Improving Nowcasts of Road Surface Temperature by a Backpropagation Neural Network

Abstract

Accurate numerical prediction of road ice is important in cutting winter road maintenance costs, reducing environmental damage from oversalting, and providing safer roads for road users. In this paper, the error of road surface temperature nowcasts (3 and 6 h ahead) by an automated numerical (Icebreak) model is regarded as a time series and is further treated by a three-layer neural network (NN) to increase accuracy of the nowcasts. The network is trained by an error-backpropagation algorithm with updated 7-day memory and progressively reduced learning rate. Its learning is based on historical and preliminary meteorological parameters measured at an automatic roadside weather station. The effectiveness of the network in improving the accuracy of numerical model forecasts was tested at both normal and problematic forecast sites in Austria, Italy, Japan, Norway, Switzerland, and England. Results of the tests show that the NN technique is able to reduce root-mean-square error of temperature forecasts and increase the accuracy of frost?ice prediction. The improvements were minor at normal sites but significant at problematic sites where complex environmental conditions and underlying nonlinear mechanisms are currently unresolvable by operational numerical models.