Parallel processing of optimization algorithms: A faster way to optimize electric power problems that utilize FEM-simulations

Abstract

The aim of this thesis was to find a way of doing optimizations based on FEM-simulations and further to parallelize the computations in order to decrease the execution time. If proven possible the difficulty setting up these simulation and the expected decrease in runtime, would to be evaluated. Optimizations of magnet field strength and torque utilizing FEM simulations were conducted by using the particle swarm optimization (PSO) algorithm and the Matlab "partial differential equation toolbox" (PDE-toolbox). Two cases were optimized with and without parallel processing, and runtimes were measured for all runs. The aim was to run all simulations 10 times on three different computers. This was achieved to the majority of the simulations, results therefore had a good statistical confidence.The optimization results were consistent with the theory: the higher number of iterations and particles used in the PSO, the better solution and smaller deviations. The runtime was found to be linear with the product of iterations and particles. This fits the expectations and the theory since the product of iterations and particles equals the total number of FEM calculations done. The FEM-simulation was the most time consuming when executing the code.The simplest case took up to 7 hours without parallel processing. The same simulation was down to 20 minutes using 12 parallels. The speedup was proportionally alike to the number of cores for the 50/50 simulation in case 1. Case 2 had a lower speedup, but this was also linear. The same tendency was found for sets with fewer particles/iterations, but in these cases deviations were significant. Setting up the model in the PDE-toolbox from command line was demanding. This may also be due to lack of example cases, even Internet searches turned out empty for similar simulation setups. But worth while when considering the reduced runtime one achieved. To prepare the optimization algorithm for parallel processing was however easy and took very little time.