Online transfer learning with an MLP-assisted graph convolutional network for traffic flow prediction: a solution for edge intelligent devices

Abstract

Traffic flow prediction is crucial for intelligent transportation and aids in route planning and navigation. However, existing studies often focus on prediction accuracy improvement, while neglecting external influences and practical issues like resource constraints and data sparsity on edge devices. We propose an online transfer learning (OTL) framework with a multi-layer perceptron (MLP)-assisted graph convolutional network (GCN), termed OTL-GM, which consists of two parts: transferring source-domain features to edge devices and using online learning to bridge domain gaps. Experiments on four data sets demonstrate OTL's effectiveness; in a comparison with models not using OTL, the reduction in the convergence time of the OTL models ranges from 24.77% to 95.32%.

1. Introduction

The integration of the Internet of Things (IoT) and intelligent transportation systems (ITSs) signifies a transformative era with unprecedented possibilities (Bojan et al., 2014). A widespread sensor network in urban traffic nodes enables real-time monitoring of variables like traffic flow and congestion, and swiftly transmits data for dynamic traffic management (Zhang H and Lu, 2020). The significance of IoT lies in processing data with advanced techniques such as deep learning, providing a scientific foundation for optimized transportation systems with heightened efficiency and reduced congestion. Accurate traffic flow prediction is crucial for intelligent transportation, empowering advanced management and optimizing resources for efficiency. However, prediction delay is a more important aspect for traffic managers and traveling vehicles (Hashemi and Abdelghany, 2015). Excessively long delays can lead to wrong decisions by traffic managers and mislead vehicles in planning their driving routes. The synergy of traffic flow prediction and IoT establishes an intelligent foundation for resilient, efficient, and secure urban transportation systems (Derawi et al., 2020).

However, some contradictions have arisen during the integration of IoT and ITSs. On one hand, IoT devices have evolved from simple magnetic and infrared sensors to the current generation of intelligent edge devices (Qadri et al., 2020). Initially, basic motion detection relied on magnetic and infrared technologies. Later, cameras were introduced to improve monitoring precision. However, integration among different sensors is still limited. Intelligent edge devices now have integrated sensors, computing abilities, and connectivity to provide advanced real-time processing and feedback functions. This integration is driving IoT towards greater intelligence and efficiency. However, such multi-functional devices are often limited by resources, especially computing resources (Asim et al., 2020). They not only perform computation, but also transmit wireless data (Cong et al., 2021) and information on road condition detection (Singh et al., 2021). Therefore, how to reduce the computing resource consumption and provide as many computing resources as possible for other aspects has become an urgent problem. In addition, collecting traffic flow data can be challenging in extreme weather conditions (Datla and Sharma, 2008).

On the other hand, as the core of traffic flow prediction problems, prediction accuracy has always been a hot topic in academic research. In the 1990s, traffic flow prediction based on micro-simulation and macro-models began to attract attention (Bando et al., 1995). During the 2010s, there was growing recognition of the significance of data-driven approaches and real-time technology (Nellore and Hancke, 2016). Currently, deep learning has emerged as an essential technology for predicting traffic flow, highlighting the superiority of neural network models in handling complex spatiotemporal relationships (Kashyap et al., 2022). Various models have emerged for spatiotemporal prediction, such as the spatial–temporal graph convolutional network (STGCN) (Yu et al., 2017), diffusion convolutional recurrent neural network (DCRNN) (Li YG et al., 2018), and graph attention method (Tang et al., 2020), all of which achieve excellent accuracy. This pursuit of higher accuracy often comes with an increase in algorithmic complexity and prediction time.