Research On Short-term Traffic Flow Prediction Model Based On Optimized Extreme Learning Machine

Short-term traffic flow forecasting has been a crucial task in the area of intelligent transportation systems(ITS),which plays a significant role in operating traffic management systems and dynamic traffic assignment effectively as well as proactively.So high prediction accuracy and low time consumption is important to transportation management system.Traditional prediction methods are mostly offline and time-consuming during training phase.Most of methods such as artificial neuron networks use smooth historical data to predict near future traffic flow without considering nonstationary condition caused by breakdown of detectors or sensors.The contributions of paper are listed as follows.In order to overcoming the shortcoming of time consumption,a novel short-term traffic flow prediction method called Real-time Sequential Extreme Learning Machine(RS-ELM)with simplified single layer feed-forward networks(SLFN)structure under freeway peak traffic condition is proposed.By quickly training historical data and incrementally updating model with new arrived data,sliding window mechanism is used.RS-ELM has the characteristics of less training time consumption and high prediction accuracy.In order to overcome the shortcoming of accuracy,feedback mechanism is used.The difference values between the prediction value of network and theoretical expectation are feedback to the input layer to improve stability and robustness.The input weights are adjusted according to the feedback difference values.In addition,ensemble mechanism is also used to optimize the stability and robustness.So a novel short-term traffic prediction model based on weighted ensemble feedback real-time sequential ELM(ERS-ELM)is proposed.The traffic data chosen in this paper are downloaded from open Caltrans performance measurement systems(PeMS version 14.0)database.Experiment results show that average mean absolute percentage error(MAPE),test root mean square error(RMSE)and training time consumption of proposed method is superior to classical Wave-NN,MLP-NN and ELM methods.