Abstract:

Accurate and timely traffic flow prediction is essential for the development and efficiency of Intelligent Transportation Systems (ITS). While recent advancements in graph-based neural networks have yielded promising results, challenges such as optimal graph construction and high computational complexity still persist. To address these issues, this paper proposes a novel Multi-Stream Feature Fusion Approach that integrates diverse features from traffic data through a data-driven adjacency matrix, replacing traditional distance-based graph structures.The adjacency matrix is initialized using the Spearman rank correlation coefficient between traffic monitoring stations and is further optimized during model training. At the core of the model lies the Multi-Stream Feature Fusion Block (MFFB), which comprises three parallel neural network streams: a Graph Convolutional Network (GCN) for spatial feature extraction, a Gated Recurrent Unit (GRU) for capturing temporal dependencies, and a Fully Connected Neural Network (FNN) for learning additional contextual features. A soft attention mechanism dynamically weights and fuses the outputs from these streams, enhancing feature representation and predictive performance.The stacked MFFB modules, followed by a convolutional and a fully connected layer, enable the model to make accurate traffic flow predictions. Experimental results on two real-world traffic datasets demonstrate that the proposed method significantly outperforms existing state-of-the-art approaches, while maintaining manageable computational complexity. Additionally, the system includes a real-time visual interface that displays traffic predictions, performance metrics, and traffic type distributions—offering valuable insights for urban traffic management, congestion control, and smart city planning.

Description:

.