Detección sin sensores del ensuciamiento fotovoltaico con un modelo físico en Node-RED: aplicación en una planta residencial de Guayaquil

Abstract

Soiling reduces photovoltaic output, yet dedicated irradiance or soiling sensors may be difficult to justify in small residential installations. This exploratory case study evaluated a sensorless cleaning-alert system implemented on Node-RED at the edge. The system combines inverter telemetry, clear-sky modeled irradiance, a temperature correction based on nominal operating cell temperature and a persistence-based threshold policy to identify sustained performance drops compatible with soiling rather than isolated anomalies. The evaluation covered a 9.99-kilowatt-peak residential plant in Guayaquil, Ecuador, using 166 days of one-minute data from January to June 2026. The retrospective classification used 23 weekly blocks and four operationally labeled positive events. The detector produced three true positives, no false positives, one false negative and 19 true negatives, corresponding to 100 % precision, 75 % recall and a precision-recall harmonic mean of 0.857 . After three water-cleaning interventions, the estimated performance ratio increased by an average of 18.77 percentage points. Relative to an assumed fixed fortnightly schedule, the policy was associated with seven avoided visits and a counterfactual estimated economic benefit of approximately USD 423. Evidence is limited to one installation, four positive events and externally sourced we ather data without an independent irradiance or soiling reference.