High-performance parallel computing-based fast power-flow solution technology for large-scale power grids

Abstract

Fast power-flow solution is a computational foundation for large-scale grid operation, renewable accommodation, contingency screening, and automated energy facilities. Conventional Newton-Raphson (NR) solvers, however, become inefficient when thousands of large-network snapshots with repeated topologies must be solved. This paper proposes a topology-aware central processing unit-graphics processing unit (CPU-GPU) parallel framework. It preserves the alternating-current (AC) NR model, reuses symbolic sparse structures, batches mismatch and Jacobian kernels, overlaps host-device transfer with numerical kernels, and redirects difficult cases to a conservative fallback. Camera, infrared, and light detection and ranging (LiDAR) inspection flags are used only as auxiliary evidence for scenario prioritization. Public IEEE 300, PEGASE 2869, ACTIVSg10k, and ACTIVSg70k cases are evaluated. With serial NR runtime normalized to 1.000, the CPU-GPU runtimes are 0.185, 0.054, 0.021, and 0.010, which yield 5.4x, 18.5x, 47.6x, and 100.0x speedups, respectively. All methods use the same 10^-4 p.u. residual tolerance and auditable convergence records. Ablation results show that removing symbolic caching, stream overlap, GPU sparse solving, or convergence masking increases the time index by 41%, 23%, 76%, and 18%, respectively. These results show that topology reuse and end-to- end heterogeneous execution can support high-throughput digital-twin power-flow services.