Enhancing Traffic Flow Prediction via Residual Spatio-Temporal Networks with Dilated Convolution

Reliable short- and medium-horizon traffic forecasts are now essential for signal timing, routing, and incident response. Yet urban flows are shaped by multi-scale spatial coupling and irregular external contexts (weather, holidays), which many models capture only partially for both smart transportation systems and urban management. Reliable prediction not only assists authorities in regulating traffic but also benefits daily commuters in planning their travel. Achieving high accuracy and efficiency in such forecasting remains a challenging research issue. With the advent of artificial intelligence and the availability of large-scale traffic datasets, deep learning methods have demonstrated superior performance compared to conventional probabilistic models. Nevertheless, two critical challenges persist: (1) excessive convolutional depth in networks often results in overfitting, and (2) urban traffic dynamics are strongly influenced by external variables, including weather conditions, temperature, and wind speed. To address these issues, we introduce an end-to-end neural architecture, termed RSTDCN, which integrates residual learning with dilated convolution. In this framework, standard convolution and dilated convolution modules are employed to capture both local and long-range spatial dependencies. Additionally, an external feature extraction module is designed to encode environmental factors into one-dimensional tensors using One-Hot Encoding.