TY - JOUR
T1 - Energy and Emission Management of Hybrid Electric Vehicles using Reinforcement Learning
AU - Hofstetter, Johannes
AU - Bauer, Hans
AU - Li, Wenbin
AU - Wachtmeister, Georg
N1 - Publisher Copyright:
© 2019, IFAC (International Federation of Automatic Control) Hosting by Elsevier Ltd. All rights reserved.
PY - 2019
Y1 - 2019
N2 - The electrification of drivetrains of conventional vehicles plays a decisive role in reducing fuel consumption. At the same time decreasing pollutant emission limits must be met also under real driving conditions. This trade-off between fuel consumption and pollutant emissions needs to be optimized, which results in powertrains with increasing complexity. A holistic energy and emission management is needed to control such systems in a way that the fuel consumption is minimized while emission limits are respected. Mathematical optimization methods are difficult to apply in real-time applications due to high computational and calibration demands. Self-learning algorithms, on the other hand, seem to be a suitable solution for such optimization problems. In this paper a control strategy for a hybrid electrical vehicle is presented, consisting of a decision-making agent, trained on different test drives with Reinforcement Learning. For these, the Proximal Policy Optimization method was applied. The strategy controls the torque-split between the combustion engine and electric motor, the power of an electrically heated catalyst and internal engine measures. The method is demonstrated in a simulation framework based on a Diesel P0-HEV with a SCR exhaust gas aftertreatment system. In comparison to a reference strategy a fuel reduction of 3.1 % averaged over the test data set was achieved.
AB - The electrification of drivetrains of conventional vehicles plays a decisive role in reducing fuel consumption. At the same time decreasing pollutant emission limits must be met also under real driving conditions. This trade-off between fuel consumption and pollutant emissions needs to be optimized, which results in powertrains with increasing complexity. A holistic energy and emission management is needed to control such systems in a way that the fuel consumption is minimized while emission limits are respected. Mathematical optimization methods are difficult to apply in real-time applications due to high computational and calibration demands. Self-learning algorithms, on the other hand, seem to be a suitable solution for such optimization problems. In this paper a control strategy for a hybrid electrical vehicle is presented, consisting of a decision-making agent, trained on different test drives with Reinforcement Learning. For these, the Proximal Policy Optimization method was applied. The strategy controls the torque-split between the combustion engine and electric motor, the power of an electrically heated catalyst and internal engine measures. The method is demonstrated in a simulation framework based on a Diesel P0-HEV with a SCR exhaust gas aftertreatment system. In comparison to a reference strategy a fuel reduction of 3.1 % averaged over the test data set was achieved.
KW - Adaptive Algorithms
KW - Automotive Control
KW - Automotive Emissions
KW - Energy Management Systems
KW - Hybrid Vehicles
KW - Multiobjective Optimization
KW - Reinforcement Learning
UR - http://www.scopus.com/inward/record.url?scp=85081574728&partnerID=8YFLogxK
U2 - 10.1016/j.ifacol.2019.12.615
DO - 10.1016/j.ifacol.2019.12.615
M3 - Conference article
AN - SCOPUS:85081574728
SN - 1474-6670
VL - 52
SP - 19
EP - 24
JO - IFAC Proceedings Volumes (IFAC-PapersOnline)
JF - IFAC Proceedings Volumes (IFAC-PapersOnline)
IS - 29
T2 - 13th IFAC Workshop on Adaptive and Learning Control Systems, ALCOS 2019
Y2 - 4 December 2019 through 6 December 2019
ER -