Temporal convolution attention–based reinforcement learning for enhanced distance protection

Distance protection is valued for stability, reliability, and fault discrimination in transmission lines. However, high-impedance fault currents can be comparable to load currents, causing distance relays to underreach and fail to operate. This paper proposes an impedance-based deep reinforcement learning (DRL) model that supplements conventional distance protection and detects resistive faults (1–100 Ω) in AC networks with distributed generation. The model combines maximal overlap discrete wavelet transform (MODWT) sub-band energies with a causal, dilated temporal convolutional network enhanced by temporal self-attention, and uses proximal policy optimization as the policy-learning algorithm. The main novelty is decision-level optimization via a selectivity-first reward that penalizes false trips and latency, avoids hand-tuned time–frequency thresholds, and improves robustness under harmonic distortion and wide ranges of fault resistance and impedance. The approach is evaluated on a modified IEEE 14-bus system, IEEE DataPort photovoltaic and wind-farm datasets, an IEEE 34-node feeder benchmark, and French transmission-grid datasets. Controller hardware-in-the-loop validation using a real-time digital simulator confirms real-time feasibility. Using local voltages and currents sampled at 32 samples per cycle, the method achieves an average detection time of approximately 27 ms and near-100% accuracy at 30 dB signal-to-noise ratio, outperforming conventional distance protection across the tested scenarios. These results indicate that the proposed scheme provides a practical and secure supplementary layer for distance protection under challenging fault conditions.