Cloud Trajectory Correction Prediction Method Combining Lightweight Tracking and Historical Information Feedback
Shuang Chen, Jinwei Liao, Wen Liu, Xiang Yan, Yi Qiao, Yi He, Lin Li, Shuting Chen · IEEE Access · 2024
Aiming at the ineffective tracking problem of photovoltaic tracking system caused by cloud shading, an efficient and high-precision cloud trajectory prediction method is proposed. Firstly, the transmission method is used to filter the associated cloud type factors to reduce the computational complexity, and a GP-DCGAN gradient penalty deep convolution generation adversarial network is proposed to realize the generalization of image data. Secondly, by combining the alpha-IoU loss function with CA and CBAM attention mechanisms, an$\alpha $CCB-YOLOv5s model is proposed to achieve accurate identification of clouds. Finally, based on the$\alpha $CCB-YOLOv5s network, the DeepSORT module is added to realize multi-target tracking. Using a series of historical location information fed back by the tracking process, the Kalman filter algorithm is used to correct and predict the moving trajectory of the tracking cloud. The experimental results show that the detection categories of cloud type factors are reduced by 50% and the computational complexity is reduced by transmission method. At the same time, compared with the original YOLOv5s algorithm in cloud detection, the precision P of$\alpha $CCB-YOLOv5s model is increased by 4.56%, the recall rate R increased by 5.48%, the mean average precision mAP increased by 4.4%, and the target detection performance is significantly improved. In addition,$\alpha $CCB-YOLOv5s is used as a detector in the prediction model. The mean absolute error (MAE) and root mean square error (RMSE) are significantly reduced, and the prediction accuracy is greatly improved, which can lay a foundation for the follow-up intelligent photovoltaic tracking strategy research.