Identification And Prediction Of Urban Expressway Traffic Status Based On Ensemble Learning

With the acceleration of urbanization,the total number of existing vehicles is approaching the saturation of road space,which makes it difficult to maintain the balance of traffic supply and demand,which has a negative impact on the traffic environment and increasingly serious problems such as traffic congestion.Therefore,in the context of smart cities,mining traffic data rules and applying big data technology to control and manage traffic have become the general trend.Among them,how to accurately detect the occurrence of traffic incidents and identify and predict road traffic conditions so as to provide effective information for traffic control departments is the focus of research in the field of traffic.Most of the existing research focuses on a single model or a combined model of machinery,that is,combining the models by improving a single model or adjusting the weights to solve the problems in the transportation field.The traditional model fusion method limits its performance.Based on this,different from the mechanical combination,this paper introduces the ensemble learning theory,in order to maximize the performance of the model,to selectively perform a better ensemble fusion of a single model of traffic event detection,traffic state recognition and short-term prediction.The main research contents are as follows:(1)Aiming at the problem that the existing single traffic event detection model cannot guarantee the detection rate and false alarm rate at the same time,a traffic event detection method based on the XGBoost ensemble algorithm is proposed,and the change of the event detection performance with the number of boosting trees of the base learner is explored.The traffic incident detection model based on MLF,Decision tree,RF and SVM is compared and analyzed.(2)A traffic state recognition method combining LSTM and selective ensemble is proposed.The LSTM traffic state recognizer is constructed to learn and train the preprocessed traffic data,and the idea of selective integration is introduced to rank the constructed series of basic learners differently.According to the difference of the input training sample data and the credibility of the basic learners,the optimal combination of basic learners is automatically determined and selected,and the sample data to be tested is identified,and the traffic state level is output,which is compared with the bagging and boosting integrated learning frameworks.(3)An indirect traffic state prediction method based on the stacking ensemble learning framework is proposed,which is to predict short-term traffic flow parameters.Construct a single traffic flow parameter prediction base learner based on SVR,RF,GRU,BP,GBDT and XGBoost,and input its output results into the LSTM traffic flow parameter prediction base learner for learning and training,and output the prediction results as MAE,RMSE and R^2 is the evaluation index to compare and analyze the prediction performance of each single model and the integrated model.The case analysis shows that,compared with the traditional single detection model,the traffic incident detection method based on ensemble learning proposed in this paper can achieve a satisfactory detection rate while ensuring that the false alarm rate fluctuates within an acceptable range,and the detection stability is enhanced;Compared with the static fusion model,the traffic state recognition method based on the selective integration idea proposed in this paper performs better in various evaluation indicators;The prediction method is more accurate for the prediction of traffic flow parameters.