Research On Traffic Flow Prediction Based On SAE-LSTM-attention Model

On the one hand,the urbanization development of China has changed from the traditional urbanization focusing on growth rate to the new urbanization focusing on quality improvement.However,in the past urbanization process,there were many problems,such as air pollution,water pollution and other environmental problems,as well as urban life problems such as roads and houses,and traffic congestion is one of the most important and urgent problems in the new urbanization construction.On the other hand,with the development of market economy and the improvement of people’s living standards,more and more people have their own private cars.According to the latest data of the public security department,in 2019,the number of cars registered nationwide increased by 25.78 million,the total number of cars in China was 260 million,and the total number of motor vehicles in China reached 348 million.However,the speed of urban road construction lags far behind the increase of cars,thus forming a serious road traffic problem.Traffic congestion has had an important impact on people’s living standards and reduced people’s travel efficiency.It has also seriously affected the social order,caused the pollution of urban living environment and consumed a lot of urban resources,which is one of the important factors restricting the new urbanization construction in China and a major problem in the long-term development of cities.Therefore,intelligent transportation system has attracted the attention of researchers,and accurate and reliable traffic flow forecast is an important part of many intelligent transportation systems,such as dynamic traffic control,intelligent route guidance and intelligent location service.However,the traffic flow data is different from the general time series data,and the fluctuation of traffic flow is affected by multiple mixed factors such as weather conditions,personal vehicle speed,road network structure,etc.However,in actual research,it is difficult to take all these factors into account in the research methods at the same time.How to effectively analyze these related data characteristics and build a highaccuracy and practical traffic flow prediction model has become an important direction in current traffic research.However,the traditional traffic flow forecasting methods are difficult to mine the complex nonlinear relationship among the related features of traffic flow data.Because of its feature extraction based on the principle of multi-layer nonlinear mapping,the deep learning model can be well applied to traffic flow forecasting.Thus,this thesis constructs an SAE-LSTM-Attention model for traffic flow prediction.According to the characteristics of large time span,high dimension and nonlinearity of traffic flow data encountered in previous research on traffic flow prediction,SAE model composed of several Autoencoder is used to extract potential high-dimensional features from massive traffic flow data,effectively reducing the data dimension and achieving the effect of data compression to a certain extent.Then,the output of the last hidden layer of SAE is used as the input of the Long Short Term Memory Networks(LSTM),and the two-layer LSTM network module is used to extract the time-dependent features in the traffic flow data,so as to reduce the influence of the time-dependent features on the prediction results,and the SAELSTM model is constructed for traffic flow prediction,which can achieve the effect of reducing the calculation amount and complexity of the prediction model.Finally,attention mechanism is added to SAE-LSTM model to give weight to the feature information of different dimensions in traffic flow data,so as to obtain the traffic flow information with high influence on the prediction value,thus further improving the prediction accuracy and fitting performance of the model.In this thesis,two commonly used data sets,Pe MSD4 and Pe MSD8,in the Caltrans Performance Measurement System(Pe MS)are also used to test the prediction effect of the two models.Due to the influence of road structure,weather conditions and multiple external factors,the traffic flow and fluctuation law of the two data sets are also obviously different,and the applicability and effectiveness of the models can be better verified by testing separately.The experimental results show that compared with LSTM model,SAES model and SAE-LSTM model,the SAE-LSTM-Attention model has further improved the prediction accuracy and fitting performance on Pe MSD4 and Pe MSD8 data sets,can effectively capture the nonlinear and periodic characteristics of traffic flow data,and can effectively predict the local abrupt change data,which proves the practicability and effectiveness of this model.