Design of an improved model for real-time fault detection and localization using hybrid transformer-CNN and graph neural networks

Accurate real-time fault detection, localization, and classification techniques are necessary to maintain grid stability and prevent faults. Traditional techniques have low accuracy rates, with high latency, and are generally incapable of handling the inherent non-linearity and dynamicity of power systems. Bearing these limitations in mind, we propose the novel Multi-Stage Deep Learning Framework, which significantly improves the process of fault management. First, it uses a hybrid Transformer-Convolutional Neural Network (CNN) in the process for fault detection. This is because transformers will capture the long-range temporal dependencies by voltage and current data, while CNNs will extract localization fault features. This approach can achieve an estimated fault detection accuracy of 99.3% with a real-time latency of less than 50 ms. For fault localization, a Graph Neural Network (GNN) is employed for modeling the power grid as a graph, hence enabling learning of spatial dependencies between nodes. This approach achieves 97.80% accuracy in the detection of exact fault location and completes within 100 ms. Next, the classification of faults uses an RNN-LSTM network to overcome temporal dependencies in the evolving nature of fault signals. The model reaches 98.20% classification accuracy by completing within 120 ms, which signifies that the model could differentiate between various kinds of faults in real-time applications. This framework improves the accuracy and speed of the fault management process, which directly contributes to the resilience and reliability of the power transmission system, thus reducing downtime period and increasing operation efficiency.