Research On Short-term Traffic Flow Forecasting

Since the end of the 20th century,with the increasingly complex road network,and the rapid growth in the number of vehicles and other means of transport,traffic flow influences more and more people’s lives.Unreasonable traffic flow guiding may lead to traffic congestion,traffic accidents or other issues.The forecast of short-term traffic flow has a positive effect on alleviating urban traffic problems,improving urban transport efficiency and building the Wisdom City.For short-term traffic flow forecasting,there are simulation,regression analysis,neural network and other methods.The simulation method requires to establishment a complex traffic network model,and it need high priori knowledge supporting.And the calculation cost may be large.So it is difficult to apply in practice.The regression analysis method mainly includes autoregressive model(AR),sliding average Model(MA)and autoregressive moving average(ARIMA)model.Neural network methods including Stacked Auto Encoder(SAE)and Deep Belief Network(DBN)are also used in traffic flow forecasting task.The main advantage of the deep learning method is that it can use the multi-layer network structure to optimize the feature space,which makes the new features more expressive.How to get more meaningful features has become the focus of the current research.These methods are generally based on historical traffic flow data to predict traffic flow for the next time period.And most of them are applied for a small amount of data collection point,but not in the entire urban area perspective.At present,the widely used intelligent transportation system(ITS),a large number of two-dimensional urban spatio-temporal traffic data has been available.And the amount of data is growing at an amazing rate.So Here is a strong data support which makes our task meaningful.Meanwhile the larger-scale data also takes more challenges for us.With the improvement of computing ability on the computer hardware,the method of deep learning has begun to get more and more exploration and application.The high-dimensional feature extraction ability of deep learning method is of great significance to the forecasting task.We propose a traffic flow forecasting method based on data grouping and convolution neural network.The main features are as follows:1.Exploring the influencing factors of traffic flow at different locations by vehicle trajectory data;2.Applying the influencing factors of traffic flow at different locations in the feature matrix of short-term traffic flow.We construct a feature matrix with time dimension and spatial dimension.3.Convolution neural network is used to predict traffic flow,and its local field principle is used to exert the influence of spatial factors on short-term flow.We have applied the model to the actual traffic data set,carried out the full experiment.The result proves the validity and the advancement of our model,It performance better than other model in the mentioned evaluation index.It has the practical application value,and can be applied in related areas.