Per-Inverter World Models for Unsupervised Fault Detection in Distributed Solar Fleets
Year: 2026
Type: preprint
Abstract: Conventional rule-based monitoring catches solar-fleet faults late and drowns operators in false positives — it cannot distinguish a failing inverter from one operating correctly under cloud cover. A per-inverter world model, built on a Joint-Embedding Predictive Architecture (JEPA) that learns each inverter's normal operating dynamics and flags deviations as prediction-error anomalies, closes that gap. Applied across seven inverters at a 2.31 MW commercial rooftop PV installation at Sibaya Casino in Durban, South Africa — 7 × 330 kW Huawei SUN2000 inverters monitored via Huawei FusionSolar SCADA telemetry at 15-minute resolution — the approach achieves area-under-ROC scores of 0.69 to 0.98 depending on data availability and fault history, and improves F1 score over a rule-based baseline by roughly 40 to 260 percent for inverters with sufficient fault history in the test window. This report covers the modeling approach, the evaluation methodology, and a field diagnosis that separated a data-collection artifact from genuine equipment degradation, without disclosing proprietary model weights, tuned hyperparameter configurations, or internal system implementation details. A retrieval-augmented layer extends detections that cross a confidence threshold into cited, natural-language troubleshooting recommendations via Nehanda v3 (https://huggingface.co/asoba/nehanda-v3-27b), an epistemically fine-tuned open-weight model, gating expensive generative inference behind the cheap detector so that reasoning cost is incurred only where the detector has already earned it. Positioned within Asoba's Intelligence Layer for distributed power operations, the model implements the Orient stage of the underlying OODA loop — it transforms raw telemetry into calibrated, per-asset anomaly signals that a downstream Decision AI layer can act on deterministically.
Author: Samudzi, Shingai
Language: English
Subfield: Control and Systems Engineering
Field: Engineering
Domain: Physical Sciences
Open Access status: green
OpenAlex ID: W7170062476
Cites:
Cited by:
Related to: