Deep learning model for efficient traffic forecasting in intelligent transportation systems

In this paper, a novel intelligent transportation framework called graph convolutional neural network for distributed machine learning (GCN-DML) is presented. Its primary application is traffic forecasting in large-scale, dynamically-evolving traffic scenarios. One unique method in traffic forecasting, called GCN-DML, is that it allows user data to be processed locally on edge devices, removing the requirement to send all data to CS (CS). Designed specifically for edge computing, this distributed machine learning framework makes it easier to train neural networks on a variety of edge devices. In addition to protecting user privacy, this method reduces the difficulties related to centralised neural network training, minimises communication delays, lowers data transfer quantities, and improves overall data processing efficiency. GCN-LSTM, the composite graph convolutional networks used in this framework, combine edge-enhanced attention mechanisms with the feature transmission capabilities of graph convolutional neural networks to effectively handle the challenging task of traffic forecasting. Even as the number of cars in the network rises, they perform exceptionally well at quickly extracting and sending interaction information between vehicles, while preserving high prediction accuracy and low time complexity. The versatility of GCN-DML is extended to a range of traffic forecasting application scenarios, whereby edge devices can adjust the kind and scale of the neural network according to available compute and storage resources, all without sacrificing performance. The usefulness of GCN-DML in traffic forecasting is demonstrated by experimental evaluations on the NGSIM public dataset, where it outperforms other models with better computation time and prediction performance. It obtains noteworthy F1 scores of 0.9473, 0.9557, and 0.9391, respectively, for recall, precision, and precision. Even when there is an increase in the number of cars, the accuracy remains high, reaching 91.21%, even with a prediction length of only 1 s. At the same time, the time complexity remains low, under 0.1 s. These findings clearly demonstrate the effectiveness and potency of GCN-DML as a traffic forecasting tool for intelligent transportation systems.