Proceedings of the 2019 Mississippi Water Resources Conference

Year: 2019 Authors: Abrokwah K., O'Reilly A.M.

Groundwater is an important resource that is extracted every day because of its invaluable use for domestic, industrial, and agricultural purposes. The need for sustaining groundwater resources is clearly indicated by declining water levels and has led to increasing demand for modeling and forecasting accurate groundwater levels. In this research, results of wavelet analysis-artificial neural network (WA-ANN) data-driven models for simulating groundwater levels are compared to the results of aquifer water levels simulated using physics-based MODFLOW models. That is, we compare the results of each model to understand how WA-ANN model results relate to real physical hydrogeologic properties, e.g. hydraulic conductivity (K) and specific storage (Ss), with the objective of using physical principles (theory) to guide development of data-driven models. These techniques are explored by modeling groundwater levels in a synthetic alluvial aquifer system consisting of a river and an unconfined aquifer, two confined aquifers, and confining layers separating each aquifer. A synthetic time series representing daily values of recharge to the unconfined aquifer, with seasonal and shorter time-scale periodicity, is used as the forcing function. Properties of K and Ss for aquifers and confining layers as well as river properties are assigned assumed values in the MODFLOW model. Simulated water levels from MODFLOW at an observation well in each aquifer are then modeled again with a WA-ANN model for each well by using various decomposition levels of the discrete wavelet transform (DWT). The DWT is used to decompose the recharge time-series data into various levels of approximate and details wavelet coefficients, which are then used as inputs for the WA-ANN models. The results for the various DWT decomposition levels of the WA-ANN models are then compared to the simulated water levels and hydrogeologic property values of the MODFLOW model. Based on this comparison, potential relationships are identified between the characteristics of the various decomposed wavelet levels of the WA-ANN models and the hydrogeologic properties of the MODFLOW model. The resulting knowledge of the underlying physics manifest by the WA-ANN models, inferred by comparison with a physics-based model, ultimately produces theory-guided data-driven models, imparting science-based consistency and interpretability into data science models.