Forecasting high road embankment settlements based on machine learning and geodetic measurement data

The paper describes the research on forecasting of road base settlements on high embankments by means of geodetic observations, and also considers the question of geodetic data representation in the form of a mathematical model which takes into account the laws of settlements development and also allows to reveal upward or downward trends of the observed deformations. The relevance and novelty of the research is that there are no requirements for making forecast models by the geodetic observation data. Whereas it is noted that building of such models is important for rational planning of repair works and more accurate determination of geodetic observations periodicity. The aim of the study is to develop a method of geodetic monitoring of foundation settlements on high embankments. The following machine learning methods were considered: linear regression, adaptive linear regression, exponential smoothing, support vectors, and random forest. After models were trained and their parameters were chosen, the forecast on a test set was carried out. The parameters that characterized the quality of the models were: root mean square error, mean absolute error percentage, mean error percentage, determination coefficient, and Theil's coefficient of variance. Minimal value of root mean squared error RMSE = 0,8 mm was obtained by the method of support vectors. The support vectors method showed the highest forecast accuracy when taking into account all indicators of model quality assessment.