Detecting Degradation in DC Fast Chargers from Single-Sided OCPP Telemetry

Abstract

Charge point operators hold years of Open Charge Point Protocol (OCPP) telemetry collected for billing, and almost none of it is used to assess the health of the hardware producing it. This is a missed opportunity, because the industry's standard health metric, uptime, is known to disagree with what drivers experience: reported uptime sits near 99 percent while roughly three in ten charging attempts fail, and more than a third of those failures occur on chargers that report themselves as operational. Degradation that is gradual rather than binary is invisible to an availability flag. This paper first establishes a negative result that constrains what any such method can claim. OCPP MeterValues reports a single side of the power conversion, so the absolute conversion efficiency of a charger is not identifiable from it, and tools reporting an efficiency figure from OCPP alone are supplying the side they cannot observe. A weaker but operationally sufficient quantity is identifiable: the change in a charger's behaviour relative to its own history, because the unknown fixed offset between reported and true quantities cancels under within-unit differencing. On that basis we define three estimators, for sustained capability, series resistance, and thermal derating burden, and show that each requires explicit treatment of a confounder that defeats the obvious approach. Most sharply, the natural resistance estimator, regressing outlet voltage on outlet current, returns a median of minus 610 milliohms where the true value is 30 milliohms, because pack voltage rises as current falls; modelling the state-of-charge term recovers 29.8 milliohms on the same data. Validation uses a simulator whose AC-side physics the estimators never observe, and which injects two failure modes the estimators do not model. On held-out seeds the method flags 25 of 25 degraded units with no false positives among 15 healthy units, and recovers injected decay rates within 0.5 percentage points. We report a detection latency of 10 to 18 days median, and quantify a false-positive inflation under repeated monitoring that practitioners must account for. Validation is in simulation only; the paper specifies the field protocol that would close this gap. The implementation is open source: https://github.com/Tech4Biz-Solutions-Pvt-Ltd/evse-drift, archived at doi:10.5281/zenodo.21969642