Data-Driven Techniques for the Prediction of Refrigerant Charge in Heat Pumps through Machine Learning Soft Sensors
The transition to natural refrigerants, such as propane (R290), presents new challenges for heat pump development, particularly in optimizing system efficiency while minimizing refrigerant charge. Within the LC150 project, over 6800 steady state operating points were experimentally recorded from 32 heat pump prototypes, with a high-resolution variation of the refrigerant charge. Each data point reflects a half hour average under stabilized conditions, offering a robust dataset for data-driven analysis. This study investigates the use of machine learning (ML) to develop a soft sensor capable of predicting the refrigerant charge based solely on typical operating variables. Seven ML models (k Nearest-Neighbours, Decision Trees, Random Forests, Support Vector Machines, Gradient Boosting, and Multi-Layer Perceptrons (MLPs), a class of feedforward artificial neural networks) were trained and evaluated. A focus was on feature importance and model accuracy, especially for a limited number of input features. When using only five input features, including pressures, temperatures, and valve positions, the highest achieved coefficient of determination (R²) was 0.93, with Gradient Boosting. When expanded to ten features, the accuracy improved to an R² of 0.98 with XGBoost. Notably, the variables identified as most influential by the models correspond to those recognized as important by domain experts, demonstrating a strong alignment between data driven and physics-based understanding. This research was conducted as part of a semester project involving four undergraduate students from Hochschule Offenburg in the “Applied Artificial Intelligence” program in collaboration with Fraunhofer ISE. The project highlights how interdisciplinary collaboration, between AI specialists and heat pump engineers, can accelerate innovation in the development of efficient, charge minimized heat pump systems using natural refrigerants. By leveraging ML for refrigerant charge estimation, this work opens the door to real time diagnostics, enhanced fault detection, and smarter system design, supporting the wider adoption of sustainable heating technologies.