Flow prediction method based on Transformer and diffusion model
A prediction method based on Transformer and diffusion model is proposed to address the problem of overfitting in existing traffic prediction methods that only consider single modal information and are prone to overfitting when there are insuffi-cient samples.Generate virtual samples with the same distribution as real samples using a diffusion model.By constructing a multi-modal Transformer model,deep feature extraction is performed on data samples from three dimensions:time,frequency domain,and correlation.Finally,the three types of feature information are fused for network traffic prediction.The experimental results show that the proposed method can significantly improve the prediction performance.Specifically,when the proposed method only utilizes 10%of the data,its prediction performance is close to that of a Transformer model trained on full data.And when the data volume increases to 50%,the effect surpasses that of the Transformer model trained on full data
flow forecastself-attention mechanismmultimodaldiffusion model
万方数据
