Time-varying weight coefficients determination based on fuzzy soft set in combined prediction model for travel time

• Published in: Expert systems with applications 2022-03, Vol.189, p.115998, Article 115998
• Main Authors: Li, Huamin, Xiong, Siyu
• Format: Article
• Language: English
• Subjects: Artificial neural networks; Back propagation networks; Coefficients; Combined prediction model; Forecasting; Fuzzy sets; Fuzzy soft set; Intelligent transportation systems; Kalman filters; Mathematical models; Neural networks; Prediction models; System effectiveness; Time-varying weight coefficient; Traffic information; Travel time; Uncertainty
• ISSN: 0957-4174; 1873-6793
• DOI: 10.1016/j.eswa.2021.115998

•Dealing with the uncertainty and fuzziness in combined prediction model.•A novel time-varying weight coefficient based on fuzzy soft set is presented.•The combined travel time prediction model with TVWDFSS is proposed. Precise prediction of travel time is a valuable reference for traffic participants to make travel choices and is important for the development of intelligent transportation systems (ITS). Combined prediction method for travel time can effectively improve prediction accuracy, but the uncertainty and fuzziness in determining the weight coefficients are inherent nature. A novel time-varying weights determination method based on fuzzy soft set (TVWDFSS11Time-varying weights determination method based on fuzzy soft set, abbreviated as TVWDFSS), that the uncertainty and fuzziness are considered, is presented, and the combined travel time prediction model with TVWDFSS is proposed. In this model, several forecasting methods are firstly selected as the component predictors, and the forecasting using these predictors are performed, respectively. Then the relative error of each predictor at the specific time point is calculated, and a tabular form of the fuzzy soft set that describes the actual time series data is constructed. Finally, the weight coefficients of all component predictors at each time point are determined according to their forecasting errors recorded in the tabular form of fuzzy soft set. In case study, three individual methods, Kalman Filter (KF), Back Propagation Neural Network (BPNN), and k Nearest Neighbor (k-NN) are selected as the component predictors of the combined travel time prediction model with TVWDFSS, and real-world traffic data is used to verify the validity and precision of this presented model. Comparing with the three individual methods and another combined approach that employs a constant weight determination method that is also based on fuzzy soft set (CWDFSS22Constant weight determination method based on fuzzy soft set, abbreviated as CWDFSS), the results confirm the validity and superiority of the proposed models.