Deep Reinforcement Learning HEV Energy Management under Consideration of Dynamic Emission Models Technische Universität Dresden

Format:

Book

Conference/Event

Author/Creator:

Fechert, Robert, author.

Contributor:

Bäker, Bernard

Liessner, Roman

Lorenz, Alexander

Conference Name:

SAE Powertrains, Fuels & Lubricants Meeting (2020-09-22 : Krakow, Poland)

Language:

English

Physical Description:

1 online resource cm

Place of Publication:

Warrendale, PA SAE International 2020

Summary:

Hybrid vehicles (HEV) contribute to reduce emissions from transportation. The energy management controls the powertrain components in HEVs. In addition to minimizing fuel consumption, improving air quality is a major opportunity for hybrid vehicles. The pollutant emissions can only be mapped with sufficient accuracy using dynamic models. A promising approach to solve the resulting high-dimensional problems is Deep Reinforcement Learning (Deep RL), which has recently been introduced in the field of HEV energy management.This paper investigates the application of a Deep RL agent for the optimal design of a diesel hybrid energy management with nitrogen oxide emissions considered in addition to fuel consumption. The dynamic nitrogen oxide model required for this is created using a supervised learning approach based on recorded measurement data. The dynamic model considers the model inputs during the last time steps to ensure sufficient model quality. Classical algorithms such as Dynamic Programming or Pontryagin's maximum principle are not even able to generate a reference solution in a time appropriate to the development process. The Deep RL, on the other hand, generates solutions for real driving situations using stochastic driving cycles. At the same time, it offers the possibility to consider other relevant system variables such as battery temperature or battery derating. Since the use of the dynamic emission model involves the violation of the Markov condition, the paper shows an approach to solve the emerging Partially Observable Markov Decision Process (POMDP) using a DDPG agent.The paper presents a new algorithm for minimizing pollutants in parallel with reducing fuel consumption in hybrid vehicle energy management. The resulting multi-criteria optimization problem is solved with a Deep RL approach suitable for dynamic models. The results show a great potential regarding the reduction of fuel consumption and nitrogen oxide emissions for real traffic

Notes:

Vendor supplied data

Publisher Number:

2020-01-2258

Access Restriction:

Restricted for use by site license