Analysis of neural network models for predicting the movements of landslide restraint structures

The contemporary machine learning capabilities offer various methods and algorithms that facilitate the development of reliable prediction models for complex phenomena such as landslides. However, predicting landslides and surrounding structure movements is challenging and requires diverse information to enhance forecasting reliability. The unstable and dynamic conditions of any landslide render the application of physically derived models somewhat problematic. A solution can be found through neural network simulation. The presented study was carried out to examine the effectiveness of various neural networks in predicting the movement of landslide restraint structures based on the results of geospatial and environmental monitoring. The subject of the study is a complex of landslide restraint structures comprised of four retaining walls. Geospatial monitoring was conducted over six months with a time interval of two weeks, providing three-dimensional displacements for 84 observation targets. The displacements were determined with excellent accuracy, approximately 2 mm. Additionally, soil temperature variations and precipitation levels were observed daily and referenced to the geospatial observation epochs. Based on these data, the simulation of neural networks was accomplished with a comprehensive assessment of their performance. The research results facilitated the development of the simulation workflow and offered recommendations for a specific neural network application.