Real-World Use Cases: ML Predictive Maintenance for Solar PV#

Summary#

Real deployments show that ML on ordinary inverter and weather data reliably finds faults and underperformance faster than people do. Predicting failures days ahead works only where there are enough recorded failures to learn from. Of 14 cases reviewed, 6 are peer-reviewed field studies, 2 are academic studies on operator data, 4 are vendor-reported and 2 are adjacent (pump monitoring at a water utility, and root-cause analysis in Thailand).

• Detection is proven. Field studies in Greece, Taiwan and Thailand detect faults with 94–97% sensitivity or accuracy, and the Taiwan case confirmed repairs on the utility's own SCADA.
• Prediction is conditional. Up to 7 days' warning is reported in Italy and Greece. A 26.6 MWp Shell plant in the Netherlands found no early signals at all, because only 7 faults were properly recorded.
• Business results are mostly vendor claims. Reported gains are 27% less unplanned downtime, 19% fewer emergency call-outs and +0.9% to +3.7% yield. None is independently audited.
• For PWA: no published case covers a water utility's solar fleet or a zero-export, self-consumption fleet. The PWA pilot would be the first such evidence.

Use cases at a glance#

#	Case (year, country)	Scale	Data	Method	Result	Evidence
1	Betti et al. (2021, Italy)	3–6 plants up to 10 MW, 60–100+ inverters	5-min SCADA + alarm logs	Self-organising map + neural net	Generic faults up to 7 days ahead, true-positive rate up to 95%	Peer-reviewed
2	Liu et al., Iowa State (2024, US)	> 700 inverters, 440 residential/commercial sites	5-min power, V, I, alert codes, weather	Siamese CNN + domain features	Next-month failure classification; one inverter failed 7 weeks after warnings	Peer-reviewed
3	Marangis et al. (2024, Greece)	Utility-scale plant	Historical performance data	XGBoost + one-class SVM + Prophet	96.9% fault detection; 92.9% sensitivity at 7 days ahead	Peer-reviewed
4	Liu, Kuo, Chen (2025, Taiwan)	120 MW, 1,292 × 100 kW inverters	Utility SCADA, DC/AC readings	PCA + one-class SVM + 2 physical rules	3 of 55 inverters flagged; after repair, SCADA showed the fault gone	Peer-reviewed
5	TU Delft / Shell Moerdijk (Netherlands)	26.6 MWp, 337 inverters	Irradiance, module temperature, active power	Elastic Net + GBM quantile regression	No early fault signals; only 7 usable fault records. OPEX cut up to 10% estimated	Master's thesis
6	Stony Brook / Ecosuite / Ecogy (2025, US)	Distributed portfolio	Inverter + weather data only	Self-supervised anomaly detection	Long-term defects flagged weeks to years before inspection; no figures published	Academic + industry
7	UiTM + Chiang Mai Rajabhat (2023, Malaysia)	Large-scale plant, central Malaysia	Plant electrical data	K-Means + LSTM	RMSE 0.7766	Peer-reviewed
8	Chulalongkorn, Thap Sakae (2025, Thailand)	5 MW, central inverters	LoRaWAN DC V/I + irradiance	Random Forest	94% fault classification, < 1 min detection	Peer-reviewed
9	Grid Dynamics client (Northern Europe)	Distributed wind + solar fleet	SCADA + sensors	ML failure-risk scoring	−27% unplanned downtime, −19% emergency dispatches in 2 quarters	Vendor claim
10	Power Factors (US)	82-inverter site	Inverter performance data	Diagnostic analytics	+0.9% / +3.7% yield on two revenue tiers	Vendor claim
11	Huawei Smart I-V (Ipoh, Malaysia)	49 MW, 9,496 strings	Inverter I-V scans	AI on I-V curves	12.1% of strings faulty; > 1.5% yield upside if fixed (projected)	Vendor claim
12	Clir client (North America)	One solar farm	Fleet benchmark from 5,400 farms	Comparative analytics	Anti-reflective coating loss identified; $250K recovered	Vendor claim
13	Southern Water (UK)	Sewage pumps network	Electrical signature of pumps	ESA + ML	63 developing failures found; £5M damage and penalties avoided	Industry (adjacent)
14	Naresuan Univ. (2022, Thailand)	6 utility plants, 5 years	Downtime logs	Root-cause analysis (not ML)	Inverter insulation 45.19%, humidity 19.44% of internal downtime	Peer-reviewed (adjacent)

Evidence map#

Only three cases are both well evidenced and close to PWA's fleet: Taiwan (same inverter class), Greece (same method) and Liu's US fleet (many small sites). Chulalongkorn is close behind but used add-on sensors, not inverter data.

Peer-reviewed and academic field deployments#

The strongest cases share three things: real operator SCADA data, a model of expected output, and a recorded fault history to check against.

Taiwan, 120 MW (closest hardware match). The plant in Tainan runs 1,292 identical 100 kW, 380 V string inverters, the same class as PWA's Growatt MAX 80–125 kW units. PCA plus a one-class SVM, with two physical rules (AC/DC ratio below 0.96, AC power drop over 20%), graded anomalies High/Medium/Low. Three inverters were flagged; O&M staff inspected and adjusted them, and the utility's SCADA then showed the fault had stopped. Limit: validated against another model (96.1% agreement), not against a confirmed fault list. Published in Applied Sciences 2025.

Greece, utility plant (closest method match). XGBoost models expected output, a one-class SVM flags faults and Prophet forecasts performance trends. It reached 96.9% detection sensitivity and gave alerts up to 7 days ahead at 92.9% sensitivity. This is the same pattern as the PWA spec's modules M1 and M3. Published in Solar RRL 2024.

Italy, 3–6 plants (Betti et al.). Up to 7 days' warning with true-positive rates up to 95%, but sensitivity fell sharply for inverters with few past failures.

US, > 700 small inverters (Liu et al.). The only large study on residential and commercial sites like PWA's. It predicts month-ahead failure and had to filter out lightning and grid "shock" failures, which give no warning.

Netherlands, Shell 26.6 MWp (negative result). A TU Delft thesis predicted inverter output with 0.98 kW mean absolute error across 337 inverters, yet found no early fault signals. The cause was poor fault tagging in SCADA: only 7 usable fault cases. It still estimated up to 10% lower operating cost from predictive maintenance.

US distributed portfolio (Stony Brook, Ecosuite, Ecogy). Self-supervised detectors trained only on inverter and weather data, chosen so the tool works on any site. The team reports it can flag long-term defects weeks or years before inspection, but published no accuracy or savings figures.

Industry and vendor deployments#

Vendor cases report operational gains in the range PWA's ROI model assumes, but none publishes a baseline, a control group or an audit.

• Grid Dynamics (anonymised wind and solar producer, Northern Europe). Telemetry pipelines (Kafka, Airflow, PostgreSQL) feed ML risk scores per asset into existing field-service workflows. Reported: 27% less unplanned downtime and 19% fewer emergency dispatches within two quarters. It took 8 engineers 5 months to reach first production, then 3 more months to cover the fleet.
• Power Factors (US, 82-inverter site). Diagnosing string-fuse failures raised yield by 0.9% and 3.7% on two revenue tiers with no added O&M spend. Its best predictive signal, coolant-pump current rising more than 30 days before failure, came from an extra sensor, not standard power data.
• Huawei Smart I-V (49 MW, Ipoh, Malaysia). Inverter-run I-V scans found 12.1% of 9,496 strings faulty. The "> 1.5% yield" figure is a projection if every fault were fixed. Growatt's MAX 100–125K units have a similar Smart I-V scan.
• Clir (North American solar farm). Benchmarking against 5,400 farms traced an output drop to aging anti-reflective coating; $250K was recovered.

The tech stack and team size in the Grid Dynamics case are close to the PWA spec (Python, PostgreSQL, scheduled pipelines). It is a useful sizing check, though PWA's lean build plans far fewer people.

Southeast Asia, Thailand and water utilities#

Thai and Malaysian work proves the methods work in tropical conditions, but all of it is on single large plants, not fleets of rooftops.

• Chulalongkorn University, Thap Sakae 5 MW (Prachuap Khiri Khan). Low-cost LoRaWAN sensors and a Random Forest classified no-fault, open-circuit and shading at 94% accuracy, detecting in under a minute. It used add-on sensors rather than inverter data, and reports no yield or downtime outcome.
• UiTM (Malaysia) with Chiang Mai Rajabhat University. K-Means plus LSTM anomaly detection on a large plant in central Malaysia, RMSE 0.7766 (Energy Reports). The authors' motivation, operators screening data and strings by hand, matches PWA's situation today.
• Naresuan University, six central-Thai plants over 5 years. Not ML, but it sets the target: inverter insulation (45.19%) and humidity (19.44%) caused most internal downtime. This is why the PWA spec ranks insulation (ISO) early warning as a top priority.
• PEA–PWA program today. PEA runs a real-time Energy Monitoring System over all 233 sites. No published analytics or maintenance outcomes yet.

Water utilities (adjacent). Southern Water (UK) used electrical signature analysis with ML on sewage pumps. A trial found 63 developing failures and avoided an estimated £5M in damage and penalties. Operators came to trust the alerts and fixed most issues in working hours. It is not solar, but it shows a water utility adopting ML-ranked maintenance alerts, and PWA's pumps could later join the same platform.

Lessons for PWA#

The evidence supports launching PWA with detection and fault ranking. Failure prediction should wait until PWA has its own fault records.

Lesson	Evidence	What PWA does
An expected-output model plus anomaly detection works on standard data	Greece, Taiwan, Stony Brook	Modules M1 + M3 first, no new sensors
Prediction fails without good fault records	Shell/TU Delft (7 usable faults); Betti (low sensitivity on rare faults)	Every alert closed with a recorded finding from day one
Simple physical rules beside ML keep alerts explainable	Taiwan (AC/DC < 0.96, power drop > 20%)	Rule checks in M0/M3, shown on each alert
Value comes from the workflow, not the model	Grid Dynamics, Southern Water	LINE alerts ranked by THB lost, contractor close-out
Sudden failures cannot be predicted	Liu et al. (lightning, grid events)	Report them separately; don't count them as missed predictions
In-house builds take months, not weeks	Grid Dynamics: 8 engineers, 5 months to first use	Keep the 3-month pilot scope small: 10 stations, P1 cases only

Gaps no case covers

• Zero-export curtailment: no study separates controller throttling from faults. PWA must solve this itself (module M0).
• Fleets of 80–575 kWp public-sector rooftops in the tropics.
• Independently audited yield or cost savings. The PWA pilot's before/after measurement would be among the first.

Sources#

• Liu, Kuo, Chen — PCA–OCSVM framework, 120 MW Taiwan plant (preprint; Applied Sciences 2025, 15(23), 12394)
• TU Delft — Predictive maintenance for utility scale solar parks (Shell Moerdijk 26.6 MWp)
• Marangis et al. — Trend-based predictive maintenance and fault detection analytics, Solar RRL 2024 (abstract)
• Betti et al. — Fault prediction and early detection in large PV plants, Sensors 2021
• Liu et al. — Deep learning failure prognostic model for PV inverters, IEEE TSTE 2024
• pv magazine USA — Stony Brook / Ecosuite / Ecogy anomaly detection (Dec 2025)
• pv magazine — UiTM and Chiang Mai Rajabhat K-Means + LSTM model (Oct 2023)
• Chulalongkorn — AI fault classification with low-cost IoT, Scientific Reports
• Naresuan — System availability of PV plants in Thailand, Energy Reports 2022
• Grid Dynamics — Predictive maintenance for a green energy producer
• Power Factors — Diagnostic analytics for PV plants
• Huawei — FusionSolar special edition (Ipoh Smart I-V results)
• Clir — Recovering $250K with anti-reflective coating insights
• Pumps & Systems — Southern Water ESA monitoring
• PEA — PEA–PWA Solar Rooftop program news