Predicting Pore Pressure at Varying Depths in a Norwegian Railway Project

Groundwater monitoring in the form ofpore pressure measurementsis widelyused toassess the impact of construction activities on the surroundings. Pore pressure datacan be used for monitoring groundwater leakage into construction pits and tunnels, as groundwater drawdown might lead to consolidation settlements and building damage depending on the magnitude, duration and spatial extent of the drawdown.
The monitoring process involves installing severalsensors some years beforethe start of groundworkto record seasonal fluctuations and thus establish sensor baselines. Additional piezometers are often installed right before or after the start of construction to cover for missing areas or to replacedamaged sensors.Data from the sensors can be used to generate triangulated/interpolated groundwater surfaces, for analyzing groundwater flow patterns in three dimensions,or for investigating areal extents of groundwater drawdown. However, with a limited number of installed sensors and large areas of interest, pore pressure estimates in areas between observation points are naturally uncertain.
New methods in the realm of machine learning canbe used for pore pressure prediction in such areas.In a Norwegian railway project with 160 installed piezometers, a pore pressure prediction model was tested using a randomforest machine learning algorithm. The model was trained on data from piezometersalong with availablespatial and geological data. Theresults show that the model has moderate but acceptable predictive accuracy with an average R² of 0.53 and a mean absolute error of 1.79 mH2O. In addition to being a possible solution forpredictingthe pore pressure at new locations and depths, such model predictions might also be used to validate data from sensors installed late in a project phase.

Keywords: Geotechnical engineering, Groundwater, Pore pressure, Machine learning, Prediction model, Site investigations.