GPU-Accelerated Optimization Solver for Unit Commitment in Large-Scale Power Grids

TL;DR

The paper addresses the computational bottleneck of solving large-scale unit commitment problems by introducing a GPU-accelerated solver based on the Primal-Dual Hybrid Gradient method. It solves the relaxed subproblems on the GPU using cuOpt, providing faster bound estimation and improved crossover and branch-and-bound convergence within a two-stage DC/AC decomposition framework. Validation on 4224-, 6049-, and 6717-bus networks shows substantial speedups (up to 2.88x) while maintaining near-perfect solution quality (mean score ~0.9999). These results demonstrate the practical potential of GPU-accelerated first-order methods for real-time grid optimization and point to further gains as GPU hardware and sparse optimization libraries evolve.

Abstract

This work presents a GPU-accelerated solver for the unit commitment (UC) problem in large-scale power grids. The solver uses the Primal-Dual Hybrid Gradient (PDHG) algorithm to efficiently solve the relaxed linear subproblem, achieving faster bound estimation and improved crossover and branch-and-bound convergence compared to conventional CPU-based methods. These improvements significantly reduce the total computation time for the mixed-integer linear UC problem. The proposed approach is validated on large-scale systems, including 4224-, 6049-, and 6717-bus networks with long control horizons and computationally intensive problems, demonstrating substantial speed-ups while maintaining solution quality.

Figures (5)

• Figure 1: Our two-stage solution framework integrates the DC and AC modules presented in TPECIEEE-TIA. In this work, we improve the computational speed of the DC module (first stage) using GPU-based optimization. The GPU-based solver accelerates computations for large-scale networks, including the 4224-, 6049-, and 6717-bus systems.
• Figure 2: Computation time distribution for the 4224-bus network (Divisions 1–3). The GPU-based solver achieves faster convergence and lower variance in Divisions 2 and 3.
• Figure 3: Computation time distribution for the 6049-bus network (Divisions 1–3). The GPU-based solver achieves faster convergence and lower variance across all Divisions.
• Figure 4: Computation time distribution for the 6717-bus network (Division 1). The GPU-based solver shows improved convergence.
• Figure 5: Comparison of computation times between GPU-based PDHG and multi-threaded CPU solvers for the 6049-bus system across two random scenarios. Each bar represents the total runtime, decomposed into three stages: (1) presolve and relaxation, (2) crossover, and (3) branch-and-bound. For the CPU solver, the relaxation stage corresponds to the barrier method, while for the GPU solver, it corresponds to PDHG. The GPU-based relaxation consistently provides a higher-quality initial solution, reducing overall computation time by accelerating convergence in the crossover and/or branch-and-bound stages.