Identification of regional groundwater net-recharge rate using expert system-a case study of Pintung plain

Groundwater parameter estimation can be classified into two categories: trial and error methods and auto-calibration methods. Trial and error methods usually are time consuming. Most auto-calibration techniques are optimization techniques which rely on sensitivity analysis or a gradient search, which become computationally difficult with increasing parameter dimension. Another problem with auto-calibration is setting up the mathematical equations for the optimization problems can be difficult and can be sensitive to initial conditions. This paper develops a robust and rapid parameter identification model, named RPIS, using the combination of the expert system model and the groundwater simulation model to reduce the computational time and increase the model applicability. The developed model is applied to identify the net recharge rate, the summation of total recharge and total extraction of the study area, of Pintung plain in southern Taiwan. The Pintung plain has an area of 78 km × 30 km composed of 3 aquifers and complex geological conditions. The study area is divided into 104 parameter zones with a planning horizon of 12 months. The study area has 1248 (104 × 12) net recharge rates to be calibrated. To solve for net recharge identification on an Intel Core2 Quad 2.66 GHz with 4GB RAM requires only 1539.9 seconds. The result shows that the proposed model can efficiently calibrate a model with high parameter dimension. The result also shows that the RMSE of each of parameters is less than 1.1 meter and the estimated net recharge rate of shallow aquifer and pumping rate of deep aquifer are consistent with the estimation result of UCODE. Moreover, in order to examine the local solution problem caused by initial guesses, this study generates 20 sets of initial guesses by using uniform random distribution and applies RPIS and UCODE to calibrate the parameters. For each of the 20 initial conditions RPIS meets the calibration criteria with less than 150 iterations, while UCODE failed to calibrate the study area for all but one of the initial conditions. This result shows that RPIS is a robust calibration method that is able to meet calibration criteria independent of the initial guess. For the one set of initial conditions that UCDOE was able to meet the calibration criteria required 6195 MODFLOW simulations compared to 121 simulations by RPIS. Based on these results RPIS is more efficient than UCODE for over parameterized models. This study demonstrates the correctness and practicability of RPIS, and RPIS can be a good solution for calibrating a large scale model with high parameter dimension.