LightCast: Efficient Traffic Flow Forecasting via an Integrated Compression Framework

Traffic forecasting plays a pivotal role in intelligent transportation systems. To enhance forecasting accuracy, existing deep learning models often feature complex structures with large computational demands for deployment. To achieve seamless traffic system operations and prompt communication between vehicles and roadway infrastructure, the best solution is to deploy models on locally operating edge devices, which are limited in hardware resources. In this realm, we consider both efficiency and effectiveness in this paper with our newly proposed Lightweight Forecasting (<sc xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">LightCast</small>) model. Specifically, we first design Spatio-temporal Global-local Former (<sc xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">STGLFormer</small>) that introduces various self-attention mechanisms comprehensively considering both global and local spatio-temporal information in traffic data to offer state-of-the-art (SOTA) forecasting accuracy. Furthermore, <sc xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">LightCast</small> involves a mix-granularity pruning strategy to remove redundant components in <sc xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">STGLFormer</small> at different granularities and an automated layer-matching distillation scheme to effectively restore the forecasting accuracy after pruning. The automated layer-matching distillation scheme resolves the issues of layer mismatching in the traditional feature distillation approach. Extensive experiments conducted on four real-world public transportation datasets demonstrate that our approach can achieve near SOTA performance at a much higher computation efficiency.