Probability prediction of pavement surface low temperature in winter based on bayesian structural time series and neural network

Reliable pavement surface temperature prediction models are important for decision making in winter road maintenance. Due to the complexity of thermal environment in winter, there is inevitably uncertainty in pavement temperature. Therefore, the prediction results of pavement surface temperature should not be a single value, but a probability distribution. This paper developed a prediction model for evaluating the probability distribution of pavement surface temperature in winter. The model consisted of two modules, namely, a Bayesian Structural Time Series module (BSTS) and a Bayesian Neural Network module (BNN). The BSTS module provided a relatively rough prediction of pavement surface temperature, while the BNN module provided refined prediction based on the results from the BSTS module. The model was validated with data collected from the field test. The impact of model structure, model parameters, and probability threshold on the predictions was analyzed. Results show that most measured pavement temperatures fell into the 95% confidence interval of the predicted values, indicating the prediction was reliable. For the BSTS module, the regression component plays the most important role in the prediction. The seasonal component also has obvious influence on the prediction results. While the influence of local linear trend component on the prediction result is not obvious. For the BNN module, although the slippery index and pavement surface condition do not directly contribute to the heat transfer process, these two factors have correlation with the pavement surface temperature. The value of probability threshold should be properly determined to balance the reliability and computational cost of the prediction model. The model presented in this paper is useful for transportation agencies in winter road maintenance