تشخيص خطا در آرايه‌هاي فتوولتاييك به كمك شبكه‌هاي عصبي

With the increasing use of clean energy, especially solar panels, and since the efficiency of this method is relatively low, obtaining the maximum power and detecting faults in solar arrays have become very important. Because undetected faults not only reduce the efficiency due to power losses but also reduce the lifespan of solar panels or even cause fire hazards. in recent years, the development of high-speed processors and the possibility of recording multiple datasets, facilitate the use of deep learning networks for that matter. In this research, four deep learning architectures like AlexNet, GoogleNet, SqueezeNet and ResNet were used to classify common faults in solar arrys. For this purpose, the common faults both simulatically and practically were implemented in the MATLAB Simulink environment and with practival setup respectively, so 3240 data including six classes were obtained. Since the input of the aforementioned networks is two-dimensional data in the form of image, continuous wavelet transformation has been used to convert numerical 1-D data into 2-D images. these 3240 data are randomly divided into two parts: Training data and validation data consisting 2592, 972 data, respectively. Each network classified common faults into 6 class and the performance of the deep learning model is eva‎luated by the Accuracy index. In order to gain better convergence. the 1-D numeric input data has been normalized to the maximum value before being converted into an image. Of course, Matlab normalizes data after being imaged. The average classification accuracy of AlexNet over the 6 classes for simulation data and real data was 78.52% and 78.5% respectively, of GoogleNet was 77.47% and 71.60% respectively, of SqueezeNet was 80.76% and 71.30% respectively, of ResNet was 81.14% and 75.41% respectively and of MLP was 75.93% and 75.41% respectively.

5