Fuzzy hybrid framework with dynamic weights for short‐term traffic flow prediction by mining spatio‐temporal correlations. Issue 2 (31st October 2019)

• Record Type: Journal Article
• Authors: Ma, Dongfang; Sheng, Bowen; Ma, Xiaolong; Jin, Sheng
• Is Part Of: IET intelligent transport systems. Volume 14:Issue 2(2020)
• Page Start: 73
• Page End: 81
• Publication Date: 2019-10-31
• Subjects: road traffic -- forecasting theory -- data mining -- fuzzy logic -- fuzzy set theory -- traffic engineering computing -- neural nets -- regression analysis -- nearest neighbour methods -- feature extraction; fuzzy logic system -- long‐short‐term memory -- fuzzy hybrid framework -- short‐term traffic flow prediction -- spatio‐temporal correlations -- intelligent transportation systems -- nonparametric regression -- deep learning -- K‐nearest neighbour -- feature extraction; Intelligent transportation systems -- Periodicals; Electronics in transportation -- Periodicals; 388.31205
• DOI: 10.1049/iet-its.2019.0287
• Languages: English
• ISSNs: 1751-956X

Abstract

Abstract : Forecasting short‐term traffic flows is crucial in designing intelligent transportation systems. There has been significant research on this topic; however, no single model can maintain optimality under all conditions. This study analysed the complementarity of non‐parametric regression and deep learning in terms of prediction accuracy. Then, a mixed prediction method, combining two sub‐models based on a fuzzy logic system, was constructed. First, the two sub‐models based on long‐short‐term memory) and K‐nearest neighbour were improved, by extracting more features that consider the spatial–temporal correlations of traffic flow and the influences of specific contextual factors to the traffic flow. Second, a fusion mechanism with dynamic weights was presented to optimise the mix predictor. A compensation term was added based on the deviations. Finally, the new method was tested using real data in Seattle, Washington. The results show that the proposed model predicts the flow for the next 5 min across the verification set, with a 6.42% mean absolute percentage error and 31.49 standard error. These results are more accurate and robust than those achieved with other state‐of‐the‐art models.