1. Introduction

The relentless march of urbanization has led to a surge in traffic congestion and related issues, demanding more sophisticated urban traffic management strategies. A crucial component of these approaches is the precise forecasting of traffic patterns across time and space, which is crucial for easing congestion, optimizing traffic patterns, and enhancing traffic efficiency [1]. Conventional traffic prediction techniques frequently overlook the intricate interactions between spatial and temporal aspects present in traffic datasets. However, the advent of deep learning has introduced a new horizon in traffic flow prediction, with models like DeepSTF [2] and so no leading the charge. These models have shown remarkable accuracy in processing multiple traffic time series and capturing both temporal dependencies and spatial correlations. As urban areas continue to grapple with complex traffic networks, the need for more refined traffic management strategies becomes increasingly apparent.

The deep learning paradigm shift has enabled the development of models that can process multiple traffic time series with remarkable accuracy, capturing both temporal dependencies and spatial correlations. For example, ST-ResNet leverages convolutional neural networks to decompose traffic data into components that model short-term, periodic, and trend-related traffic dynamics, providing a more nuanced view of traffic patterns [3]. This approach has shown to be particularly effective in urban areas with complex traffic networks, which often require more refined traffic management strategies.

Furthermore, incorporating sophisticated network designs—including those with attention mechanisms and graph neural networks—has demonstrated promise in enhancing the precision and adaptability of traffic flow forecasting models. These architectures can better capture the intricate patterns within traffic data, leading to more robust forecasting models. The work by Veličković on Graph Attention Networks for traffic flow forecasting is a notable example, highlighting the potential of attention mechanisms in capturing complex dependencies within traffic data [4].

In this paper, we review recent advances in deep learning-based spatio-temporal traffic flow prediction methods. We analyse the current state of these methods, their performance in different scenarios, and discuss future trends that may shape the field. As deep learning continues to evolve, its potential to transform traffic management is becoming increasingly apparent, promising improved traffic conditions and a deeper understanding of urban mobility systems.

2. Deep Learning-Based Spatial-Temporal Traffic Flow Prediction Methods

2.1. Convolution-Based Models

In spatio-temporal traffic forecasting, deep learning stands out as an effective technique for delivering precise and detailed forecasts. Notably, convolution-based models have demonstrated significant potential. They utilize CNNs to detect spatial correlations among various regions and effectively model traffic flow across different locations.

2.1.1. ST-ResNet

ST-ResNet, a deep learning model, tackles the challenges of predicting traffic flow across both space and time. The model decomposes spatial-temporal data into multiple sub-networks, each dedicated to modeling different aspects of the data. By utilizing a residual network framework, ST-ResNet addresses the challenges associated with training deep networks and enhances the model's prediction accuracy. Its core ideas include spatial-temporal decomposition, residual networks [5], convolutional layers [6], dynamic aggregation, and integration of external factors. The ST-ResNet model employs a three-tiered approach to the analysis of spatial-temporal data, comprising short-term, periodic, and trend sub-networks. This enables the model to discern the immediate, cyclical, and long-term temporal characteristics of traffic flow, respectively. The model employs convolutional layers to capture the spatial dependencies between different regions. It also dynamically aggregates the outputs of the three sub-networks, assigning different weights to different branches and regions to adapt to the different characteristics of traffic flow in different areas. By integrating external factors such as weather conditions into the final prediction, ST-ResNet further improves its accuracy. The model's high prediction accuracy, good generalization performance, and strong extensibility make it suitable for traffic flow prediction in different cities and for various types of traffic data.

Building upon the foundation laid by ST-ResNet, researchers have explored the potential of capturing multi-scale features within traffic data. Shi, Zhang, and Fang introduced the Multi-scale Spatial-Temporal Graph Convolutional Networks (MSTGCN), which captures traffic data's spatial-temporal correlations at various scales. This new approach significantly enhances the accuracy of traffic flow forecasting by leveraging the data's multi-scale features [7].

2.1.2. STGCN

The STGCN model [8], another prominent example in the convolution-based category, employs Graph Convolutional Neural Networks (GCNs) for spatial feature extraction and Gated Convolutional Networks for temporal feature analysis within traffic flow data [9][10]. These features are then fused through a Spatio-Temporal Convolutional Block (ST-Conv block) to model and predict the spatio-temporal relationships of traffic flow. The ST-Conv module, which is central to the model, comprises two layers of gated convolution and one layer of graph convolution. The role of the gated convolutional layers is to capture temporal characteristics, whereas the graph convolutional layer is tasked with capturing spatial characteristics. These features are then fused through residual connections and bottleneck strategies. The output of the ST-Conv block undergoes fully connected operations to obtain the final prediction results. STGCN's ability to utilize graph structures to model traffic networks and extract spatial features through graph convolution, combined with its gated convolutional operations on the temporal dimension to extract temporal features, makes it effective at capturing spatiotemporal dependencies. However, its complex model structure can lead to higher computational costs.

2.2. Models based on attention mechanism

2.2.1. GMAN

GMAN, which stands for Graph Multi-Attention Network, utilizes an encoder-decoder architecture to dynamically capture long-term dependency relationships within spatiotemporal sequences [11]. The model excels at identifying relationships across different moments and spatial points, efficiently seizing long-term dependencies vital for forecasting traffic patterns in environments prone to constant fluctuations.

In two real-world traffic forecasting experiments—forecasting traffic volume and speed—the GMAN model showed its superiority. In particular, for predictions one hour ahead, GMAN outperformed state-of-the-art methods by up to 4% improvement in the Mean Absolute Error (MAE) measure. The source code for GMAN is available for further investigation and application to traffic prediction tasks.

In a similar vein, the application of attention mechanisms in traffic prediction has been further explored by Veličković, Young, Hamill, and Fernandes. They proposed the use of Graph Attention Networks (GATs) for traffic flow forecasting, leveraging the model's ability to learn complex dependencies between different locations in a traffic network. This research highlights the potential of attention mechanisms in capturing the nuanced relationships within traffic data, which can be critical for accurate traffic flow forecasting.

2.2.2. AGCRN and STEGNN

The AGCRN (Adaptive Graph Convolutional Recurrent Network) model represents an advancement over traditional graph convolutions through the incorporation of an adaptive mechanism, which is integrated into a Laplacian current network to facilitate the capture of spatiotemporal correlations [12]. AGCRN utilizes an adaptive mechanism to dynamically adjust the weights of graph convolutions, better capturing the correlations between different spatial locations.

The STSGCN (Spatio-Temporal Synchronous Graph Convolutional Network) model captures complex localized spatiotemporal correlations through a spatiotemporal synchronous modeling mechanism [13]. STSGCN uses this mechanism to consider both temporal and spatial information, effectively capturing localized spatiotemporal correlations.

The STFGNN (Spatio-Temporal Fusion Graph Neural Networks) model incorporates both an STFGN module and a novel gated CNN module [14]. The model is capable of capturing hidden spatial dependency relationships through the utilisation of data-driven graphs, which are then subjected to further fusion with the given spatial graphs. The STFGNN employs data-driven graphs to identify and extract hidden spatial dependency relationships, utilising a gated CNN module to facilitate the extraction of spatiotemporal features.

3. Model comparison

3.1. Comparison of Strengths and Weaknesses

When evaluating these methods, we've noticed that ST-ResNet excels at capturing spatial dependencies, while GMAN is particularly good at capturing long-term dependencies. STGCN is effective at capturing spatiotemporal dependencies, but its complex model structure can lead to higher computational costs. In contrast, models like AGCRN, STSGCN, and STFGNN are better at handling spatially structured data with heterogeneity, although their training times are generally longer.

3.2. Comparison of Suitable Scenarios

ST-ResNet and STGCN are suitable for traffic flow prediction scenarios with complex spatiotemporal dependencies, especially when it is necessary to consider both spatial and temporal factors at the same time. GMAN is suitable for scenarios with variable traffic flow predictions, particularly when capturing long-term dependencies is crucial. Models such as AGCRN, STSGCN, and STFGNN are more suitable for traffic flow prediction scenarios with spatially structured data that exhibit high heterogeneity, especially when the data's heterogeneity is significant.

4. Future Trends

The field of spatiotemporal traffic prediction is poised for significant advancements, with future research expected to focus on refining the granularity of spatiotemporal decompositions. This could involve breaking down time into finer increments such as hours and minutes, or segmenting space into more detailed geographic units. Such refinements aim to enhance the precision of traffic flow forecasts.

In addition, the development of more efficient network architectures is an important area of research. Innovative approaches, such as the incorporation of attention mechanisms and graph neural networks, show promise for improving the predictive accuracy and generalisation capabilities of model. These advanced architectures can better capture the intricate patterns within traffic data, leading to more robust forecasting models.

Furthermore, the integration of comprehensive fusion mechanisms is another frontier in traffic prediction. By amalgamating diverse data sources—ranging from traffic and meteorological data to social media information—models can achieve higher prediction accuracy [15]. The synergy of data from multiple sources provides a more holistic view of traffic patterns, which is essential for developing accurate and responsive traffic management strategies.

5. Conclusion

Forecasting traffic patterns over space and time is vital for traffic management, helping to alleviate congestion, streamline flow, and enhance efficiency. This review has synthesized the current state of deep learning-based spatiotemporal traffic prediction methods and analyzed their development trends. As deep learning technology continues to evolve, it is anticipated that spatiotemporal traffic prediction methods will achieve further breakthroughs, offering more effective support for urban traffic management. The integration of sophisticated algorithms with expansive data sources will undoubtedly pave the way for a new era of traffic prediction, one that is characterized by unprecedented accuracy and adaptability.