Optimizing Sparse Multi-GPU Architectures for Large-Scale Power System Analysis

Power Flow (PF) calculation is a critical tool for efficiently analyzing power systems, determining voltage magnitudes and phase angles in the context of evolving energy supply networks. The increasing sparsity and complexity of large-scale power systems have made solving these problems in a parallelized manner a significant challenge for power engineers. To address this bottleneck, this paper presents a sparse multi-GPU Fast Decoupled (FD) method to accelerate PF calculations. The proposed approach leverages data parallelism to enhance scalability while maintaining load balancing across multiple GPUs. Furthermore, GPUDirect technology is utilized to minimize communication overhead among GPUs, significantly improving performance. Experimental results demonstrate that the proposed method achieves nearly a $4 x$ speedup on power systems with more than 10,000 buses, compared to the optimized MATLAB-based library MATPOWER, which showcases its effectiveness for large-scale power system analysis.