Spatial–temporal multi-feature fusion network for long short-term traffic prediction
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 ( minutes) and long-term prediction ( 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 , while their mutual influence becomes stronger from to . 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.
Section snippets
Related work
This section reviews related work from three aspects: traffic prediction, spatial–temporal graph neural networks and attention mechanism.
Problem definition
A traffic network is represented as a graph , where is the set of nodes, is the set of edges between two nodes, and is the adjacency matrix that reflects the relationship between two nodes. In real applications, a node usually denotes a sensor located at the road segment. Traffic prediction is a typically spatial–temporal prediction problem, this paper focuses on predicting traffic speed. The other features such as traffic volume and density can be treated
Methodology
This section details the STMFFN framework and all components proposed in Fig. 2.
Experiments
In this section, extensive experiments are conducted to verify the superiority of the proposed model.
Conclusion
In this paper, a new novel framework was proposed to solve the traffic prediction problem. In particular, multiple adaptive graph structures were constructed to model varying correlations between road segments. Moreover, multi-scale temporal correlations were captured simultaneously. Finally, a multi-feature fusion layer integrated the spatial and temporal features with an attention mechanism to enhance prediction performance. The extensive experiments on two real-world datasets demonstrated
CRediT authorship contribution statement
Yan Wang: Conceptualization, Methodology, Software, Validation, Writing – original draft, Formal analysis, Investigation, Data curation, Writing – review & editing, Visualization. Qianqian Ren: Writing – review & editing, Supervision, Project administration. Jinbao Li: Writing – review & editing, Resources.
Declaration of Competing Interest
The authors declare that they have no known competing financial interests or personal relationships that could have appeared to influence the work reported in this paper.
Acknowledgments
This work was supported by National Key R&D Program of China under Grant No. 2020YFB1710200, the innovative research projects for graduate students of Heilongjiang University YJSCX2022-231HLJU, the China Postdoctoral Science Foundation under Grant No. 2022M711088, the National Natural Science Foundation of China under Grant No. 62172243.
References (46)
Performance evaluation of density-based clustering methods
Information Sciences
(2009)- et al.
A dynamic-inner LSTM prediction method for key alarm variables forecasting in chemical process
Chinese Journal of Chemical Engineering
(2023) - et al.
Short-term bitcoin market prediction via machine learning
The Journal of Finance and Data Science
(2021) - et al.
Time series forecasting using ensemble learning methods for emergency prevention in hydroelectric power plants with dam
Electric Power Systems Research
(2022) - et al.
Short-term performance degradation prediction of a commercial vehicle fuel cell system based on CNN and LSTM hybrid neural network
International Journal of Hydrocarbon Engineering
(2023) - et al.
Aggregate investor attention and bitcoin return: The long short-term memory networks perspective
Finance Research Letters
(2022) - et al.
Dynamic-LSTM hybrid models to improve seasonal drought predictions over China
Journal of Hydrology
(2022) - et al.
A dynamic spatial-temporal deep learning framework for traffic speed prediction on large-scale road networks
Expert Systems with Applications
(2022) - et al.
A3T-GCN: Attention temporal graph convolutional network for traffic forecasting
ISPRS International Journal of Geo-Information
(2021) - et al.
A graph convolutional stacked bidirectional unidirectional-LSTM neural network for metro ridership prediction
IEEE Transactions on Intelligent Transportation Systems
(2021)
Learning dynamics and heterogeneity of spatial-temporal graph data for traffic forecasting
IEEE Transactions on Knowledge and Data Engineering
A hybrid GLM model for predicting citywide spatio-temporal metro passenger flow
ISPRS International Journal of Geographical Information Science
Link traffic speed forecasting using convolutional attention-based gated recurrent unit
Applied Intelligence
Adaptive spatial-temporal graph attention networks for traffic flow forecasting
Applied Intelligence
Modeling long-and short-term temporal patterns with deep neural networks
Dynamic graph convolutional recurrent network for traffic prediction: Benchmark and solution
Traffic flow prediction with big data: A deep learning approach
IEEE Transactions on Intelligent Transportation Systems
Multi-step wind speed forecasting based on hybrid multi-stage decomposition model and long short-term memory neural network
Energy Conversion and Management
Cited by (14)
Modeling vehicle U-turning behavior near intersections: A deep learning approach based on TCN and multi-head attention
2024, Expert Systems with ApplicationsSpatial-temporal graph convolution network model with traffic fundamental diagram information informed for network traffic flow prediction
2024, Expert Systems with ApplicationsAn adaptive deep multi-task learning approach for citywide travel time collaborative estimation
2024, Expert Systems with ApplicationsSTGWN: Enhanced spatiotemporal wave forecasting using multiscale features
2024, Applied Ocean ResearchA decentralized federated learning-based spatial–temporal model for freight traffic speed forecasting
2024, Expert Systems with ApplicationsGSTC-Unet: A U-shaped multi-scaled spatiotemporal graph convolutional network with channel self-attention mechanism for traffic flow forecasting
2023, Expert Systems with Applications