ANR AI4HP

Heat pumps constitute an effective solution to mitigate the energy consumption and environmental impact of buildings. They have a high potential to use renewable energies for converting power to heat. However, the actual on-field performance of heat pumps does not always rise to expectations. High heat losses occur, and energy efficiency is reduced due to unsuitable system layout but also mal parametrization of heat pump controls and undetected operation deficits. The current practice is to set the control parameters by the installer once and for all, however, the over-time variability, e.g. of occupancy behaviour, building’s age, weather conditions, requires an adaptive heat pump control and supervision.
The subject of the AI4HP project is therefore the development of novel artificial intelligence (AI) methods based on incremental learning with artificial neural networks (ANN) for adaptive heat pump control and monitoring. AI methods provide a “self-learning” capability allowing to create and adapt automatically models and predictions just from measurement data. The resulting low-effort/low-cost adaptive AI methods will improve the operational performance of heat pumps by a) automatically adapting controller settings to varying boundary conditions and b) detecting heat pump and system mal-functioning. This leads to the development of a new generation of “smart heat pumps”, which integrate new functionalities and interactions with a changing environment in order to provide the best energy-efficiency and comfort for the user, make maintenance operations easier and avoid performance degradation by fault detection. The project addresses both dual service heat pumps and heat pumps serving domestic hot water only, which make up a substantial part of the French market.

ANNs can improve heat pump operation by system modelling and predicting future developments based on measurement data. They can capture the complexity of the heat pump system – e.g., there is not a one-to-one relationship between a specific fault and a single variable. However, in most cases ANNs and other machine learning methods lead to large errors when confronted with significantly different or new data. Another problem is the continuous generation of measurement data. The memory size and computing power are limited and prevent a retraining using the complete data set. However, if the system is trained only on the new data, catastrophic forgetting or interference occurs. Consequently, for real-time use in leading edge technologies the ANN system must continuously learn new knowledge without forgetting previous knowledge, which requires more logic than the existing methods alone. In this project, the consortium consisting of experts in the fields of ANN research, energy research, heat pump manufacturing and energy supply will develop adaptive ANNs based on methods of incremental learning, which are suitable for real-time use in heat pump operation with continuous measurement data acquisition. The adaptive AI pipeline is being developed for the three use cases adaptive heating curve control, adaptive control based on load forecasts and fault detection and diagnosis (CEA List, LPNC, Fraunhofer ISE), implemented in a heat pump controller (Stiebel Eltron) and validated in laboratory tests and a pilot system (EDF, Stiebel Eltron).

Using the advanced AI methods, we expect up to 20% energy savings and CO2-emissions reduction for domestic hot water and space heating without comfort violation. If the project aims can be successfully validated in the pilot demonstrations, the industrial partners plan to initiate a project development in a short to middle term after the end of the project to include the AI routines in their portfolio, either by an implementation directly in the heat pump controller or as a service in their heat pump cloud platform.

See the publications in the HAL-ANR