⚡ electrical engineering

Building Power Grid Models from Open Data: A Complete Pipeline from OpenStreetMap to Optimal Power Flow

This paper presents a fully open-data pipeline that constructs realistic, OPF-solvable transmission grid models for all 48 contiguous U.S. states and six multi-state regions by integrating OpenStreetMap infrastructure data with EIA and Census statistics, successfully validating the approach through convergence in AC-OPF simulations and cost metrics consistent with real wholesale markets.

Original authors: Andrea Britto, Thiago Spina, Weiwei Yang, Spencer Fowers, Baosen Zhang, Chris White

Published 2026-05-07

📖 5 min read🧠 Deep dive

CC BY 4.0

Original authors: Andrea Britto, Thiago Spina, Weiwei Yang, Spencer Fowers, Baosen Zhang, Chris White

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, intersections, and traffic lights. To understand how this city handles traffic (electricity), engineers need a detailed map. However, in the real world, the most accurate, up-to-date maps of the US power grid are locked away in a vault. They are treated as "critical infrastructure secrets" to prevent hackers or terrorists from finding weak spots. This leaves researchers trying to study the grid without a map, forced to use tiny, toy-sized models that don't look anything like the real thing.

This paper presents a clever solution: building a realistic, giant-scale map of the US power grid entirely from public, open-source data.

Here is how the authors did it, explained through a simple construction analogy:

The Problem: The "Secret Blueprint"

Think of the real power grid as a high-security building. The blueprints (detailed data on wire thickness, voltage, and exact locations) are classified. Researchers can't see them. They can't buy them (it costs hundreds of thousands of dollars), and they can't share them even if they get them. Without these blueprints, you can't test how the building would handle a massive crowd or a storm.

The Solution: The "Five-Stage Construction Crew"

The authors built a pipeline (a step-by-step assembly line) that constructs a working model of the grid using only things anyone can see or find online.

Stage 1: The Satellite Scout (Data Extraction)
Instead of stealing blueprints, the team uses OpenStreetMap (OSM). Think of OSM as a giant, collaborative Google Maps where volunteers draw roads, buildings, and power lines. The team sends a digital "drone" to scan the US and download every power line, substation, and power plant they can find.

The Catch: The map is incomplete. Some lines are missing, and the map doesn't say how thick the wires are or how much power they can carry.

Stage 2: The Architect (Topology Reconstruction)
The raw map is messy. A single power line might be drawn as 20 tiny, disconnected segments. The team's software acts like a smart architect who stitches these fragments together.

Filling in the blanks: If a line has no voltage tag, the software looks at its neighbors. If all the neighbors are "345kV lines," it assumes this one is too.
Sorting the traffic: It figures out which lines connect to which substations and identifies special "bridges" (transformers) that change voltage levels.
Result: A clean, connected network map of the grid.

Stage 3: The Engineer (Parameter Estimation)
Now they have the shape of the grid, but not the physics. They need to guess how much electricity the wires can carry and how much they resist the flow.

The Rulebook: Since they can't measure the wires, they use standard engineering textbooks. If a line is 345kV, they look up the "average" wire used for that voltage in the US and assign those properties to the model.
The "Parallel" Trick: Real power corridors often have 2, 3, or even 4 wires running side-by-side, but the map only shows one. The team uses a "scaling factor" (a mathematical guess) to pretend there are more wires, ensuring the model can handle the right amount of traffic.

Stage 4: The Demographer (Demand Allocation)
A grid is useless without knowing where people are using power. The team takes national data on how much electricity a whole state uses and spreads it out like sprinkles on a cake.

The Proxy: They assume that where there are more people (based on Census data), there is more electricity usage. They assign the power demand to specific "intersections" (buses) on the map based on local population density.

Stage 5: The Traffic Cop (Optimal Power Flow)
Finally, they run a simulation. They ask the computer: "If we turn on all the power plants and turn on all the lights at 4 PM on a hot summer day, what happens?"

The Goal: The computer tries to find the cheapest way to power the whole state without overloading any lines or causing blackouts.
The Safety Net: Because the map is an estimate, the simulation might get stuck. The team built a "progressive relaxation" strategy. If the strict rules cause a crash, the computer gently loosens the rules (like allowing a slightly higher speed limit) until it finds a solution that works. This tells them exactly how "rough" the data is in different areas.

The Results: A Working City

The team tested this on all 48 contiguous US states and huge regional chunks (like the entire Eastern or Western power grids).

Success Rate: 88% of the states solved perfectly without loosening any rules. The rest solved with minor adjustments.
Realism: The cost to generate electricity in their model ($22/MWh on average) and the amount of power lost as heat (1%) matched real-world market data almost exactly.
The Big Picture: They successfully modeled the Eastern Interconnection (21,697 intersections) and the Western Interconnection (5,076 intersections).

Why This Matters

The authors didn't just build a model; they built a public library of grid models.

No Secrets Needed: Anyone can now study how the grid reacts to AI data centers, extreme weather, or renewable energy without needing classified data.
Reproducible: Because the code and data are open, other scientists can check their work or improve it.
The "What-If" Machine: Researchers can now simulate disasters (like a hurricane knocking out a line) or future scenarios (like doubling the number of electric cars) to see how the grid holds up.

In short, the authors took a locked vault of secrets and replaced it with a public, open-source construction kit that is accurate enough to be useful for serious research, but built entirely from things you can see from your car window or find on a free website.

Technical Summary: Building Power Grid Models from Open Data

Problem Statement

Access to realistic, high-fidelity transmission grid models is a critical bottleneck for power systems research. In the United States, detailed network data—including topology, line impedance, and operating parameters—is classified as Critical Energy Infrastructure Information (CEII) by the Federal Energy Regulatory Commission (FERC) and the North American Electric Reliability Corporation (NERC). Consequently, researchers cannot access real-world models for planning, reliability assessment, or machine learning training without navigating restrictive legal processes or paying prohibitive licensing fees ( $50,000–$ 500,000).

Existing alternatives are insufficient for modern research needs:

Standard Test Cases (e.g., IEEE 14/30/118-bus): These are pedagogical or algorithmic benchmarks that lack geographic realism, do not reflect large-scale renewable integration, and miss continental-scale structural characteristics.
Synthetic Grids (e.g., TAMU): While statistically realistic, these are generated to resemble real grids rather than being derived from them, lacking specific geographic correspondence and the ability to evolve with actual infrastructure changes.

This data inaccessibility hinders the study of critical challenges such as AI-driven demand growth, renewable integration, and resilience against extreme weather events.

Methodology

The authors present a five-stage pipeline that constructs complete, Optimal Power Flow (OPF)-solvable transmission network models entirely from publicly available data. The pipeline transforms raw OpenStreetMap (OSM) data into solver-ready models using the following stages:

1. Data Extraction

The pipeline queries a local Overpass API instance to extract power infrastructure (substations, lines, generators, converters) for specific US states. Data is normalized into GeoJSON. The system relies on OSM's power=* tagging schema, which includes overhead lines, cables, substations, and plants.

2. Topology Reconstruction

Raw OSM data is converted into a structural bus-branch model through a sequence of inference and cleaning steps:

Voltage Inference: Since ~15–30% of OSM lines lack voltage tags, an iterative neighbor-consensus algorithm infers missing voltages based on connected segments and substation tags.
Line Merging: Fragmented OSM ways representing a single physical corridor are stitched together using a union-find data structure.
Circuit Classification: The system distinguishes between inter-facility branches (retained), loops, taps, and isolated segments.
Transformer & HVDC Detection: Transformers are inferred where buses of different voltage levels coexist at a substation. HVDC links are identified via specific tag patterns (e.g., frequency=dc, bipolar voltage tags) or proximity to converter nodes.
Bus Creation: Spatial clustering assigns line endpoints to substations or creates ad-hoc buses for isolated endpoints.

3. Parameter Estimation

Since OSM lacks electrical parameters, the pipeline assigns them using lookup tables (LUTs) calibrated with US Energy Information Administration (EIA) data:

Impedance & Ratings: Line resistance ( $R$ ), reactance ( $X$ ), and thermal ratings are derived from voltage-class LUTs based on standard conductor types (e.g., bundled conductors for EHV).
Scaling Factors: To compensate for OSM's tendency to map multi-circuit corridors as single ways, the pipeline applies Topology Factors ( $N_T$ ) and Capacity Factors ( $N_C$ ). These factors scale impedance and thermal limits to match EIA-reported circuit-miles.
Generator Costs: Quadratic cost curves are generated using EIA-923 heat rates and fuel prices (Henry Hub for gas). If specific plant data is missing, default fuel-type LUTs are used.

4. Demand Allocation

Hourly demand data from EIA-930 (Balancing Authority level) is distributed to individual buses. The pipeline uses US Census tract population as a spatial proxy to allocate load, ensuring that demand distribution correlates with residential and commercial density.

5. Optimal Power Flow (OPF)

The final models are solved using PowerModels.jl with the Ipopt solver.

Progressive Relaxation Strategy: Recognizing that OSM-derived models may have parameter imprecisions that cause infeasibility under strict constraints, the pipeline employs an automatic relaxation strategy. It attempts to solve the problem at a strict level (L0) and progressively loosens constraints (angle limits, thermal headroom, voltage bounds, load shedding) through levels L1–L5 until convergence is achieved.
Reactive Power Compensation: To address missing shunt data in OSM, the pipeline injects compensating shunt capacitors/reactors at buses with reactive power deficits identified during a DC-OPF warm-start.

Key Contributions

End-to-End Open Data Pipeline: The first complete pipeline to generate AC-OPF solvable models for the entire US continental grid using only public data (OSM, EIA, Census).
Validation at Scale: The pipeline was validated on all 48 contiguous US states and six multi-state regions, including the full Western (5,076 buses) and Eastern (21,697 buses) Interconnections.
Public Release of 54 Models: The authors release all 54 solved models (48 single-state, 6 multi-state) as PowerModels-compatible JSON files, providing the research community with ready-to-use benchmarks derived from real infrastructure.
Progressive Relaxation Framework: A novel solver strategy that automatically handles model infeasibility due to data gaps, providing a transparent metric (relaxation level) for model quality.

Results

The pipeline successfully generated models for all 48 contiguous states and six regional interconnections.

Convergence:
- DC-OPF: 45 of 48 single-state models converged at the strictest level (L0).
- AC-OPF: 42 of 48 single-state models (88%) converged at L0; 44 (92%) converged at L0 or with minimal relaxation (AC1) during off-peak hours.
- Regional Models: The full Eastern Interconnection (21,697 buses) converged at relaxation level L3 in 47 minutes.
Economic Metrics:
- Median dispatch costs were $22.1/MWh, consistent with real wholesale market outcomes.
- System losses had a median of 1.0%, aligning with physical expectations for high-voltage transmission.
- The AC-OPF cost premium over DC-OPF was a median of 1.8%, reflecting resistive losses and reactive constraints, consistent with literature for well-characterized networks.
Geographic Fidelity: The models preserve real geographic correspondence, capturing actual transmission corridors and generation sites, unlike synthetic test cases.

Significance and Claims

The paper claims that grid data is foundational infrastructure for the energy transition, yet it remains inaccessible to most researchers. This work addresses that gap by demonstrating that usable, structurally, and electrically plausible transmission models can be constructed entirely from open data.

The authors emphasize that the goal is not to recover the exact operational grid (which is impossible without proprietary data) but to generate models that:

Converge under AC-OPF.
Preserve real geographic correspondence.
Reproduce system-level statistics (costs, losses) within realistic ranges.

By releasing these models, the authors aim to lower the barrier to entry for power systems research, enabling students, policymakers, and researchers to study grid behavior, train machine learning models, and simulate cascading failures without relying on restricted proprietary data or expensive commercial licenses. The methodology is presented as generalizable to other regions with adequate OSM coverage and public statistics.