Long-term Traffic Volume Prediction Based On Combined LSTM

Traffic flow long-term prediction plays an important role in urban road planning and traffic policy making.However,traffic flow is affected by many factors such as climate,economy and travel.It can be seen as a random and non-stationary process with randomness,volatility and non strict periodicity.This leads to the difficulty of long-term traffic flow prediction.Note that previous time series prediction models are mostly ideal traffic volume forecasting schemes in relatively stable traffic conditions.Drastically varying traffic conditions require excessive calculation and may yield lower forecasting accuracy.This paper proposes a combined forecasting model which combines the three common algorithms of Markov chain,Type-2 fuzzy sets and Long Short-Term Memory(LSTM)neural network to analyze and predict the traffic flow.On the basis of informed researches,traffic flow forecasting method is improved to further improve the prediction accuracy.The main works of this paper is as follows:First of all,this paper selects the LSTM neural network which is good at dealing with time series highly related problems as the core prediction algorithm,and constructs the three-layer network structure of LSTM.The LSTM prediction model constructed in this paper is a three-layer network structure of one input layer,one hidden layer and one output layer,including 288 nodes in the input layer,18 nodes in the hidden layer and 288 nodes in the output layer.The initial weight is generated randomly,and the number of iterations is 3000.Secondly,considering the long training time of single LSTM neural network,higher requirement of hardware facilities,this paper selects Markov chain to extract the periodic characteristics of traffic flow.Firstly,Fuzzy C-Means(FCM)clustering is used to generate three states,and then Markov chain is used to generate one-step transition probability matrix to extract the periodic state transition characteristics of traffic flow.Finally,this paper uses the Type-2 fuzzy sets to extract the proximity feature of traffic flow so as to mine the longitudinal feature and enhance the interpretability of neural network.Firstly,the Adaptive Network-based Fuzzy Inference System(ANFIS)neural network is used to verify that the traffic flow at the current time is most closely related to the first 15 minutes,and then the Type-1 fuzzy set of every three five minutes is calculated and the Type-2 fuzzy sets is obtained.The upper and lower boundary values of the type-2 fuzzy sets are used to update the weight of the input gate of LSTM,so as to make full use of the proximity characteristics of traffic flow.