Optimizing Autonomous Aerial Monitoring of Photovoltaic Power Plants via an Integrated Software Package and a Digital Twin Based Simulation Environment

This paper presents an advanced embedded software package tailored for the autonomous aerial monitoring (AAM) of photovoltaic (PV) plants. It employs an encoder-decoder deep learning model to accurately pinpoint the boundary points of the PV plants. Additionally, a unique path-planning algorithm guarantees comprehensive coverage of the monitoring area. A highly precise neural network is also utilized to analyze images in real time, enabling automatic fault detection. To improve performance during inspections, custom decision-making, and maneuvering algorithms are designed to adapt to various flight conditions. To showcase and optimize the software's performance and test autonomous PV monitoring flights and missions, a virtual environment is also presented. This innovative platform allows for the examination of different scenarios and configurations of PV power plants, assessing their impact on the AAM process. It features tools for generating data that enable the development of intelligent monitoring and inspection models. The creation of this platform involved building a digital twin of a PV plant using Unreal Engine, simulating drone flight with AirSim, and expanding the application programming interfaces (APIs) to adapt to various scenarios for evaluating smart monitoring models and collecting datasets. Additionally, a dataset of aerial images was compiled from this platform to train a segmentation model aimed at identifying bird droppings on PV panels.