Report from Selected Technical Sessions
Digitalisation ran through the Vienna programme as a cross-cutting theme, but the dedicated sessions showed a field that has moved past demonstrating that machine learning can be applied to heat pumps and is now asking harder questions about data quality, deployability and what a digital method must do to earn a place in a commercial product.
Each section is based on a report contributed by the organiser or session chair and has been edited by the HPT Magazine editor for length and consistency.
From pre-screening to installed plant
Report contributed by Benjamin Zühlsdorf, Danish Technological Institute
Fault detection and diagnosis
Report contributed by Windholz Bernd, Austrian Institute of Technology
Fault detection attracted the largest concentration of papers. The problem is well defined: soft faults such as refrigerant leaks and heat exchanger fouling leave a system still able to meet the thermal load while quietly degrading its efficiency, so they are rarely noticed and rarely fixed. Mauro presented results from the EU-funded BEYOND project, which aims to detect and quantify refrigerant leaks and fouling before they significantly affect performance. An important early finding was that physics-based models provide more robust fault evaluation than purely data-driven methods, a result that recurred in several other contributions and that tempers the assumption that more data alone will solve the problem.
A two-stage artificial neural network tool for detection and diagnosis of soft faults in vapour compression systems was presented, separating the detection and diagnosis steps. Machine learning based fault detection was also demonstrated on an R290 residential unit, reflecting the sector-wide move to hydrocarbon refrigerants. The most methodologically interesting contribution came from Leroy Roger, who proposed peer-benchmarking-based fault detection across a residential geothermal heat pump fleet. Applied to data from 199 units, the unsupervised approach identified five significant outliers, three of which were linked to confirmed faults such as short cycling or improper sizing. The appeal of the method is that it requires no labelled fault data, which is precisely the resource the field lacks.
Virtual sensors and the economics of instrumentation
Cost is the constraint that shapes domestic monitoring, and several papers attacked it from different directions. Nicholson presented physics-informed neural networks as virtual sensors for domestic heat pumps, combining machine learning with thermodynamic knowledge to estimate refrigerant mass flow without additional hardware, reporting around 93 percent accuracy. Tran and Speerforck examined the usability of a virtual sensor in a domestic application from a digital twin perspective. Fazaa and colleagues presented data-driven techniques for predicting refrigerant charge through machine learning soft sensors, drawing on more than 6,800 steady-state operating points from the LC150 project and achieving prediction accuracies up to an R-squared of 0.98, with applications in commissioning, leakage detection and diagnostics.
The alternative route is to make the hardware cheap rather than to eliminate it. Barandier presented a low-cost data acquisition system built around a microcontroller and ten digital temperature sensors, able to monitor key operating parameters continuously while keeping hardware cost low, and suitable both for standalone field diagnostics and for integration into connected monitoring platforms. Related work presented intermittent refrigerant mass flow measurement for enhanced performance monitoring. Between them these approaches sketch a plausible path to instrumenting the installed base rather than only new premium equipment.
Control, flexibility and the grid
Model predictive control appeared in several forms, and notably in field rather than simulation settings. A novel model predictive control framework was presented as a replacement for traditional on-board heat pump controls, and insights were reported from a field test of a predictively controlled air-source heat pump in single-family homes using non-linear optimisation, motivated by the observation that efficiency falls in winter precisely when heating demand peaks. Yang presented energy optimisation of a solar house using model predictive control for an integrated air-source heat pump water heater and radiant floor system, tested in two cells at Concordia University in Montréal, where predictive control notably reduced energy expenses compared with rule-based control while achieving slightly better comfort. Guo presented a multi-scale hybrid physical and data-driven HVAC model for optimal demand response control, explicitly addressing two barriers to deployment: the degradation of data-driven accuracy under small-sample conditions, and computational cost.
Flexibility was framed around grid need. One contribution evaluated heat pump enabled building flexibility via deep reinforcement learning with an explicit trade-off analysis, noting that grid operators increasingly seek flexibility from underused resources such as the thermal inertia of residential buildings. Another experimentally evaluated the impact of circulation pump control on performance and energy flexibility in heat pumps with thermal storage, a reminder that flexibility claims depend on auxiliary components that are often ignored. Monitoring results were also presented from the SunStore project, where phase change material capsules in a buffer tank are being tested under real conditions. Dynamic controls for hybrid and efficient rooftop units extended the flexibility discussion into the commercial sector.
Data quality, tools and infrastructure
The most quietly consequential paper in this area may have been Uhl’s evaluation of publicly available heat pump specifications for digital planning and design tools. The study assessed the quality and completeness of standardised product data and found substantial differences between manufacturers, along with gaps that currently limit automated design, three-dimensional planning and installation workflows. Digital services depend on product data that is consistent and machine-readable, and at present it is neither.
Velte-Schäfer and co-authors surveyed the application of AI tools along the heat pump value chain, from component design through performance-map generation to operational supervision, and Furtwengler presented neural network estimation of parameter values in installed systems where technical parameters such as pipe characteristics are often unknown or undocumented. Jensen presented the Digital Heat Pump Lab at DTU, a laboratory enabling real-time two-way communication between physical heat pumps and digital models, supporting research on advanced control, fault detection, hardware-in-the-loop testing and digital twins, and designed explicitly to bridge laboratory work and real-world digital services. Modelling contributions included dynamic simulation of an absorption heat pump integrated with mechanical vapour recompression and a temperature and pressure based polynomial regression model for predicting heat transfer rate and coefficient of performance across different system types.
Conclusions
The conclusion from these sessions is that the enabling technologies are largely in place and the constraints have moved elsewhere: to the availability of labelled data, to the consistency of manufacturer product information, to the cost of instrumentation in the domestic segment, and to the question of whether a method validated on one fleet transfers to another.
All conference papers are available from the searchable HPT TCP database: https://heatpumpingtechnologies.org/publications
