Abstract
Automated detection of EVs from smart meter data can provide important insights for DSOs about spatiotemporal EV charging patterns. However, smart meters typically provide only hourly measurements of consumption while most load disaggregation techniques require at least minute level data. We use machine and deep learning methods to detect EV signatures in hourly smart meter data. Models are trained and evaluated on labelled data, before being tested on unlabelled field data. While balanced models catch about 75% of EVs at false positive rates of 35%, tuned models detect up to 90% of EVs with 10% false positives. When using models to detect EVs on unlabelled Norwegian smart meter data, detections are in line with EV fractions from the national registry as well as expected spatiotemporal patterns. However, models may be confused by baseline consumption patterns. Collection and inclusion of labelled EVs is therefore the next step.
