A Physics Enhanced Residual Learning (PERL) Framework and Its Application to Vehicle Trajectory Prediction
CME Department Seminar
March 7, 2025
11:00 AM - 12:00 PM America/Chicago
Presenter: Xiaopeng (Shaw) Li, PhD, University of Wisconsin-Madison
Location: ERF, Room 1047
Abstract: Physics models and data-driven neural network (NN) models are two predominant methodologies for a general prediction problem. However, each approach presents its own set of challenges: physics models fall short in predictability, while NN models lack interpretability. Addressing these identified shortcomings, this paper proposes a novel framework, the Physics-Enhanced Residual Learning (PERL) framework. PERL integrates the strengths of physics and the NN model for predictions. This framework melds the foundation of a physics model with a residual learning model, producing predictions by summing the outputs of the physics model with a corrective predicted residual. In this way, it preserves the interpretability inherent to physics models and has reduced data requirements compared to NN models. We demonstrate the PERL model in vehicle trajectory prediction using a real-world trajectory dataset. The physics model in this case is the Intelligent Driver Model (IDM) and the residual learning model is a Long Short-Term Memory (LSTM) model. We compare this PERL model with the pure physics model, NN model, and other physics-informed neural network (PINN) models. The result reveals that PERL achieves better prediction with a small dataset, compared to the physics model, NN model, and PINN model. Besides, the PERL model showed faster convergence during training, offering comparable performance with fewer training samples than the NN model and PINN model. Sensitivity analysis proves the comparable performance of PERL using different methods of building the residual learning model and the physics car-following model.
Speaker Bio: Dr. Xiaopeng (Shaw) Li is currently Harvey D. Spangler Professor in the Department of Civil and Environmental Engineering and an affiliate in the Department of Electrical and Computer Engineering at the University of Wisconsin-Madison (UW-Madison). He directs the USDOT Rural Autonomous Vehicle Program for Passenger Transportation. He established the Connected and Automated Transportation Systems Lab that developed a multi-scale CAV testbed including multiple full-scale and reduced-scale CAVs and associated system units. Before joining UW-Madison, he served as the director of one USDOT national university transportation center, the National Institute for Congestion Reduction (NICR). He is a recipient of a National Science Foundation (NSF) CAREER award. He has led a number of federal (e.g., NSF, USDOT, USDOE), state, and industry grants, mostly as the PI, with a total budget of over $35 million in addition to over $25 million local matching. He has published over 120 peer-reviewed journal papers. He is a fellow of ASCE and a senior member of IEEE. He serves as an area or associate editor for multiple peer review journals, such as ASCE Journal of Urban Planning & Development, ASCE-ASME Journal of Risk and Uncertainty in Engineering Systems, Transportation Research Part E, IEEE TITS, IEEE ITSM, ASCE . He also chairs the IEEE ITSS Emerging Transportation Technology Testing committee. His major research interests include modeling and field experiments of connected, electric, and automated vehicles, transportation and infrastructure systems analysis. He received a B.S. degree (2006) in civil engineering from Tsinghua University, China, an M.S. degree (2007), and a Ph.D. (2011) degree in civil engineering along with an M.S. degree (2010) in applied mathematics from the University of Illinois at Urban-Champaign, USA.
Date posted
Feb 27, 2025
Date updated
Feb 27, 2025