Using graphics processing units for the efficient dynamic simulation of wind farms

Dynamic simulations are crucial for proper power system operation. However, these can be computationally expensive. This article introduces an efficient CUDA kernel for GPU acceleration, that achieves parallelization without the need to solve many different power systems at once. The proposal unifies generator and grid execution within a single CUDA kernel, ensuring accurate dynamic simulations of wind farms. A custom solution from scratch was developed, that in-tegrates the power system network solver with connected devices. Wind turbines are modeled as state-space linear systems, while the grid is parallelized through subunit partitioning and small submatrix inverses to reduce computational costs. The resulting kernel extracts parallelism in grids with limited inherent parallelism, outperforming previous methods such as column-level parallelism. Simulations on an NVIDIA GeForce RTX 3090 show a 26.39x speedup with respect to the C++ version, and the cooperative groups' synchronization primitives provide optimal performance while simulating up to 2464 wind turbines.