Deep Learning-Based Faulted-Phase Selector for Transmission Networks with High Renewable Penetration Level

CIG shows reduced current magnitudes, higher harmonic content and faster variations of sequence components, which compromises the reliability of conventional phase selection algorithms used in protection schemes based on stable phasor relationships.

This paper presents the design, training and real-time implementation of a deep learning-based faulted-phase selector, which operates directly on instantaneous voltage and current signals without explicit phasor estimation. The objective is to provide a robust identification of the phase or phases involved in the fault, even under adverse conditions such as low fault current contribution, harmonic distortion or high fault impedance.

The methodology combines electromagnetic transient simulations (EMT) in two complementary systems: (i) an IEEE 14-bus network adapted with type-IV wind generation, and (ii) a radial network with photovoltaic generation executed in real time. From these environments, more than 40000 fault scenarios were generated, with systematic variation of fault type, impedance, location and loading conditions. Voltage and current signals were sampled at 4 kHz and segmented into 10 ms windows (40 samples), including residual voltage and current components computed in the time domain as the sum of phase quantities, without explicit symmetrical component transformation.

Three deep neural network architectures were evaluated: a pure Transformer model, a hybrid

CNN–Transformer, and a hybrid LSTM–Transformer. The latter provided the best performance in terms of accuracy and stability under varying conditions, by combining local temporal feature extraction with global attention over the input window. The selected model was exported to

ONNX format and integrated into an edge computing platform, executing embedded inference in C++ within a digital substation architecture compliant with IEC 61850.

The validation was carried out by means of Hardware-in-the-Loop (HIL) tests, using IEC 61850-9-2LE Sampled Values for analogue inputs and GOOSE messages for the phase selection output. The system was evaluated using fault scenarios not included during the training stage, achieving high classification performance together with inference times in the order of a few milliseconds, which are compatible with the time constraints typically required by protection-related functions.

The results confirm that deep learning architectures applied to short windows of instantaneous signals represent a viable alternative for faulted-phase selection in digital environments, particularly in networks with high penetration of converter-interfaced renewable generation.

Thanks to its seamless integration with IEC 61850, the proposed selector can operate as an auxiliary function to enhance distance and directional protection schemes in modern substations.