Transformer-Based Fault Section Identification in DER-Integrated Distribution Systems Under Communication Failures

(DERs), fault management has become increasingly complex from an operational perspective.

In particular, fault section identification plays a critical role in post-fault decision-making, as it directly determines isolation boundaries and restoration strategies. Conventional fault section identification methods rely on deterministic, rule-based logic using Fault Indicators (FIs), assuming radial configurations and unidirectional fault current flow. However, in DERintegrated distribution systems, reverse current contributions, unbalanced fault conditions, and inverter control characteristics frequently violate these assumptions, leading to ambiguous or unreliable FI behavior.

This paper proposes an operationally oriented fault section identification framework for DERintegrated distribution systems based on Transformer-based attention modeling. Rather than replacing existing protection schemes, the proposed approach is designed to support post-fault operational decision-making under incomplete and uncertain information. Fault section identification is formulated as a sequence-based decision problem, where each FI and its associated steady-state fault current magnitudes are treated as tokens in an ordered sequence representing the feeder configuration after fault clearance. The physical order of switches along the feeder is preserved through positional encoding, enabling the model to capture spatial relationships among FIs without relying on explicit topology-dependent rules.

The core of the framework is a Transformer encoder employing multi-head self-attention mechanisms to learn contextual dependencies among FI tokens. This allows the model to assign adaptive importance to individual FIs based on the overall fault pattern, even when FI data are partially missing or incorrectly activated. By focusing on global relationships rather than local deterministic rules, the framework enhances robustness against FI misoperation, data incompleteness, and frequent feeder reconfiguration. The Transformer output is mapped to a probability distribution over candidate fault sections, providing interpretable decision support for operators.

To evaluate the proposed framework, extensive case studies were conducted using large-scale synthetic datasets generated through distribution system simulations. The datasets encompass diverse operating conditions, including multiple feeder topologies, load levels, DER capacities, fault types, fault locations, and fault impedances. Importantly, scenarios involving FI communication failures and partial observability were explicitly considered to reflect realistic operating environments. Unlike approaches that rely on waveform data, synchronized phasor measurements, or directional FI installations, the proposed method operates solely on steadystate fault current magnitudes and basic feeder layout information, improving its practicality and scalability for integration into Distribution Management Systems (DMS).

The results demonstrate that the proposed framework improves the reliability of fault section identification under DER-integrated conditions, particularly in scenarios where conventional

FI-based logic fails. By reformulating fault section identification as an operational decision problem and leveraging attention-based modeling, this work provides a practical pathway toward more robust fault management in modern distribution systems.