⚡ electrical engineering

Geospatial sensitivity of transmission-constrained ACOPF to generator retirement

This paper presents an HPC-enabled framework that utilizes k-nearest-neighbors models to efficiently map the geospatial sensitivity of voltage magnitude and angle in transmission-constrained ACOPF to over 8,000 generator retirement scenarios on a large-scale synthetic grid, addressing the computational challenges of resource adequacy planning.

Original authors: Willa Gutowski, Ulyana Shyrokaya, Nicholson Koukpaizan, Joshua Hambrick, Slaven Peles, Eve Tsybina

Published 2026-06-16

📖 4 min read☕ Coffee break read

CC BY 4.0

Original authors: Willa Gutowski, Ulyana Shyrokaya, Nicholson Koukpaizan, Joshua Hambrick, Slaven Peles, Eve Tsybina

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 US power grid as a massive, complex web of roads connecting thousands of towns (cities) to power plants (factories). Right now, this web is under stress for two main reasons:

More Traffic: New factories (like giant data centers) are popping up, demanding huge amounts of electricity.
Old Cars: The old power plants are wearing out and need to be retired, just like an old car that can't keep up with highway speeds anymore.

The big question for grid managers is: Which old power plants can we safely turn off, and which ones are too important to lose? If we turn off the wrong one, the "traffic" (electricity) might get stuck, causing blackouts or dangerous voltage spikes in nearby towns.

The Problem: Too Many Possibilities, Not Enough Time

In the past, figuring this out was like trying to test every single car in a fleet of 8,000 vehicles by driving them one by one on a test track. It would take forever. The math required to simulate the grid (called ACOPF) is incredibly heavy, and running thousands of "what-if" scenarios used to be practically impossible for a whole country.

The Solution: A Super-Fast "Traffic Simulator"

The researchers at Oak Ridge National Laboratory built a super-fast computer framework (using High-Performance Computing) that acts like a high-speed traffic simulator. Instead of testing one car at a time, they simulated turning off 8,107 different power plants across a massive, 70,000-node model of the Eastern US grid.

They did this so quickly that they could rank every single generator in under 30 minutes.

How They Measured the "Shock"

When you turn off a power plant, it's like removing a pillar from a building. The researchers wanted to see how much the building wobbles. They measured two things:

Voltage Magnitude: Think of this as the "pressure" in the electrical pipes. If the pressure drops too low, lights flicker; if it spikes, equipment breaks.
Voltage Angle: Think of this as the "timing" or rhythm of the electricity. If the rhythm gets out of sync, the whole system can crash.

They looked at how far this "wobble" traveled:

Tier 1 (Immediate Neighbors): The towns right next to the power plant.
Tier 10 (Far Neighbors): Towns far away across the grid.

Key Findings (The "Aha!" Moments)

1. The "Local Neighborhood" Effect
The study found that turning off a power plant mostly hurts the immediate neighborhood.

Analogy: Imagine a loudspeaker in a room. If you turn it off, the people sitting right next to it notice the silence immediately. People across the room barely notice.
The Data: The impact on voltage was strongest in the first few "neighborhoods" (buses) around the plant and faded away quickly. By the time you get 6 neighborhoods away, the effect is usually gone.

2. Not All Plants Are Created Equal
Most power plants are like small, local generators. Turning one off causes a tiny ripple (less than 2% change). However, a small group of "Super Plants" (about 500 to 1,000 of them) are critical.

The Danger: Turning off these specific plants caused voltage swings of up to 13%. That's like a sudden, violent earthquake in the grid. These are the ones that need to stay online until new power is built.

3. The "Voltage Class" Rule
The researchers realized that not all roads are the same.

High Voltage Roads (The Interstate): These carry power over long distances. They are less sensitive to a single plant going offline because they have many other sources feeding them.
Low Voltage Roads (The Local Streets): These are closer to homes and businesses. They rely heavily on nearby plants for stability. If a plant near a low-voltage area goes offline, the local "pressure" drops significantly.

4. The "Upward" Regulator
Most power plants in this grid act like a "pressure booster," pushing voltage up. When you retire one, the pressure in the neighborhood drops. This means the grid loses a vital service (voltage support) that other devices can't always replace immediately.

The Bottom Line

This paper doesn't just say "we need more power." It provides a map of danger. It tells grid operators exactly which power plants are the "keystone arches" of the bridge. If you remove those specific arches, the bridge collapses. If you remove the others, the bridge might wobble a bit but will hold.

By using supercomputers to run these thousands of scenarios quickly, the US can now make smarter decisions about which old plants to keep running and which can safely be retired, ensuring the lights stay on while the grid transitions to new energy sources.

Technical Summary: Geospatial Sensitivity of Transmission-Constrained ACOPF to Generator Retirement

Problem Statement
The US electric grid faces a critical resource adequacy challenge driven by the simultaneous addition of new loads at unprecedented scales (particularly from the digital economy and datacenters) and the retirement of aging generating assets, many of which were built in the 1970s and 1980s. In transmission-constrained grids, determining which units can be safely retired versus those requiring lifetime extensions is computationally prohibitive using traditional methods. Historically, a thorough analysis using Transmission-constrained Alternating Current Optimal Power Flow (ACOPF) requires evaluating hundreds or thousands of scenarios, a process that is too time-consuming for nationwide planning. While previous studies have examined retirement consequences for system reliability and cost, they have been limited to smaller grids or aggregate metrics, failing to address the local and regional nuances of voltage stability in large-scale, transmission-constrained systems.

Methodology
The authors present a High-Performance Computing (HPC) enabled framework designed to rapidly assess the geospatial effects of generator retirement on steady-state voltage magnitude and angle. The study utilizes a 70,000-bus synthetic Eastern Interconnection model (ACTIVSg70k) containing 8,107 candidate generators.

The core methodology involves:

Scenario Generation: The framework simulates the retirement of each of the 8,107 generators individually. Three distinct ACOPF scenario sets are processed in parallel:
- Main: Enforces MVA line limits with no load shedding.
- Shed Load: Enforces MVA limits but allows load shedding.
- Unlimited: Does not enforce MVA limits and prohibits load shedding.
Neighbor Analysis: The impact of retirement is evaluated across 1 to 10 "neighbor bus tiers" (graph distance). Two modeling approaches are used:
- A simple graph model where all branches are treated equally.
- A voltage-class-adjusted model that subdivides neighbors into three categories based on voltage levels: distribution/sub-transmission (≤34 kV), lower transmission (69–230 kV), and high-voltage transmission (345–765 kV).
Metrics: For convergent cases, the framework calculates voltage magnitude deviation ( $\Delta V_m$ ) and voltage angle deviation ( $\Delta V_\theta$ ). Generators are flagged as high-impact if their retirement causes significant deviations (up to three neighbor tiers, per PJM standards) or if the ACOPF solution fails to converge.
Computational Implementation: The simulations are executed using the IPOPT optimization package with the MA57 linear solver on the Oak Ridge Leadership Computing Facility. The process utilizes parallel batching (427 CPU nodes) to solve 24,321 cases (8,107 retirements + 3 baselines) in under 30 minutes total, with data post-processing and ranking taking approximately 30 seconds via DuckDB.

Key Results

Convergence and Impact: In the "Main" scenario (most constrained), 86.58% of cases converged. Non-convergent cases typically indicate that retiring a specific unit renders the ACOPF problem infeasible, flagging them as critical.
Voltage Magnitude: The average impact of retirement on voltage magnitude ranges from 0.129% to 0.469% in the first neighbor tier. This impact attenuates as distance increases, with significant changes dissipating into background variation by tier 6.
Voltage Angle: Angle deviations also decrease with neighbor tier but show different scenario ordering; the "Unlimited" case produced the largest mean absolute angle deviations, while the "Main" case produced the smallest.
Generator Ranking: The study identified a small subset of high-impact generators (approximately 500–1,000 units) capable of causing significant grid disturbances. In the "Main" scenario, the top-ranked units caused voltage magnitude changes of up to 7.82%. Notably, while most generators regulate voltage "up" (causing a drop upon retirement), about 30% of retirements resulted in higher voltage profiles on neighboring buses.
Geospatial and Voltage Class Findings: High-impact outliers are predominantly found in lower voltage class grids (69–230 kV). Distribution buses (≤34 kV) show less sensitivity at close tiers due to their distance from generators but become more sensitive at larger neighbor distances. High-voltage buses (345–765 kV) exhibit smaller, more stable distributions of voltage deviations.
Geographic Distribution: The geospatial mapping of impactful generators highlights specific concentrations in Texas, Alabama, Florida, and South Carolina within the synthetic model.

Significance and Claims
The paper claims that this HPC-enabled framework successfully addresses the computational bottleneck that has historically prevented detailed, nationwide generator retirement analysis. By reducing the assessment time for over 8,000 generators to under 30 minutes, the authors demonstrate that comprehensive generation portfolio assessments can be conducted with a speed and fidelity comparable to existing industry-grade interconnection studies.

The study provides a systematic way to detect not only units whose retirement causes immediate ACOPF infeasibility but also those whose retirement causes significant local voltage deviations without causing total system failure. This allows planners to prioritize which units require lifetime extensions or which areas need new generation before retirement occurs. The authors note that while the current work focuses on single-unit retirements (N-1), future work will address non-convergent cases more systematically and evaluate multiple unit retirement scenarios (N-k contingencies).