In the past ten years, many artificial intelligence methods have been successfully applied to the fault detection and diagnosis of photovoltaic arrays, but most of them rely on complex manual feature extraction and a large amount of actual monitored fault data. Because the scale of the actual photovoltaic power plant is too large, it is difficult to obtain a large amount of data. In the case of small training data, it will severely limit the performance of the fault diagnosis model. In response to these problems, this paper first proposes a multi-element time series waveform based on maximum power point tracking as the fault feature. The interference data is filtered through the change point algorithm to obtain the fault feature with absolute high quality and complete information. We also designed a simulation model of photovoltaic array based on Simulink, and obtained simulation data and experimental data through simulation and actual experiment. In addition, this paper proposes a fault detection and diagnosis method based on transfer learning for densely connected convolutional neural networks (DenseNet), using simulated data for pre-training, and then using experimental data for fine-tuning. The comparison of experimental results proves that the fault diagnosis method can achieve better results in terms of accuracy, generalization performance, and reliability in the case of small samples.
Keywords Fault diagnosis, DenseNet, Transfer learning, Time series