⚡ electrical engineering

Tight Formulations for Unit Commitment with Different Levels of Details -- Part I: Models and Theoretical Insights

This paper addresses the computational challenges of the unit commitment problem by defining models with varying levels of detail, deriving convex hull-based formulations for each, and providing new theoretical proofs on the tightness of ramping and start-up/shut-down constraints to guide the selection of optimal formulations for different generator types.

Original authors: Maaike B. Elgersma, Karen I. Aardal, Mathijs M. de Weerdt, Germán Morales-España

Published 2026-07-09

📖 5 min read🧠 Deep dive

CC BY 4.0

Original authors: Maaike B. Elgersma, Karen I. Aardal, Mathijs M. de Weerdt, Germán Morales-España

Original paper licensed under CC BY 4.0 (http://creativecommons.org/licenses/by/4.0/). ✨ This is an AI-generated explanation of the paper below. It is not written or endorsed by the authors. For technical accuracy, refer to the original paper. Read full disclaimer

Imagine you are the conductor of a massive orchestra made up of hundreds of different instruments (power generators). Your job is to decide exactly when each instrument should start playing, how loudly it should play, and when it should stop, all while ensuring the music (electricity) perfectly matches the audience's demand every single second.

This is the Unit Commitment (UC) problem. It's a math puzzle that power companies solve to keep the lights on.

The Big Problem: Too Many Choices

The puzzle is incredibly hard because every instrument has rules. Some take a long time to warm up (start-up costs). Some can't play too quietly or too loudly (minimum/maximum output). Some can't switch from a loud note to a quiet one instantly (ramping limits). And some, once they start playing, must keep playing for at least 30 minutes before they can stop (minimum up time).

To solve this, computers use a method called Mixed Integer Linear Programming (MILP). Think of this as a giant spreadsheet where the computer tries every possible combination of "on" and "off" switches to find the cheapest way to run the orchestra.

The Catch: In a large power grid, there are so many generators and so many time slots that the number of combinations is astronomical. It's like trying to find a specific grain of sand in a desert by checking every grain one by one. The computer gets stuck, takes too long, or gives up.

The Paper's Solution: "Tightening" the Net

The authors of this paper (Part I of a two-part series) aren't trying to make the computer faster by buying better hardware. Instead, they are trying to make the mathematical map of the problem more accurate.

They use a concept called the "Convex Hull."

The Analogy: Imagine you have a bag of marbles scattered on a table. The "feasible region" is the space where the marbles can legally sit.
A Loose Map: If you draw a giant, sloppy circle around all the marbles, you include a lot of empty space where no marbles actually exist. When the computer tries to solve the problem, it wastes time exploring these empty spaces.
A Tight Map (The Convex Hull): If you stretch a rubber band tightly around the marbles, you create the smallest possible shape that still contains every single marble. This is the "Convex Hull." It leaves out all the empty space.

The paper argues that if you give the computer a "tight" map (a formulation based on the convex hull), it can find the solution much faster because it doesn't have to waste time looking in empty areas.

The Levels of Detail

The authors realized that not every generator needs the same level of detail. A simple solar panel doesn't have the same rules as a massive coal plant. So, they created a "menu" of models with different levels of detail:

Level 1 (The Simple View - 1bin): This is like looking at the orchestra from far away. You only see if an instrument is "On" or "Off." You don't worry about the exact moment it starts or stops. This is fast but less accurate.
Level 2 (The Detailed View - 3bin): This is like sitting right next to the instruments. You track not just "On/Off," but also the specific moments of "Starting Up" and "Shutting Down." This is more accurate but creates a bigger, more complex map.

What They Actually Did

The paper focuses on Part I: The Theory. They didn't run massive simulations here; instead, they did the heavy mathematical lifting to prove which "maps" are the tightest for specific situations.

They proved the "Perfect Fit": For certain types of generators (like those with simple ramping rules), they showed that a specific set of mathematical rules creates the perfect "rubber band" (convex hull). This means the computer can solve these specific problems instantly without guessing.
They compared the "Loose" vs. "Tight" versions: They showed that while some older, simpler formulas are smaller (fewer lines of math), they leave gaps (empty space) that slow down the computer. Their new "tight" formulas fill those gaps.
They handled the "Start-Up" and "Shut-Down" costs: They proved that you can model the cost of turning a generator on or off in a way that is mathematically perfect, ensuring the computer doesn't make mistakes about how much it costs to switch gears.

The Takeaway

This paper is the instruction manual for building the best possible map.

The authors say: "Here are the different types of generators. Here is the most mathematically perfect (tightest) way to describe each one without making the map too huge. If you use these specific formulas, your computer will have the best chance of solving the puzzle quickly."

They leave the actual testing of how fast these models run in real-world scenarios for Part II of their research. For now, they have provided the theoretical proof that these specific "tight" formulations are the gold standard for accuracy and efficiency.

Technical Summary: Tight Formulations for Unit Commitment with Different Levels of Details – Part I

1. Problem Statement

The Unit Commitment (UC) problem is a critical Mixed-Integer Linear Programming (MILP) challenge in power system management, aiming to determine the optimal operational schedule for generators (thermal, nuclear, renewable, etc.) over a short-term horizon (typically 24–48 hours). The problem involves minimizing system-wide operational costs while satisfying demand and respecting unit-specific operational constraints, such as ramping limits, minimum up/down times, and start-up/shut-down capabilities.

Despite its importance, the UC problem faces severe computational limitations in large-scale applications, particularly when integrated into investment models or stochastic/robust frameworks. The presence of binary variables (representing unit commitment status) causes the problem complexity to grow exponentially. Current approaches to mitigate this often involve:

Reducing Model Fidelity: Clustering similar units or using representative days, which can introduce errors by overestimating flexibility or underestimating costs.
Relaxing Integrality: Solving the Linear Programming (LP) relaxation instead of the MILP, which can lead to large optimality gaps and infeasible solutions in practice.

There is a lack of a clear, a priori overview regarding which formulation offers the best balance between tightness (the gap between the MILP feasible region and its LP relaxation) and model size (number of variables and constraints) for different generator types and levels of detail.

2. Methodology

This paper (Part I) focuses on the theoretical development and analysis of UC formulations. The authors adopt the following methodological approach:

Classification by Detail Level: The authors categorize UC models based on the number of binary variables per time period:
- 1bin Models: Contain only the commitment variable ( $u_{gt}$ ).
- 2bin/3bin Models: Include additional variables for start-up ( $v_{gt}$ ) and shut-down ( $w_{gt}$ ).
Convex Hull Analysis: The primary goal is to identify formulations based on the convex hull of the feasible region. The tightest formulation corresponds to the convex hull. However, since finding the full convex hull often requires exponentially many constraints, the authors focus on subsets of facets that:
1. Describe the convex hull (or a tight approximation) for specific subproblems (e.g., ramping, start-up capabilities).
2. Maintain a model size similar to the smallest valid formulation.
Theoretical Proofs: The paper provides new proofs and theoretical insights regarding the tightness of well-known formulations. It utilizes techniques such as Fourier-Motzkin elimination and disjunctive programming (Balas' method) to derive and verify convex hull descriptions for single-unit problems over two time periods.
Scope Limitation: All theoretical insights regarding tightness are derived for a single-unit operational problem, excluding the system-wide energy balance constraint. The objective function is generally considered irrelevant for convex hull determination unless specified.

3. Key Contributions

The paper makes four specific contributions, with contributions 1 and 2 detailed in Part I:

Structured Overview of UC Models: The authors present a structured taxonomy of UC models with varying levels of detail (from basic generation limits to complex start-up trajectories). For each level, they propose specific formulations that meet a trade-off between tightness and size.
Theoretical Insights and Proofs:
- Ramping: New proofs are provided for the tightness of ramping constraints. The paper establishes that specific formulations (Model I-E-T) constitute the two-period convex hull for ramping up and down.
- Start-up/Shut-down Costs: The paper analyzes two common formulations for modeling start-up and shut-down costs (Model II and Model II-S). It proves that while Model II-S has fewer constraints, both formulations define the same solution space in the direction of the objective function, making them equally tight for optimization purposes.
- Capabilities: The paper presents the 1bin and 3bin convex hull formulations for start-up and shut-down capabilities (maximum output during start-up/shut-down), proving that specific combinations of constraints (e.g., Model I-E2-T and Model II-E2) define the convex hull for two-period problems.
Comparison of Formulations: The paper directly compares different formulations for the same problem (e.g., 1bin vs. 3bin, or different facet sets) to highlight theoretical differences in tightness and size.
Foundation for Part II: This work lays the theoretical groundwork for Part II, which numerically compares these models in large-scale investment and operational contexts.

4. Results and Formulations

The paper defines several specific models, ranging from basic to extended:

Level I (1bin Models):
- Model I: Basic generation limits. Proven to be the tightest possible formulation for its scope.
- Model I-E: Adds ramping constraints.
- Model I-E-T: A tighter version of Model I-E including additional facets for ramping up/down. Proven to be the two-period convex hull.
- Model I-E2: Includes ramping and start-up/shut-down capabilities.
- Model I-E2-T: The tightest version, proven to be the two-period convex hull for ramping and capabilities.
Level II (3bin Models):
- Model II: Includes start-up/shut-down variables and costs. Proven to be the tightest formulation for this scope.
- Model II-S: An alternative formulation using continuous cost variables instead of binary start-up variables. Theorem III.4 proves it is equally tight as Model II in the direction of the objective function, though it defines an unbounded convex hull without the upper bound constraint $u_t \leq 1$ .
- Model II-E: Adds minimum up/down time constraints. Proven to be the tightest formulation for this scope (linear size).
- Model II-E2: The most detailed 3bin model, incorporating ramping, start-up/shut-down capabilities, and minimum up/down times. Theorem III.5 proves it defines the convex hull for two periods.
- Model II-E3: Extends Model II-E2 to include start-up and shut-down trajectories (multi-period ramps to reach minimum output), proven to define the two-period convex hull.

Key Theoretical Findings:

For single-unit problems, the LP relaxations of Models I, II, and II-E describe the convex hull of their respective feasible spaces.
For two-period problems, Models I-E-T, I-E2-T, and II-E2 describe the convex hull.
Relaxing the integrality of start-up/shut-down variables (from binary to continuous) does not change the feasible set of the MIP, but modern solvers may benefit from keeping them as integers for branching strategies.
3bin models can be rewritten as 2bin models (by eliminating the shut-down variable) without losing tightness, though the number of constraints may vary.

5. Significance and Claims

The paper claims that its primary significance lies in providing a structured, theoretically grounded framework for selecting UC formulations. By defining models with different levels of detail and proving their tightness properties, the authors enable modelers to:

Select the most appropriate level of detail for specific generators or problem scales without sacrificing computational efficiency unnecessarily.
Understand the trade-offs between model size and tightness a priori, rather than relying solely on a posteriori computational testing which can be solver-dependent and non-generalizable.
Incorporate high-fidelity operational details into large-scale investment models (as demonstrated in Part II) while staying within computational limits.

The authors emphasize that while finding the full convex hull is often intractable due to exponential growth, their proposed formulations offer the "best" balance by utilizing subsets of facets that are similar in size to the smallest valid formulations but significantly tighter than standard relaxations. This work aims to guide the selection of UC models that are computationally efficient yet accurate enough to prevent the overestimation of system flexibility and underestimation of costs often found in simplified investment models.