Research On Traffic Speed Prediction Based On Improved Informer Model

The traffic speed is an important indicator for assessing the traffic condition.Traffic speed prediction can provide real-time information about traffic congestion,helping people make better travel decisions.Currently,there are two main types of traffic speed prediction models:single-section prediction models and multi-section prediction models.However,in singlesection traffic speed prediction,the direct prediction of noisy traffic speed is not taken into account.In multi-section traffic speed prediction,only the time characteristics of traffic speed data are often considered,while the spatial characteristics are ignored,and the long-term prediction accuracy needs improvement.To address the existing problems in traffic speed prediction models,this paper proposes two improved models based on Informer: Waveformer and Graghformer.The main research is as follows:(1)This paper introduces the Waveformer model to address the problem of noisy traffic speed prediction in single-section models.The Waveformer model incorporates a wavelet decomposition module into the encoder and decoder of the Informer network.It performs window decomposition on the sequence,removes high-frequency noise signals,and reconstructs the remaining high-frequency signals.By using self-attention mechanism,it captures temporal characteristics of the high-frequency signals and combines them with lowfrequency signal features.This approach achieves denoising and prediction in a single step,eliminating the need for separate denoising preprocessing and improving prediction accuracy.(2)To solve the problem of low prediction accuracy in multi-section traffic speed prediction,this paper proposes the Graghformer model by comprehensively considering the spatial and temporal characteristics of traffic speed.The model combines graph convolutional networks with Informer networks to capture the spatial characteristics of the road network and use the Informer model to extract the time characteristics of dynamic time series,thereby improving the prediction accuracy.This paper conducts extensive experiments on two proposed prediction models using realworld datasets.The results demonstrate that the Waveformer model outperforms existing models in predicting traffic speed with noise in a single section,achieving a minimum 11%reduction in MAE.It exhibits superior prediction accuracy,along with strong stability and fast convergence speed.Additionally,the Graghformer model performs well in short-,medium-,and long-term traffic speed forecasting for multiple sections.