Edge AI for real-time anomaly classification in solar photovoltaic systems

Abstract

This study details the development and deployment of a real-time anomaly classification system on edge AI devices for solar systems. We used a neural network model, fine-tuned using the keras-tuner library, resulting in an average accuracy of 97.95%. Our optimal model demonstrated a robust performance with an accuracy of 97.84% and a small size (31.531 kB). We applied quantization as a model reduction technique, substantially decreasing the model size to 7.455 kB while maintaining similar accuracy. The reduced model was successfully implemented on various edge AI platforms, with STM32F767 Nucleo-144 proving to be the most cost-effective and energy-efficient. The study suggests further research on different solar systems and a comprehensive cost-effectiveness analysis for large-scale deployment.