Abstract

In this paper, a new deep fuzzy network structure called Fuzzy Temporal Graph Neural Intelligence Network (FT-GNIN) is designed for traffic patterns mining and intelligent transportation prediction. The presented method involves adaptive fuzzy membership functions, multiscale temporal features extraction, learning the spatial dependencies with graphs and attention guided neural fusion. First of all, observed traffic data is normalized and converted into time-series. Afterwards, using fuzzy adaptive inference, uncertain traffic data is converted into fuzzy membership functions representation. In parallel, spatial dependencies are captured by the graph neural module, while temporal dependencies are identified by the temporal learning module. Attention guided fusion is applied for merging the three types of information and generating the prediction of traffic flow, congestion and travel times. At the evaluation step 30, FT-GNIN has the following performance metrics: MAE=2.58 mph, RMSE=4.42 mph, MAPE=7.35%, R2=0.94, and congestion classification F1-score=0.97. Compared with Graph WaveNet, our approach achieves relative improvements of 8.51% in MAE, 6.95% in RMSE and 8.13% in MAPE, as well as an absolute improvement of 0.03 in R2 and 0.01 in F1-score.

Authors

Mohammad Abdur Rasheed
Shaqra University, Saudi Arabia

Keywords

Deep Fuzzy Neural Network, Traffic Pattern Mining, Intelligent Transportation Systems, Spatiotemporal Prediction, Adaptive Congestion Forecasting