PaperlayerView exact revision on arXiv

Source-linked AI summary

Traffic Prediction using Artificial Intelligence: Review of Recent Advances and Emerging Opportunities

Maryam Shaygan, Collin Meese, Wanxin Li, Xiaolong Zhao, Mark Nejad

arXiv:2305.19591v2cs.LGcs.AI

TL;DR

Traffic prediction remains challenging because congestion has substantial costs and forecasting depends on complex spatial, temporal, nonlinear, and external-factor relationships. This survey synthesizes AI-based methodologies, with emphasis on multivariate traffic time series, covering data, preprocessing, models, applications, challenges, and future directions. Its supported conclusion is that deep learning has improved prediction accuracy and scale, but substantial work remains.

  • Problem

    Traffic forecasting remains difficult because traffic data involve complex external factors, nonlinearities, and inherent spatial and temporal dependencies, while non-recurrent events receive comparatively limited attention.

  • Method

    The paper conducts a broad survey of AI-based traffic prediction, organizing traffic data resources, preprocessing methods, prediction models, applications, research challenges, and future directions.

  • Results

    Deep learning models have made great strides in improving the accuracy and scale of traffic prediction methods.

  • Takeaways & Limitations

    The survey identifies emerging research opportunities involving more efficient models, broader data use, and traffic prediction for connected and autonomous vehicles.

  • Takeaways & Limitations

    Current traffic prediction research still faces unresolved model-selection and scalability challenges, including computational complexity and the need for lightweight architectures.

Abstract

from arXiv · show

Traffic prediction plays a crucial role in alleviating traffic congestion which represents a critical problem globally, resulting in negative consequences such as lost hours of additional travel time and increased fuel consumption. Integrating emerging technologies into transportation systems provides opportunities for improving traffic prediction significantly and brings about new research problems. In order to lay the foundation for understanding the open research challenges in traffic prediction, this survey aims to provide a comprehensive overview of traffic prediction methodologies. Specifically, we focus on the recent advances and emerging research opportunities in Artificial Intelligence (AI)-based traffic prediction methods, due to their recent success and potential in traffic prediction, with an emphasis on multivariate traffic time series modeling. We first provide a list and explanation of the various data types and resources used in the literature. Next, the essential data preprocessing methods within the traffic prediction context are categorized, and the prediction methods and applications are subsequently summarized. Lastly, we present primary research challenges in traffic prediction and discuss some directions for future research.

1. Introduction

Traffic congestion has widespread costs, while accurate prediction can improve infrastructure use and support more effective traffic management. This survey reviews AI-based traffic prediction, especially multivariate traffic time series, alongside data, preprocessing, methods, applications, challenges, and future directions.

  • 8.8 billion hours of additional travel time and 3.3 billion gallons of increased fuel consumption cost the United States about $179 billion in 2019.
  • Accurate traffic prediction can improve network capacity utilization, help alleviate congestion, and support traffic management centers and road operators.
  • Non-recurrent events account for about 50% of traffic congestion, yet most studies focus on recurrent prediction, especially during rush hours.
  • Emerging technologies such as autonomous vehicles, the Internet of Things, novel machine learning techniques, and secure connectivity may substantially affect traffic prediction and management.
  • Traffic prediction research must address complex multivariate time-series properties, including non-stationary relationships, spatial locality, and variability across multiple time frames.
  • The survey reviews AI-based methodologies by organizing traffic data resources, preprocessing methods, prediction models, applications, open challenges, and emerging technologies.

2. Prediction aspects assessed in previous reviews

Previous traffic prediction surveys cover varied subsets of methods, data, applications, and forecasting settings. Their fragmentation across technical areas and publication venues motivates a broader review connecting recent advances with practical applications and future directions.

  • Earlier surveys examined traditional short-term prediction, data-driven and model-driven approaches, machine learning methods, long-term forecasting, and traffic forecasting taxonomies.
  • Other reviews focused on classical and early deep learning models, transportation applications, congestion prediction, urban flow, robustness, and benchmark evaluation.
  • Several surveys organized public traffic datasets, reviewed deep learning methods, studied hybrid models, or examined traffic classification, generation, and forecasting.
  • Recent work also assessed deep learning for spatiotemporal data mining, caveats of using deep learning for every short-term application, and actionability gaps.
  • Existing surveys are generally specialized, do not always connect advances to applications and future directions, and are dispersed across disciplines and venues.

3. Traffic Data

Traffic data research spans diverse spatio-temporal sources, external information, resolutions, and preprocessing challenges. The survey categorizes these resources and highlights data-quality, coverage, temporal-resolution, and benchmarking issues affecting traffic prediction.

  • Data types and resources: The survey categorizes traffic prediction data into spatio-temporal sequence data, external data, and related public resources.It reviews available data types, resolutions, and preprocessing approaches for converting raw data into actionable information.
  • Spatio-temporal sequence data: Fixed-position sensors provide readily usable traffic variables but have limited spatial coverage and may experience outages.Common devices include inductive loops, cameras, radar, and LIDAR; widely used datasets include PeMS and METR-LA.
  • Spatio-temporal sequence data: Moving sensors capture detailed trajectories and road connections, but require sufficient penetration and preprocessing to address sampling and measurement errors.GPS, cellular, AVI, and floating-car data can cover locations with few fixed sensors, while fixed sensors avoid trajectory-location ambiguity.
  • External data: External data such as pandemic and urban-systems information can reflect changing traffic conditions, but collection and integration remain difficult.The survey identifies COVID-19 movement changes and correlations between utility usage and transportation patterns as relevant examples.
  • Temporal and spatial resolution: Data resolution must match the forecasting horizon and step size because longer horizons or larger steps generally reduce prediction accuracy.High-resolution data can support dynamic decisions but exhibits strong fluctuations and imbalance that challenge forecasting models.
  • Data preprocessing: Preprocessing must address map-matching errors, missing values, outliers, data imbalance, and cold-start conditions.Long-period missing data are especially difficult because limited surrounding context makes substitution inaccurate; no historical trajectories also hinder predictor construction.

4. Traffic Prediction Methods and Models

Traffic forecasting is formulated as spatio-temporal time-series prediction over one or more traffic variables. The reviewed methods increasingly emphasize multivariate spatial and temporal dependencies, deep learning, and hybrid model designs.

  • Problem formulation: Traffic forecasting predicts future traffic conditions from historical and current traffic variables across one or more time series.For a single series, the input sequence length determines how many historical time steps are used, while the prediction horizon specifies the forecast span.
  • Model categories: Traffic forecasting models are broadly classified as univariate, using one sensor series, or multivariate, using multiple network sensors.The survey focuses on multivariate traffic time-series modeling.
  • Spatio-temporal modeling: Precise traffic prediction depends on jointly modeling road-network topology and traffic dynamics over time.Recent approaches consider how historical trends at networked nodes relate to the future state of a given node.
  • Deep learning: Deep learning models have become popular because they can represent complex relationships and have outperformed many traditional approaches when sufficient data are available.Their effectiveness is conditioned on data availability.
  • Hybrid models: Hybrid approaches combine multiple models, often weighting and fusing their predictions to capture localized and network-wide relationships.These designs address differing model trade-offs through ensemble-style prediction fusion.
  • Survey organization: The survey reviews multivariate forecasting models according to their model-driven scientific contributions and distinctive approaches.This organization is intended to improve understanding of existing traffic forecasting methodologies.

4.1. Statistical Methods

Statistical methods offer transparent, computationally simple approaches for traffic forecasting, but their limited representation of spatial dependencies and strong modeling assumptions constrain broader traffic applications.

  • Statistical models frame forecasting as regression and include KNN, historical average, VAR, Bayesian networks, Kalman filtering, and ARIMA variants.
  • ARIMAX, KARIMA, and Vector ARIMA have been reported to improve traffic flow prediction accuracy.
  • These methods suit smaller datasets with few time series and short observation periods because of their simplified, transparent computational structure.
  • Statistical approaches generally focus on temporal data and lack sufficient complexity to model inherent spatial dependencies in traffic.
  • ST-ARIMA and VARMA extend time-series modeling to spatial and temporal interactions, but explicit parameterization and strong assumptions limit real-world suitability.
  • Hybrid statistical and deep-learning methods can incorporate exogenous variables, long-term dependencies, and feature selection, but the cited case study considers only temporal exogenous variables.

4.2. Traditional Machine Learning

Traditional machine-learning methods improve flexibility over classical models by learning complex traffic relationships, but their categories have distinct strengths and limitations in uncertainty and nonlinear generalization.

  • Compared with classical methods, machine-learning models provide more robust generalization and adapt better to changing traffic-network conditions.
  • Traditional machine-learning traffic prediction methods comprise feature-based models, Gaussian-process models, and state-space models.
  • Feature-based methods train regression models using human-engineered traffic features and are easy to implement.
  • State-space models capture latent data structures and system uncertainty naturally, which is desirable for traffic prediction applications.
  • State-space models have difficulty generalizing nonlinear relationships, limiting their suitability for some traffic prediction settings.

4.3. Deep Learning for Multivariate Traffic Prediction

Deep-learning models are increasingly used for multivariate traffic prediction because they capture nonlinear relationships, with basic architectures offering different feature-learning mechanisms and practical trade-offs.

  • Deep-learning approaches have demonstrated strong applicability to multivariate traffic prediction and nonlinear modeling.
  • MLP: MLPs are feed-forward networks of fully connected neurons trained with backpropagation, but they do not model local traffic dependencies effectively.
  • MLP: CNN-based traffic prediction outperformed MLP models in a reported comparison, while MLPs were more prone to noise.
  • AE: Autoencoders use an encoder-decoder bottleneck to learn compressed representations of unlabeled input data for imputation, compression, and feature extraction.
  • AE: Autoencoders require preprocessing and may involve pretraining, long training times, poorly understood representations, and application-specific tuning of hidden units.
  • RBM & DBN: DBNs can learn features unsupervised but struggle to map complex traffic-system relationships; integrating SVR was proposed to address this limitation.

4.3.2. CNN

CNN-based methods model localized spatial or temporal traffic correlations, while hybrid and modified convolutional architectures extend coverage to joint spatio-temporal, multiscale, and regional relationships.

  • CNN architecture: CNNs use convolution, ReLU, pooling, and fully connected layers to extract and map spatial or temporal features.
  • CNN architecture: Pooling reduces feature-map dimensionality and parameters by subsampling while retaining important information.
  • CNN limitations: Deeper CNNs capture higher-level information but add parameters that can increase overfitting and training time.
  • Traffic applications: CNN traffic models capture localized dependencies, but long-range interlink relationships require greater network depth and computational demand.
  • Hybrid models: Conv-LSTM combines CNN and LSTM sequentially to model short- and long-term spatio-temporal trends jointly, outperforming standalone LSTM and SVM models in a case study.
  • TCN and DeepGLO: TCNs use causal and dilated one-dimensional convolutions to expand the receptive field while processing historical samples.
  • TCN and DeepGLO: DeepGLO applies a TCN-based hybrid model to 228 traffic sensors and considers global network patterns during prediction.
  • Emerging convolutional models: Convolutional approaches show promise across spatial granularities, while newer hybrids include Conv-LSTM, TCN, and DeepGLO.

4.3.3. RNN

RNNs model multivariate traffic time series by processing inputs sequentially and maintaining hidden-state memory, while gated variants address long-term dependencies. Recent hybrid approaches add spatial feature extraction through KNN, GCN, and correlation-based selection, but results are difficult to compare across datasets.

  • RNN foundations: RNNs process sequential inputs while updating hidden states, providing a memory component suited to time series data.Conventional feedforward networks lack this memory component and perform poorly on sequences.
  • Gated RNNs: LSTM uses gated memory to selectively retain long-range dependencies, whereas GRU offers similar performance with fewer learnable parameters and more efficient training.The LSTM forget and output gates regulate cell state and visibility; GRU uses reset and update gates without a separate memory block.
  • Hybrid RNN models: KNN-LSTM extracts spatio-temporal correlations between sensing stations and weights predictions from the K nearest stations using a rank-exponent method.Its effectiveness depends heavily on similarity in the magnitude and distribution of station time series.
  • Hybrid RNN models: T-GCN combines graph convolution with GRU-inspired reset and update gates, outperforming standalone GCN and GRU while sometimes predicting poorly at local minima and maxima.Graph convolution extracts transportation-network spatial features before recurrent processing, but smoothing can obscure small changes.
  • Hybrid RNN models: STFSA selects temporally correlated history and six spatially correlated detectors before using GRU to predict future traffic flow.The method uses Pearson correlation and produced the best performance among the compared standalone GRU and CNN models.
  • Open limitations: Across RNN-based traffic prediction studies, differing datasets prevent direct cross-comparison, leaving the existence of an ideal MTP architecture unresolved.Backpropagation through time can also be expensive for high-dimensional series with several timesteps.

4.3.4. GCN

GCNs represent roadway networks as graphs to capture spatial relationships that Euclidean convolutions can distort, and later models jointly model spatial and temporal traffic dependencies. Extensions address directed diffusion, dynamic connectivity, multiscale regions, and temporal propagation, while original GCN assumptions limit applicability.

  • GCN foundations: GCNs perform convolution on structured graphs, matching roadway node-link relationships more directly than grid-based Euclidean convolution.Euclidean representations can distort meaningful connections, such as nearby but unconnected roadways.
  • GCN foundations: The original GCN simplified spectral graph convolution with a localized first-order approximation, reducing parameters and improving efficiency.Its undirected-graph and equal-locality assumptions do not represent directed traffic diffusion or differing roadway characteristics well.
  • Spatio-temporal extensions: STGCN jointly captures spatial and temporal trends without an RNN by placing a GCN layer between two gated, fully convolutional temporal layers.This ST-Conv block combines graph convolution with one-dimensional convolution.
  • Spatio-temporal extensions: DCRNN extends graph convolution to directed and undirected graphs by modeling traffic as diffusion and using forward and reverse graph processes.Diffusion convolution is integrated into GRU cells within an encoder-decoder architecture, with scheduled sampling for long-term forecasting.
  • Dynamic and multiscale graphs: Graph WaveNet dynamically learns an adjacency matrix and combines graph convolution with dilated causal one-dimensional convolution.Other approaches learn multiple adjacency matrices tied to specific past time intervals and encode road-segment features such as capacity and length.
  • Dynamic and multiscale graphs: HGCN models transportation networks at both micro and macro layers to capture localized connections and higher-level regional relationships.Ignoring macro regions can obscure important trends when interdependent regions are far apart.

4.3.5. WNN

WNN integrates wavelet transformations with artificial neural networks to model complex signal relationships for short-term traffic flow prediction. Recent variants improve optimization and convergence, but remain limited for non-recurrent events and poorly documented comparisons.

  • Architecture: WNN combines wavelet transformations and artificial neural networks, using a nonlinear wavelet basis function in its hidden layer.Scale and shift factors improve approximation accuracy.
  • Architecture: The hidden layer uses wavelet activation, while gradient descent adjusts connection weights and the wavelet’s shift and scale factors.The Morlet wavelet is generally selected because it extracts magnitude and phase information.
  • Applications: WNNs are attractive for traffic prediction because they provide nonlinear processing, self-organization, self-adaptation, and learning capabilities.Their use is motivated by similarities between traffic time series and signals.
  • Optimization: WNN with improved particle swarm optimization achieves the highest accuracy and fastest convergence compared with standard WNN and WNN using traditional PSO.The improved algorithm addresses stochastic-gradient-descent slow convergence and local optimality problems.
  • Limitations: The improved WNN approach is ineffective during non-recurrent network events and does not consider external factors such as weather and precipitation.These constraints limit its coverage beyond recurrent traffic-flow conditions.
  • Optimization: Alternative optimization methods such as improved PSO and harmony search improve convergence speed and optimal-solution reliability in reported case studies.However, the datasets in two studies are poorly documented, limiting cross-comparison of optimization approaches.

4.3.6. Attention-based

Attention-based methods enhance multivariate traffic prediction by dynamically weighting spatial and temporal information. Studies apply attention to recurrent, convolutional, graph, and encoder-decoder architectures, with performance depending on prediction horizon and dataset.

  • Attention with recurrent models: Attention mechanisms dynamically learn the relative importance of adjacent time series for high-dimensional traffic prediction.This improves spatio-temporal feature extraction and learning efficiency.
  • Attention with recurrent models: Clustering time series by similar distributions before training improves convergence time and prediction accuracy in an attention-based bi-directional LSTM approach.The model captures spatial and temporal traffic features in high-dimensional data.
  • Attention with recurrent models: Attention-based LSTM models learn the importance of input segments instead of compressing all inputs into a fixed-length representation.Reported experiments use different datasets and study areas, limiting direct cross-comparison.
  • Attention with convolutional models: ASTGCN combines graph and standard convolution with attention to capture spatio-temporal traffic trends and dynamically learn the GCN adjacency matrix.The architecture extends attention-based modeling beyond recurrent designs.
  • Attention with encoder-decoder models: GMAN uses transform attention between encoder and decoder to model relationships between future and previous time steps while limiting error propagation.It also stacks spatio-temporal attention blocks that dynamically weight vertex correlations and temporal influences.
  • Results: GMAN outperforms DCRNN and Graph Wavenet at 30- and 60-minute horizons but performs worse than Graph Wavenet for 15-minute predictions.The comparison indicates that attention-based performance varies with prediction horizon.

4.3.7. Limitations of the Deep Learning-based Methods

Deep learning methods for traffic prediction face limitations in transferability, computational efficiency, model selection, interpretability, and uncertainty quantification. The survey identifies ongoing research needs for scalable, generalizable, and interpretable systems.

  • Computational complexity and efficiency: Deep learning models require greater computing power because their increasingly complex structures raise training-resource demands.The survey calls for scalability analysis and lightweight architectures that reduce computational complexity while preserving performance.
  • Computational complexity and efficiency: Model pruning can reduce inference computation costs while maintaining good performance, but it does little to improve training efficiency.Using pruned representations for new sparse models could potentially improve training efficiency.
  • Optimal model choice: An appropriate traffic-prediction model must capture spatial dependency, temporal correlation, and spatio-temporal correlation.Model choice remains application- and data-specific.
  • Interpretability and uncertainty: Deep learning models have limited interpretability and often cannot quantify uncertainty for stochastic, noisy traffic data.Further work is needed because traffic forecasts support high-stakes decisions such as signal timing and route guidance.
  • Interpretability and uncertainty: A path-based bi-directional LSTM maps the deep learning network to the physical roadway network, but its generalizability to other applications remains unexplored.The survey calls for improved interpretability and methods that use explanatory insights to enhance predictive capabilities.
  • Limited transferability: Deep learning traffic predictors are generally network-specific because learned spatial relations depend on sensor-network topology.Transfer learning research remains limited, although recent approaches demonstrate transfer between selected transportation networks.
  • Limited transferability: A DCRNN transfer-learning approach generalizes across Los Angeles and San Francisco highway networks, but it excludes external data sources and focuses only on highways.RegionTrans also depends on matching a data-scarce city with a topologically similar data-rich city.

4.3.8. Deep Learning Evaluation Methods

Traffic-prediction evaluation commonly uses error metrics whose behavior depends on training-set size. The survey also emphasizes that training size and feature count must be considered alongside overfitting and underfitting.

  • Evaluation metrics: MSE, MAPE, and RMSE are among the most commonly used metrics for evaluating deep learning traffic-prediction accuracy.These metrics calculate prediction accuracy or error.
  • Training data: Prediction error generally decreases as more instances are added to the training set.The survey links larger training sets with more accurate prediction capabilities.
  • Generalization: Training-set size and feature count should be selected while considering both overfitting and underfitting.The survey identifies the overfitting problem in deep learning traffic prediction as requiring further research.
  • Generalization: Recent research challenges the assumption that increasing model parameters and training data necessarily produces poor test-data generalization.The survey notes that this relationship does not always hold in practice.

5. Common Traffic Prediction States and Associated Applications

Traffic prediction is organized around forecasting network metrics at different spatial scopes, from detector and link levels to regions and entire networks. AI and graph-based methods address complex spatio-temporal dependencies, but scalability, generalizability, data density, and computational efficiency remain challenges.

  • Traffic forecasting predicts future network conditions from one or more traffic-variable time series under dynamic spatio-temporal conditions.
  • Regional Flow Forecasting: Regional flow prediction supports urban planning, although regional modeling remains immature and requires additional external factors beyond weather and region-specific features.
  • Network Flow Forecasting: Network flow forecasting estimates flow through roadways and intersections, enabling traffic management, route guidance, and traveler information systems.
  • Network Flow Forecasting: Deep learning models increasingly combine fully connected, recurrent, convolutional, attention, and graph networks to capture spatial and temporal traffic features at larger scales.Examples include a 45-minute forecast across 33 detectors and an attention-based model covering 1,250 road segments.
  • Speed Forecasting: Traffic speed studies often use limited road-specific scopes, while TCHA dynamically captures correlations between target and nearby roads to improve prediction.
  • Travel Time Forecasting: Network-specific spatial features can generalize poorly across transportation networks, while probe-data sparsity complicates path inference for travel-time models.
  • OD Travel Time Forecasting: A GCN-GRU model for origin-destination travel time reduced MAPE error by upwards of 1.78% versus state-of-the-art baselines, but required considerable training and inference time.

6. Open Research Issues and Future Research Directions

The survey identifies open challenges spanning practical evaluation, distributed traffic prediction, anomaly response, multimodal trip planning, mixed traffic, data access, and benchmarking. Future work must address operational constraints, heterogeneous data, and realistic deployment conditions.

  • Existing studies often stop at predictive accuracy and do not assess real-world benefits for applications such as vehicle routing.
  • Traffic Prediction For Connected and Autonomous Vehicles: Connected-vehicle systems face latency, bandwidth, cloud-storage, and onboard energy constraints, motivating energy-efficient computation offloading across edge resources.
  • Federated Learning for Traffic Prediction: Federated learning preserves crowdsourced CAV-data privacy and reduces distributed-data transmission, but heterogeneous devices create skew, imbalance, missing classes, features, and values.
  • Distributed IoV Applications: Blockchain-based distributed IoV applications remain constrained by excessive energy consumption, low throughput, and high resource consumption, requiring improved scalability and efficiency.
  • Anomaly Traffic Event Detection and Mitigation: Rapid anomaly detection is needed because communicating updates and coordinating responses takes time, while deep learning anomaly classifiers remain relatively uncommon.
  • Anomaly Traffic Event Detection and Mitigation: Prediction under perturbation is difficult because rerouting changes travelers’ routes and can temporarily alter network topology, hindering existing route-guidance models.
  • Traffic Prediction in Mixed Traffic Transportation Networks: Integrating multimodal, time-dependent OD and path travel-time prediction with shortest-path algorithms could support complete-trip scheduling across walking, bicycling, transit, and automobile modes.
  • Traffic Prediction in Mixed Traffic Transportation Networks: Mixed traffic is underrepresented because interactions among human vehicles, autonomous vehicles, bicycles, pedestrians, and public transportation remain poorly understood.

7. Conclusion

The survey reviews traffic prediction data, preprocessing, methods, tasks, and applications, emphasizing multivariate traffic time-series modeling and outlining challenges and future directions.

  • Traffic forecasting remains challenging because of complex relationships, external factors, non-linearity, and inherent spatial and temporal dependencies.
  • The survey covers traffic data types and limitations, preprocessing, classical and current methods, prediction-task taxonomies, and multivariate time-series applications.
  • The paper concludes by discussing significant research challenges and future research directions.

Author Statement

The author statement lists investigation, original-draft writing, review and editing, and supervision responsibilities among the contributors.

  • Maryam Shaygan and Collin Meese contributed investigation and original-draft writing.
  • Wanxin Li and Xiaoliang Zhao contributed review and editing.
  • Mark Nejad contributed supervision, review, and editing.
Loading 2305.19591v2…