Spatial–temporal multi-feature fusion network for long short-term traffic prediction

Highlights

• Capturing complex spatial–temporal correlations for short and long term prediction.

• A multi-scale attention is presented to capture temporal corrections.

• A gated graph convolution is proposed to model time varying spatial correlations.

• A multi-feature fusion is used to fuse the extracted spatial and temporal features.

Abstract

Exploiting deep spatial–temporal features for traffic prediction has become growing widespread. Accurate traffic prediction is still challenging due to the complex spatial dependencies and time varying temporal dependencies, especially for long-term prediction tasks. Existing studies usually employ pre-defined spatial graphs or learned fixed adjacency graphs and design models to capture spatial and temporal features. However, the pre-defined or fixed graph cannot accurately model the complex hidden structure. In this paper, a novel deep learning framework called Spatial–Temporal Multi-Feature Fusion Network (STMFFN) is proposed to address these challenges. Specifically, a multi-scale attention module with temporal convolution is designed to capture the temporal dependencies from different scales. Then, a gated graph convolution module is proposed, which constructs adaptive adjacency matrices, and integrates graph convolution and graph aggregation modules to capture spatial dependencies from different ranges. Moreover, a multi-feature fusion layer is presented to fuse the extracted spatial and temporal dependencies by obtaining the attention vectors of temporal and spatial features. Experimental results on real-world datasets show a consistent improvement of 6%–9% over state-of-the-art baselines.

Introduction

Traffic prediction is a classical spatial–temporal prediction problem, which aims to predict future traffic conditions (e.g. traffic volume or speed) in traffic networks based on historical observations. It has been found helpful in many real-world applications such as dynamic route planning, smart traffic management, and location-based applications (Li et al., 2021, Li et al., 2018).

Traffic prediction is commonly divided into short-term prediction ($\le 30$ minutes) and long-term prediction ($>30$ minutes). The methods for the time series such as Kalman filtering and autoregressive integrated moving average (ARIMA) (Williams & Hoel, 2003) perform well on prediction tasks. However, they fail to deal with complex non-linear spatial–temporal correlations. Recently, deep learning based methods have shown the ability to improve the accuracy of traffic prediction. Existing deep learning approaches often utilize recurrent neural network (RNN) (Song et al., 2016, Ye et al., 2022, Yu et al., 2017) such as gated recurrent unit (GRU) and long short-term memory model (LSTM) (Bai et al., 2023, Jaquart et al., 2021, Moreno et al., 2020, Ribeiro et al., 2021, Stefenon et al., 2022, Sun et al., 2023, Wang et al., 2022, Wu et al., 2022) to extract temporal features that can establish temporal correlations and conserve memory through gated control mechanisms and employ convolution neural network (CNN) (Lai et al., 2018, Zhang et al., 2017) to extract spatial features in the grid-based traffic network. However, traffic networks such as road networks are often graph-structured, these convolution based approaches may not obtain satisfactory results.

Traffic networks can be represented as graphs, in which traffic sensors are represented as nodes in a graph structure, and the graph neural network (GNN) can be employed to extract graph structure information (Tang et al., 2020, Wu et al., 2020, Zhao, Song, et al., 2019, Zheng et al., 2022). By incorporating GNN into a deep learning architecture, GNN based networks have significantly improved traffic prediction. Though GNN based prediction methods have achieved remarkable success, there are still some fundamental problems that need to be further addressed: (1) The spatial–temporal correlations of road segments are varying over time. As shown in Fig. 1, the state of node A is weakly influenced by the state of node B at $t_1$, while their mutual influence becomes stronger from $t_2$ to $t_3$. Existing studies usually utilize pre-defined or fixed graphs and overlook the fact that the graph structure needs to be updated adaptively, which cannot solve such complex correlations. (2) The correlations of short-term patterns can be different from that of long-term patterns. The correlations between road segments at different scales should be considered. Therefore, it is necessary to further explore graph neural networks for these problems.

In a word, accurate traffic prediction is still challenging due to the complex spatial dependencies and time-varying temporal dependencies, especially for long-term prediction tasks. Existing studies usually employ pre-defined spatial graphs or learned fixed adjacency graphs and design models to capture spatial and temporal features. However, the pre-defined or fixed graph cannot accurately model the complex hidden structure. To solve this problem, we propose a novel spatial–temporal multi-feature fusion network with a multi-scale attention mechanism.

This paper aims to address above two problems and proposes a novel deep learning framework named Spatial-Temporal Multi-Feature Fusion Network (STMFFN) for traffic forecasting. Specifically, a hierarchical architecture is designed to capture the multi-scale temporal correlations and dynamic spatial correlations simultaneously. Moreover, instead of maintaining a fixed graph structure, an adaptive adjacency matrix is constructed for each scale. Finally, the multi-feature fusion layer utilizes an attention-augmented mechanism to fuse the multi-scale spatial and temporal features. Compared with existing traffic prediction methods, our method focuses not only on capturing multi-scale temporal correlations for both long and short-term prediction but also on the influence of dynamic network topology for spatial-correlations extraction. In particular, our work has the following contributions:

• We propose a novel spatial–temporal fusion framework (STMFFN) to solve both long and short-term traffic prediction tasks. To the best of our knowledge, this is the first study on traffic prediction based on multi-scale attention with dilated convolution, which captures the dynamic features from different scales and offers a more accurate long-term prediction.

• We design a gated graph convolution module (Gated GCN) to model time-varying spatial correlations among nodes that have similar patterns in geographical locations. In particular, Gated GCN aggregates information not only from physically adjacent nodes but also from nodes that are similar in geographical locations.

• We propose a multi-feature fusion layer to fuse the extracted spatial and temporal features with an attention mechanism to further improve prediction performance.

• We empirically conduct experiments on two real-world datasets for traffic flow prediction tasks to evaluate STMFFN and show its superior performance.

The rest of the paper is organized as follows. Section 2 presents related work. Section 3 gives the problem definition. Section 4 elaborates on the proposed model. Section 5 presents an extensive experiment evaluation. Section 6 concludes the paper.