smart-edge engineer details PI-based solar fault detection pipeline

Smart-edge, a company developing EMS and energy-market systems for the Japanese market, writes the piece as a practical field guide rather than a product pitch. Its central argument is that the most common solar monitoring metric, PR, is easy to calculate and standardized, but leaves temperature, low-light conditions and inverter output caps in the signal. The result, the author writes, is a flood of false alarms in summer and on cloudy days, with real partial faults buried inside. The proposed alternative models expected output from irradiance and cell temperature, then compares actual output to that expectation as PI. A key implementation caution is to freeze the baseline model trained on healthy data rather than retraining daily, because a model that follows recent performance would learn slow degradation and soiling as normal. Degradation is estimated with a year-over-year method, and alerts are managed as states with hysteresis rather than fired on every threshold crossing. The author frames AI monitoring as a complement to PR reporting, not a replacement.