Optimizing the design and topology of electrical drive units regarding performance, efficiency and weight has become an integral part in the early phase of vehicle design. Present approaches sacrifice either computational speed or precise physical behavior in their component models. To overcome this, a computationally inexpensive, high-fidelity approach is proposed, that surrounds a neural network model of the electrical machine with an analytical inverter model and an interpolation model for the transmission. All three models share the common feature, that their inputs are parameters describing the geometry, materials, or electric design of the component, to maintain a link between the components design and its properties. The outputs of the models are maps and curves of their efficiency and performance, which can be combined in a modular way to account for interactions between the components e.g., shifted load points in the machine due to variations in the transmission ratio or the inverter switching frequency. For the electrical machine that is facilitated by multiple linked neural networks, predicting scalar properties, the torque-speed curve and loss maps. The entire model is compared with present approaches and assessed regarding its accuracy, training requirements and limitations. Finally, the computational performance and improvement in design of the approach is evaluated in an exemplary optimization of a drive unit.
Session:
Electric Powertrains I
|
| 14:00 - 14:30