Traffic flow prediction is very important to urban traffic management and public safety, and is also a key link in the development of intelligent transportation system. However, the long-term traffic flow data has complex time characteristics, which brings severe challenges to the traffic flow forecasting research. Deep learning methods are mainly used in time series prediction tasks such as traffic flow prediction. Recently, Transformer-based approaches have been able to improve traffic flow forecasting results, but there are still limitations such as high computational costs and difficulty capturing overall trends in traffic flow data. We propose a multi-scale sensing neural network (MSPFormer) based on Transformer, which uses multi-layer perceptrons to achieve more efficient feature extraction in step prediction, while designing a progressive decomposition architecture to preserve historical information authenticity as much as possible. Detailed experimental studies have shown that our approach can improve the accuracy of traffic flow prediction in Transformer architecture by 4% to 24% without consuming additional computing resources.
Long-term traffic flow prediction based on multi-scale perceptual change
Seventh International Conference on Traffic Engineering and Transportation System (ICTETS 2023) ; 2023 ; Dalian, China
Proc. SPIE ; 13064
2024-02-20
Aufsatz (Konferenz)
Elektronische Ressource
Englisch
Progressive-based large-scale short-term traffic flow prediction method
Europäisches Patentamt | 2023
|Long-term and short-term traffic flow prediction model and method
Europäisches Patentamt | 2020
|Traffic flow prediction method based on long-term and short-term Transform model
Europäisches Patentamt | 2022
|A Short-Term Traffic Flow Prediction Method Based on Long Short-Term Memory Network
British Library Conference Proceedings | 2019
|