21 March 2024 in IEEE Transactions on Intelligent Transportation Systems
Traffic congestion is a pressing issue worldwide, and machine learning (ML) methods are increasingly being used in Intelligent Transportation Systems (ITS) to address this problem. Deep hybrid models, in particular, have emerged as an efficient solution for traffic flow prediction. Among these models, Long Short-Term Memory (LSTM) and Bidirectional LSTM (Bi-LSTM) have been widely used to capture the temporal and periodic features of traffic data. The advancements in RNNs provide an opportunity to enhance the performance of existing models. Therefore, this work proposes a BiGRU-BiGRU model with two modules to extract temporal and periodic features from traffic data. Recurrent Neural Networks (RNNs) have proven to perform well using attention mechanism. However, there is a need for attention mechanism that strictly focuses on traffic dynamics and nearby data from the prediction points. Thus, a novel confined attention mechanism is proposed and incorporated into the first module to improve the model’s performance by focusing only on the recent relevant information in the traffic flow sequence. Furthermore, the external features are integrated to improve the model’s prediction performance. The proposed model is evaluated on the publicly available real-world dataset and compared with several baseline state-of-the-art methods. As an outcome, the model offers a reduction in the average value of RMSE, MAE, and MAPE for all the prediction horizons, that is ranged from 5.1% – 20.4%, 7.3% – 27.3%, and 6.1% – 56.6%, respectively.
13 July 2024 in Engineering Applications of Artificial Intelligence
12 June 2024 in IEEE Transactions on Computational Social Systems
Traffic flow prediction is essential for alleviating congestion and enhancing traffic management. While previous efforts have mainly concentrated on road-level predictions, the advancement of technology and the growing presence of automated vehicles have encouraged lane-level forecasting for more efficient driving and routing decisions. Therefore, this work presents a hierarchical model using recurrent neural networks (RNNs) that takes the lane-level information as the input and makes multiscale (lane level and road level) traffic flow predictions. Initially, the lane-level predictions are made using confined attention enabled RNN (CAERNN), built using bidirection long short-term memory (Bi-LSTM) and bidirection gated recurrent unit (Bi-GRU), and separately learning the periodic and temporal characteristics. Further, the features extracted by each lane are used to predict the traffic flow on the entire road. A novel confined attention mechanism is integrated into the Bi-LSTM module to improve the model’s performance by focusing on recent, more relevant information in the traffic flow sequence. It is observed that the confined attention mechanism performs better than the conventional attention mechanism. Furthermore, the external features are integrated to improve the model’s performance. The proposed model is evaluated on publicly available real-world data sets (in normal and COVID-19 scenarios), and the compared results with several state-of-the-art methods are evidence for the CAERNN’s effectiveness.
4 August 2025 in IEEE Transactions on Computational Social Systems
Traffic flow prediction is crucial for efficient transportation systems, allowing better traffic management and decision-making. However, it’s a complex challenge due to its intricate spatial and temporal dependencies and fine-grained variations. Various network architectures, particularly graph neural networks (GNN) and transformers, have been proposed to handle these complexities. Nevertheless, as network architectures advanced, their performance improvements started to slow down due to increased complexity. This shift in focus led to an emphasis on input embeddings, which offer a promising way to build simpler, more effective models. This work introduces the spatial–temporal long-term embedding transformer (STLTEformer) framework for precise traffic prediction. A 5, 15, and 30-min interval embedding is designed to capture short-term fluctuations, intermediate trends, and long-term patterns, respectively. These embeddings equip the STLTEformer with the ability to comprehend intricate traffic patterns. Extensive experiments were conducted on four real-world datasets. The results demonstrate that STLTEformer achieves an improvement of up to 13.6% in MAE, 9.4% in RMSE, and 14% in MAPE over popular models based on GNN, and up to a 5% improvement in MAE, 1.9% in RMSE, and 7.5% in MAPE over popular transformer-based methods with good computational efficiency.
29 October 2025 in IEEE Transactions on Computational Social Systems
Intelligent transportation systems (ITS) have witnessed significant advancements, particularly with the emergence of deep hybrid neural networks in artificial intelligence. Accurate traffic flow prediction is essential for optimizing ITS and mitigating urban congestion. While deep learning models have advanced traffic prediction, effectively capturing the intricate relationship between spatial, temporal, and periodic traffic patterns remains challenging. Additionally, the existing methods often lack the ability to focus on critical input segments, hindering prediction accuracy. To address these limitations, we propose a dual-stream convolutional gated recurrent unit (conv-GRU) model augmented with a fuzzy-inspired attention mechanism. Here, one stream is used to capture the temporal characteristics, whereas another stream is employed to extract the periodic features in the entire network. The convolution layers mine the spatial correlations in both streams. Further, selective emphasis on the crucial input segments through a fuzzy-inspired attention mechanism significantly improves the prediction. External factors like the day of the week and holidays are also incorporated to enhance the model’s adaptability and forecasting capabilities. Rigorous evaluation on a publicly available dataset from PeMS demonstrated the model’s superior performance, with significant reductions in mean absolute error (MAE), mean absolute percentage error (MAPE), and root mean squared error (RMSE) ranging from 4.4%–59.1%, 2.8%–64.9%, and 2.6%–54%, respectively, compared with state-of-the-art models.