Put a SCADA energy total beside a settlement statement and the two rarely tie out. That is usually not a fault. It is the predictable result of two systems measuring the same energy at different points, under different rules, on different clocks. Chasing the difference to zero leads nowhere. The useful goal is to explain it, bound it, and alert only when it leaves expectations.
Real-time telemetry is provisional by design; settlement data is revised by design. Reconciliation holds both as true at once.
Why the numbers diverge
Real-time feeds over Modbus, OPC UA or MQTT are provisional: a power register sampled every few seconds and integrated locally into energy, with no later correction. Settlement runs on validated half-hourly volumes from fiscal metering, delivered days later and revised again afterwards. The two also measure at different boundaries. SCADA usually reports at the inverter or feeder, while the settlement meter sits at the grid connection point, so transformer and cable losses and any auxiliary load fall between them.
Two quieter mismatches do the most damage. Summation conventions differ: import and export may be netted, signed, or held on separate registers, and a single sign error inverts a whole site. Timestamps are the other trap. Telemetry often arrives in device-local or UTC time while settlement is keyed to market-local intervals, so a naive join across a DST transition double-counts or drops an hour. Reconcile on UTC instants and treat the displayed interval label as presentation only.
Building the reconciliation layer
A reconciliation layer normalises both sources onto a common grain before comparing anything. We land raw telemetry in a time-series store, re-aggregate it to the settlement interval on UTC boundaries, apply a documented loss-adjustment factor to bring it to the metering point, then compare against each settlement vintage. Because settlement is revised, every figure carries its version and ingest time, so any reconciliation can be replayed and you can see which revision moved a number.
Express the comparison as tolerance bands, not exact equality. A small percentage plus a fixed kilowatt-hour floor absorbs rounding, sub-interval edge effects and normal loss variation; anything inside the band reconciles silently. Classify breaches by signature rather than just flagging them: a constant offset, a single bad interval and a DST-shaped error each point to a different root cause and a different fix.
Reconciliation is less about making two numbers identical than about making their difference legible: versioned, time-correct, loss-adjusted, and bounded. Done well, it turns a recurring argument over whose figure is right into an audit trail that settles the question itself. We build this kind of data and integration work for energy operators; if it sounds familiar, get in touch.