Enhancing Traffic Flow Forecasting with Deep-Learning and Noise Reduction Techniques

Traditional short-term traffic volume forecasting approaches make it difficult to predict the highly spatiotemporally coupled short-time traffic. To tackle the problem, this paper first proposes a variational modal algorithm (GWO-VMD) based on the optimization of the gray wolf search algorithm. It a...

Full description

Saved in:
Bibliographic Details
Published inDiscrete dynamics in nature and society Vol. 2024
Main Authors Zhao, Jinqiu, Yu, Le, Wang, Shuhua, Zhang, Zhonghao
Format Journal Article
LanguageEnglish
Published New York John Wiley & Sons, Inc 26.08.2024
Hindawi Limited
Subjects
Accuracy
Algorithms
Analysis
Decomposition
Deep learning
Error reduction
Forecasting
Forecasts and trends
Geospatial data
Graph neural networks
Machine learning
Methods
Neural networks
Noise control
Noise prediction
Noise reduction
Optimization techniques
Search algorithms
Spatiotemporal data
Traffic flow
Traffic volume
Online AccessGet full text

Cover

More Information
Summary:Traditional short-term traffic volume forecasting approaches make it difficult to predict the highly spatiotemporally coupled short-time traffic. To tackle the problem, this paper first proposes a variational modal algorithm (GWO-VMD) based on the optimization of the gray wolf search algorithm. It aims to decompose and reduce the noise of short-time traffic flows. Meanwhile, it reduces the intricacy of data sequences and enhances the regularity pattern. To address the insufficient utilization of spatiotemporal features, this paper presents an innovative deep-learning traffic prediction framework based on the stacking of multiple temporal trend-aware graph attention (TGA) layers and gated temporal convolution (GTC) layers, which are called trend-aware temporal graph neural network (TTGAN). TGA dynamically models the space-time relationships of traffic data, and GTC models the temporal characteristics of traffic data. The experimental findings demonstrate that the MAPE model, as presented, achieves a reduction of 9% and 2% compared to the AGCRN and GWNET models, respectively, in the domain of deep spatiotemporal graph modeling. Data decomposition and noise reduction are necessary to achieve accurate results. This model has superior performance in terms of mean absolute error (MAE), coefficient of determination (R2), and explained variance score (EVAR).
ISSN:1026-0226
1607-887X
DOI:10.1155/2024/1928189