Comparing numerical and machine learning algorithms for optimized operation points of an electrical machine

Abstract

Abstract This work compares the Lookup Table (LuT) based numerical method with neural network (NN) based, and reinforcement learning (RL) based methods, for finding the optimal operating point of an Interior Permanent Magnet Synchronous Machine. Commonly, numerical methods are used to search for Maximum Torque Per Ampere (MTPA) points, which, although relatively accurate, often require significant computation time and generate large amounts of output data to obtain precise operating points. In this thesis, a simple approach was employed to establish a three-dimensional LuT based on nonlinear data, which is used as a baseline for comparing machine learning models. By comparing multiple metrics, it was verified that the presented NN-based method can quickly, efficiently, and accurately fit the LuT data, making it suitable for data reduction and addressing the issue of large output data of LuT. The RL-based method offers a simple model that is not dependent on data and can essentially achieve MTPA control, providing new inspiration for finding operating points. Finally, based on the comparative results, the advantages and challenges of the proposed different models are presented