Feature-Normalized k-Nearest Neighbor Classification for Real-Time Photovoltaic Maintenance Monitoring

Abstract

Photovoltaic (PV) module performance is significantly affected by environmental factors, particularly solar irradiation, panel surface temperature, and soiling, requiring a monitoring and maintenance system responsive to varying PV operating conditions. This study proposes a maintenance condition classification system integrating real time sensor measurement with a feature-normalized k-Nearest Neighbor (k-NN) algorithm to identify four panel conditions (Neglected, PV Normal, Dusty PV, and Hot PV) requiring specific maintenance intervention. The methodological contribution comprises three aspects: an explicit labeling procedure with thresholds derived from empirical analysis of 250 field samples, avoiding circularity between label determination and classification features; a Min-Max normalization scheme ensuring each feature contributes proportionally to the Euclidean distance calculation in k-NN; and stratified 10-fold cross validation providing an unbiased estimate of generalization performance while revealing overfitting risks otherwise hidden in a simple train-test split. Experimental results show that the k=3 configuration achieves 95.2% ± 2.99% accuracy with superior consistency, balancing high accuracy and stability for deployment on power constrained microcontrollers such as the ESP32. The system integrates a Nextion graphical interface for real-time monitoring and condition notification, providing a platform for condition based maintenance decision making in community scale and industrial photovoltaic installations.