An adaptive framework to enhance microscopic traffic modelling: an online neuro-fuzzy approach

Because of various environmental factors (e.g. road type and traffic congestion) and the involvement of human action (e.g. drowsiness and consciousness level), the time-variant nature of the car-following process necessitates the use of adaptive modelling approaches. In contrast with the existing car-following models with a fixed structure, this paper proposes an adaptive framework based on an online local linear neuro-fuzzy model, supported by a recursive singular spectrum analysis signal-processing technique, to model the time-variant car-following behaviour in a microscopic traffic flow. The online local linear neuro-fuzzy model is initially trained by a set of offline data and then is adapted to the car-following data by means of an adaptive weighted least-squares technique. Furthermore, the recursive singular spectrum analysis technique is employed to decompose the traffic data in an online manner and then to remove useless components (e.g. the measurement noise) to produce well-behaved data. The proposed synergistic approach is applied to real-world car-following data, collected at the Hollywood freeway section of the US 101 Highway. The empirical results demonstrate that the developed approach successfully describes the car-following behaviour while conventional offline models fail in the case of large variations in the traffic data or congestion in the traffic flow.