Skip to content

Power systems for quantum researchers

You know QAOA, HHL, and kernel methods. This page gives you the power system vocabulary you need to work on grid problems without misreading them — and, more importantly, the constraints that decide whether a formulation is credible to power engineers.

The object of study

A transmission network is a graph. Buses (nodes) carry generation and load in megawatts (MW). Branches (edges) are lines and transformers with an impedance; power flows over them according to physics, not routing — you cannot send power along a chosen path. Generators attach to buses and have cost curves, typically quadratic: \(c_2 p^2 + c_1 p + c_0\) in $/h with \(p\) in MW.

The community's standard data format is the MATPOWER case: three matrices (bus, branch, gen) with fixed column meanings. QuGrid's Network keeps those exact semantics, which is why a researcher's own case file loads without translation:

import qugrid as qg

net = qg.cases.case9()          # bundled WSCC 9-bus test system
net.bus.shape, net.branch.shape  # MATPOWER columns, unchanged

The IEEE test cases (9, 14, 30, 39, 57, 118 buses, all bundled) are the community's shared benchmark instances — the equivalent of MaxCut on 3-regular graphs, except each one is a real physical system with published provenance.

The four problems QuGrid encodes, and why they matter

Power flow. Given injections, find bus voltages such that Kirchhoff's laws hold. The DC approximation is a linear system \(B'\theta = P\) (sparse, symmetric, condition numbers typically \(10^2\)–\(10^4\)); the full AC problem is solved by Newton's method, a sequence of such linear systems. This is the entry point for quantum linear solvers — and the reason condition_number() is on every LinearSystemProblem: HHL's cost scales with it.

Unit commitment (UC). Decide which generators are on in each hour (binary) and their output (continuous), minimizing cost subject to meeting demand. The mixed binary-continuous structure is why it becomes a QUBO only after discretizing power output — an approximation with a measurable cost that QuGrid reports separately (continuous_reference()).

Controlled islanding. After a disturbance, split the grid into self-sufficient islands to stop cascading failures. Graph partitioning with a power-balance constraint — the cleanest grid-native QUBO, and QuGrid's default demo.

Security screening. Operators check thousands of operating points against the N-1 criterion: the system must survive any single component outage. Each check is cheap; the volume motivates learned classifiers — the natural grid application for quantum kernels.

What makes a grid formulation credible

These are the standards power system reviewers apply; violating them is why quantum papers get rejected at power venues.

  1. Engineering units end to end. A result is a dispatch in MW and a cost in $/h, not an Ising energy. QuGrid's decode() exists for this.
  2. Constraint violations quantified. “The ground state was feasible” is not enough; report the balance error in MW. Penalty encodings make this non-optional.
  3. The baseline is the field's actual tool. DC power flow is solved by sparse LU in microseconds; UC by mixed-integer programming at national scale nightly. A quantum comparison against random guessing or unoptimized enumeration convinces nobody.
  4. Test systems are standard. Results on IEEE cases are checkable; results on a hand-made 4-bus graph are not.
  5. N-1 thinking. Any operational claim must survive single-outage scrutiny; it is the field's default notion of robustness.

Scale reality

Problem Realistic research instance Variables / qubits Production instance
DC power flow 9–118 bus IEEE cases system size 8–117, needs \(\lceil \log_2 n \rceil\) + clock qubits in HHL 10k–80k buses
Unit commitment 2–5 units, 2–4 periods as QUBO 10–22 binary vars 1k units, 36–48 periods
Islanding 9–39 bus cases one var per bus 10k buses
Screening 3–8 load features one qubit per feature hundreds of features

The gap between columns two and four is the honest headline: research-scale quantum experiments inform algorithm design; they do not run grids. QuGrid keeps you productive in column two and precise about the distance to column four.

Where to go next

  1. Tutorial 02 — from a MATPOWER case to a QUBO: the full encoding pipeline, including what penalty weights do.
  2. Example zoo: ten runnable studies, each with a classical reference.
  3. Honest benchmarking: the reporting rules this library enforces.