Abstract
In this work, a drone-based system is implemented for autonomous real-time fault detection in solar farms for improved reliability of power generation. To capture and analyze high-dimensional features from aerial images of solar panels, three powerful feature-extracting models are used. Then, for feature optimization, a new modified hybrid metaheuristic called Red Panda-Dhole Optimization (RPDO) is proposed. For accurate classification, a residual boosting-based Long Short-Term Memory (LSTM) architecture is proposed. The developed drone, based on the STM32F427VG microcontroller, follows a predefined flight path autonomously and captures images of the solar farm at regular intervals. These images are transmitted to a standalone system where they are processed and classified into different fault categories using a trained model. The proposed model is trained and tested on inspection videos recorded with drones in solar farms. The model achieved the highest accuracy of 96.2% for the classification of solar panel status into Bird-drop, Clean, Dusty, Electrical-damage, and Physical-damage

This publication has 33 references indexed in Scilit: