machine learning for data from smart electricity meters to enable smart grid applications
As a PhD candidate in the Bits to Energy Lab at ETH Zurich, I develop machine learning techniques for energy applications to promote sustainability and for the digitalization of the power grid. In particular, my work focuses on analyzing data from smart electricity meters.
In mid of 2023, my colleague Markus Kreft and I published a hands-on Python tutorial on smart meter data analytics at the International Conference on Learning Representations (ICLR 2023) as part of the Workshop on Tackling Climate Change with Machine Learning. Further, Climate Change AI featured the tutorial in the program of their 2023 summer school with more than 6000 virtual participants around the globe.
Example: Time series data in 15-minute resolution of a smart electricity meter.
Example: Offline change point detection algorithm applied to a smart meter data time series.
References
2024
Predictability of electric vehicle charging: explaining extensive user behavior-specific heterogeneity
Smart charging systems can reduce the stress on the power grid from electric vehicles by coordinating the charging process. To meet user requirements, such systems need input on charging demand, i.e., departure time and desired state of charge. Deriving these parameters through predictions based on past mobility patterns allows the inference of realistic values that offer flexibility by charging vehicles until they are actually needed for departure. While previous studies have addressed the task of charging demand predictions, there is a lack of work investigating the heterogeneity of user behavior, which affects prediction performance. In this work we predict the duration and energy of residential charging sessions using a dataset with 59,520 real-world measurements from 267 electric vehicles. While replicating the results put forth in related work, we additionally find substantial differences in prediction performance between individual vehicles. An in-depth analysis shows that vehicles that on average start charging later in the day can be predicted better than others. Furthermore, we demonstrate how knowledge that a vehicles charges over night significantly increases prediction performance, reducing the mean absolute percentage error of plugged-in duration predictions from over 200% to 15%. Based on these insights, we propose that residential smart charging systems should focus on predictions of overnight charging to determine charging demand. These sessions are most relevant for smart charging as they offer most flexibility and need for coordinated charging and, as we show, they are also more predictable, increasing user acceptance.
@article{BRUDERMUELLER/EVCharging,author={Kreft, Markus and Brudermueller, Tobias and Fleisch, Elgar and Staake, Thorsten},title={Predictability of electric vehicle charging: explaining extensive user behavior-specific heterogeneity},year={2024},journal={Applied Energy},volume={370},pages={123544},issn={0306-2619},doi={https://doi.org/10.1016/j.apenergy.2024.123544},keywords={Electric vehicles, Smart charging, Demand response, Demand prediction, Real-world data},}
Identifying Electric Water Heaters from Low-Resolution Smart Meter Data
Despite an increasing share of heat pumps, electric water heaters are still widely used in residential applications. Their high connected load and energy consumption combined with their thermal inertia make them ideal candidates for demand response and energy saving programs. However, due to missing or outdated information about installation locations, it is difficult to run large and targeted campaigns. Thanks to an increased roll-out of advanced metering infrastructure, smart meter data is widely available, opening new opportunities to generate the missing information. In our work, we identify electric water heaters from lowresolution smart meter data with a 15-minute sampling rate and estimate their consumption using a trainingfree detection method that is easy to interpret and adapt. On a real-world data set with measurements from 1,962 meters over one year, we achieve 89.4% accuracy in detecting households with electric water heaters. We predict heating capacity with a mean absolute error of 0.9 kW. We also find that the installed electric water heaters lead to a peak demand that is 51% higher than in a setting without electric water heating, highlighting the importance of appropriate load management. Our method can be directly incorporated into existing demand response applications.
@inproceedings{BRUDERMUELLER/EWHDisaggregation,title={Identifying Electric Water Heaters from Low-Resolution Smart Meter Data},booktitle={2024 IEEE Conference on Technologies for Sustainability (SusTech)},year={2024},volume={},number={},pages={128-135},author={Kreft, Markus and Brudermueller, Tobias and Anderson, Tyler and Staake, Thorsten},doi={10.1109/SusTech60925.2024.10553590},keywords={Meters, Accuracy, Runtime, Heat pumps, Water heating, Smart meters, Demand response, Electric Water Heater, Smart Meter Data, Low Resolution, Demand Response, Peak Shaving, Load Shifting},}
2023
Large-scale monitoring of residential heat pump cycling using smart meter data
Heat pumps play an essential role in decarbonizing the building sector, but their electricity consumption can vary significantly across buildings. This variability is closely related to their cycling behavior (i.e., the frequency of on–off transitions), which is also an indicator for improper sizing and non-optimal settings and can affect a heat pump’s lifetime. Up to now it has been unclear which cycling behaviors are typical and atypical for heat pump operation in the field and importantly, there is a lack of methods to identify heat pumps that cycle atypically. Therefore, in this study we develop a method to monitor heat pumps with energy measurements delivered by common smart electricity meters, which also cover heat pumps without network connectivity. We show how smart meter data with 15-minute resolution can be used to extract key indicators about heat pump cycling and outline how atypical behavior can be detected after controlling for outdoor temperature. Our method is robust across different building characteristics and varying times of observation, does not require contextual information, and can be implemented with existing smart meter data, making it suitable for real-world applications. Analyzing 503 heat pumps in Swiss households over a period of 21 months, we further describe behavioral differences with respect to building and heat pump characteristics and study the relationship between heat pumps’ cycling behavior, energy efficiency, and appropriate sizing. Our results show that outliers in cycling behavior are more than twice as common for air-source heat pumps than for ground-source heat pumps.
@article{BRUDERMUELLER/2023121734,title={Large-scale monitoring of residential heat pump cycling using smart meter data},journal={Applied Energy},volume={350},pages={121734},year={2023},issn={0306-2619},doi={https://doi.org/10.1016/j.apenergy.2023.121734},author={Brudermueller, Tobias and Kreft, Markus and Fleisch, Elgar and Staake, Thorsten},keywords={Smart meter data, Heat pump, Cycling behavior, outlier detection, Residential buildings, Energy efficiency, Appliance monitoring, Real-world operation, Machine learning},}
Disaggregation of Heat Pump Load Profiles From Low-Resolution Smart Meter Data
As the number of heat pumps installed in residential buildings increases, their energy-efficient operation becomes increasingly important to reduce costs and ensure the stability of the power grid. The deployment of smart electricity meters results in large amounts of smart meter data that can be used for heat pump optimization. However, sub-metering infrastructure to monitor heat pumps’ energy consumption is costly and rarely available in practice. Non-intrusive load monitoring addresses this issue and disaggregates appliance-level consumption from aggregate measurements. However, previous studies use high-resolution data of active and reactive power and do not focus on heat pumps. In this context, our study is the first to disaggregate heat pump load profiles using commonly available smart meter data with energy measurements at 15-minute resolution. We use a sliding-window approach to train and test deep learning models on a real-world data set of 363 Swiss households with heat pumps observed over a period of 8 years. Evaluating our approach with a 5-fold cross-validation, our best model achieves a mean R2 score of 0.832 and an average RMSE of 0.169 kWh, which is similar to previous work that uses high-resolution measurements of active and reactive power. Our algorithms enable real-world applications to monitor the energy efficiency of heat pumps in operation and to estimate their flexibility for demand response programs.
@inproceedings{BRUDERMUELLER/10.1145/3600100.3623731,author={Brudermueller, Tobias and Breer, Fabian and Staake, Thorsten},title={Disaggregation of Heat Pump Load Profiles From Low-Resolution Smart Meter Data},year={2023},isbn={9798400702303},publisher={Association for Computing Machinery},address={New York, NY, USA},doi={10.1145/3600100.3623731},booktitle={Proceedings of the 10th ACM International Conference on Systems for Energy-Efficient Buildings, Cities, and Transportation},pages={228–231},numpages={4},keywords={Non-Intrusive Load Monitoring, Low Resolution, Load Disaggregation, Smart Meter Data, Heat Pump Optimization, Energy Efficiency},location={Istanbul, Turkey},series={BuildSys '23},}
Smart meter data analytics: Practical use-cases and best practices of machine learning applications for energy data in the residential sector
To cope with climate change, the energy system is undergoing a massive transformation. With the electrification of all sectors, the power grid is facing high additional demand. As a result, the digitization of the grid is becoming more of a focus. The smart grid relies heavily on the increasing deployment of smart electricity meters around the world. The corresponding smart meter data is typically a time series of power or energy measurements with a resolution of 1s to 60 min. This data provides valuable insights and opportunities for monitoring and controlling activities in the power grid. In this tutorial, we therefore provide an overview of best practices for analyzing smart meter data. We focus on machine learning applications and low resolution (15-60 minutes) energy data in a residential setting. We only use real-world datasets and cover use-cases that are highly relevant for practical applications. Although this tutorial is specifically tailored to an audience from the energy domain, we believe that anyone from the data analytics and machine learning community can benefit from it, as many techniques are applicable to any time series data. Through our tutorial, we hope to foster new ideas, contribute to an interdisciplinary exchange between different research fields, and educate people about energy use.
@inproceedings{BRUDERMUELLER/2023ICLR,title={Smart meter data analytics: Practical use-cases and best practices of machine learning applications for energy data in the residential sector},author={Brudermueller, Tobias and Kreft, Markus},booktitle={International Conference on Learning Representations (ICLR 2023) Workshop on Tackling Climate Change with Machine Learning},year={2023},organization={Climate Change AI},video={https://iclr.cc/virtual/2023/workshop/12828#collapse-sl-13527},}
2022
Automatic Differentiation of Variable and Fixed Speed Heat Pumps With Smart Meter Data
With the increasing prevalence of heat pumps in private households, the need for optimization is growing. At the same time, the growing number of active smart electricity meters generates data that can be used for remote monitoring. In this paper, we focus on the automatic differentiation between fixed speed and variable speed heat pumps using smart meter data. This distinction is relevant because it is necessary for evaluating the state or cyclic behavior of a heat pump. In addition, identifying fixed speed heat pumps is important because they are known to be the less efficient systems and therefore may be preferred targets in energy efficiency or replacement campaigns. Our methods are applied to electricity data from 171 Swiss households with a resolution of 15 minutes. In this setting, a K-Nearest Neighbor model achieves a mean AUC of 0.976 compared to 0.5 of a biased random guess model.
@inproceedings{BRUDERMUELLER/9961055,author={Brudermueller, Tobias and Wirth, Florian and Weigert, Andreas and Staake, Thorsten},booktitle={2022 IEEE International Conference on Communications, Control, and Computing Technologies for Smart Grids (SmartGridComm)},title={Automatic Differentiation of Variable and Fixed Speed Heat Pumps With Smart Meter Data},year={2022},volume={},number={},pages={412-418},doi={10.1109/SmartGridComm52983.2022.9961055},}
Smart meter data analytics: Practical use-cases and best practices of machine learning applications for energy data in the residential sectorIn International Conference on Learning Representations (ICLR 2023) Workshop on Tackling Climate Change with Machine Learning, 2023
