Teaching Mobility Patterns to Machines: A Deep Learning Based Approach

Kurzfassung

The integration of renewable energy sources and electric vehicles presents transformative opportunities but also significant challenges for understanding and managing mobility data. In Germany, datasets like the Mobilitätspanel (MoP) provide invaluable insights into travel behaviors and energy consumption. However, the proprietary nature of such datasets, combined with strict privacy regulations, limits their accessibility, hindering research, urban planning, and the development of sustainable transport solutions. Addressing this data scarcity is essential to support the effective integration of electric vehicles into the energy grid and to design data-driven strategies that align with sustainability goals. This thesis proposes a privacy-compliant synthetic data generation framework that replicates real-world mobility patterns while preserving individual privacy. Leveraging a neural network-based synthesizer, with a focus on autoregressive architectures such as decoder-only transformers, the framework captures complex temporal and statistical dependencies inherent in mobility behaviors. The proposed model is evaluated through quantitative and qualitative analyses, demonstrating its efficacy in producing high-fidelity, privacy-compliant datasets that closely mirror the statistical properties of real-world data. By enabling researchers and policymakers to analyze mobility trends, evaluate transportation policies, and design sustainable infrastructure, this work bridges the gap between data accessibility and privacy compliance. Ultimately, this research fosters innovation in mobility analytics, supports data-driven decision-making, and contributes to sustainable urban and energy planning in Germany.