A spatial-temporal trend-event decoupling dual-channel framework for traffic flow prediction

Abstract

Accurate traffic flow prediction is crucial for urban traffic control, route planning, and congestion detection. However, traffic data is influenced by spatial-temporal relationships and exhibits significant distribution drifts. This phenomenon can be attributed to volatile events in the traffic network, which make periodic trends ambiguous and difficult to learn. Consequently, traffic signals can be seen as a combination of fluctuating event signals and stable trend signals, both possessing rich and distinct spatial-temporal characteristics. Although recent methods have achieved considerable performance, most of them still roughly treat the traffic flow as a whole without considering the interactions between trend and event factors from a decoupled perspective. To address this issue, we propose a Spatial-Temporal Trend-Event Decoupling Dual-Channel Framework (TEDDCF) for traffic forecasting. TEDDCF first decomposes traffic flow into trend and event signals, and constructs a Dual-Channel Signal Encoder (DCSE) to model each signal independently. Temporally, DCSE uses multi-head attention and causal convolution to learn long-term trends and short-term event features. Spatially, we design two novel dynamic fusion graph convolutional modules-Trend-GCN and Event-GCN-to capture the independent spatial characteristics of each signal. In the decoder, the complete spatial-temporal representation of traffic flow is obtained through a Trend-Event Interactive Fusion (TEIF) module for prediction. Experiments on six traffic datasets show that TEDDCF outperforms state-of-the-art baseline models in prediction performance while significantly reducing computational costs. The source code is available at https://github.com/XYHSMU/TEDDCF.