Combined Physical and ANN-Based Engine Model of a Turbo-Charged DI Gasoline Engine with Variable Valve Timing

Abstract

<div class="section abstract"><div class="htmlview paragraph">High-efficient simulations are mandatory to manage the ever-increasing complexity of automotive powertrain system and reduce development time and costs. Integrating AI methods into the development process provides an ideal solution thanks to massive increase in computational power. Based on an 1D physical engine model of a turbo-charged direct injection gasoline engine with variable valve timing (VVT), a high-performance hybrid simulation model has been developed for increasing computing performance. The newly developed model is made of a physics-based low-pressure part including intake and exhaust peripheries and a neural-network-based high-pressure part for combustion chamber calculations. For the training and validation of the combustion chamber neural networks, a data set with 10.5 million operating points was generated in a short time thanks to the parallelizable combustion chamber simulations in stand-alone mode. The data set covers wide variation ranges of boundary and operating conditions in the combustion chamber including variable valve timings. A special neural network structure was configurated, which consists of five interconnected gated recurrent unit (GRU) sub-networks for calculating mass fuel burned, pressure values, the peak pressure position, NO emissions as well as knock condition. To form a whole working cycle simulation within GT-SUITE, the neural networks were converted into a functional mock-up unit (FMU), which is connected with the physics-based low-pressure part through FKFS RapidCylinder®. A performance evaluation of the hybrid engine model shows that, the mean deviations of brake torque, MFB50 and NO emissions compared to the physics-based reference mode are respectively 0.529%, 0.048°CA and 32.67ppm over the whole engine characteristic map, indicating an equal calculation quality. While maintaining the calculation accuracy, the neural networks with FMU connections realize a combustion chamber calculation 5- to 10-fold faster than real-time and an acceleration of the whole engine calculation of up to 80% compared to the completely physics-based simulation.</div></div>