Sensor-Independent One-Hour-Ahead Forecasting and Anomaly Detection of Grid-Connected PV Inverters Using an Interpretable Random Forest Framework

¹ Associate Professor, Department of Mechanical Engineering, Vinayaka Mission's Kirupananda Variyar Engineering College, Vinayaka Mission’s Research Foundation (Deemed to be University), Salem, Tamilnadu - 636 308
² Assistant Professor and Head, PG Department of Computer Applications, St. Joseph's College of Arts and Science (Autonomous), Cuddalore, Tamilnadu, India - 607001
³ Associate professor,Department of Mechanical Engineering, Saveetha School of Engineering, Saveetha Institute of Medical and Technical Sciences, SIMATS, Chennai, Tamil Nadu, India - 602105.
⁴ Department of Medical Electronics, Sengunthar Engineering College and Tiruchengode, Namakkal, Tamilnadu - 637205
⁵ Department of Biomedical Engineering, Velalar College of Engineering and Technology, Thindal, Erode, Tamilnadu - 638012.
⁶ Department of Manufacturing Engineering, Vinayaka Mission's Kirupananda Variyar Engineering College, Salem, Tamil Nadu, 636 308 Email ID: 1 This email address is being protected from spambots. You need JavaScript enabled to view it. , This email address is being protected from spambots. You need JavaScript enabled to view it. ,This email address is being protected from spambots. You need JavaScript enabled to view it. , This email address is being protected from spambots. You need JavaScript enabled to view it. , This email address is being protected from spambots. You need JavaScript enabled to view it. , This email address is being protected from spambots. You need JavaScript enabled to view it.

Published online: 16 April 2026

Abstract

Dependable tracking of grid-linked photovoltaic (PV) systems is also unfeasible because it relies on meteorological sensors and rule-based management, especially in sensorlimited situations. The study constructs a machine learning model interpretable to humans based on electrical measurements of a PV plant of 30 kW and 40 kW inverters, and a five-minute resolution dataset of active and reactive power, phase voltages and phase currents and time dependent features during January 2025. An application of the regression, classification, and Z-score anomaly detection method was conducted using the Random Forest and conditioned on the timestamp alignment, feature engineering, percentile-based categorization of the output, and a one-hour-ahead label shifting, and the method was validated with the aid of the temporal splits and 5-fold Time Series cross-validation. The regression model was found to have great predictive accuracy (MAE = 0.12 kW, R² = 0.995), and cross-validation performed showed great robustness (R² = 0.9802, MAE = 0.1024 kW). The Z-score analysis (z = 3) revealed the presence of anomalous samples (2.77%). Though the accuracy of a static classification was 86 percent, time-varying forecasting lowered the accuracy to 33 % ( macro F1 = 0.33) which points to the impact of dynamic environmental variability. The suggested light and interpretable structure allows predicting the performance of inverters in real-time, early detecting anomalies, and intelligent planning of PV systems maintenance without the use of external meteorological devices.