A charger fails, a driver notices, someone reports it, a technician is dispatched. Predictive maintenance tries to reverse that order: spot the drift before the failure. It is an operating practice, not a legal obligation, and its value depends entirely on the quality of the data you collect.
What the AFIR requires, and what it does not
The AFIR has applied since 13 April 2024. Two of its requirements shape maintenance work:
The text sets no numeric availability threshold and no mandatory reliability report for operators. It only mentions the uptime of services as a goal in a recital. A 99 % reliability threshold does exist in the United Kingdom for networks of 50 kW and above, according to a parliamentary answer relayed by Parallel Parliament, but it is not an AFIR rule and does not apply in the EU.
So predictive maintenance is a business choice. What the AFIR gives you is a duty to publish status data, which makes any gap between your records and reality visible.
Uptime is not the same as successful sessions
A 2025 report by ChargerHelp, based on United States data, found declared uptime of 98.7 to 99 % while only 71 % of sessions succeeded. The company sells maintenance services, so read it as an indication, not a benchmark. Its lesson still holds: a charger can answer the network and fail the driver. The causes cited were software and firmware compatibility, ageing hardware and fragmented systems.
Track the session success rate per charger model next to availability. That is the number a driver experiences.
Where interventions come from
A vendor study of European sites attributed call-outs to power supply (24 %), payment terminals (20 %), charging speed or failed charges (20 %), hardware (18 %) and communication (13 %). It reports that 70 % of faults were resolved remotely and 30 % needed an engineer on site. This is supplier data, not an independent measurement, but the direction is useful: a large share of faults is visible and sometimes fixable from the back end, before anyone drives to the site.
What a CSMS can watch
Over OCPP, the CSMS receives status changes, error codes, meter values and connection events. Useful signals to trend:
Start with simple rules and thresholds on this history. Statistical or machine-learning models only help once you have clean, labelled failure records, and few operators have them on day one. Note also that OCPP 1.6 and 2.0.1 are not compatible, and manufacturers implement them unevenly. Check what each model actually sends before designing alerts around it.
Turn alerts into actions
A predictive alert is only worth the process behind it. Define who receives it, in which order to act (remote restart first, then a scheduled visit), and how the result is recorded. Measure the share of faults closed remotely, repeat visits to the same site and session success by charger model. Do not claim savings before you have these figures for your own network.
This week
GFO Omni is a CSMS for OCPP 1.6, 2.0.1 and 2.1 that supervises charge points.