Minimal Sensor Deployment for Cost-Effective Monitoring
By relying on as few as two to four temperature sensors and up to two pressure sensors, the methodology avoids the expensive and invasive installation of mass flow meters. This cost-effective sensing strategy makes advanced refrigerant mass estimation feasible in real-world applications, potentially including residential as well as commercial HVAC systems.
Environmental and Operational Impact
Precise refrigerant mass estimation supports early detection of leaks, which is crucial given the high global warming potential of many refrigerants. This technology aids in balancing the benefits of HVAC systems for human comfort and health with the imperative to reduce environmental harm, especially in large commercial installations where refrigerant leakage can be significant.
Additional Context and Future Directions
The webinar also touched upon the broader opportunities for integrating machine learning with physics-based HVAC models to simplify calibration, create surrogate models, and enable cloud-based system monitoring. Such integration could facilitate the development of district-level or grid-interactive HVAC solutions, advancing energy efficiency at community and utility scales.
Moreover, the Julia language’s modern computational framework enables seamless extension of these models with machine learning techniques such as universal differential equations, opening the door to hybrid modeling approaches that combine first principles with data-driven insights.
In conclusion, Mitsubishi Electric’s experience with JuliaSim demonstrates that modern, equation-oriented programming languages coupled with advanced numerical methods can overcome longstanding challenges in HVAC modeling, enabling the creation of industrial-grade digital twins that are adaptable, transparent, and highly accurate. This paves the way for more intelligent, efficient, and environmentally responsible HVAC system design and operation in the future.
