# Dual-Head Physics-Informed Graph Decision Transformer for Distribution System Restoration

arXiv: https://arxiv.org/abs/2508.06634 · HTML: https://mlsanity.com/paper/2508.06634/

## Key facts

- No citing papers in the map yet (data through 14 Oct 2026).
- Placed in Decision transformers and sequence modeling, one of 13 directions in Offline RL and imitation learning from data.
- Builds on 7 of its 19 references that are in the map.
- First posted on arXiv 8 Aug 2025.

## Placed in

- Offline RL and imitation learning from data / [Decision transformers and sequence modeling](https://mlsanity.com/offlinerl/decision-transformers-and-sequence-modeling/)

## Builds on (most cited first, 7 of 7)

- [Proximal Policy Optimization Algorithms](https://mlsanity.com/paper/1707.06347/) (2017)
- [Decision Transformer: Reinforcement Learning via Sequence Modeling](https://mlsanity.com/paper/2106.01345/) (2021)
- [Prompting Decision Transformer for Few-Shot Policy Generalization](https://mlsanity.com/paper/2206.13499/) (2022)
- [A Time-Series Distribution Test System Based on Real Utility Data](https://mlsanity.com/paper/1906.04078/) (2019)
- [Batch-Constrained Reinforcement Learning for Dynamic Distribution Network Reconfiguration](https://mlsanity.com/paper/2006.12749/) (2020)
- [Multi-objective Cross-task Learning via Goal-conditioned GPT-based Decision Transformers for Surgical Robot Task Automation](https://mlsanity.com/paper/2405.18757/) (2024)
- [Resilient Load Restoration in Microgrids Considering Mobile Energy Storage Fleets: A Deep Reinforcement Learning Approach](https://mlsanity.com/paper/1911.02206/) (2019)

## Abstract (arXiv)

Driven by recent advances in sensing and computing, deep reinforcement learning (DRL) technologies have shown great potential for addressing distribution system restoration (DSR) under uncertainty. However, their data-intensive nature and reliance on the Markov Decision Process (MDP) assumption limit their ability to handle scenarios that require long-term temporal dependencies or few-shot and zero-shot decision making. Emerging Decision Transformers (DTs), which leverage causal transformers for sequence modeling in DRL tasks, offer a promising alternative. However, their reliance on return-to-go (RTG) cloning and limited generalization capacity restricts their effectiveness in dynamic power system environments. To address these challenges, we introduce an innovative Dual-Head Physics-informed Graph Decision Transformer (DH-PGDT) that integrates physical modeling, structural reasoning, and subgoal-based guidance to enable scalable and robust DSR even in zero-shot or few-shot scenarios. DH-PGDT features a dual-head physics-informed causal transformer architecture comprising Guidance Head, which generates subgoal representations, and Action Head, which uses these subgoals to generate actions independently of RTG. It also incorporates an operational constraint-aware graph reasoning module that encodes power system topology and operational constraints to generate a confidence-weighted action vector for refining DT trajectories. This design effectively improves generalization and enables robust adaptation to unseen scenarios. While this work focuses on DSR, the underlying computing model of the proposed PGDT is broadly applicable to sequential decision making across various power system operations and other complex engineering domains.