quantum-rl-scuc-qsample

name	quantum-rl-scuc-qsample
description	Hybrid quantum-classical reinforcement learning for Security-Constrained Unit Commitment (SCUC). Uses Bernoulli hybrid soft actor-critic (HSAC) with quantum-sampled feature augmentation for economic dispatch in power systems. Activation: quantum RL SCUC, unit commitment optimization, quantum-sampled features, hybrid soft actor-critic, economic dispatch quantum, power grid optimization quantum
metadata	{"arxiv_id":"2606.26345","published":"2026-06-24","authors":"George Dimas, Amin Masoumi, Mert Korkali","tags":["quantum","reinforcement-learning","economics","power-systems","unit-commitment"]}

Quantum-RL SCUC with Quantum-Sampled Features

Description

Hybrid quantum-classical RL framework for Security-Constrained Unit Commitment (SCUC). Couples a Bernoulli hybrid soft actor-critic (HSAC) policy for binary commitment decisions with quantum-sampled auxiliary feature augmentation, then recovers dispatch/security via native SCUC MILP solver. Tested on 14-, 57-, 118-bus power systems.

Activation Keywords

• quantum RL SCUC
• unit commitment quantum
• quantum-sampled features
• hybrid soft actor-critic
• economic dispatch quantum
• power system quantum optimization
• SCUC reinforcement learning
• 安全约束机组组合量子
• 量子采样特征

Core Concepts

SCUC-to-RL Interface

SCUC couples binary commitment, economic dispatch, reserves, and network security over multiperiod horizon. The RL agent proposes hourly commitments; a limited enforcement cap fixes subset of binaries; MILP recovers remaining variables.

Three-Layer Architecture

1. Bernoulli HSAC Policy: Proposes hourly unit commitments via exploratory Bernoulli actor
2. Quantum-Sampled Auxiliary Channel: Augments state representation with quantum-sampled features
3. Native SCUC MILP Recovery: Solves dispatch/security after partial commitment enforcement

Key Findings

• 14-bus: Stable, low-cost recovery
• 57-bus: Very low screen-rejection rate → learned feasibility generalization
• 118-bus: Coverage bottleneck when enforcement cap spans < complete commitment period; accepted episodes show tightly clustered runtimes → policy captures repeatable recovery pattern

Feasibility-Optimality Trade-off

The dominant limitation: amount of useful commitment information reaching recovery model under exploratory Bernoulli actor + small enforcement cap governs scalability.

Methodology

Step 1: Formulate SCUC as MDP

• State: system load forecasts, unit availability, network conditions + quantum-sampled features
• Action: binary commitment decisions per unit per hour (Bernoulli distribution)
• Reward: negative of total cost (fuel + startup + shutdown) if feasible, large penalty if infeasible

Step 2: Quantum-Sampled Feature Augmentation

• Generate auxiliary features via quantum sampling process
• Concatenate with classical state features before policy network input
• Purpose: enhance policy expressiveness for capturing complex feasibility patterns

Step 3: HSAC Policy Training

• Use hybrid soft actor-critic with Bernoulli output layer for discrete commitments
• Maximize expected reward + entropy regularization for exploration
• Train on smaller systems (14-bus) first, transfer to larger (57-, 118-bus)

Step 4: MILP Recovery

• Fix subset of commitment decisions from RL policy (enforcement cap)
• Solve remaining SCUC variables via native MILP solver
• Screen for feasibility; reject and re-sample if infeasible

Usage Patterns

Pattern 1: Economic Dispatch Optimization

Use when optimizing generator commitment schedules with quantum-enhanced RL for improved feasibility generalization.

Pattern 2: Scalable Unit Commitment

Apply to multi-period unit commitment problems where classical solvers are expensive at realistic system sizes.

Pattern 3: Quantum Feature Augmentation

Use quantum-sampled features to augment classical RL states for any constrained combinatorial optimization problem.

Pitfalls

Enforcement Cap Size

• Enforcement cap must span at least one complete commitment period for scalability
• 118-bus case shows clear bottleneck when cap is too small
• Increase cap or use more informed (less exploratory) actor for larger systems

Bernoulli Actor Exploration

• Pure Bernoulli exploration may propose infeasible commitments frequently
• Consider informed initialization or curriculum learning for large systems
• Screen-rejection rate increases with system size under small cap

Quantum Sampling Overhead

• Quantum-sampled features add computational overhead
• Validate that augmentation improves policy quality vs. classical-only baseline
• Feature dimension should be tuned to avoid overfitting

Transfer Learning Limits

• Policy trained on 14-bus generalizes to 57-bus but not directly to 118-bus
• Retrain or fine-tune on intermediate system sizes for best results

References

• arXiv: 2606.26345 - "Feasibility-Aware Security-Constrained Unit Commitment via Hybrid Soft Actor-Critic with Quantum-Sampled Features"
• Related: distributionally-robust-control (DRDDPC for building energy management)
• Related: vqa-dynamic-portfolio-optimization (VQA for dynamic portfolio optimization)