CNN-based photovoltaic fault diagnosis using normalized I–V curves with Explainability analysis

• Measured PV string I–V curves were normalized with a simulation model. • Normalized I–V curves were classified as normal or faulty using key features. • The trained CNN reached 99.3% training and 99.4% validation accuracies. • Fault diagnosis model was analyzed via layer-wise analysis and occlusion sensitivity. • Fault diagnosis of PV strings was validated through test data evaluation. Countries worldwide are expanding the adoption of renewable energy to achieve carbon neutrality by 2050. Among the renewable sources, solar energy has experienced the fastest growth and largest deployment. As the number of photovoltaic (PV) plants increases, the operation and maintenance market expands, along with fault-diagnosis technologies that integrate traditional methods with artificial intelligence. This study proposes a fault-diagnosis technique that utilizes normalized current–voltage (I–V) curves of PV strings and a convolutional neural network (CNN). Measured I–V curves were normalized using a simulation model considering irradiance, module temperature, and degradation rate. The normalized curves were labeled as normal or as one of six fault types based on patterns and electrical parameters. A CNN trained with these data achieved training and validation accuracies of 99.34% and 99.39%, respectively. Layer-wise and occlusion sensitivity analyses were performed to interpret the classification process of CNN. Additionally, in a PV string where normal and faulty conditions were simulated, the trained CNN classified measured I–V curves with an average accuracy of 98.3%. When evaluated at an operational PV plant, the PVDF model prioritized the dominant I–V curve pattern for fault classification and successfully classified faults with subtle patterns.