Open Spatio-Temporal Foundation Models for Traffic Prediction

In intelligent transportation systems (ITS), various detectors, such as the Sydney coordinated adaptive traffic system (SCATS), the global position system (GPS), and the microwave detector, are deployed in the urban road network. Traffic data detected from these detectors are generally heterogeneous. In this paper, the heterogeneous multi-sensor traffic data will be integrated and formally represented to provide a data resource to be effectively shared by various ITS applications. In this framework, the directive road segment is expressed as the basic element, and a three-dimensional spatio-temporal data domain is established to formalize the data representation. In the data domain, the heterogeneous traffic data are transformed to extract the spatio-temporal information as well as the traffic state features, such as speed, flux, and queuing time, therefore constituting a vector as the basic data element of the domain. Moreover, as an attachment to the data domain, a related sparse matrix is constructed to further demonstrate the relations among different directive road segments. As a case, the traffic data detected by SCATS and GPS are respectively transformed into flux of road segments, and the flux of next time is predicted by combining the related sparse matrix and the flux matrix of previous time.

Towards explainable traffic flow prediction with large language models

With the exponential increase in the urban population, urban transportation systems are confronted with numerous challenges. Traffic congestion is common, traffic accidents happen frequently, and traffic environments are deteriorating. To alleviate these issues and improve the efficiency of urban transportation, accurate traffic forecasting is crucial. In this study, we aim to provide a comprehensive overview of the overall architecture of traffic forecasting, covering aspects such as traffic data analysis, traffic data modeling, and traffic forecasting applications. We begin by introducing existing traffic forecasting surveys and preliminaries. Next, we delve into traffic data analysis from traffic data collection, traffic data formats, and traffic data characteristics. Additionally, we summarize traffic data modeling from spatial representation, temporal representation, and spatio-temporal representation. Furthermore, we discuss the application of traffic forecasting, including traffic flow forecasting, traffic speed forecasting, traffic demand forecasting, and other hybrid traffic forecasting. To support future research in this field, we also provide information on open datasets, source resources, challenges, and potential research directions. As far as we know, this paper represents the first comprehensive survey that focuses specifically on the overall architecture of traffic forecasting.

The increasing complexity of urban transportation systems and the growing volume of vehicles have made traffic congestion a persistent challenge in modern cities. Efficient traffic flow prediction is essential for mitigating congestion, improving road safety, optimizing traffic signal control, and enhancing overall transportation efficiency. In recent years, artificial intelligence (AI) has emerged as a transformative tool in the field of traffic management, offering sophisticated algorithms capable of modeling, analyzing, and predicting complex traffic patterns with high accuracy. The application of AI in traffic flow prediction leverages vast amounts of real-time and historical data to generate precise forecasts, supporting data-driven decision-making by urban planners and traffic control authorities. The prediction of traffic flow involves analyzing time-series data that exhibit nonlinear, dynamic, and often stochastic behavior. Traditional statistical models, such as autoregressive integrated moving average (ARIMA), have proven to be limited in handling the high dimensionality and variability inherent in traffic systems. In contrast, AI algorithms possess the capacity to learn and adapt from complex data inputs without the need for explicit programming, making them particularly suitable for traffic-related applications. AI algorithms used in traffic flow prediction can be broadly categorized into machine learning (ML) and deep learning (DL) approaches. Machine learning algorithms such as k-nearest neighbors (KNN), support vector machines (SVM), decision trees, and random forests have demonstrated effectiveness in short-term traffic prediction tasks. These algorithms are capable of identifying hidden patterns in traffic data and adjusting to changes in traffic behavior over time. Ensemble methods, which combine the strengths of multiple learning models, further enhance prediction accuracy and robustness. Deep learning algorithms, a subfield of AI inspired by the human brain’s neural architecture, have shown exceptional performance in capturing spatial-temporal dependencies in traffic data. Recurrent neural networks (RNNs), particularly long short-term memory (LSTM) networks and gated recurrent units (GRUs), are widely used for their ability to process sequential data and retain information over extended time intervals. Convolutional neural networks (CNNs) are employed to extract spatial features from traffic sensor data or road network imagery. Hybrid models that integrate CNNs with RNNs have achieved high levels of predictive precision by simultaneously learning spatial and temporal correlations. In addition to supervised learning methods, unsupervised and reinforcement learning techniques are also applied in traffic flow prediction. Clustering algorithms, such as k-means and DBSCAN, assist in identifying traffic patterns, while reinforcement learning models optimize adaptive traffic signal control systems by learning optimal actions through environmental interaction. This study explores the different types of AI algorithms used in traffic flow prediction, examining their theoretical foundations, structural differences, and practical applications. It aims to evaluate the comparative advantages of various algorithms in addressing the challenges of real-time traffic prediction in increasingly complex transportation networks. Keywords: Machine Learning, Deep Learning, Neural Networks, Regression Models, Reinforcement Learning

Dynamic multi-scale spatial–temporal graph convolutional network for traffic flow prediction

Traffic prediction is an elemental function of Intelligent Transportation Systems, and accurate and timely prediction is of great significance to both traffic management agencies and individual drivers. With the development of deep learning and big data, deep neural networks (DNN) achieve superior performances in traffic prediction. Developing DNN prediction models needs large scale and diverse data, however, it is costly to collect large volume of accurate traffic data. In this paper, we propose to use small volume of real traffic data and large volume of synthetic traffic data to developing traffic prediction models. The evolving of parallel system paradigm for traffic prediction and the algorithm to incrementally train traffic data generation models and traffic prediction models are presented. We use an improved generative adversarial networks to generate traffic data, and a stacked long short-term memory model for traffic prediction. Experimental results on a real traffic dataset demonstrate that our method can significantly improve the performance of traffic flow prediction.

Long-term traffic flow forecasting is crucial for intelligent transportation systems and also helps in the dynamic management of traffic flow. The goal of this research is to estimate future road traffic conditions using neural network models to reduce road congestion. Reducing prediction error is the most difficult task in traffic prediction. To predict traffic flow, this paper uses traffic data from Paris, France road dataset with normalized preprocessing of time series data. In this work, a novel CNN-LSTM hybrid model is built using MATLAB to predict traffic flow. The preprocessed data is used to train the CNN model, LSTM model and hybrid CNN-LSTM model. Finally, the predicted traffic flow data is evaluated using three evaluation indicators: coefficient of determination (R2), mean absolute error (MAE) and root mean square error (RMSE). The results show that the proposed hybrid CNN-LSTM model can well capture and learn the long-term temporal characteristics of traffic flow and can predict traffic flow quickly and effectively.

Traffic flow prediction is an essential task in the intelligent transportation system (ITS). However, extracting spatial and temporal features from complex traffic data is a critical challenge. Some deep learning-based methods have been applied to perform the traffic flow prediction problem, including convolutional neural network and recurrent network. Recently, the graph-based network has received a lot of attention from researchers, which could model the complex road network as a graph structure. With the rapid growth of graph neural network, it has become a state-of-the-art technique to perform the traffic prediction tasks. However, these deep learning-based traffic prediction approaches require a huge amount of traffic data, while many road segments suffer from the lack of historical traffic data. Meanwhile, some traffic data are missing due to unexpected issues such as malfunction of detectors. The severe data paucity problems will greatly degrade the performance of prediction. To address this challenge, this paper proposes a domain-adversarial-based neural network for traffic flow prediction. Contrast experiments have been conducted on several open-source traffic datasets, which have well demonstrated that the domain adaptation techniques could transfer the learned features from a source road network to a target network with data paucity problem. It also illustrates the effectiveness of our proposed method and shows the importance of domain adaptation method in traffic prediction problem.

Large and expanding cities suffer from a traffic congestion problem that harms the environment, travelers, and the economy. This paper aims to predict short term traffic congestion on a road section of expressway in Delhi city. For this purpose, we first propose a traffic congestion index based on traffic speed and flow. Clustering techniques and the Greenshield’s model were used for the derivation of the congestion index. Using this congestion index, congested time intervals of each day and each location of a weekday were identified. This study also introduces a feature series long short-term memory neural network (FSLSTMNN), which links a long short-term memory (LSTM) layer to each feature. It is trained using the many heterogeneous traffic features data collected in Delhi city for the next five minutes of traffic flow and speed prediction. FSLSTMNN achieved the good capability to learn feature series data. We also trained several traditional and deep-learning models using the same traffic data. The FSLSTMNN reduces mean absolute error 12.90% and 17.13%, respectively, in speed and traffic flow prediction compared to the second good-performance long short-term memory neural network (LSTMNN). Finally, traffic congestion is predicted classwise (light, medium, and congested) using the developed congestion index and traffic speed and flow predicted by the FSLSTMNN. Predicted results are consistent with the measured field data. Study results confirm that the developed congestion index and FSLSTMNN can be used successfully to predict traffic congestion.

ABSTRACTProviding reliable travel time prediction is very much needed for commuters for their upcoming trips to reduce travel time and relieve traffic congestion. This article proposes an integrated model for path and multi-step-ahead travel time prediction on freeways using both historical and real-time heterogeneous traffic and weather data. The model's performance is investigated in a case study under various traffic scenarios. Results indicate that the proposed model provides satisfactory prediction results in various performance tests. For practical purposes, general guidelines for selecting the model's parameter sets as well as the efficient size of historical data are also presented.

Reliable traffic prediction is critical to improve safety, stability, and efficiency of intelligent transportation systems. However, traffic prediction is a very challenging problem because traffic data are a typical type of spatio-temporal data, which simultaneously shows correlation and heterogeneity both in space and time. Most existing works can only capture the partial properties of traffic data and even assume that the effect of correlation on traffic prediction is globally invariable, resulting in inadequate modeling and unsatisfactory prediction performance. In this paper, we propose a novel end-to-end deep learning model, called ST-3DNet, for traffic raster data prediction. ST-3DNet introduces 3D convolutions to automatically capture the correlations of traffic data in both spatial and temporal dimensions. A novel recalibration (Rc) block is proposed to explicitly quantify the difference of the contributions of the correlations in space. Considering two kinds of temporal properties of traffic data, i.e., local patterns and long-term patterns, ST-3DNet employs two components consisting of 3D convolutions and Rc blocks to, respectively, model the two kinds of patterns and then aggregates them together in a weighted way for the final prediction. The experiments on several real-world traffic datasets, viz., traffic congestion data and crowd flows data, demonstrate that our ST-3DNet outperforms the state-of-the-art baselines.

Timely, effective, and accurate traffic movement prediction is highly important in developing and implementing the intelligent traffic prediction system. Given the overwhelming traffic in recent years, traffic data have been increasing as we enter the era of high traffic data. Existing traffic prediction methods use the traditional prediction models, which remain unsatisfactory for real-world applications. Sharp non-linearities, such as free flow, breakdown, recovery, and free congestion cause challenges in forecasting the traffic flow. This condition motivates us to reconsider the traffic forecast model based on the deep learning with the high traffic data. To predict non-linearities spatio-temporal effects, we have employed the deep learning technique with non-parametric regression. The first layer in the deep learning algorithm identifies spatio-temporal relationships between predictors and other non-linear relationships. The deep learning technique with non-parametric regression is significantly better compared with other models. Experimental results show that the proposed technique for the traffic flow forecast has a better-quality performance.

Traffic prediction is an important part of modern intelligent transportation systems (ITS), which helps transportation management and city planning. However, it is a very challenging task for modeling complex spatio-temporal dependencies, since the traffic data belongs to highly periodic multivariate time series which makes it hard to model accurate spatial dependencies only from time series and observed geolocation information of road segments. The existing research mainly focuses on finding ways of capturing dynamic spatial dependencies of road segments while neglecting the importance of periodicity, and few studies have explored a pure embedding-driven method that is robust to corrupted data to model periodicity. In this paper, we propose an embedding-driven multi-hop spatio-temporal attention network for traffic prediction (), which mainly focuses on leveraging the multi-scale periodicity of traffic data. Specifically, the proposed network applies a designed Fourier-series-based embedding, to capture the periodicity, which is more in line with real-world facts. Driven by the designed embedding, both local and global temporal dependencies are modeled properly by combining the attention-based methods and the convolution-based methods. Besides, we implement a trial that can hardly be seen in the existing traffic prediction works to combine the graph self-attention mechanism with a multi-hop diffusion process to explore the large-scale structural information on a designed set of graphs. Experiments on two real-world traffic datasets which contains traffic speed data for months show the effectiveness of our proposed methods. The experiments also suggest the methods can provide stable reasonable and smooth predictions for completely corrupted data.

Urban traffic status (e.g., traffic speed and volume) is highly dynamic in nature, namely, varying across space and evolving over time. Thus, predicting such traffic dynamics is of great importance to urban development and transportation management. However, it is very challenging to solve this problem due to spatial-temporal dependencies and traffic uncertainties. In this article, we solve the traffic dynamics prediction problem from Bayesian meta-learning perspective and propose a novel continuous spatial-temporal meta-learner (cST-ML), which is trained on a distribution of traffic prediction tasks segmented by historical traffic data with the goal of learning a strategy that can be quickly adapted to related but unseen traffic prediction tasks. cST-ML tackles the traffic dynamics prediction challenges by advancing the Bayesian black-box meta-learning framework through the following new points: (1) cST-ML captures the dynamics of traffic prediction tasks using variational inference, and to better capture the temporal uncertainties within tasks, cST-ML performs as a rolling window within each task; (2) cST-ML has novel designs in architecture, where CNN and LSTM are embedded to capture the spatial-temporal dependencies between traffic status and traffic-related features; (3) novel training and testing algorithms for cST-ML are designed. We also conduct experiments on two real-world traffic datasets (taxi inflow and traffic speed) to evaluate our proposed cST-ML. The experimental results verify that cST-ML can significantly improve the urban traffic prediction performance and outperform all baseline models especially when obvious traffic dynamics and temporal uncertainties are presented.

Urban traffic status (e.g., traffic speed and volume) is highly dynamic in nature, namely, varying across space and evolving over time. Thus, predicting such traffic dynamics is of great importance to urban development and transportation management. However, it is very challenging to solve this problem due to spatial-temporal dependencies and traffic uncertainties. In this paper, we solve the traffic dynamics prediction problem from Bayesian meta-learning perspective and propose a novel continuous spatial-temporal meta-learner (cST-ML), which is trained on a distribution of traffic prediction tasks segmented by historical traffic data with the goal of learning a strategy that can be quickly adapted to related but unseen traffic prediction tasks. cST-ML tackles the traffic dynamics prediction challenges by advancing the Bayesian black-box meta-learning framework through the following new points: 1) cST-ML captures the dynamics of traffic prediction tasks using variational inference; 2) cST-ML has novel designs in architecture, where CNN and LSTM are embedded to capture the spatial-temporal dependencies between traffic status and traffic related features; 3) novel training and testing algorithms for cST-ML are designed. We also conduct experiments on two real-world traffic datasets (taxi inflow and traffic speed) to evaluate our proposed cST-ML. The experimental results verify that cST-ML can significantly improve the urban traffic prediction performance and outperform all baseline models.