Hybrid physics-data-driven online modelling: Framework, methodology and application to electric vehicles

This paper proposes a novel hybrid physics-data-driven framework for system modelling by integrating a physical model and an online learning data model to improve model accuracy, interpretability, and generalization. Taking an in-wheel Motor Driven Vehicle (IMDV) as an example, two hybrid representations, i.e. the Dynamic Linearization Data Model (DLDM) and Recurrent High-Order Neural Network (RHONN) are introduced for the planar dynamics modelling of the electric vehicle. However, it is difficult to obtain the statistical information of the operation process and measurement noise when the weight vectors of the data-driven model is updated online. To address this issue, a H infinity-based learning algorithm is adopted. The stability and convergence rate are elaborated and compared with an existing Extended Kalman Filter (EKF)-based method. Finally, we compare four methods, including the physics-based, data -based and two hybrid models, to evaluate their performances of modelling the IMDV's dynamics. The feasibility test and comparison studies are conducted in simulations and on a Hardware-in -the-Loop (HiL) test rig. The results demonstrated that the proposed H infinity-based hybrid method, which does not make any assumption on measurement noise, has better generalization ability and robustness in practical implementations, compared to other baseline methods.

Automated summary

This paper proposes a hybrid physics-data-driven framework to improve system modelling by integrating a physical model and an online learning data model. Two hybrid representations and a H infinity-based learning algorithm are introduced. The proposed hybrid method shows better generalization ability and robustness in practical implementations compared to other baseline methods, as demonstrated through simulations and experiments.