Urban Regional Traffic Flow Prediction Based On Neural Network And Clustering

The rapid development of China's economy and the acceleration of urbanization have increased the employment needs of cities,and more people choose to gather work and live in cities.Urban population growth has made tremendous contributions to the promotion of urban development.On the other hand,it has also brought some negative impacts on the city.The most obvious problems are traffic congestion and air pollution.Solving the problem of traffic congestion has been regarded as one of the priorities of many cities,and has become a hot research topic in academia.Traffic congestion is affected by factors such as transportation infrastructure and population density,and is also affected by various factors such as the location of the school,densely populated enterprises,public activities,climate,and meteorological conditions.Therefore,predicting and solving traffic congestion is a huge challenge.This paper discusses the urban regional traffic flow forecasting problem based on spatio-temporal correlation,and proposes a regional traffic flow forecasting method based on neural network and clustering.The main research contents include:(1)A medium and long term traffic flow prediction method based on periodic features is proposed.The main idea of this method is to use the periodic characteristics of traffic flow time series data(such as a certain area,the traffic flow at each time point of the working day has a great similarity,etc.)to establish the correlation between the data,Thereby constructing a targeted prediction model.This paper starts with the establishment of Long-Short Term Memory(LSTM)traffic flow prediction model,selects the clustering granularity according to the variation range of regional traffic flow,clusters the original traffic data,and then pairs it.The post-class data is trained and predictively verified.The purpose is to ignore the fluctuations caused by small changes in regional traffic flow data,and to pay more attention to the trend changes of the data,so as to better match the periodic characteristics of regional traffic,Thereby improving the accuracy of prediction,especially for medium and long-term time traffic flow prediction.(2)Propose a short-term traffic flow prediction method based on proximity characteristics.The main idea of this method is to use the proximity characteristics of the traffic flow time series data(for example,for a specific area,the traffic flow changes in the past few hours have a greater impact on future short-term traffic flow changes)to establish correlation between data.To construct a targeted prediction model.This paper first constructs a predictive model based on proximity features and LSTM.Secondly,it clusters the original data,establishes a predictive model based on dynamic cluster proximity and LSTM,trains and predicts the clustered data.Then,the prediction results of the two models are integrated.According to the Root Mean Squared Error(RMSE),the predicted values of the two models are assigned appropriate weights,and the prediction bias is appropriately modified,That is,a model with a small RMSE assigns a larger weight,and a larger RMSE assigns a smaller weight.Multi-model integration realizes the complementarity of each model and can improve the accuracy of prediction.(3)A traffic flow prediction method based on correlation region clustering and CNN is proposed.The main idea of this method is to take advantage of the spatial correlation characteristics of the traffic flow data(because the vehicle enters the target area or leaves the target area through some of its adjacent areas,the target area has a very close spatial correlation with its neighboring areas)establish correlations between data to build targeted predictive models.In this paper,the regional traffic correlation metrics are used to select relevant regions,and the computational load of model training is reduced.Then,the traffic flow data of the target region and related regions are clustered,and then the clustering model based on correlation region clustering and CNN is used.The relevant regional data is trained and predicted,which improves the training accuracy and efficiency of CNN.This paper uses the data set provided by Microsoft Research Asia to train and predict the proposed traffic flow prediction model,and analyze the experimental results.