A spatio-temporal multi-scale dual-pooling neural network for traffic predictions

Abstract

The problem of traffic predictions is playing an increasingly important role in intelligent transportation systems. While recent progress has enhanced the accuracy of traffic prediction to some extent, numerous existing methods still suffer from significant limitations. Specifically, they often fail to comprehensively explore the intricate spatio-temporal dependencies across different scales, which are essential for accurately understanding and forecasting traffic patterns. Moreover, these methods tend to struggle when confronted with noise or abrupt fluctuations in traffic data, which are common occurrences in real-world traffic scenarios. To address these challenges, this paper introduces a Spatio-Temporal Multi-scale Dual-Pooling neural network (STMDP) for traffic predictions. It is composed of three key components: 1) a dual-pooling layer that combines average pooling and median pooling, leveraging median pooling’s resilience to outliers to preserve critical features amid noise, 2) multi-scale parallel encoders designed to effectively model spatio-temporal dependencies at various temporal scales and 3) a multi-scale fusion decoder that utilizes these dependencies to generate accurate future predictions. Together, these components significantly improve the capacity of STMDP to capture spatio-temporal dependencies in traffic data, particularly in handling sudden fluctuations. Extensive experiments were conducted on three real-world traffic datasets. The results clearly demonstrate that the STMDP approach outperforms other existing models, highlighting its superiority in traffic prediction tasks and indicating its great potential for practical applications in intelligent transportation systems.