PV module fault detection technique based on convolutional neural network

A group of researchers led by the University of Sharjah in the UAE proposed to use the convolutional neural network (CNN) technique to detect temperature and shading-induced faults in PV modules. CNN is a deep learning algorithm that enables extracting and learning features from visual data.

“CNN extracts feature maps from data sets using kernels and filters that slide over input features,” the researchers said. “Transfer learning is applied to train a baseline model in which certain features were extracted to train our main model due to the limited experimental data.”

The team first used CNN and transfer learning models to train and test the faults of external solar data and then artificially decreased and tested these data under different numbers of epochs. Transfer learning is a machine learning technique where a model trained on one task is re-purposed on a second related task.

Their dataset classified seven operating conditions: PV with no shading and temperature effect; PV working under the effect of one cell shading in one string; PV working under the effect of two cells shading in one string; PV working under the effect of three cells shading in two strings; PV working under the effect of four cells shading in two strings; PV working under operating temperature 25 C; and PV working under operating temperature 50 C.

The scientists created a first model using only CNN and a second model based on the more extensive Mendeley database, which is not directly applicable to the fault detection system. This database includes two measurements – maximum power point (MPPT) and intermediate power point tracking (IPPT) – under the above-mentioned seven operating conditions. The latter includes a fault in the inverter, a fault in a feedback sensor, two conditions of array mismatch, a grid anomaly, a fault in the MPPT/IPPT controller, and a fault in the boost converter controller.

“One could use transfer learning and train a data-driven model from external data, i.e., data not directly applicable to the fault detection system, which is then used to classify the faults of the experimental data,” explained the researchers. “Transfer learning is a method within deep learning where the goal is to train a base model from a large data set, which is then used to train a model on novel data; the model uses the knowledge from the base model and, as such, does not need to acquire said knowledge by itself.”

Both models were tested with few data variations – all available data, half of the data, a third of the data, and one-fourth of it. They were also tested with different epochs – the number of complete passes on the given database – of 50, 25, and 10.

Per the results, in the case of 50 epochs, the transfer learning model got an average accuracy of 96.6%, and the new model got 97.1%. For the first ten epochs, they achieved an accuracy of 93.9% and 89.8%, respectively.

“The main goal of this study is to show how the transfer learning model built based on the large data set from solar data from Mendeley has an accuracy of 73.9%, 81.9%, and 93.9% for fourth, half, and full training data, respectively, using ten epochs. In comparison, using ten epochs, the new model has an accuracy of 26.7%, 60.8%, and 89.8% for fourth, half, and full training data,” the researchers highlighted.

“The use of transfer learning is a great asset in analyzing data sets with limited data or when computational complexity is inhibited,” they concluded. “The increased classification accuracy from using transfer learning models instead of training a new model can reach >170%. Therefore, operating with transfer learning models is highly recommended when the experimental is sparse and access to similar data is available.”

The novel PV module fault detection methodology was presented in “Detecting the faults of solar photovoltaic module due to the temperature and shading effect by convolutional neural network,” published in the International Journal of Thermofluids. The research team also included academics from Egypt’s Minia University, the United Kingdom’s Aston University, and Sweden’s Linköping University.