Electric machine co-optimization for EV drive technology development: Integrating Bayesian optimization and nonlinear model predictive control

Electric powertrains are becoming increasingly prevalent in various mobile propulsion applications, not only due to legislations for lower CO 2 emissions and local pollution, but also due to growing sustainable consciousness. However, conceptualizing those systems, consisting of component and controller design processes, is a complex task. The complexity itself arises from the amount of requirements for design objectives and use-cases, which can be met inside a multidimensional parameter space. Additionally, system design and evaluation are inherently tied to coupled component and system control strategy optimization. In this context, the paper presents a fully automated active machine learning methodology applied for a combined optimization of electric machine and system controller design, considering system performance, durability, and energy consumption. During this iterative approach a stochastic optimization of a permanent magnet synchronous machine (PMSM) takes place, constrained from a nonlinear model predictive control in a model-in-the-loop system environment. The active learning is covered by a Bayesian optimization algorithm with a Gaussian process regression to determine the most suitable parameter set in terms of exploration and exploitation. To demonstrate the feasibility of this novel methodology, a thermal subsystem from an electrified state-of-the-art powertrain has been used and further optimized regarding PMSM scaling and final gear ratio. Different real-world drive scenarios from highway to city were taken into account to cover typical sport utility vehicle use-cases. It could be shown that the electric machine losses of the optimized system are reduced by up to 32 . 7 % , which equals a consumption of − 0 . 43 k W h 100 k m compared to the reference vehicle. Due to slightly worse operating conditions of the inverter the whole system consumption has been minimized by − 0 . 35 k W h 100 k m . Three parameter studies with fixed iteration count have been executed to find the optimal machine diameter to be increased by 25 % and the length slightly reduced by 16 % . Moreover, the total gear ratio was adjusted by − 31 % to shift the load points of highest energy conversion into the machine’s optimal efficiency area. • Multi-layer co-optimization approach for electric machine design and thermal control strategy. • Combination of numerical (nonlinear model predictive control) and machine learning (Bayesian optimization) methods. • Electric machine right-sizing, considering electromagnetic and thermal performance, as well as system energy consumption. • System optimization of electric benchmark SUV, regarding real-world city and highway use-cases. • Active design learning and temperature control minimized electric machine losses by 32.8%, while maintaining sufficient performance. • Total energy consumption improvements of up to 0.35 kWh/100km.