A Compact 1D-CNN for Photovoltaic Fault Diagnosis With Leakage-Aware Validation: A Comparative Study

Abstract

Reliable photovoltaic fault diagnosis depends on classifiers that maintain high accuracy under varying environmental conditions while remaining suitable for lightweight implementation. This study presents a one-dimensional convolutional neural network (1D-CNN) for six-class photovoltaic fault detection using normalized electrical features and compares its performance with alternative methods including a linear multiclass support vector machine, inverse-distance k -nearest neighbors, a compact feedforward neural network and a wavelet-assisted decision tree. A photovoltaic array test system was simulated using 39 months of irradiance and temperature data comprising 105,213 records. Six operating states, namely normal operation, small line-to-line fault, large line-to-line fault, open-circuit fault, partial shading and bypass-diode anomaly, were generated using 46,334 operating points obtained after daylight filtering with G ≥ 80 W/m2. To prevent data leakage, samples derived from the same irradiance–temperature condition were assigned exclusively to a single subset among training, validation, or testing. The proposed 1D-CNN achieved an accuracy of 98.95% and a macro-F1 score of 98.95% on the held-out test set, with a serialized model size of 0.162 MB. These results indicate that compact one-dimensional convolution can effectively support the proposed weather-driven photovoltaic fault test system. However, validation using measured field-fault data remains necessary before establishing long-term deployment viability.