Document Type

Thesis

Degree Name

Master of Applied Computing

Department

Physics and Computer Science

Program Name/Specialization

Applied Computing

Faculty/School

Faculty of Science

First Advisor

Dr. Sukhjit Singh Sehra

Advisor Role

Thesis Supervision

Second Advisor

Dr. Dariush Ebrahimi

Advisor Role

Thesis Co-supervision

Abstract

Intelligent transportation systems (ITS) depend on accurate traffic prediction to support congestion management, infrastructure planning, and real-time operational decisions. Despite substantial progress in data-driven forecasting, several challenges continue to limit practical deployment: traffic data is distributed across independent regional authorities, making centralized aggregation infeasible, standard federated aggregation strategies ignore traffic-specific characteristics that meaningfully affect model quality, and existing models produce only numerical outputs without interpretable reasoning that urban planners can act upon. This thesis addresses these challenges through four contributions that collectively advance privacy-preserving, explainable, and scalable traffic forecasting.

The first contribution provides a systematic review of 129 peer-reviewed publications, establishing the methodological landscape of large language model (LLM) applications in ITS, categorizing integration strategies and connecting architectural choices to empirical performance. Domain-adapted models consistently outperform prompt-only approaches, while interpretability and privacy-preserving deployment emerge as the two most consequential open challenges. These findings directly motivated the subsequent contributions. Building on this foundation, the second contribution proposed a traffic-aware federated learning (FL) framework that replaces standard weighting averaging with a dynamic client scoring mechanism driven by six domain-specific metrics. When combined with traffic-informed client clustering and an attention-augmented architecture, the framework predicts both traffic flow and speed across multiple future horizons, demonstrating consistent improvements over standard federated and centralized baselines under heterogeneous non-IID conditions.

The third contribution designs a domain-adapted LLM fine-tuned on structured traffic prompts encoding sensor metadata, spatial neighbour dynamics, demand statistics, and temporal context, producing both numerical predictions and natural language explanations. The fourth contribution integrates domain-adapted LLM with federated training into a unified federated LLM (FedLLM) framework. This is, to the best of our knowledge, the first of its kind implementation for freeway traffic flow prediction. Independent clients train locally and exchange only lightweight adapter parameters, substantially reducing communication overhead while keeping raw data private. The framework outperforms several centralized and federated baselines and generalizes to unseen geographic regions without retraining.

Collectively, these contributions demonstrate that privacy-preserving decentralized training, human-interpretable predictions, and cross-region generalization are simultaneously achievable within a single deployable framework, offering transportation agencies, city planners, and mobility researchers a practical foundation for next-generation traffic management systems that respect data ownership, adapt to local conditions, and explain their outputs.

Convocation Year

2026

Convocation Season

Spring

Download

Share

COinS