How to monitor a C&I solar portfolio in 2026: SCADA, inverter APIs and AI
A practical guide to what commercial and industrial solar monitoring looks like now: which data sources to use, which metrics matter, and where AI genuinely helps.
In 2026 a C&I solar portfolio is monitored from data it already produces: inverter vendor APIs, data loggers, and SCADA exports, with no extra hardware. The stack that finds money has four layers: reliable data acquisition, weather-corrected KPIs, per-inverter analytics for soiling and faults, and alerting that turns findings into tickets.
Start from the data you already have
Ten years ago portfolio monitoring meant buying loggers, sensors, and a SCADA integration project per site. In 2026 most C&I fleets already emit everything a monitoring platform needs: five-minute AC power per inverter, plane-of-array or satellite irradiance, and temperature. Huawei, Sungrow, SolarEdge and the other major vendors all expose cloud APIs for the data their inverters already report.
- Inverter vendor cloud APIs: the default source for C&I. One credential covers every site on that vendor.
- Data loggers and gateways: fill gaps where a site is off the vendor cloud.
- SCADA or CSV exports: the fallback that always works, and enough for a first assessment.
- Satellite irradiance services: replace or sanity-check on-site pyranometers, which drift.
The four KPIs that actually move money
Dashboards fail when they track everything and rank nothing. Four weather-corrected numbers cover most of the economics of a C&I fleet.
| KPI | What it catches | Typical cadence |
|---|---|---|
| Performance Ratio (PR) | Whole-plant underperformance vs irradiance | Daily |
| Specific yield (kWh/kWp) | Site-to-site and inverter-to-inverter ranking | Daily |
| Availability | Downtime, comms loss, tripped devices | Continuous |
| Soiling ratio | Recoverable dust and dirt losses, cleaning timing | Daily |
The trap is plant-level averaging. A fleet-level PR of 82 percent can hide one inverter at 60 percent behind nineteen healthy ones, and inverter clipping can mask soiling entirely during the hours that matter. Per-inverter granularity is what separates monitoring that reports from monitoring that recovers revenue.
Where AI genuinely helps, with honest numbers
Machine learning earns its place in two jobs: classifying faults from electrical signatures, and estimating losses that have no direct sensor, such as per-inverter soiling. The accuracy is real but fault-specific. On the public Lazzaretti PV dataset a gradient-boosted classifier reaches 0.998 macro-F1 across five fault classes, while the harder 8-class GPVS-Faults set drops to 0.766, with thermal and sensor-drift faults hardest to separate. Remaining-useful-life models flag fast faults within a day but miss slow degradation by nearly a week.
The practical rule for 2026: trust AI for triage and ranking, verify with physics. A digital twin that corrects for weather, temperature, and curtailment tells you what the plant should have produced; the ML layer explains the gap. Platforms that skip the physics step produce alerts nobody trusts, and alert fatigue is the main reason monitoring tools get ignored.
A deployment checklist
- Inventory data sources per site: vendor API, logger, or export. Multi-vendor fleets need a platform that reads all of them.
- Backfill at least one seasonal cycle of history so baselines and soiling patterns are learned before live alerting starts.
- Set per-inverter baselines, not plant-level ones.
- Route findings into a ticketing workflow with validation, so every alert either becomes work or improves the model.
- Review the alert precision monthly: an alert stream the O&M team mutes is worse than no alerts.
Frequently asked questions
See also
See this on your own plants
NuraVolt turns your SCADA and BMS data into early fault detection, degradation-aware BESS analytics, and audit-ready reporting. A fixed-scope audit shows you what we’d find on your portfolio.