Research On Adaptive Multi-model Traffic Flow Prediction Method In Big Data Environment

With the development of China's transportation industry,real-time traffic data can be more easily collected through the Internet of Things technology,and the use of realtime traffic data has brought tremendous changes to the transportation industry.The real-time data in the big data environment is different from the traditional static data,which puts higher requirements on the efficiency and accuracy of data processing.How to predict the real-time data in the big data environment is the primary challenge of the intelligent transportation system at this stage.In the current big data environment,faced with a larger data scale and more complex traffic scenarios,the traditional model based on historical data has a problem of drastically decreasing efficiency,and it is difficult to cope well with situations with drastic changes in traffic flow.To solve the above problems,this paper proposes an adaptive multi-model prediction method.By extracting features from real-time traffic flow sequences,sample construction is performed to reduce the problem of data storage pressure in the traffic system.On this basis,the traffic flow combined forecasting model is designed,and use the dynamic weight to realize the adaptive to the changing traffic environment.The main contributions of this paper include:(1)A real-time traffic flow feature extraction algorithm based on statistical properties of data is proposed.By monitoring the time series features in the real-time state to identify the occurrence of stationary non-stationary phenomena in the time series,the stationary features of the traffic flow sequence are extracted,and the realtime traffic flow implied trend change information is realized,and the early adaptation to the traffic scene changes is realized.At the same time,based on the extracted data features,the corresponding traffic flow feature model is designed to better describe the true characteristics of traffic flow.(2)A sample construction method suitable for big data environment is proposed.Aiming at a large amount of traffic flow data and low-value density in the big data environment,based on the extracted traffic flow characteristics,the similarity test for traffic flow sequences is realized.Resampling the original massive data reduces training GPU consumption without reducing the accuracy of the model training.(3)An adaptive weight combination prediction model based on the traffic scene transition probability is proposed.Based on the extracted traffic flow characteristics,the traffic flow is divided into different traffic scenarios.The online incremental sub-models in different scenarios are trained separately,and the traffic scene transition probability is taken as the model weight combination sub-model.By maximizing the approach to real traffic flow scenarios,the prediction accuracy is improved,and the negative impact of traffic conditions on traffic flow prediction is reduced.In this paper,the proposed sample construction method and model fusion method is comprehensively tested by using the data in the PEMS database.Experiments show that the proposed method performs well in the big data environment and can improve the prediction accuracy by reducing the training time of the model.The method Can provide excellent technical support for subsequent transportation decisions.