Application of Deep Neural Networks for Fault Detection in Modern Power Systems

For this purpose, this paper proposes a robust co-simulation methodology leveraging deep neural networks (DNN) for accurate and rapid fault location detection and classification. The methodology for developing a DNN model, consists of four key stages. The first stage involves modelling the power system network within a simulation environment, where essential system parameters such as line impedances, bus configurations, and load characteristics are initialised. In the second stage, fault scenarios are simulated based on probabilistic distributions, accounting for various fault types, fault impedances, and fault locations to reflect realistic conditions. This stage is also dedicated to data acquisition, where the simulated fault conditions are applied to the test system model to generate measurement data, such as voltage and current waveforms during fault events. The third stage focuses on developing the DNN model using the acquired data, which includes preprocessing the input features, training the model, and performing hyperparameter tuning to optimise its performance. Finally, the fourth stage involves validating the trained DNN model against theoretical two - end impedance - based fault location methods.

Experimental results demonstrate the advantages of the DNN approach in terms of speed, accuracy, and adaptability, underscoring its potential as a transformative tool and redundancy system for fault detection in complex power systems. The DNN model exhibits a low Mean

Absolute Error (MAE) across three short-circuit scenarios when comparing predicted and actual fault locations. Specifically, the MAE for the single line-to-ground, three-phase, and two-phase scenarios is 0.24, 0.03, and 0.04, respectively. In contrast, the impedance-based method achieves an impressive MAE of 0.00, indicating perfect accuracy for the scenarios it can calculate. However, under the test conditions the DNN model demonstrated a significantly faster execution time—processing 10,000 test samples in just 0.51 seconds— compared to 9.38 seconds required by the theoretical method. Although the DNN model exhibits slightly lower accuracy, its rapid processing enables operators to identify faults much more quickly, offering substantial advantages in time-critical scenarios.

While numerous traditional methods, such as the two-end impedance-based method, can provide precise fault locations under specific conditions, there is a growing imperative to incorporate AI approaches to enhance the capabilities of modern control centres as a powerful redundancy system to prepare for the complexities of future grid operations.