TY - GEN
T1 - Detection and characterization of domestic heat pumps
AU - Ray, Guillaume Le
AU - Christensen, Morten Herget
AU - Pinson, Pierre
PY - 2019/6/1
Y1 - 2019/6/1
N2 - Smart meters allow utilities to gain access to tremendous amount of metering data, from which they can improve their knowledge on customers. In this paper, we propose a method to detect and characterize domestic heat pumps from household power consumption data. Appliance detection is not a trivial task, owing to the variability in heat pump load dynamics and to the distortion of their power consumption signature by other appliances. Compared to State-of-the-art methodologies that relies on energy disaggregation with supervised learning and high-resolution data (e.g., 10 seconds), our novel approach uses lower resolution data. Also, it is semi-supervised and relies on a Bayesian framework allowing to continuously learn as new data becomes available. To do so, the overall power consumption signals are decomposed and approximated into a dictionary of boxcar functions using sparse signal approximation to isolate heat pumps activity. The learning phase consists then to generate distributions summarizing power consumption, operation time and frequency of heat pumps activation events summarized as boxcar functions after approximation. During the test phase, the distributions are used as prior to calculate the likelihood that a boxcar function is generated by a heat pump. Using standard classification performance measure and an application to data from the EcoGrid EU project, the methodology reaches high performance in heat pump detection.
AB - Smart meters allow utilities to gain access to tremendous amount of metering data, from which they can improve their knowledge on customers. In this paper, we propose a method to detect and characterize domestic heat pumps from household power consumption data. Appliance detection is not a trivial task, owing to the variability in heat pump load dynamics and to the distortion of their power consumption signature by other appliances. Compared to State-of-the-art methodologies that relies on energy disaggregation with supervised learning and high-resolution data (e.g., 10 seconds), our novel approach uses lower resolution data. Also, it is semi-supervised and relies on a Bayesian framework allowing to continuously learn as new data becomes available. To do so, the overall power consumption signals are decomposed and approximated into a dictionary of boxcar functions using sparse signal approximation to isolate heat pumps activity. The learning phase consists then to generate distributions summarizing power consumption, operation time and frequency of heat pumps activation events summarized as boxcar functions after approximation. During the test phase, the distributions are used as prior to calculate the likelihood that a boxcar function is generated by a heat pump. Using standard classification performance measure and an application to data from the EcoGrid EU project, the methodology reaches high performance in heat pump detection.
KW - Bayesian framework
KW - Energy analytics
KW - Heat pump detection
KW - Semi-supervised learning
U2 - 10.1109/PTC.2019.8810930
DO - 10.1109/PTC.2019.8810930
M3 - Article in proceedings
T3 - 2019 IEEE Milan PowerTech, PowerTech 2019
BT - Proceedings of 2019 IEEE Milan PowerTech
PB - IEEE
T2 - 2019 IEEE Milan PowerTech, PowerTech 2019
Y2 - 23 June 2019 through 27 June 2019
ER -