Research On Traffic Flow Prediction Model Based On Time-Series Enhancement

Traffic flow is closely related to traffic efficiency,and accurate traffic flow prediction is helpful to save social transportation cost and shorten travel time.Traffic flow is a non-stationary time series influenced by multiple factors,and many road sections in reality do not have the conditions for spatio-temporal prediction,so it is difficult to design an accurate and reliable time-series traffic flow prediction model.In this paper,we propose a traffic flow prediction model based on time-series enhancement and analyze the reliability of the model.The main original works of this thesis are as follows.(1)To improve the accuracy of the traffic flow prediction model,this paper proposes a prediction model based on wavelet transform and convolutional neural network.The WT-2DCNN model is constructed based on the convolutional neural network by integrating wavelet noise reduction and two-dimensional time series enhancement,and the WTRes Net model is proposed based on the idea of residual network introduced.The experimental results show that both WT-2DCNN and WT-Res Net models have better accuracy than the current mainstream models,and the training time is significantly shortened.(2)In order to realize the lightweight of traffic flow prediction model,this paper proposes a prediction model based on hybrid feature engineering and lightweight convolution.Combining empirical modal decomposition for multi-channel timing enhancement,the Mix-Light Net and MixGroup Net models are constructed from two perspectives of lightweight and balance.After experimental analysis,the computational volume of MixLight Net model is reduced by 1/2,and Mix-Group Net better balances the model volume and prediction accuracy.(3)To verify the reliability of the traffic flow prediction model,the robustness and explanatory analysis of the Mix-Group Net model is conducted as an example.The robustness of the model is assessed by introducing disturbances,and the prediction logic of the model is explained in terms of both feature contribution and model structure.The experimental results show that the Mix-Group Net model structure improves the model robustness and demonstrates the ability of the Mix-Group Net model to extract data cycle features.