Continuous Machine Learning for Cooperative, Connected and Automated Mobility applications

Abstract

Urban mobility is inherently non-stationary: shifts in weather, demand, and network conditions alter data distributions and degrade model accuracy over time. This diploma thesis designs, implements, and evaluates a drift-aware platform for continuous machine learning in Cooperative, Connected and Automated Mobility (CCAM). The system combines a modular microservice architecture for model serving, online monitoring, multi-detector drift consensus, and automated retraining with hot-swap deployment. A reproducible SUMO-based dataset for central Athens is constructed under normal and adverse-weather scenarios. Concept drift is induced via friction reduction to emulate rain, preserving network topology while yielding measurable shifts (average speed −13.68%, average trip duration +16.74%). Evaluation centers on Estimated Time of Arrival (ETA) prediction with a LightGBM model enriched by domain-specific features, exercised in a time-lapse experiment with an abrupt drift at the midpoint, due to rain conditions. Under drift, baseline errors increase markedly (MAE 30.36 s → 56.93 s and MAPE 13.20% → 19.02%). After detecting drift, the model is retrained using one hour of post-drift data and hot-swapped, recovering performance, when compared to the baseline model (MAE −11.00 s, −25.42% and MAPE −2.34 pp, −13.85%). The platform’s flexible architecture is further evidenced by the evaluation of Fuel Consumption and Number of Stops models, while its dashboard and user interface provide real-time interpretability. Overall, the results demonstrate a practical, closed-loop approach to detecting, responding to, and mitigating concept drift in realistic urban traffic settings.