Formulation comes before solver comparison
QuantumGrid explores binary encodings and variational solvers for generator scheduling. The useful question is whether the mathematical problem survives each transformation—not whether two solvers produce visually similar numbers. The current implementation is exploratory and does not demonstrate quantum advantage.
Commitment and dispatch are different decisions
Commitment chooses which generators are on; dispatch chooses their output subject to generation limits, demand and operating constraints. The classical MILP can choose continuous output below maximum capacity. A binary encoding that forces every active generator to full capacity solves a different problem.
Classical formulation · Binary encoding
QUBO energy is not a monetary objective
The existing QUBO and classical formulations are not yet a like-for-like benchmark. The QUBO penalizes distance from demand using binary full-capacity dispatch; the MILP allows continuous dispatch. QUBO energy contains penalty terms and omits constant offsets, so it is not a monetary generation cost. The scaling entry point also constructs different problem instances for its classical and quantum paths.
Dropping a constant leaves the minimizer unchanged, but changes the numeric value of the energy. Penalty terms also carry units chosen for the encoding. Recompute physical generation and startup costs from a feasible decoded schedule before displaying a monetary cost.
QAOA optimizes the supplied encoding
Alternating cost and mixer layers define a parameterized circuit. A classical optimizer adjusts those parameters using the encoded Hamiltonian’s expectation. The result is meaningful only relative to that Hamiltonian; sampling a low-energy bit string does not independently certify demand satisfaction.
VQE uses the same binary scheduling experiment
The VQE path uses a parameterized ansatz over the QUBO-derived Ising model. It is not an implementation of continuous optimal power flow. Comparing it with QAOA requires the same encoded instance, comparable optimization budgets and a common decoding/evaluation procedure.
Preserve baseline status and feasibility
MILP, greedy dispatch and simulated annealing expose different trade-offs. Retain solver status, feasibility, runtime and any optimality bound instead of labeling every MILP return “exact.” For comparisons, each solver must receive the same fleet, demand and constraints.
The reviewed scaling script constructs different instances for its classical and quantum paths and does not include the advertised MILP baseline in that loop. Its outputs cannot support a like-for-like ranking. Inspect the scaling script.
A larger penalty cannot repair the wrong constraint
A squared demand residual encourages equality at the encoded dispatch levels. With one generator of capacity 100 and demand 20, full-capacity ON has residual 80 while OFF has residual −20. Squaring those residuals favors OFF, even though it fails demand. Increasing the penalty strengthens that preference.
This simple counterexample is enough to reject a feasibility guarantee based only on penalty magnitude. The encoding needs a justified treatment of dispatch or slack variables, plus an independent checker.
Validate tiny instances before scaling
Use an instance small enough to enumerate every commitment vector and independently solve or check dispatch. Compare feasible schedules, physical costs, QUBO energies and Ising energies with their offsets. Persist the complete instance and results before increasing circuit size.
No persisted benchmark outputs in the reviewed repository establish the earlier scaling table. Statevector simulation also measures a classical simulation workload; its runtime is not evidence of quantum-hardware speedup.
A common evaluation contract
instance = fixed_fleet_and_demand
for solver in solvers:
candidate = solver(instance)
feasibility = independently_check(candidate, instance)
physical_cost = evaluate_cost(candidate, instance) if feasibility else None
save(instance, candidate, feasibility, physical_cost, solver_status, runtime)Protocol sketch for the next implementation milestone. Preserve encoded energy as a separate diagnostic field, never as a substitute for physical cost.
The next useful result
A verified tiny-instance oracle and a shared evaluation contract would make this project materially stronger. Real ENTSO-E ingestion, a continuous OPF formulation and quantum advantage remain unestablished. The current loader generates synthetic data even when given an API key; the data source should remain explicit in every experiment.