Urban Road Speed Prediction Model Based On Bi-directional Long And Short Term Memory Network With Multi-source Feature Fusion

As an important part of intelligent transportation system,traffic speed prediction can dynamically grasp the development and change trend of traffic flow,which is the premise and foundation for traffic guidance and travel route planning.It is of great significance to urban management,traffic scheduling,road resource allocation and even economic development.Urban road speed prediction using deep learning method is a current research hotspot.The urban traffic state is affected by a variety of characteristic factors,and it is an urgent problem to comprehensively consider the related characteristic variables affecting the traffic state to improve the model prediction performance.In this thesis,the time-series data set of road speed in Chengdu calculated based on Didi open-source algorithm was modeled and analyzed.By comprehensively considering multiple characteristic variables related to traffic state,the temporal and spatial distribution of road speed in Chengdu road network was analyzed,and a two-way LSTM deep learning framework based on multi-source feature fusion was constructed.At the same time of time series prediction of road speed,the influence of relevant characteristic factors in the forward and backward periods on the prediction model is captured,which is expected to make better prediction of road speed.The main research contents include:(1)A method based on computer simulation to drive Web pages is proposed,which realizes automatic batch capture and analysis of meteorological data in Web browser,and improves the data quality and acquisition efficiency for model prediction.In this thesis,a road speed grid cell data set transformation method is proposed to express the spatial distribution of road network speed data reasonably.(2)The temporal and spatial distribution of road speed in Chengdu network and the influence of rainfall on the change of urban road speed are analyzed.In terms of space,the road speed decreases as it gets closer to the city center,and the road with lower driving speed tends to be around the tourist attractions and large commercial centers.In terms of time,the variation pattern of traffic speed on weekdays and weekends is different.Under different road grades,the change of driving speed is not stable.Rainfall has a significant influence on the driving speed of urban roads,and the expressway is more sensitive to the rainfall weather,and the influence degree is significantly different on different grades of roads.(3)A multi-source Feature bi-directional Long and Short Term Memory Network(MF-BILSTM)was constructed.Model considering the time and space distribution characteristics of the traffic network,the introduction of state of the weather,air quality,and time attribute variables,two-way network structure can capture the traffic velocity variation characteristics of forward and backward time,can effectively capture the meteorological factors,environmental changes and time correlation characteristics influence on urban traffic running state.(4)To Chengdu road speed sequential data set modeling analysis,respectively,selected the ARIMA,CNN,RNN,single LSTM and multi-layer LSTM network five benchmark model validation experiments,the results show that the MF-Bi LSTM in a standard weeks each time period were significantly better than the benchmark model,under the different road grade can perform at a higher stability,compared with the benchmark model prediction accuracy by 2.93%,1.10%,2.45%,1.67% and 1.02%.At the same time,MFBILSTM can show more accurate and stable prediction performance under different weather conditions and air quality grades.(5)Based on the traffic network time series data set of Chengdu and the two-way MFBILSTM traffic velocity prediction model integrating multi-source features,an intelligent traffic velocity prediction system based on Web GIS is designed and implemented by adopting the secondary development of hypergraph GIS and Huawei cloud platform technology.The MF-BILSTM prediction model is integrated into the system,and four main function modules are realized: "road speed prediction","road condition information query","road speed temporal and spatial dynamic analysis" and "three-dimensional visual analysis of traffic congestion".The system can display the prediction effect of each road model more intuitively.