Forecasting Slope Stability with a Digital Twin: Performance Evaluation

This study details a case from Norway, where a digital twin of a slope was developed utilizing real-time monitoring, numerical modelling, and machine learning (Piciullo et al., 2024). Moreover, a performance evaluation of the digital twin in predicting the Factor of Safety (FoS) is presented. This digital twin is deployed on a web-based platform called NGI Live, functioning as a cloud service to daily forecast the stability of the modelled slope. The FoS values are projected for the subsequent three days, using monitored hydrological variables (Heyerdahl et al., 2018) and meteorological data sourced from publicly available datasources. NGI Live aggregates data from diverse sources, storing and displaying it in real time via online dashboards. Hydrological parameters, including volumetric moisture content (VWC) and pore water pressure (PWP), are monitored at various slope locations, and these data were used to validate the numerical model of the slope and calculate FoS for different time frames. The modelled FoS values, alongside the monitored hydrological variables and meteorological conditions, were employed to train two distinct data-driven models: one using Polynomial Regression (PR) and another utilizing Random Forest (RF). These pre-trained models serve as proxies for the numerical model within the cloud service to forecast FoS values using forecasted hydrological and meteorological variables. Meteorological forecasts are obtained from publicly accessible sources, while hydrological variables are forecasted using a Python-based package called Pastas (Collenteur et al., 2019), which utilizes the historical relationship between monitored variables and meteorological conditions. A performance evaluation has been carried out to compare the FoS forecasted with the machine learning models (PR and RF) and the GeoStudio numerical model, assessing their accuracy and reliability in predicting slope stability. This method is readily exportable and holds potential for implementation at additional sites for landslide early warning.

Keywords: Internet of Thing; landslide; digital twin; forecast; real-time; hydro-geotechnical modelling.