Abstract: In order to reduce the prediction error of occasional traffic congestion dissipation time, based on the whole process of the accident, the congestion dissipation time is divided into dwell time, disposal time and recovery time, and the regression models of dwell time and disposal time are established from the driver's personality characteristics, accident grade characteristic value and initial speed. The weight of RBF neural network is optimized by using linear decreasing time-varying weight and speed limiting improved standard particle swarm optimization algorithm. The RBF neural network model of occasional congestion recovery time is established by taking Transmodeler simulation data and measured data as training samples. The simulation results show that the average absolute error of the model is 245.3 s. The relative error of improved PSO-RBF network for prediction of recovery time is 11.2%, the root mean square error is 102.3, the average relative error is 38.1% and 23.8% lower than that of single RBF network and standard PSO-RBF network respectively.