Optimized lightweight ANN with domain-driven feature engineering for fault-diagnosis in grid-connected photovoltaic systems

Reliable fault-diagnosis in grid-connected photovoltaic (PV) systems requires models capable of distinguishing subtle variations in electrical measurements, where the separability between certain fault conditions is limited. To address this challenge, this work employs domain-driven feature-engineering through the derivation of physically meaningful features to enhance class separability and develops an optimized lightweight Artificial Neural Network (ANN) architecture incorporating Layer Normalization, SiLU activation, dropout regularization, and scaled residual connections, coupled with focal loss to improve learning from difficult-to-distinguish fault patterns. These architectural components enhance gradient stability, strengthen feature representation, and maintain computational efficiency. Experiments conducted on a laboratory-scale PV dataset comprising seven fault scenarios demonstrate that the proposed ANN achieves superior F1 score, faster convergence, and reduced misclassification compared with baseline ANN models trained using standard cross-entropy or focal loss alone. The results indicate that the synergistic integration of feature-engineering, architectural refinements, and focal-loss-based optimization substantially improves fault discrimination under highly overlapping fault conditions. The proposed lightweight ANN has only 36,488 trainable parameters, highlighting its potential suitability for future resource-constrained photovoltaic monitoring applications.