Data Incremental Learning for Fault Detection and Classification in Photovoltaic Arrays

In this paper, a data incremental learning algorithm known as Learn++ is embedded in a very simple Multi-Layer Perceptron for fault detection and classification in Photovoltaic arrays. Usually in PV arrays, classifiers are limited to classical batch setting where data are provided before training, and thus optimization of meta-parameters and the selection of model become dataset dependent. Moreover, during training it is assumed that the dataset and its underlying structure are static and since in PV systems constantly new batch of data arrives over time, the classifiers are likely to misclassify the previously unseen data. In an attempt to learn new information, completely stable classifiers lose previously acquired knowledge which is very undesirable. Moreover, completely plastic classifiers learn new information but cannot retain their previous knowledge. Learn++ algorithm is employed in this paper to address the stability-plasticity dilemma of feedforward MLP as base classifier to learn incrementally without undergoing catastrophic forgetting. Using MATLAB simulations, the incremental learning of the MLP based on Learn++ is demonstrated by classifying line-ground and line-line faults in two PV arrays system with 97% classification performance on a test set, TEST , containing 300 instances.