On Parameter Identification for Better Predictions of Dam Behaviour

Sammanfattning: Numerical modelling is often needed as a tool to predict the behaviour and assess the safety of dam structures. Embankment dam structures analyses are quite complex and potential failures are hazardous. Predictions of dam behaviour by numerical modelling rely on knowledge about the mechanical properties of the materials the dam is constructed with. The materials included in a dam vary significantly because zones in the dam have different functions. In order to conduct reliable modelling, parameter values defining the stress-strain relationship of the materials are needed to be assigned. Obtaining information about the mechanical behaviour in already existing embankment dams is usually challenging. As many dams are old, there might be a limited amount of information available of the materials used, construction methods and mostly about the stress-strain relationship of the soil. Traditionally, field sampling is performed in order to obtain such information. However, conventional field sampling might negatively affect the dam body and thereby the performance as well as the safety of the dam. This is of special importance if sampling is performed in the impervious (core) part. Since traditional sampling might harm the dam body, use of non-destructive methods would be advantageous to utilise for obtaining information about the stress-strain relationship and the strength in a dam structure. An option for a non-destructive method is parameter identification by inverse analysis. The idea of inverse analysis is to calibrate finite element models towards field measurements. In the calibration process, the input for a stress-strain relationship (constitutive model) is modified until the discrepancy between the output of the numerical model and the associated chosen field measurement is minimised. The agreement between output from the numerical model and reality is measured by an objective function that will calculate the error. In order to automatically search for the minimum a search algorithm is utilised in the optimisation process. When the objective function is minimised, the calibration of the material parameters is done. In previous research at Luleå University of Technology, the method of inverse analysis was applied to an embankment dam. The finite element program PLAXIS was used in combination with an optimisation code. The optimisation code includes an objective function (for error evaluation) and a search algorithm. The genetic algorithm was employed as search algorithm, since it is known for its robustness and efficiency as well as the fact that it provides a set of solutions instead of one unique answer. This is beneficial from a geotechnical point of view, since engineering judgement can be included in the final choice of solution.    The first study in the present thesis deals with a case study of an embankment dam, where a simple model calibration was performed. This was a part of a larger study, at the ICOLD Benchmark Workshop in 2017, where the work presented here was forming one of the contributions. In order to have a model response similar to reality, the contributors were asked to choose constitutive models and calibrate them. The calibration was done by manually changing the input for the constitutive model chosen. While the response of the numerical finite element model was capturing the trends of measured total stresses and pore pressure in the dam quite well, there were difficulties in capturing the long term deformations of the dam. This was a challenge for all contributors. An idea for improving the model response, is to run a more advanced calibration by inverse analysis. In the second study in the thesis, predictions are presented for the embankment dam that inverse analysis was previously conducted for at LTU. Strengthening actions in form of a new berm were performed at the dam. With identified material parameter values from the inverse analysis, predictions were conducted both before and after the strengthening measures. The predicted deformations were compared to deformation data from inclinometer measurements.  A reasonably well agreement was obtained with the real deformations. The trend of the deformations was replicated and the magnitudes of the deformations were in the right order. The study is indicating that predicting future dam behaviour based on results from inverse analysis can be done reasonably well. In the third and final study in the thesis, effects of random measurement error on the performance of the genetic algorithm for soil parameter identification are assessed. Also here, with the application to the embankment dam used in previous research at LTU. Optimisations were performed against inclinometer measurements. To be sure that the constitutive model can find the correct solution, synthetic (i.e. numerically generated) inclinometer data was utilised. Perturbations were randomly generated within chosen intervals of error and added to the numerically generated deformations. The genetic algorithm showed its robustness, by continuing to search for solutions without breaking down even if the field data was substantially perturbed. Considering usual errors for inclinometer measurements, the genetic algorithm can deliver good solutions. The inclinometer errors used were taken from literature, and thereafter related to the perturbations of the numerically generated data. Dealing with errors that are becoming gradually larger than what can be considered as usual, problems are faced by the genetic algorithm. In this cases it is difficult to find a solution, and if solutions are found they might significantly deviate from the unperturbed optimum solution. The three studies handled in this thesis are treating aspects of back analysis of embankment dams; from a simple calibration, to predictions based on material parameters from advanced inverse analysis and finally effects of errors on the genetic algorithm. It been shown that using inverse analysis for already existing embankment dams is very beneficial for the material characterisation and is forming a step towards better predictions of future dam behaviour.