Jul 2026· Frontiers in Computing and Intelligent Systems· Vol 17, pp. 62-67· 0 citations· 9 references
TL;DR
A multimodal spatiotemporal deep learning model fusing Multi-view Graph Convolutional Network, Transformer and Temporal Convolutional Network is proposed to realize binary traffic risk prediction at the 10×10 grid level to meet the actual business needs of urban traffic risk prediction.
Abstract
Aiming at the problem that single-modal features are difficult to depict the complex spatiotemporal evolution law in urban traffic risk prediction, a multimodal spatiotemporal deep learning model fusing Multi-view Graph Convolutional Network (GCN), Transformer and Temporal Convolutional Network (TCN) is proposed to realize binary traffic risk prediction at the 10×10 grid level. Based on the traffic dataset of Area C in New York City (NYC), the study integrates grid time series, static features and multi-type adjacency matrix data. Strategies such as outlier processing, dynamic threshold binarization and feature normalization are adopted to solve the problems of data quality and imbalance between positive and negative samples. Meanwhile, gradient clipping, weighted loss and early stopping strategies are combined to ensure the stability of model training. Experimental results show that the model achieves an F1 score of 0.745, an AUC-ROC of 0.798 and an AUC-PR of 0.756 on the test set. It can effectively capture the spatial correlation and temporal dependence features of traffic risks, maintain good prediction performance in the unbalanced data scenario, and meet the actual business needs of urban traffic risk prediction.
Traffic flow prediction is of great significance for improving the operation efficiency of the transportation system, optimizing travel experience and reducing traffic congestion. Traditional traffic flow prediction methods are difficult to capture the spatio-temporal nonlinear characteristics of traffic flow due to its simple model and insufficient feature extraction ability. Therefore, an intelligent traffic flow prediction system based on deep learning is proposed, constructs a deep learning model based on graph convolution and fusion of attention mechanism LSTM. Based on this, a traffic flow prediction system is implemented. Experiments show that, on the PeMSD4 and PeMSD4 datasets, the error of the model in RMSE and Mae indicators is significantly reduced compared with the traditional methods, which provides an efficient solution for traffic flow prediction and congestion analysis, and has both theoretical innovation and engineering practical value.
Zhan Tang, Xiaoyu Lu, N. Yang et al.· SAE technical paper series· 0 citations
A prediction model that incorporates multiple attention mechanisms with spatiotemporal graph convolutional networks (HASTGCN) and designs a spatiotemporal map convolution module to collaboratively model the dynamic spatiotemporal connection of traffic flow collaboratively model is used.
Chu-xia Chen· Proceedings of the 3rd Inter...· 0 citations
Accurate short-term traffic prediction is a critical component of intelligent transportation systems (ITS), yet it remains challenging due to nonlinear temporal dynamics, evolving spatial dependencies, and uncertainty in real-time urban traffic data. This paper proposes a novel uncertainty-aware deep ensemble spatiotemporal forecasting framework integrating Dynamic Graph Convolutional Networks (DGCN), Temporal Transformers, and CNN–LSTM hybrid models. A confidence-guided ensemble fusion strategy dynamically weights individual predictions using Bayesian uncertainty estimation. Experiments conducted on real-time Bhopal city traffic data demonstrate significant improvements over state-of-the-art baselines, achieving up to 90% performance gains during peak and abnormal traffic conditions.
Accurate forecasting of vessel traffic flow (VTF) is essential for modern maritime and port management, as it improves route-planning efficiency, reduces congestion and collision risks, and optimizes port operations. This study proposes a novel deep learning framework, namely, the Bidimensional Empirical Mode Decomposition–Nocal Convolutional Neural Network–Transformer (BEMD–NocalCNN–Transformer), for high-precision VTF prediction. The proposed framework first applies the BEMD algorithm to decompose the original time-series data into high- and low-frequency components. The NocalCNN module is then employed to extract spatial features from each component, while the Transformer module captures temporal dependencies and predicts future traffic-flow trends. The final predictions are obtained by aggregating the outputs of the high- and low-frequency components. Sensitivity analyses are conducted on key parameters, including input sequence length, learning rate, number of iterations, and convolution kernel size, to optimize the model configuration. To comprehensively evaluate the proposed framework, SVM, BPNN, RNN, LSTM, GRU, Transformer, WVMA-LSTM, and NocalCNN–Transformer were implemented and evaluated using the same CFD and Wuhan datasets, data preprocessing procedures, training–testing partitions, prediction settings, and evaluation metrics. The experimental results demonstrate that the proposed model outperforms the benchmark models and achieves substantially lower prediction errors for both the Caofeidian Promontory (CFD) and Wuhan waterways. These findings demonstrate consistent prediction performance of the proposed framework and provide a robust technical foundation for intelligent maritime traffic management and port operation optimization.
Chao Zhang, Bi-Yu Chen, Zehao Yuan et al.· Journal of Marine Science an...· 0 citations
Traffic accidents remain a persistent challenge for urban safety, particularly in Medan City, Indonesia, where heterogeneous road networks, mixed traffic flows, and varying environmental conditions complicate risk prediction. Existing statistical and classical machine learning methods often fail to capture the nonlinear spatial–temporal dependencies inherent in such environments, creating a gap in accurate, context-specific accident risk forecasting for developing countries. This study addresses this gap by proposing a Hybrid CNN–LSTM Deep Learning Model that jointly learns spatial features from georeferenced accident maps, road networks, and traffic density heatmaps, and temporal dependencies from historical accident logs, GPS traces, meteorological data, and road surface conditions. Unlike prior models, the proposed framework is explicitly tailored to localized urban traffic patterns in Medan, enabling robust performance under heterogeneous and data-imbalanced conditions. Data preprocessing included cleaning, normalization, and categorical encoding, followed by model training with an Adam optimizer and tuned hyperparameters. Experimental evaluation against baseline models—pure CNN, pure LSTM, and Random Forest—demonstrated statistically significant improvements (p < 0.05), with the hybrid CNN–LSTM achieving an accuracy of 96.8% (95% CI: 96.5–97.1%), precision of 96.5%, recall of 96.7%, and F1-score of 96.6%, outperforming baselines by up to 5% in predictive accuracy. The model effectively identified high-risk spatial clusters and peak accident periods, offering actionable intelligence for targeted safety interventions. These findings highlight the model’s potential for integration into intelligent transportation systems to support real-time monitoring, proactive policymaking, and enhanced urban traffic safety management.
Rusmin Saragih, Suria Alamsyah Putra, Togu Harlen Lbn Rajab et al.· JOIV: International Journal...· 0 citations
This study proposes a deep-learning-based approach for short-term traffic-state classification using real-world traffic data collected during 2022 at the Alésia intersection in Paris, and demonstrates that recurrent architectures substantially outperform the ANN baseline, highlighting the importance of temporal dependencies in traffic-state classification.
Chaymae Chouiekh, Ali Yahyaouy, M. A. Sabri et al.· Vehicles· 0 citations