Research On Short-term Traffic Flow Prediction Method Based On Deep Learning

With the rapid development of the social economy,the number of vehicles has increased rapidly,and the problem of traffic jam has become increasingly serious.The most effective way to solve this problem is to use the Intelligent Transportation System(ITS),and one of the core elements of the ITS is short-term traffic forecasting.Accurate traffic flow prediction models play an important role in reducing traffic congestion,improving air quality and supporting government decision-making.This paper summarizes the existing methods of traffic forecasting,and improves and innovates on the basis of previous studies.For different situations,models with higher prediction accuracy are proposed.This paper has done the following research based on the existing work:Firstly,this paper propose an LSTM-based approach.An improved method(LSTM+)that uses the attention mechanism to capture the high-impact values of a very long sequence and connect it to the current time step,giving the LSTM a super long-term memory function.When the number of sensors in some areas is small or isolated,and the sensors cannot form an effective topological relationship.In this paper,long short-term memory network(LSTM)is used for traffic prediction.However,due to the limited si e of LSTM s cell and he con in o s updating of cell status,LSTM has a lack of extremely long-term dependence on capture.Recently,researchers have noticed that adding features on multiple time scales can help improve the long-term dependence of the recurrent neural network(RNN),inspired by this,this paper improves LSTM by the way of attention mechanism.Experimental results show that the improved LSTM+ model has certain competitiveness in isolated sensor prediction.Secondly,this paper propose a densely connected spatial-temporal graph convolutional network(DSTGCN),which captures temporal relationships in data through convolutional networks(CNNs),captures spatial relationships in data through a graph convolutional network(GCN),and finally uses densely connected networks to enhance the propagation of these two relationships.In order to reduce the input of redundant data,a special mask convolution layer is also designed.This kind of network is suitable for the situation that there are many sensors and it can form spatial topological relationship,The experimental results show that the proposed model can extract the hidden spatial-temporal relationship more fully,and the model outperforms the state-of-the-art baselines in the sensor prediction with topological relationship.