Fault diagnosis in renewable-integrated distribution systems using EMD-GAF and ANN

The increasing integration of distributed renewable energy sources and dynamic loads has made fault detection in modern distribution systems significantly more challenging. Traditional protection schemes often fail to accurately distinguish between faults and non-fault disturbances such as switching events, islanding, or power quality anomalies, which can lead to delayed or incorrect responses. This paper proposes a fast and reliable fault diagnosis technique integrating Empirical Mode Decomposition (EMD), Gramian Angular Fields (GAF), and Artificial Neural Networks (ANN) to detect, classify, and locate faults in renewable-integrated distribution networks. Three-phase current and voltage signals are first decomposed using EMD to extract low-frequency residues, that are then transformed into two-dimensional GAF visual patterns. Cosine similarity compares these patterns against reference healthy conditions for fault detection. For fault localization, an ANN is trained using statistical features from four levels of EMD residues. The proposed method achieves over 99.5% accuracy in fault detection and classification using only 0.25 cycles of post-fault data and single-point current and voltage measurements at the substation, even under noisy (20 dB SNR) and high-impedance (up to 5 Ω ) conditions. It outperforms existing signal-analysis-based and visual-pattern-based techniques by accurately distinguishing faults from switching and islanding events, making it a robust and scalable solution for real-time smart grid protection. Furthermore, the method achieves up to 99.04% bus-level fault localization accuracy and reduces distance-to-fault errors by over 25% compared to existing techniques, further enhancing suitability for protection and precise fault location. • Ultra-Fast Fault Detection with Minimal Data: Achieves over 99.5% detection and classification accuracy using only 0.25-cycle post-fault data using EMD-GAF based hybrid fault detection and classification technique; Requires single-point measurement only, significantly reducing hardware/sensor complexity; Outperforms existing methods that need longer durations ( ⩾ 0.5 cycles) or multi-point data. • High Fault Localization Accuracy Across Systems: Delivers 95.87% location accuracy in the IEEE 13-bus system and 99.04% in the IEEE 34-bus system (with DG); Outperforms Existing Decision Tree based classifiers by up to +3.57% improvement in location accuracy under comparable conditions. • Superior Distance-to-Fault Estimation: Reduces distance-to-fault estimation error by 20%–30% compared to existing DTR-based methods; Performs well even for buses near the substation or with high fault resistance (up to 5 Ω ). • Robust Performance in Noisy and High-Impedance Environments: Maintains high detection and location accuracy at low SNR (20 dB) and fault resistance up to 5 Ω ; Uses cosine similarity of GAF-encoded visual patterns to reliably detect and classify faults even in low signal-to-noise conditions. • Discrimination from Non-Fault Events: Accurately distinguishes faults from switching transients, islanding, and power quality disturbances; Similarity indices for non-fault events remain well below the 0.55 fault threshold, e.g., 0.21–0.32 for switching events, ensuring zero false positives. • Scalability and Generalization: Validated on both IEEE 13-bus and 34-bus systems, with and without Distributed Generation; Model scales effectively without requiring fault-type-specific thresholds, and works across a wide range of test scenarios (e.g., > 80,000 cases simulated across both systems).