Abstract
In recent years, damage caused by heavy snow has been seen in various places. Various plans for road maintenance work such as snow removal work are currently being formulated based on Disaster Prevention Weather Information announced by the Meteorological Agency and based on the rules of thumb of experienced engineers. However, Disaster Prevention Weather Information covers a wide area and is strictly a forecast. Even with the empirical rules of experienced engineers, it is difficult to estimate the condition of the road surface at key monitoring points. If the road surface temperature can be predicted with a high degree of accuracy, the optimum antifreeze spray time will become clear, leading to the prevention of slip accidents. For the purpose of optimizing the application timing of road deicing agents, data obtained from IoT sensors on roads that are intensively monitored by road administrators and Disaster Prevention Weather Information Data published daily are combined and used for predicting road surface temperature up to 26 hours ahead. And we aim to improve the accuracy of the learning model. We construct Feed-forward Neural Network and Recurrent Neural Network to confirm the prediction accuracy of the road surface temperature. And we also confirm the effect of ensemble learning. As a result, it was found that the air temperature and wind speed have an effect on the prediction of the road surface temperature. In the Feed-forward Neural Network, the combination of road surface temperature, air temperature, wind speed, and Disaster Prevention Weather Information Data resulted in the prediction with the smallest error. In the Recurrent Neural Network, the prediction with the smallest error was obtained when using GRU. It was found that the prediction accuracy can be improved by taking the average of both Neural Network models.
