Efficient Newton-Raphson Power Flow with Parallel Jacobian Construction

This paper presents a novel node-level data localization approach to reduce the time spent on the Jacobian building step of the Newton-Raphson Power Flow problem on CPUs. By leveraging the sparsely connected graph structure of the electrical network, we package the data of a network bus and its immediate neighbors and structure the computations to reuse data. This reduces the overhead associated with random memory accesses on GPU and CPU architectures. We demonstrate the effectiveness of our method on large-scale power networks, achieving a 16.8x speedup for Jacobian matrix building compared to Matpower on a 25,000-bus network.