Groundwater levels and pore pressure distributions are important input data for many engineering calculations in civil engineering and mining. Direct measurements and numerical models are the main tools for obtaining this input, each having different strengths and drawbacks. Models provide a continuous distribution in space and allow forecasting but are associated with a model-to-measurement misfit. Direct observations are usually very accurate and allow real-time monitoring but are spatially sparsely distributed.
Consequently, optimal knowledge of hydraulic heads requires combining both methods. The standard method is model calibration. This method can - if performed wisely – reduce the model predictive error associated to a minimum. At the end of the process however, the model’s output still contains a bias that propagates into further engineering decisions and can reduce their robustness. Also, assimilation of new measurement data into the model that is continuously collected in the field traditionally requires recalibration of the model. This process still requires high effort of labour and time, making it less suitable for real-time monitoring and fast decisions.
To face these challenges, this presentation demonstrates how the results from a calibrated model are augmented using a bias correction method based on geostatistical principles. Based on past performance of the model, the likely spatial and temporal distribution of the model error is estimated at unsampled or forecasted locations. By correcting for this bias, we can estimate the most likely “real” system state based on Bayesian principles including uncertainty margins.
The process is illustrated using case studies on generating design water levels and pore pressure distributions for civil engineering, pit dewatering and slope stability projects.
The presentation shows why augmented estimates are generally more accurate than those of the calibrated model or interpolated data alone, especially in complex hydrogeologic environments like mining areas. Additionally, we obtain confidence intervals for the estimates, which is useful for choosing appropriate safety factors and proposing new measurement locations.