Multi-Step Ahead Prediction of Freezing Depth via Deep Learning with Long Short-Term Memory

Accurate predictions of freezing and thawing depths could help set up an accurate time for seasonal load restriction (SLR), a load limitation policy that transportation agencies apply to reduce road damages during spring-thaw seasons. Mechanistic or empirical models have been used to predict the freezing and thawing depths. However, these model-based approaches cannot efficiently utilize the rapidly increasing data. This study attempts to apply recurrent deep learning to develop a data-based approach for the prediction of freezing depth. For this purpose, one frequently used algorithm in multivariate time series predictions, that is, long short-term memory (LSTM), is employed considering its ability to preserve historical information and dependencies. An LSTM network was constructed and implemented using the Keras deep learning library in Python for multi-step predictions of freezing depth with weather information as input. This new data-driven approach was tested on available data from road weather information sites in Michigan. Based on the training and validations with different model parameters, the optimum LSTM architecture was obtained for the application. This study showed that the proposed method can obtain satisfactory results for multi-step ahead predictions of the freezing depth.