A novel solar tracking algorithm based on robotic object recognition techniques

Objective: The solar position tracking accuracy of solar systems with trackers and adapting system to both weather and seasonal conditions straightly affect the amount of output energy and the system stability. This study aims to develop a robotic technology-based computer vision algorithm to implement on a parabolic solar dish system. Material and Methods: The scope of the study includes deep learning, image processing, deep neural networks and automatic control theory since the solar tracking systems are dynamic systems consisting different types of sensors, sky image cameras and control devices such as automated valves, encoders, switches, irradiation and wind sensors etc. The proposed algorithm is based on the sky images taken by the camera which will be coded on a small sized circuit board computer, that has a microprocessor, memory and random-access memory. Results: The percentage error rate which indicates the accuracy of the algorithm is calculated both x- and y-axes. The error rates are 0.081502% and 0.533543% for x and y-axes, respectively. The proposed algorithm detects the position of the sun and drives tracking motors with less than %1 error. The processed images of the algorithm were validated by different trained Tensorflow object detection models such as CenterNET ResNet50, SSD ResNet101 and EfficientDet D1. The trained models have proved to detect the sun in the range of 86% to 97%. According to these results, sun detection and image processing of the detected images were successfully performed in order to find the center of the circle and move the mechanism to the center of the sun.