Two-Level Neural Network Boosts Freeway Traffic Forecasting

Traffic forecasting on freeways has long relied on deep learning models, especially graph neural networks that blend graph theory with deep learning. These models are great at learning complex time and space patterns. However, traditional graph convolutional networks (GCNs) sometimes struggle with long-range spatial correlations, which can make long-term predictions less accurate. To solve this problem, a new model called the Two-level Resolution Neural Network (TwoResNet) has been introduced. It enhances interpretability by featuring two resolution blocks. The first block focuses on big-picture regional traffic patterns, while the second block, using a GCN, zeroes in on smaller, local spatial correlations. This dual approach allows the model to blend both nearby and distant traffic data, leading to better long-term forecasting. TwoResNet also shines in situations with noisy or incomplete data, offering clearer insights.