EridanusQ/UC_Trajectory_Dataset

Unit Commitment Trajectory Dataset (UCTD)

1. Overview

This dataset is generated using a customized version of the UnitCommitment.jl framework, specifically designed for Machine Learning for Optimization (ML4Opt) research. It provides Unit Commitment (SCUC) optimization problems ranging from small IEEE test systems to large-scale national grids.

Core Innovation: Unlike standard datasets, UCTD incorporates Power Trajectories for generator startup and shutdown. This provides a high-fidelity physical representation of power system operations, making it a challenging benchmark for modern optimization solvers and ML models.

2. Case Statistics

The dataset contains 464 .mps files across three grid models:

• Case14: Basic test system (14-bus, 67 days).
• Case30: Medium-scale system (30-bus, 45 days).
• Case2383wp (Challenge Set): Large-scale Polish national grid (2383-bus), used for testing scalability.

3. Model Variants

For each day, 4 modeling variants are provided:

1. hourly_noline: 1-hour resolution, unit constraints only.
2. hourly_withline: 1-hour resolution, including full network constraints (SCUC).
3. subhourly_noline: 15-minute resolution, unit constraints only.
4. subhourly_withline: 15-minute resolution, including full network constraints (SCUC).

4. File Naming Convention

Format: {case}_{date}_{granularity}_{variant}.mps Example: case30_2017-01-01_s_withline.mps (Case30, Sub-hourly, with network constraints).

5. Key Features

• High-Fidelity Physics: Includes power output trajectories during generator startup/shutdown phases.
• Multiple Resolutions: Covers both traditional 1-hour scheduling and modern 15-minute sub-hourly scheduling.
• Standardized Format: Uses the industry-standard .mps format, compatible with Gurobi, CPLEX, HiGHS, and more.

6. Use Cases

• Supervised Learning: Predict optimal commitment status or production levels.
• End-to-End Optimization: Train Graph Neural Networks (GNN) to map problem instances directly to solutions.
• Solver Benchmarking: Test the performance and scalability of modern MIP solvers on large-scale power grid problems.

7. Citation

If you use this dataset in your research, please cite the original UnitCommitment.jl paper and acknowledge the source of this trajectory-enhanced version.