AI-powered ground surface temperature forecasting for cold regions geotechnical applications

ABSTRACT: Ground surface temperature is an essential variable in cold region geotechnical engineering. Physics-based long-term simulations of surface energy budgets are associated with complexity, high variance, and computational intensity. This study proposes an alternative data-informed framework based on long short-term memory (LSTM) networks to predict ground surface temperatures from meteorological variables. The LSTM model was evaluated using monitoring data from a permafrost site in the Canadian Arctic and a mid-latitude non-permafrost site. Various aspects of the machine learning problem were studied using a series of sensitivity analyses. Long-term projections of ground surface temperature were presented for the two sites under both moderate and extreme climate change scenarios. Data scarcity was found to be one of the major challenges for the proposed framework. However, the growing number of stations and more reliable instrumentation will be in favor of data-driven methods. Provided that suitable training data are available, the data-driven framework shows several advantages over the physics-based simulations in forecasting ground surface temperature, and potentially other related variables.