Real-Time Fault Detection and Classification in Power Grids Using Hybrid TCN-LSTM

In the modern era, real-time fault detection is crucial for ensuring the stability and safety of power grid systems due to the dynamic integration of Distributed Generation (DG) sources into conventional electricity systems. The traditional models suffer significantly when faced with incremental variations in operational and environmental conditions. This research therefore proposed a hybrid deep learning architecture based on Temporal Convolutional Network (TCN) and Long Short-Term Memory (LSTM) to provide both accurate and efficient diagnosis of faults. More specifically, the proposed methodology includes the generation of IEEE 6-bus and 9-bus models to first demonstrate the fault scenario and generate time-series data of 3-phase voltage and current signals. The collected noisy data is then normalized using Min-Max normalization to accelerate the trend change towards zero during training. Then, three voltage and three current signals are converted into 3-dimensional time-series tensor signals then passed into temporal convolutional layers for extraction of temporal-spatial features using causal dilative convolutions. The output of the TCNs is further fed into LSTM layers to ensure discovery of long-term dependencies in the data and the change of fault patterns over time. A fully connected dense layer followed by a SoftMax classifier to produce predictions of fault type, faulty line and fault location. The proposed TCN-LSTM attained better results when compared Convolutional Neural Network-LSTM with accuracy (99.89%), precision (99.91%), recall (99.79%) and F1- Score (99.81%) respectively.