Development of a Recursive Prediction Model for Non-Stationary/Non-Gaussian Aquifers through History Curves Matching
Wednesday, 17 December 2014
Deficiency of direct information has been the main obstacle in the subsurface characterizations, which may be supplemented by the secondary information. In the present study, a recursive subsurface prediction algorithm based primarily on direct lithological information and transient pressure changes is proposed by combining a non-parametric geostatistical model and groundwater flow simulation under Bayesian framework. The general performance of the proposed model is comparatively analyzed against that of ensemble Kalman filer (EnKF), which is one of the most popular model for predicting subsurface properties (e.g., permeability) recursively through the assimilation of transient measurements (e.g., groundwater head with time) and the corresponding physical model. For the demonstration of the proposed method, an assumed geologic map containing four different lithologies with the high categorical non-stationarity is employed and a hypothetical hydraulic conductivity map is formulated based on the lithological map. In the predictions, it is hypothesized that the information only at 45 locations are available and the rests are unknown. In the comparisons of two competing models, 48, 96 and 148 conditional realizations are generated to examine the model sensitivities to the number of the realizations. Additionally, a hypothetical pumping test was performed based on the assumed hydraulic conductivity to acquire history curves at a few observation wells. Both models commonly show gradual improvements in the prediction quality. However, the proposed model shows overall outperformance in terms of the improvement rate and the convergence, especially, in the predictability for spatial connectivity of the highly permeable structures. Besides, the predictions from the proposed model shows less sensitivity to the number of input realizations indicating computational advantages over EnKF. Therefore, it may be concluded that the predictions of the proposed model supplements those from the EnKF in various ways with relatively trivial computational burden. Due to the modularity, it is expected that the proposed model can be applied to many practical problems which have resemblance to the demonstrated one once the physical models are provided.