Icing Time Prediction Model of Pavement Based on an Improved SVR Model with Response Surface Approach

Abstract

Pavement icing imposes a great threat to driving safety and impacts the efficiency of the road transportation system in cold regions. This has attracted research predicting pavement icing time to solve the problems brought about by icing. Different models have been proposed in the past decades to predict pavement icing, within which support vector regression (SVR) is a widely used algorithm for calibrating highly nonlinear relationships. This paper presents a hybrid improved SVR algorithm to predict the time of pavement icing with an enhancement operation by response surface method (RSM) and particle swarm optimization (PSO). RSM is used to increase the number of input data collected onsite. Based on that, the optimal SVR model is established by optimizing the kernel function parameters and penalty coefficient with the particle swarm optimization (PSO) algorithm. The hybrid improved SVR is compared with SVR, PSO-SVR, and RSM-PSO for coefficient of determination (R2), mean absolute error, mean absolute percentage error, and root mean square error to check the effectiveness of PSO and RSM in optimizing SVR. The results show that the combination of two methods in the hybrid improved algorithm has a better optimization capability with R2 of 0.9655 and 0.9318 in a train set and test set, respectively, which outperforms PSO-SVR, RSM-SVR, and SVR. In addition, the R2 of the hybrid improved SVR and PSO-SVR both reach the optimal fitness value approximately at the iteration of 20, which suggests that convergence capacity remains relatively constant with the predictive accuracy being improved.