Accelerating nonlinear model predictive control through machine learning with application to automotive waste heat recovery

  • Beschleunigung nichtlinearer modellprädiktiver Regelung mittels maschinellem Lernen mit Anwendung auf Abwärmerückgewinnung in Fahrzeugen

Vaupel, Yannic; Mitsos, Alexander (Thesis advisor); Lucia, Sergio (Thesis advisor)

Aachen (2020, 2021)
Book, Dissertation / PhD Thesis

In: Aachener Verfahrenstechnik series - AVT.SVT - Process systems engineering 12 (2020)
Page(s)/Article-Nr.: 1 Online-Ressource : Illustrationen, Diagramme

Dissertation, Rheinisch-Westfälische Technische Hochschule Aachen, 2020


Waste heat recovery (WHR) from heavy-duty (HD) diesel trucks is a viable option for reducing the carbon footprint of the transport industry. Among the various available technology options for WHR, using a bottoming organic Rankine cycle (ORC) with the exhaust gas as heat source is considered the most promising. The ORC system in a HD diesel truck is subject to strong heat source fluctuations, which is in contrast to ORC operation in established processes. This poses substantial challenges for safe and efficient operation of the WHR system. In this thesis, we address these challenges using model-based methods. We first develop a dynamic ORC model for WHR and validate it with measurement data from a test rig. Next, we extend our dynamic model to a switching model that it is capable of accounting for start-up and shutdown situations. We compare two popular modeling approaches for the heat exchangers, identifying their perks and weaknesses. With our model established, we use dynamic optimization to understand how the system is best operated and we find that it can be beneficial to temporarily increase workfing fluid superheat in certain situations.From our findings, we derive a control structure for model-based control of the process. We apply this structure in silico to nonlinear model predictive control (NMPC) and to a PI controller with feedforward term. Our findings indicate good control performance of NMPC but excessive computational demand for on-board application. The PI controller achieves similar control performance at insignificant computational demand. Next, we apply a machine-learning (ML) based method for NMPC to the ORC system. While this achieves a drastic reduction in online computational demand, constraint satisfaction cannot be guaranteed. Hence, as a final contribution, we develop methods that use ML to reduce the computational demand of NMPC while promoting constraint satisfaction.