Sistema de localización automático de fallas por el método de ondas viajeras

Abstract

This research presents a methodology for the fault location in high-voltage transmission lines, combining traveling wave analysis with deep learning models. The proposed approach integrates electromagnetic simulations performed in ATPDraw, automated event generation using Python, and a robust signal processing pipeline that includes high-frequency filtering and modal transformation. The resulting current waveforms—extracted from both terminals of the transmission line—serve as input to a deep neural network composed of convolutional and recurrent layers (CNN-LSTM), trained to classify the fault position along the line with one-percent granularity. To evaluate the effectiveness of different model configurations, a comparative study was conducted involving eight architectures with varying levels of complexity. Each model was trained under identical conditions using a large dataset of simulated fault events spanning a wide range of fault types, positions, and impedances. Metrics such as global accuracy, macro-averaged F1-score, and per-class performance were used to guide the model selection. The final architecture, which includes one convolutional layer followed by two LSTM layers and dense output neurons, achieved a validation accuracy close to 99%. The training dataset was generated with a sampling rate of 1 MHz and a temporal window of 1.5 milliseconds, sufficient to capture both the initial wavefronts and relevant reflections at the remote terminals. High-pass filtering with a Butterworth response was applied to preserve the transient components of interest, while the Clarke transformation facilitated the decoupling of phase interactions by projecting the signals into the α–β frame. These preprocessed signals were then standardized and used to train the model in a supervised manner. Model performance was analyzed in detail, not only in terms of overall metrics, but also through the inspection of the confusion matrix and class-specific errors. The analysis revealed that most misclassifications occurred between neighboring classes, typically within a ±2% error margin, which remains operationally acceptable in most fault location scenarios. Nonetheless, a group of underperforming classes—mostly located in the central portion of the line was identified and examined in a reduced confusion matrix. These difficulties are attributed to the intrinsic similarity of the waveforms and the physical proximity between adjacent fault points. The proposed methodology demonstrated strong generalization capabilities and consistent behavior throughout the entire classification domain. Its design allows for future extensions, such as incorporating voltage signals, adapting the approach to more complex network topologies, or deploying the trained model on embedded hardware for real-time applications. Overall, the study provides a reliable and scalable solution for automated fault location in transmission systems, leveraging the strengths of both physical modeling and data-driven learning.