A comparative study of AI-based and signal processing techniques for condition monitoring of power transmission networks

Abstract

Transmission lines are essential for transporting power over long distances as far as it is not always effective and it may also fail thus putting the flow of power at risk and the people who depend on power at risk as well. This is why the maintenance of their reliability and efficiency is of paramount importance in energy infrastructure of the contemporary world. This research aims at exploring the detection and classification of faults in power networks. Here, a novel hybrid model, DWT- CNN-BiLSTM, is presented which combines signal processing and deep learning and compares it with the classic machine-learning and deep learning approach. A dataset of 11,000 current and line-voltage samples that were provided in MATLAB SimPowerSystem evaluated the proposed model. Wavelet transform breaks down the signals by time-frequency, which represents transient faults as well as attenuating noise. The resulting coefficients are then inputted into a CNN, which produces spatial features, which are then inputted into a BiLSTM that yields bidirectional time information. One last Softmax layer allows fault detection and classification of multiple classes reliably. It is a robust, high-accuracy end-to-end pipeline that can be deployed in real time, and the accuracy of fault-detection is 99.86 %, and the accuracy of classification is 93.75%. Further proof of the effectiveness of the hybrid model is that the hybrid model surpasses single-method machine-learning and Deep learning baselines in terms of precision, accuracy and recall. Such findings show that the model can be applied in practice in real-time as a way of enhancing system stability in power-generation, predictive maintenance, or operational resiliency.