Short-term Traffic Flow Prediction Based On DFT-KNN-LSTM And Orthogonal Parameter Optimization

The Intelligent Transportation Systems is recognized as an effective way to improve the road safety,alleviate traffic jams,reduce traffic pollution,shorten travel time,and improve the road network traffic efficiency.Real-time and accurate short-term traffic flow prediction is one of the key technologies of the Intelligent Transportation Systems,also the research focus in the field of traffic flow prediction,and the basis of traffic guidance and control,travel route planning and so on,which plays an extremely important role in traffic planning and traffic optimization management of highway network.To solve the problem that existing prediction models cannot extract the internal rules of the traffic flow in massive traffic big data,and fail to consider the influence of different components of traffic flow sequence on prediction performance,a short-term traffic flow prediction model is proposed based on Discrete Fourier Transform(DFT),K-nearest neighbor(KNN),long short term memory(LSTM)and orthogonal parameter optimization in this paper after making full use of the spatiotemporal correlation characteristics of traffic flow sequences.Firstly,the DFT is used to decompose the traffic flow data into trend component and residual component,and the influence of the trend component on the prediction result is removed.Secondly,the KNN algorithm based on distance weighting is used to select K stations which are related to the test station in the road network.The spatiotemporal correlation data sets are constructed from the traffic flow of the selected K stations and input into the LSTM model for short-term traffic flow prediction.The parameters of LSTM model are optimized by orthogonal experiment,and the optimal combination of model parameters is determined with the minimum prediction error as the objective function.Finally,the model proposed in this paper is verified by the real traffic data from the Transportation Research Data Lab in USA.The real experimental results show that:(1)The prediction performance can be improved by removing the trend component.When the optimal spectral energy threshold is set,the average prediction error is reduced by 28.64%;(2)The KNN based on distance weighting can effectively optimize the data set and make full use of the spatiotemporal correlation characteristics of traffic flow data;(3)After using the orthogonal experiment for parameter optimization,the average prediction accuracy is improved by 4.09%,and the average running time is reduced by 77.18%;(4)The prediction performance of the proposed DFT-KNN-LSTM model based on orthogonal parameter optimization in this paper is better than other models.Compared with the existing ARIMA,SVR,WNN,DBN-SVR and LSTM models,the average prediction accuracy of the proposed model is improved by 32.82%,7.96%,13.03%,6.64% and 5.19%.