Online fault diagnosis of PV array considering label errors based on distributionally robust logistic regression

Abstract

This paper proposes a robust diagnosis method of photovoltaic (PV) array faults considering label errors in training data. First, the online data of PV systems, including the sequences of voltages, currents, and output power at maximum power points, are used to establish the input data of fault diagnosis. Second, a data processing method is used to extract fault features from electrical signals under fluctuating ambient conditions. Third, the parameter estimation of the regression-based fault diagnosis model is formulated as a stochastic optimization problem. To hedge against label errors, an ambiguity set of probability distributions is established from training data, and a distributionally robust logistic regression method is proposed to minimize the expected log-loss function under the worst-case probability distribution for obtaining model parameters of fault diagnosis. Finally, the proposed method is tested on real-world PV arrays under diverse conditions and scenarios. Data processing increases diagnosis accuracy by 18.4% when training data is error-free. The diagnosis accuracy is higher than 98% when the label error rate is smaller than 4%.