Non-Intrusive Residential Electric Vehicle Detection Using Positive-Unlabeled Contrastive Learning

Zahrasadat Sajjadifar, Antonio Hernandez Espin, Bart De Moor

Keywords: Contrastive Learning, Positive Unlabeled Learning, Clustering, Electric Vehicle Detection, Low-voltage Grid.

Abstract:

Detecting electric vehicle (EV) charging on residential connections
is essential for distribution system operators (DSOs)
to manage grid load, forecast demand, and plan infrastructure
upgrades. However, the widespread availability of labeled
data for such detection remains limited, especially in
non-intrusive settings. Notably, while there is often a reliable
positive label for EV presence through EV registration,
the negative labels tend to be unreliable due to the presence
of non-registered EVs. This paper addresses the problem of
EV charging detection as a positive and unlabeled (PU) contrastive
learning task, where only a subset of the measurements
is positively identified based on the households registered
with EV, and the rest of the data remains unlabeled.
Based on contrastive learning, we learn representations that
pull together examples likely to contain EV charging signatures
and push apart background load patterns and non-
EV signatures, while explicitly accounting for the label uncertainty
inherent in PU data. We propose an approach that
works with raw, aggregated electricity load data at the household
level, without relying on intrusive metering or extensive
manual labeling. Our primary dataset consists of quarterhourly
household electricity consumption data provided by
Fluvius, the distribution system operator (DSO) in Flanders,
Belgium, and we additionally validate our method using other
publicly available open datasets. Our results highlight the feasibility
of scalable EV detection with minimal supervision,
offering critical observability for DSOs aiming to monitor EV
adoption and manage localized grid impacts.

Submission Number: 27