Research On Short-term Traffic Flow Prediction Model Based On Time Series Decomposition

With the advancement of technology and the development of a well-off society,the number of car ownership in the country has increased dramatically,bringing convenience to life but also infinite traffic pressure on limited transport facilities.Intelligent Transport Systems(ITS)are the medium that connects roads,vehicles and managers,and are an important technical tool to relieve traffic pressure and protect the ecological environment.As a key technology of ITS,the task of short-term traffic flow prediction is to predict the number of vehicles passing on the target road section in the next 5 to 15 minutes,and provide the prediction results to the sub-programs of ITS for analysis and making corresponding decisions,including traffic signal control,tidal lane control,etc.It can even be connected to the navigation system of each traffic participant to make reasonable adjustments for their next travel plans This enables integrated monitoring and management of traffic and roads.The traffic network is very complex and most of the traffic data is interdependent,so it is difficult to accurately predict future traffic flows by relying on historical data from a single sampling point.Most of the existing traffic flow forecasting models rely on the influence of historical data from a particular sampling point to make predictions,which often make it difficult to respond to the changing trend of traffic flow in a timely manner,resulting in lagging prediction values.In this paper,a short-time traffic flow prediction model combining spatiotemporal feature capture and time series decomposition is proposed to address this problem.Details of the work are shown below:(1)A short-time traffic flow prediction model based on EMD-LSTM-SVR is constructed to address the lag problem that arises when a single model prediction relies only on historical data.The traffic flow sequence is decomposed by the empirical modal decomposition(EMD)algorithm,and the permutation entropy of each component is calculated.Then,each subsequence is divided into low,medium,and high frequency components and constructs the prediction model separately,while the component with the highest entropy value does noise reduction using the wavelet threshold noise reduction(WTD)algorithm.Finally,the support vector regression(SVR)prediction model is constructed and different kernel functions are fused to predict the low-frequency and mid-frequency components,while the high-frequency component relies on the long short-term memory neural network(LSTM)to construct the prediction model.It is demonstrated that the decomposition prediction model can effectively improve the prediction accuracy by constructing a targeted prediction model for each component.(2)To simplify the complexity of the prediction model,an EMD-DFT-based short-time traffic flow prediction model is constructed.The model decomposes the traffic flow sequence into trend component,cycle component and fluctuation component according to the unique characteristics of each component.Among them,the trend component is the residual component obtained after decomposition by the EMD algorithm.The periodic component is obtained by using the discrete Fourier transform(DFT)algorithm on the detrended component,which has a standard periodicity.The fluctuating component is the residual component by removing the first two components.As the most difficult part of the prediction is the fluctuation component,the MI algorithm needs to be fused to capture the spatio-temporal characteristics of the traffic flow and thus improve the accuracy of the overall prediction model.Finally,the SVR and LSTM prediction models are constructed for the trend and fluctuation components,and the final prediction results are obtained by predicting and superimposing them respectively.(3)In order to meet the requirements of dynamic traffic data prediction,a short-time traffic flow prediction model based on dynamic time series decomposition is constructed.The trend component,seasonal component,and fluctuation component can be obtained when the inner and outer cycles of the seasonal trend decomposition(STL)algorithm are finished.In order to be able to forecast dynamic sequences,the standard period of this component is calculated using the DFT algorithm and extended over the entire length of the sequence as the period component for that sampling point.Similarly,SVR and LSTM prediction models are constructed for the trend and fluctuation components for training.In the face of new traffic data,the trend and fluctuation terms are obtained by LOESS algorithm by completing the inner and outer cycles,and the state vector is reconstructed according to the newly obtained components and input to the corresponding prediction model,and then the final prediction results are obtained by superimposing the component prediction values to achieve dynamic traffic flow sequence prediction.