Real-Time Short-Term Intersection Turning Movement Flows Forecasting Using Deep Learning Models for Advanced Traffic Management and Information Systems

Abstract

Traffic congestion remains a persistent challenge in urban transportation systems, causing excessive travel delays, increased fuel consumption, and severe environmental pollution. To address these issues, Advanced Traffic Management and Information Systems (ATMIS) have been developed, integrating real-time traffic monitoring, adaptive control strategies, and data-driven decision-making to enhance overall traffic efficiency. A crucial component of ATMIS is the real-time forecasting of intersection Turning Movement Flows (TMFs), which provides essential data for optimizing signal timings, improving vehicle routing, and implementing proactive congestion mitigation strategies. By leveraging accurate TMFs predictions, transportation agencies can dynamically adjust traffic signals, enhance intersection operations, and reduce delays, ultimately improving urban mobility and minimizing environmental impacts. While numerous traffic forecasting models exist, they face significant limitations in capturing the complex spatial and temporal patterns inherent in intersection-level TMFs, as they primarily rely on historical traffic data without adequately modeling these dependencies. Moreover, most existing approaches fail to incorporate exogenous factors, such as weather conditions, road characteristics, and other time-dependent variables, which significantly influence traffic flow but are often ignored. These shortcomings lead to poor generalization performance when applied to hold-out intersections (few-shot) and unseen regions (zero-shot), making them less effective in real-world dynamic traffic environments. To overcome these challenges, this study systematically develops and evaluates a deep learning-based TMFs forecasting framework designed for improved generalization and interpretability. First, we employ a Parallel Bidirectional LSTM (PB-LSTM) with multilayer perceptron (MLP) to capture both long-term seasonality and spatial dependencies, thereby enhancing the model's transferability across different locations, improving performance across hold-out intersections. Second, we integrate an encoder-decoder architecture using Deep Autoregressive (DeepAR) model, which enables probabilistic forecasting and quantifies uncertainty, ensuring robust predictions under varying traffic conditions. Third, we leverage the Temporal Fusion Transformer (TFT) to assess the relative importance of external covariates, such as weather conditions and road characteristics, improving interpretability and model reliability by identifying speed zone, road category, hour of the day, and temperature as key influential factors. Finally, we explore the potential of TimesFM, a decoder-only model, to enhance zero-shot learning capabilities, demonstrating strong performance in previously unseen intersections and new city datasets, particularly when enhanced with EMD and RF. To evaluate model performance, we conduct a series of experiments, including hold-out intersection tests, cross-city generalization assessments, and evaluations under extreme weather conditions, to assess robustness and adaptability. Experimental results highlight the effectiveness of integrating exogenous factors and hybrid modeling approaches in improving real-time TMFs forecasting accuracy, generalizability, and robustness under dynamic conditions. These insights provide valuable contributions to the development of scalable and interpretable deep learning models for intersection-level traffic flow prediction, supporting more adaptive and data-efficient traffic management strategies.