Digital design

This article tackles the complex challenge of predicting magnetic iron losses in high-frequency magnetic components by introducing a novel conditional generative adversarial network model. Diverging from traditional loss prediction methodologies that often overlook intricate interactions of factors, our conditional generative adversarial network framework is designed to comprehensively incorporate diverse aspects such as material properties, geometrical variations, and environmental conditions. To facilitate this advanced approach, a specialized four-wire measurement kit was employed, which significantly enriched the training dataset with a wide range of measurements. When benchmarked against conventional deep neural network models, the conditional generative adversarial network not only achieves faster convergence but also demonstrates markedly superior accuracy in predicting iron losses. This superiority is particularly notable in scenarios that extend beyond the training data's range, underscoring the model's robustness and adaptability. Such advancements in predictive accuracy and efficiency represent a significant leap forward in the design and optimization of high-frequency magnetic components.

The study introduces a proof-of-concept self-adaptive neural network model predictive control (SA-NNMPC) system, which uses a neural network as main component of the prediction model, for the anti-jerk control of electric vehicles. Through the adaptation mechanism of the network and cost function weights during vehicle operation, which is activated when the plant behaves significantly differently from its digital twin, the SA-NNMPC architecture adjusts to the progressive vehicle aging, or to the replacement of hardware parts, which is expected to be an important feature of next-generation vehicles. Validation tests and simulations show that the neural network accurately replicates the drivetrain dynamics of the considered electric vehicle, and, for nominal conditions, already leads to a performance improvement of the NNMPC implementation – which can run in real-time on a rapid control prototyping unit – with respect to a benchmarking nonlinear model predictive anti-jerk controller. Moreover, the preliminary simulation results confirm the potential of the proposed architecture in terms of: i) adaptability to operating conditions not covered in the original training, and variations of vehicle parameters; and ii) auto-tuning of the algorithm when applied to different vehicles.

Description of the virtual sensing for electric motor and inverter control developed in the EU funded project EM-TECH.