Photovoltaic Module Hot Spot Identification Based On Convolutional Capsule Network

The world is facing problems such as the depletion of traditional energy sources and the deteriorating environment and climate.At the same time,in order to solve the imbalance of energy supply and demand in our country and achieve sustainable social development,it is necessary to transform the economic structure and take the road of green development.Taking into account various factors,solar energy is one of the most suitable new energy sources for development at the current stage.The utilization of solar energy is very rich,among which photovoltaic power generation is the most widely used.However,the problem of hot spot fault in the process of power generation has become one of the important factors restricting the safety of power generation.Finding an efficient and accurate method to timely detect hot spot faults during the operation of photovoltaic power plants is of great significance for maintaining the safe operation of power plants,improving power generation efficiency,and saving maintenance costs.In this paper,the infrared images of photovoltaic arrays collected during the actual operation of photovoltaic power plants are taken as the research object,and a method for hot spot detection of photovoltaic modules based on convolutional capsule network is proposed.And improved on this basis,an attention capsule network hot spot detection method based on transfer learning is proposed.The specific research contents are as follows:(1)Photovoltaic module infrared image processing,including three aspects: first,image preprocessing,the original image has a lot of noise,and median filtering,top hat transformation,and erosion expansion are used for denoising;second,image segmentation is carried out,and an optimized Canny edge detection algorithm is proposed to suppress the isolated weak edges,making the results more accurate and the component outline clearer.Finally,data enhancement is performed,and after perspective transformation of the segmented components,histogram equalization is used to increase the hot spot.Contrast of the area.Secondly,the components are divided into battery slices,naming rules are formulated to facilitate subsequent positioning,and then the data set is expanded by traditional methods.(2)For the processed data set,a hot spot detection model based on shallow convolutional neural network and VGG-19 network is built respectively.According to the recognition results,the disadvantages of convolutional neural network are analyzed.Aiming at these drawbacks,a hot spot identification method based on convolutional capsule network is further proposed,and then the training results and test results are analyzed.(3)Combining transfer learning with attention capsule network structure,a hot spot detection method of attention capsule network based on transfer learning is proposed.In the case of a small amount of data,the generalization ability of the model is improved,the occurrence of overfitting is prevented,and the accuracy of hot spot detection is improved.Finally,according to the naming rules of the cell,the location of the hot spot is located and marked,and the visualization of the detection result is realized.Experiments show that the hot spot detection method proposed in this paper can efficiently and accurately identify and locate the hidden hot spot area,and has good application and reference value for the detection of hidden hot spots or faults in photovoltaic power generation sites.