[A-HW25] Near Surface Investigation and Modeling for Groundwater Resources Assessment and Conservation

[AHW25-P04] Developing a Long-term Groundwater-level Prediction Model Using Data Mining and Conceptual Model

*You Cheng Chen¹, Liang Cheng Chang¹, Chin Wei Huang¹ (1.Department of Civil Engineering, National Yang Ming Chiao Tung University, Taiwan)

Keywords:Data mining, Conceptual model, Groundwater recession rate, Long-term prediction model

The groundwater can supplement the dry season's water supply since the groundwater resources are hydrologically more stable than the surface water. Moreover, impacted by climate change, the streamflow difference between dry and wet seasons will increase, and groundwater management is more critical than before. A reliable long-term groundwater prediction model is a vital tool for increasing groundwater management efficiency.
This study integrated data mining methods with groundwater conceptual models to develop a long-term groundwater level prediction model and applied the proposed model to the groundwater-level prediction in Pingtung Plain. The data mining methods include the time-frequency analysis to classify the groundwater-level recession mechanism and the groundwater-level retreat regression analysis. The groundwater-level recession is mainly caused by human pumping and natural groundwater retreat. Most of the wells' groundwater-level recession in Pingtung plain were either nature retreat dominated or pumping dominated. The groundwater-level rises due to rainfall recharge were modeled by applying the series linear reservoirs concept as simulating the catchment rainfall-runoff. The complete prediction model was the integration of the groundwater-level recession model and the linear reservoirs model. The prediction models for the nature retreat dominated wells were a single model, and those were monthly models for pumping dominated wells. The model parameters for each prediction model are obtained by the parameter optimization method to minimize the sum of the square errors between the simulated groundwater level and the observed groundwater level.
The developed long-term groundwater-level prediction model was applied to simulate the groundwater level in Pingtung Plain. This study applied time-frequency analysis to classify the recession mechanism for the groundwater level of each groundwater well as either nature retreat dominated or pumping dominated. Based on the classification, the single model or monthly model was used. The 2014–2018 data served as the training set for the calibration of the parameters, and the 2019 data (January to December) was the test set. The wells' predicted groundwater level in the unconfined aquifer upstream of the alluvial fan showed a high degree of conformity with the observed groundwater level. In Fangshan (1), Wanlong (1), and Xipu (1), the groundwater-level recession was nature retreat dominated, and the groundwater-level prediction accuracy achieved a high score of 0.899, 0.906, and 0.971. Some wells in the aquifers at the midstream and downstream of the alluvial fan exhibited pumping-dominated groundwater recession, such as Kanding (1), Meinong (1), and Dazhuang (1). The groundwater-level prediction accuracy achieved favorable results of 0.786, 0.778, and 0.901, respectively.
The model developed in this study accurately predicted the groundwater level of unconfined aquifer wells in the upstream alluvial fan that have abundant groundwater resources in general. For the groundwater levels in midstream and downstream regions, the model also has acceptable prediction accuracy. Therefore, the developed model is a valuable tool for regional groundwater management.