Heat pumps have undergone enormous development in recent years. Due to environmental concerns and the high prices of fossil fuels, the number of heat pumps is increasing year on year. According to forecasts, two million new heat pumps will be connected to the power grid in Europe alone by 2025. Thanks to their ability to transfer heat from a hot source to a cold source, they naturally consume less energy than traditional heating solutions such as underfloor heating.
According to a study by EnergieSchweiz [1] from 2021, the average efficiency of heat pumps in Switzerland is between 3 and 4.5. This means that a heat pump heats a household 3 to 4.5 times more than a conventional electric boiler while consuming the same amount of electricity. Although they are highly efficient, heat pumps are still major energy consumers and therefore a significant factor in reducing electricity costs for end users and in terms of grid stability for utility companies.
Active monitoring of heat pump usage is therefore crucial for both end customers and utility companies. Based on the monitoring data, several improvements can be made. Energy suppliers can use smart meters to predict and forecast load more accurately, allowing them to better estimate the amount of electricity to be purchased or produced and gain a better understanding of the grid. End consumers can track their heating habits and adjust them based on this information. In the event of operational anomalies, notifications can help to take timely action. In the future, monitorable loads will also be the first step toward smart grid initiatives and flexibility solutions (e.g., SmartGridready), which will enable greater use of renewable energies and energy-saving measures.
Monitoring can be carried out using an electricity meter such as the CLEMAP Energy Monitor, which is connected directly to the heat pump, or via a smart meter. As part of the Social Power Plus(SSP) research project, we asked ourselves whether it would be possible to monitor heat pumps using existing hardware such as a smart meter. Normally, a smart meter records the electricity consumption of the entire household, not just that of the heat pump. Non-Intrusive Load Monitoring (NILM) is a machine learning technique that makes it possible to break down the loads running in a household using a single metering point. With NILM, we can therefore use existing smart meters and only need to add a software component, which greatly reduces setup costs.
Within SSP, CLEMAP developed a NILM algorithm for the disaggregation of heat pumps, which aims to raise people's awareness of their energy consumption. The algorithm works by processing electrical time series data for active (P) and reactive (Q) power—summed over the three phases—with a temporal resolution of 15 minutes. Heat pumps have specific electrical profiles and can therefore be identified using statistical methods. The machine learning algorithm developed by CLEMAP was evaluated using a reference data set. To this end, various electrical data were collected over several months: data from smart meters that measure the entire household and from special smart meters for heat pumps in 20 households in eastern Switzerland that are equipped with different types of heat pumps.
The following metrics can be evaluated for a machine learning model:
The accuracy of heat pump detection: How well can we predict whether the heat pump is on or off at a given point in time? (Classification problem)
Errors in energy breakdown: How accurately can the proportion of total energy consumed by the heat pump be broken down when considering energy consumption over a specific period of time? (Regression problem)
These metrics were evaluated for the entire data set with the following results:
Current research reports ([2], [3]) indicate an accuracy of 90-95% for the detection of large loads such as heat pumps—under laboratory conditions and sometimes with a higher sampling rate. Our accuracy of 86% in field tests can therefore be considered a success.
To better understand the result, we can look at a box plot showing the metrics for individual households and the statistical indicators for the entire data set. Here we look at the accuracy of heat pump detection (balanced_accuracy_state_ts[%]).
The scattered blue dots on the left show the accuracy of each household. One dot stands out with a poor accuracy of around 40%. This outlier was investigated and it was found that its electrical profile differs significantly from that of the other heat pumps, which led to this result. It was decided to leave the outlier in this benchmark to show that, while the model performs well, fine-tuning may be necessary for special cases. In the future, further analysis based on other data sets and over longer periods of time will further improve the performance of the algorithm.
Heat pumps are developing rapidly and offer both environmental and cost benefits. Monitoring them is crucial for end users and energy suppliers as well as for heat pump manufacturers (e.g., the field performance of their installed base). This monitoring can be done in a non-invasive way using machine learning algorithms and existing smart meter electricity data to keep costs low. CLEMAP has developed a powerful algorithm that detects heat pumps with 86% accuracy. In the future, this will enable the implementation of solutions for energy reduction, grid serviceability, and flexibility management. The solution is currently being introduced via existing customer interfaces, where customers have the option of explicitly using their data. As the result of improved efficiency is obvious, regulatory authorities must adapt and allow such solutions to be effectively implemented and used in the power grid as the volume of data from smart meters grows.
Sources:
[1] SwissEnergy, report "Field measurements of heat pump systems, heating season 2020/21"
[2]“Results from Non-intrusive Load Monitoring”, A. Hutter et al., CSEM Scientific and Technical Report 2016
[3] Salani, M., Derboni, M., Rivola, D. et al. Non-intrusive load monitoring for demand side management. EnergyInform 3, 25 (2020). https://doi.org/10.1186/s42162-020-00128-2