⚡ electrical engineering

PGLib-CO2: A Power Grid Library for Real-Time Computation and Optimization of Carbon Emissions

This paper introduces PGLib-CO2, an open-source library that extends standard power grid test cases with generator-level carbon emission data and proposes advanced differentiable and multiparametric programming methods to enable accurate, real-time computation of carbon emission metrics for sustainable power system optimization.

Original authors: Young-ho Cho, Min-Seung Ko, Hao Zhu

Published 2026-03-20

📖 5 min read🧠 Deep dive

CC BY 4.0

Original authors: Young-ho Cho, Min-Seung Ko, Hao Zhu

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 the electrical grid as a massive, complex city of roads. In this city, power plants are the factories producing goods (electricity), and our homes and businesses are the shops buying them. For decades, the city planners (grid operators) have only cared about two things: how cheaply they can get the goods to the shops and how reliably the roads stay open.

They didn't really care about the smoke coming out of the factories.

This paper introduces a new toolkit called PGLib-CO2 to fix that. It's like giving the city planners a new set of maps that not only show traffic jams and toll prices but also color-code every factory based on how much pollution it creates.

Here is a breakdown of what the paper does, using simple analogies:

1. The Problem: The "Blind" Map

Until now, the standard maps used by researchers (called PGLib-OPF) were great for calculating costs, but they were "blind" to carbon.

The Old Way: If you asked the map, "How much does it cost to light up this house?" it would give you a dollar amount. But if you asked, "How much carbon did that cost?" the map would just say, "I don't know, we didn't track that."
The Result: Researchers had to guess the pollution levels or make up their own numbers, which made it hard to compare different ideas fairly.

2. The Solution: PGLib-CO2 (The "Pollution-Tagged" Map)

The authors created PGLib-CO2, which is an upgrade to the standard map.

The Upgrade: They went through every power plant in the test cases and attached a "pollution tag" to it.
- A coal plant gets a big, heavy red tag (high pollution).
- A solar farm gets a green tag (zero pollution).
- A natural gas plant gets a yellow tag (medium pollution).
Why it matters: Now, when the grid runs, it can calculate exactly how much carbon is being emitted for every single decision it makes. It turns the grid from a "cost-only" machine into a "cost-and-cleanliness" machine.

3. The Hard Part: Calculating the "Carbon Price" of a Single Lightbulb

The paper tackles a very tricky math problem: LMCE (Locational Marginal Carbon Emissions).

Think of it this way: If you turn on a lightbulb in New York, which power plant actually turns on to power it?

If the grid is calm, maybe a clean wind farm turns on. (Low carbon cost).
If the grid is congested (traffic jam on the wires), maybe a dirty coal plant has to turn on to fill the gap. (High carbon cost).

The "Carbon Price" changes depending on where you are and when you turn on the light. Calculating this is like trying to predict exactly which car will move first in a traffic jam when one new car enters the road. It's incredibly complex.

The paper offers two new ways to solve this puzzle:

Method A: The "Super-Brain" (Differentiable Programming)

The Analogy: Imagine a super-smart student who has memorized the rules of the game. Instead of guessing, they use a powerful calculator (a technique called "differentiable programming") to instantly see exactly how the pollution changes if you nudge the system by a tiny bit.
The Benefit: It is perfectly accurate. It works even if the rules of the game are complicated (non-linear costs).
The Downside: It's like asking that student to solve a math problem from scratch every single time. It's accurate, but it takes a few seconds to think. In a real-time grid, a few seconds is too long.

Method B: The "Cheat Sheet" (Multiparametric Programming)

The Analogy: This is the real-time hero. Imagine the student spent all night studying and created a giant Cheat Sheet (a lookup table).
- The sheet divides the world into different "zones" (Critical Regions).
- Zone A: "If the load is here, the answer is X."
- Zone B: "If the load is there, the answer is Y."
How it works: When a real-time event happens, the computer doesn't do any math. It just looks at the current situation, finds the zone on the cheat sheet, and reads the answer.
The Benefit: It is instantly fast (sub-millisecond). It's like looking up a word in a dictionary instead of writing a dictionary definition from scratch.
The Result: This allows the grid to track carbon emissions in real-time, faster than a human can blink.

4. The Results: Speed and Accuracy

The authors tested these methods on standard electrical grids (the "14-bus" and "118-bus" systems, which are like test cities).

Accuracy: The "Super-Brain" method was proven to be mathematically perfect, even better than existing tools.
Speed: The "Cheat Sheet" method was thousands of times faster than the old ways.
- Old way: Takes seconds or minutes to calculate.
- New way: Takes less than a millisecond.

Why Should You Care?

This isn't just about math; it's about the future of energy.

Better Decisions: Grid operators can now see exactly how dirty their decisions are and choose cleaner options without breaking the bank.
Real-Time Tracking: Just like your phone tracks your location, this tool allows the grid to track its "carbon footprint" in real-time.
Fair Markets: It helps design electricity markets where you pay for the actual pollution you cause, encouraging everyone to use clean energy when it's available.

In short: The authors built a new, pollution-aware map for the power grid and invented a super-fast way to read it. This helps us transition from a grid that just cares about money to a grid that cares about the planet, all while keeping the lights on instantly.

1. Problem Statement

The transition toward sustainable power systems requires the explicit integration of carbon emissions into grid modeling and optimization. However, significant gaps exist in current research infrastructure:

Lack of Standardized Data: Existing open-source power system test cases (e.g., PGLib-OPF) lack generator-level carbon profiling. Researchers often rely on ad-hoc or unverified emission assumptions, hindering the benchmarking of carbon-aware strategies.
Computational Limitations in Carbon Metrics: The Locational Marginal Carbon Emission (LMCE) is a critical metric for evaluating the emission impact of load changes at specific nodes.
- Numerical Perturbation: Traditional methods re-solve Optimal Power Flow (OPF) with small load increments. This is computationally expensive, unstable near congestion boundaries, and sensitive to step-size selection.
- Analytical Derivation: Closed-form sensitivity analysis is often limited to specific cost functions (e.g., linear) and fails for general convex costs.
- Real-Time Latency: Existing rigorous methods are too slow for real-time carbon tracking and market signaling.

2. Methodology

The authors propose a two-pronged approach: a standardized data library and advanced computational algorithms for LMCE.

A. PGLib-CO2 Library (Data Enrichment)

The authors extend the widely adopted PGLib-OPF library to include generator-level carbon attributes.

Fuel-Type Expansion: They expanded the limited fuel categories in PGLib-OPF to include a comprehensive list (e.g., Anthracite coal, Bituminous coal, Natural Gas, CCGT, ICE, Nuclear, Renewables).
Emission Factors: Each generator is assigned specific CO2 and CO2-equivalent (CO2e) emission intensity factors (t/MWh) based on established research.
Integration: The library is packaged for both Python (pandapower) and Julia (PowerModels.jl), allowing users to seamlessly embed emission data into OPF formulations.
Carbon-Aware Optimization: The library supports adding a carbon cost term ( $c(P_G) = (\tau \cdot e^G)^\top P_G$ ) to the standard OPF objective, where $\tau$ is a carbon tax rate.

B. Algorithm 1: Differentiable LMCE Computation (Exact Sensitivity)

To compute exact LMCE for general convex cost functions without numerical perturbation:

Approach: Treats the OPF problem as a differentiable layer within a computational graph (using Deep Learning frameworks like PyTorch).
Mechanism: Utilizes Automatic Differentiation (AD) and the Implicit Function Theorem (IFT).
- Forward Pass: Solves the OPF to find the optimal dispatch and identifies the set of active constraints.
- Backward Pass: Constructs a linear system based on the KKT conditions of the active set to compute the exact Jacobian ( $J = \partial P_G / \partial P_D$ ).
Result: Provides rigorous, exact sensitivity information for any convex cost function, eliminating the step-size trade-offs of numerical methods.

C. Algorithm 2: Real-Time LMCE via Multiparametric Programming (MPP)

To achieve sub-millisecond computation for real-time applications:

Approach: Uses Multiparametric Programming (MPP) to shift the optimization burden to an offline phase.
Offline Phase:
- Reformulates the DC-OPF as a multiparametric linear program where load ( $P_D$ ) is the parameter.
- Partitions the load space into Critical Regions (CR). Within each region, the set of active constraints remains constant.
- Derives an explicit affine dispatch function ( $P_G^* = J_k P_D + g_k$ ) for each region.
- Pre-computes the sensitivity matrix $J_k$ and region boundaries.
Online Phase:
- Identifies which critical region the current load falls into (a geometric point-location problem).
- Retrieves the pre-computed $J_k$ and calculates LMCE via simple matrix-vector multiplication: $e^{LMCE} = (e^G)^\top J_k$ .
Benefit: Reduces online complexity from $O(N^3)$ (solving LP) to $O(N)$ (checking inequalities and matrix multiplication).

3. Key Contributions

Standardized Carbon-Enriched Dataset: PGLib-CO2 provides the first systematic, open-source extension of PGLib-OPF with generator-level CO2/CO2e intensity factors, enabling reproducible benchmarking.
Differentiable LMCE Computation: A novel algorithm using differentiable optimization to compute exact marginal carbon sensitivities for general convex cost functions, serving as a "ground truth" for validation.
Real-Time MPP-Based Identification: A framework that enables sub-millisecond retrieval of exact LMCE signals by leveraging offline critical region mapping, making real-time carbon tracking feasible.

4. Numerical Results

The authors validated their methods on IEEE 14-bus and 118-bus systems.

Accuracy Validation (Differentiable Method):
- Compared against ElectricityEmissions.jl (a standard tool).
- Results: Mean deviation was negligible (<0.021). The differentiable method correctly handled boundary operating points where numerical perturbation methods often fail or produce outliers.
- Generalization: Successfully computed LMCE for quadratic cost functions, a scenario unsupported by existing libraries.
Efficiency Validation (MPP Method):
- IEEE 14-bus:
  - Differentiable: Mean time ~37.7 ms; Max time ~81.2 ms.
  - MPP: Mean time 0.17 ms; Max time 0.21 ms.
  - Accuracy: Estimation errors remained below 0.2%.
- IEEE 118-bus:
  - Differentiable: Mean time 7.23 seconds; Max time 25.12 seconds (processing 1,000 scenarios took ~130 mins).
  - MPP: Mean time 0.14 ms; Max time 0.16 ms.
  - Observation: The MPP online time was even faster for the 118-bus system than the 14-bus system because the number of critical regions decreased (15 $\to$ 10), proving that online speed depends on the complexity of the region search tree, not the system size.

5. Significance and Impact

Bridging Data and Computation: PGLib-CO2 fills a critical gap by providing the necessary data infrastructure (fuel types/intensity) to support carbon-aware research.
Enabling Real-Time Decarbonization: The MPP-based approach solves the latency problem, allowing market participants and grid operators to access instantaneous carbon signals. This is crucial for:
- Real-time carbon tracking and reporting.
- Designing carbon-aware electricity markets.
- Enabling consumers to infer carbon intensity based on public Locational Marginal Price (LMP) signals without needing access to internal dispatch data.
Scalability: The method scales efficiently to larger systems, making it viable for future high-resolution, real-time grid management.

In conclusion, PGLib-CO2 provides a foundational toolkit for the next generation of sustainable power system operations, combining high-fidelity data with computationally efficient algorithms to make carbon-aware optimization a practical reality.