Short-term Traffic Flow Forecasting Of Urban Road Network Based On Multi-feature LSTM

Many cities and towns in China are faced with the following problems after urbanization: while the urban transportation infrastructure is becoming more and more perfect,the per capita motor vehicle ownership is gradually rising,the urban road network is under great pressure,resulting in urban traffic congestion,people travel efficiency reduction and other problems gradually aggravating.Intelligent Traffic System(ITS)has become one of the most effective ways to alleviate Traffic congestion and satisfy travelers’ wishes with ITS advanced scientific and technological means.Its core function is to realize Traffic guidance and risk avoidance.Real-time and high accuracy short-term traffic flow prediction is the premise to improve the ability of guidance and road control.The research of short-term traffic flow prediction based on urban road network has very important practical significance.On the basis of summarizing the traffic flow time series clustering,traffic flow prediction method and the method of spatio-temporal feature extraction of road network,this paper puts forward a short-term traffic flow prediction method of urban road network based on multi-feature long and short-term memory network.The main work of this paper is as follows:1.In order to solve the inevitable missing problems in the process of traffic flow data collection,a fault data repair model based on the correlation analysis of fault data date attributes is proposed,and weather and social attributes are introduced into the model as important basis for data repair.A data set containing missing data was constructed at random time points,and the effectiveness of the method was demonstrated by using the data set.The results show that the proposed model performs better than other existing repair methods.2.Aiming at the shortcomings of existing methods of spatial feature extraction of road network,such as dependence on actual topology structure,high application cost but poor effect,a spatial feature extraction algorithm of urban traffic flow based on time series and clustering is proposed.The algorithm proposes two models to fit the time series with a long reference period and a short reference period respectively.Both models used road network output time series to represent road sections and found the spatial correlation between road sections by clustering these time series.On the basis of the self-organizing mapping network,the former introduces the multi-feature extractor,which effectively extracts the fluctuation features of time series from multiple angles.After amplification and enhancement,the self-organizing mapping network is used for clustering operation,so as to get the grouping of the closely connected sections within the road network.Through the effective processing of the eigenmatrix,not only can improve the clustering effect,but also effectively save the time of algorithm.The latter is to improve the density-based clustering algorithm by introducing the maximum number of information to measure the correlation between time series and improve the accuracy of clustering.Experimental results show that both algorithms have good performance on data sets,and the clustering effect is improved by more than 5%.3.In view of the current one-dimensional LSTM and the classical two-dimensional LSTM neural network,which can only refer to the same time series information,or can not better meet the traffic flow prediction requirements on the road network,a new two-dimensional LSTM structure is proposed.Different from the three gate switches of one-dimensional LSTM neurons,the new structure adds a layer to the neuron to collect the immediate state unit of the remaining neurons in the same group at the previous moment.In other words,the output of the neuron is not only transmitted to the neighboring neurons,but also transmitted along time in the matrix.The experimental results show that the above method has good performance in terms of accuracy and timeliness.Compared with LSTM,the error reduction rate is more than 10% in all three evaluation dimensions,and the test time is reduced by about 11.5%.