Merit-Order State Preparation for Budget-Efficient MA-QAOA: A Factorial Study of Initialisation, Training, and Noise Resilience in Power Systems Unit Commitment
Аннотация
We present a factorial study of Multi-Angle Quantum Approximate Optimisation (MA-QAOA) applied to the Unit Commitment (UC) problem in power systems, formulated as a binary optimisation over the Ising Hamiltonian. Forty-two QAOA variants are evaluated across four design dimensions: (1) physics-informed merit-order state preparation (MOSP), which biases each qubit toward on or off according to generator cost rank, versus uniform Hadamard initialisation; (2) four parameter initialisation schemes; (3) three training schedules, including a newly introduced layerwise progressive method; and (4) power-orbit symmetry reduction. Experiments span 5- and 10-generator instances across three evaluation budgets from 50 to 9,000 circuit evaluations. MOSP with uniform parameter initialisation achieves a 98.5% success rate (SR) at 50 evaluations, versus 95.0% for the best Hadamard baseline and 24.0% for Simulated Annealing (SA). However, MOSP paired with any other initialisation scheme collapses to 22–40% SR without symmetry reduction, revealing a strict compatibility requirement between state preparation and parameter initialisation. Power-orbit symmetry reduction partially rescues this collapse, lifting the physics-informed and inverse variants to 49–63% while leaving the random-beta ones unchanged. At 20-qubit scale, the top MOSP variant retains the highest mean approximation ratio (0.892 vs SA 0.588); this lead persists at 25 qubits, though its advantage over uniform Hadamard initialisation no longer widens with system size. Under simulated hardware noise, Hadamard-initialised variants prove substantially more resilient, with the best showing only 8.3% SR degradation at very-heavy noise, while MOSP-initialised variants degrade by 68–73%.