Highlights
- Implemented an MPI-enabled interface between the simulation and optimization software layers
- Obtained weak and strong scalability for a realistic test case of wind farm optimization on a 6528 core machine
- Identified bottlenecks for the use of PARA-OPT in Alpaqa
| Keywords | Energy, Environment/Climate/Weather, Mechanical Engineering, Sequential Quadratic Programming (SQP), Data assimilation |
| Technologies used | SP-Wind, Alpaqa4Scale, ParaOPT |
Challenge
Wind farm control applications pose unique challenges because of the huge scale of the nonlinear PDE-constrained optimization problems involved, and due to the real-time requirements of feedback control. Parallelization in the three space dimensions alone has diminishing returns, caused by communication overhead, necessitating further parallelization in the time dimension. However, parallel-in-time multiple shooting techniques introduce additional constraints into the optimization problem, which requires constrained solvers that usually take more iterations to converge to a feasible solution. Additionally, these numerical optimization solvers need to be integrated efficiently with the large-eddy simulation code. The large-scale nature of the problems calls for matrix-free methods with efficient and parallelizable preconditioners.
Research Topic
The goal of the project is to make an important step towards real time simulation and control of wind farms. Matrix-free optimization is required to enable many constraints, and parallelization both in space and time levels is used to fully exploit the capacity of large supercomputers. In addition, current state-of-the art numerical libraries are not easily accessible, as they do not work with the complex (and efficiently parallelized) data structures of SP-Wind. The current project aims to close these gaps, leading to a potential order of magnitude speed-up of optimal control of wind farms.
Solution
To overcome the limitations of parallelization in the space dimensions only, this project implemented a parallel-in-time multiple shooting formulation of a velocity tracking problem, using SP-Wind for the large-eddy simulations. The simulation code for this problem interfaces with Alpaqa4Scale, a matrix-free numerical optimization library designed for large-scale distributed memory computing, using a flexible and reusable interface. The novel PowerALM- based optimization solver in Alpaqa4Scale reduces the number of iterations required to reach constraint satisfaction and simplifies tuning of the algorithmic hyperparameters. Finally, preconditioners based on the ParaOpt framework were studied in order to efficiently compute Newton-Krylov directions in the inner optimization solvers.