[home] [Personal Program] [Help]

---

Session: Session 7D: Simulation methods & control
Session starts: Wednesday 11 September, 11:10
Presentation starts: 11:30
Room: Kallirhoe

---

Yannic Vaupel (RWTH Aachen University - Process Systems Engineering (AVT.SVT))
Adrian Caspari (RWTH Aachen University - Process Systems Engineering (AVT.SVT))
Nils C. Hamacher (RWTH Aachen University - Process Systems Engineering (AVT.SVT))
Wolfgang R. Huster (RWTH Aachen University - Process Systems Engineering (AVT.SVT))
Adel Mhamdi (RWTH Aachen University - Process Systems Engineering (AVT.SVT))
Ioannis G. Kevrekidis (Johns Hopkins University)
Alexander Mitsos (RWTH Aachen University - Process Systems Engineering (AVT.SVT))

Abstract:
Recovering waste heat from the exhaust gas of heavy-duty diesel trucks using a bottoming organic Rankine cycle (ORC) is a promising option to reduce fuel consumption. In contrast to most other applications of ORCs, e.g., geothermal or solar-thermal, the heat source in automotive applications is subject to strong fluctuations with limited predictability. Consequently, controlling the ORC system to maintain safe and efficient operation is a challenging task. Nonlinear model predictive control (NMPC) has been proposed for ORC systems and showed promising in silico results. It suffers, however, from a high computational expense and real-time capable implementation on vehicle hardware is questionable. Several methods are available that aim at shifting the majority of the computational load to the design phase of the controller, reducing on-line resource demand. In this work, we apply artificial neural networks (ANN) in silico to learn the control law of the NMPC controller off-line. We obtain training data from various NMPC scenarios with different initial conditions and heat source conditions using our in-house dynamic optimization tool DyOS. Subsequently, we apply the ANN-based controller in silico to different scenarios of transient heat source conditions. We compare the results to the NMPC solution obtained with DyOS and findings indicate that performance loss associated with the ANN-based controller is marginal while the control policy can be obtained at negligible computational cost.