Research On Detection Method Of Photovoltaic Cluster Hot Spot Based On Deep Learning

Hot spot is one of the main reasons that affect the power generation efficiency of photovoltaic modules.However,conventional hot spot detection methods have low accuracy,slow detection speed,and poor robustness,which are difficult to meet the requirements of various types of photovoltaic power station hot spot detection.Therefore,it is of great significance to design a method with high precision,fast detection speed,strong robustness,and can meet the real-time detection requirements of various types of photovoltaic power stations.This thesis takes the infrared video of different types of photovoltaic power plants in the actual operation of UAV inspection as the research object.The specific research contents are as follows:（1）A photovoltaic string identification method based on improved YOLOX is proposed.Aiming at the problem that traditional image processing methods such as binarization and morphological processing can only solve the problem of PV string recognition in a single scene and have poor robustness,this deep learning target detection method is proposed,which is added to YOLOX’s backbone feature extraction network Dark Net53 The coordinate attention module and the self-attention module are used to capture the position information and channel information on the low-resolution feature map,which strengthens the feature extraction ability of the model for the low-resolution feature map,and conducts experiments for different types of photovoltaic power plants.The results demonstrate the effectiveness of the scheme.（2）A two-stage hot spot detection method is proposed.Aiming at the problems of low accuracy and high false detection rate in the traditional local extremum-based watershed algorithm for hot spot detection,this method is proposed.There are two types of hot spot faults,circular hot spot and rectangular defect.Based on the DenseNet-121 network structure,the network structure is simplified,and the Focal Loss loss function is introduced to solve the problem of unbalanced hot spot fault samples.Then,the improved DenseNet classifier is used to identify the hot spot fault and the interference of light and current collector,and the identified hot spot is marked in the infrared image of the photovoltaic string,while the interference of light and current collector is eliminated.The algorithm has certain accuracy and robustness,and can basically meet the requirements of different types of photovoltaic string hot spot detection.（3）A hot spot detection method based on improved YOLOv5 s is proposed.Aiming at the problem that the two-stage hot spot detection method proposed by the previous method is slow in accurate measurement and difficult to achieve real-time detection,this method is proposed.Under the basic network framework of YOLOv5 s,the backbone feature extraction network is replaced by a lightweight network Mobile Netv3 At the same time,an ultra-lightweight attention mechanism ECANet is introduced,and testing and analysis are carried out for different types of photovoltaic strings and several other target detection algorithms.On the basis of further improving the accuracy of hot spot detection,this method greatly improves the Improved hot spot detection speed.