Detecting and classifying faults reliably in smart grid infrastructure is one of those problems that looks straightforward on paper but gets progressively harder the closer you look at it. Modern distribution networks are no longer the passive, radial systems that conventional protection theory was built around — they carry bidirectional power flows from distributed generators, host large numbers of inverter-interfaced resources, and are monitored by sensing equipment that produces far more data than any human operator can interpret in real time. Standard overcurrent and impedance relays, designed for an era of predictable fault currents, struggle badly with this landscape, particularly when it comes to high-impedance faults whose signatures can be almost invisible to threshold-based detection logic.

This paper presents CNN-BiLSTM-Attn, a hybrid deep learning architecture that brings together one-dimensional convolutional neural networks, bidirectional long short-term memory units, and a multi-head self-attention mechanism in a single end-to-end trainable model. Rather than treating fault detection, fault type classification, and fault localization as three separate problems, we cast them as a unified multi-task learning objective and train a single network to solve all three simultaneously. The model was evaluated on IEEE 13-bus and IEEE 34-bus test feeder datasets extended with PSCAD electromagnetic transient simulations, giving a total of 128,000 labeled waveform samples spanning ten fault categories. On a held-out test partition of roughly 19,000 samples, CNN-BiLSTM-Attn achieved an overall classification accuracy of 99.41%, a macro-averaged F1-score of 0.9933, and a mean fault localization error of 0.87% — improvements of between 2.3 and 8.7 percentage points in accuracy over six comparison methods. Crucially, inference on quantized embedded hardware runs in under 2 ms, leaving ample margin within the 80 ms response budget required by IEEE protection standards.

Index Terms—Smart grid, fault detection, convolutional neural network, bidirectional LSTM, attention mechanism, power system protection, deep learning, fault classification, transient analysis.