- Titel
- Deep Learning based framework for Predicting Motion Dynamics in Complex Systems
- Zitierfähige Url:
- https://nbn-resolving.org/urn:nbn:de:bsz:14-qucosa2-1017072
- Datum der Einreichung
- 17.03.2025
- Datum der Verteidigung
- 24.03.2025
- Abstract (EN)
- Understanding and predicting the behavior of complex systems is critical towards performance optimization, ensuring safety and enhancing the controllability of such systems. A transportation network is also such a system, and existence of it relies on many different variables. Within this domain, the prediction of the motion of vehicles in the road is vital consideration that belongs to microscopic simulation. In this context, the ability to predict the motion of subject vehicle with relation to its neighbor vehicles and environment is crucial. However, this motion is two dimensional as longitudinal and lateral, which in general considered as two domains of studies. However, recently combination of both of these two movements into a single modeling domain has been popularized. Some of the advantages of that are, applicability to lane free traffics and providence of holistic idea about the vehicle movements through a single model. The modeling of such motion has generally been achieved through kinematic models. However, with the advent of deep learning an alternative data-driven modeling approach, assessing the ability of such a model to depict this behavior is crucial. Hence, this study attempt to fill this gap. To this end, a deep learning based model has been developed in order to predict the motion of subject vehicles. The performance is compared with a kinematic based two dimensional model. Finally, the output of this study is presented with advantages and disadvantages of both models while presenting the future research directions.
- Freie Schlagwörter (EN)
- Car Following Model, Deep Learning, Graph Neural Network
- Klassifikation (DDC)
- 006
- 380
- Klassifikation (RVK)
- WC 7722
- ZO 4600
- GutachterIn
- Prof. Dr. rer. pol. Ostap Okhrin
- Dr. rer. nat. Martin Treiber
- BetreuerIn Hochschule / Universität
- M.Tech Ankit Anil Chaudhari
- Den akademischen Grad verleihende / prüfende Institution
- Technische Universität Dresden, Dresden
- Version / Begutachtungsstatus
- angenommene Version / Postprint / Autorenversion
- URN Qucosa
- urn:nbn:de:bsz:14-qucosa2-1017072
- Veröffentlichungsdatum Qucosa
- 20.01.2026
- Dokumenttyp
- Masterarbeit / Staatsexamensarbeit
- Sprache des Dokumentes
- Englisch
- Lizenz / Rechtehinweis
CC BY 4.0- Inhaltsverzeichnis
List of Figures iv List of Tables v List of Abbreviations vii 1 Introduction 3 2 Literature Review 7 2.1 Car Following Models . . . . . . . . . . . . . . . . . . . . . . . . . . . . 7 2.2 Lane Changing Models . . . . . . . . . . . . . . . . . . . . . . . . . . . . 9 2.3 Mixed Traffic Models . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 10 2.4 Data Driven Models . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 11 2.5 Viable Networks for Deep Learning (DL) Modeling . . . . . . . . . . . . 13 3 Methodology 15 3.1 Data Description . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 15 3.1.1 Neighbor Selection . . . . . . . . . . . . . . . . . . . . . . . . . . 16 3.1.2 Handling Missing Data . . . . . . . . . . . . . . . . . . . . . . . . 18 3.1.3 Feature Selection . . . . . . . . . . . . . . . . . . . . . . . . . . . 18 3.1.4 Lane Changing Behavior of the Dataset . . . . . . . . . . . . . . 19 3.2 DL Model Development . . . . . . . . . . . . . . . . . . . . . . . . . . . 20 3.2.1 Long Short Term Memory (LSTM) . . . . . . . . . . . . . . . . . 20 3.2.2 Categorical Embedding . . . . . . . . . . . . . . . . . . . . . . . 21 3.2.3 Graph Convolution Network (GCN) . . . . . . . . . . . . . . . . 21 3.2.4 Model Development Process . . . . . . . . . . . . . . . . . . . . . 23 3.3 Proposed DL Model . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 24 3.3.1 Hyperparameter Optimization . . . . . . . . . . . . . . . . . . . 26 3.4 Intelligent Agent Model (IAM) . . . . . . . . . . . . . . . . . . . . . . . 26 3.4.1 IAM: Longitudinal Acceleration . . . . . . . . . . . . . . . . . . . 26 3.4.2 IAM: Lateral Acceleration . . . . . . . . . . . . . . . . . . . . . . 28 3.4.3 IAM: Floor Fields . . . . . . . . . . . . . . . . . . . . . . . . . . 29 3.5 Intelligent Driver Model (IDM) . . . . . . . . . . . . . . . . . . . . . . . 29 3.6 IDM and IAM Calibration . . . . . . . . . . . . . . . . . . . . . . . . . . 30 iiiContents 3.7 Simulation Environment . . . . . . . . . . . . . . . . . . . . . . . . . . . 31 3.7.1 Evaluation Metrics . . . . . . . . . . . . . . . . . . . . . . . . . . 32 4 Results and Discussion 33 4.1 Model Training - Acceleration Prediction . . . . . . . . . . . . . . . . . 33 4.2 Model Training - Position Prediction . . . . . . . . . . . . . . . . . . . . 34 4.3 IDM Calibration . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 35 4.4 IAM Calibration . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 37 4.5 DL Model Selection for Simulation . . . . . . . . . . . . . . . . . . . . . 39 4.6 Simulation Results Comparison . . . . . . . . . . . . . . . . . . . . . . . 40 4.6.1 Deviations . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 40 4.6.2 Lane Changing Behavior . . . . . . . . . . . . . . . . . . . . . . . 42 4.6.3 Time Space Diagrams . . . . . . . . . . . . . . . . . . . . . . . . 43 4.7 Discussion of the Results . . . . . . . . . . . . . . . . . . . . . . . . . . . 43 5 Conclusion and Future Research Directions 45 Bibliography 45 Appendix 51