Regional groundwater sequential forecasting using global and local LSTM models

Study region: Choushui River alluvial fan in central Taiwan. Study focus: Groundwater contributes significantly to various socio-economic developments worldwide. Unique hydrogeological characteristics at a wider spatial scale and unavailability of adequate hydro-meteorological or pumping datasets have made its prediction a challenging prospect over regions. This study proposes and evaluates a Long Short-Term Memory (LSTM) model to predict daily fluctuations of regional groundwater patterns using observations across multiple monitoring wells in Choushui River Alluvial Fan located in Central Taiwan. Global models calibrated over all monitoring well data were compared with the local LSTM models each trained on the local well data for regional groundwater sequential forecast. The global LSTM model can ensure great precision in groundwater forecasting; however, the global model may not quite be necessary if the piezometric observations have a nearly similar temporal autocorrelation in groundwater sequence in the study area. New hydrological insights for the region: The proposed local LSTM models that were trained on local well data to predict groundwater over the rest of the wells exhibited nearly identical performance when compared to the global model. However, an in-depth spatial assessment showed that the results from the local model were adversely affected when trained on data from coastal wells indicating that these areas require special attention for regional forecasting. Eventually, a fine-tuning through Transfer Learning (TL) scheme was further proposed to show considerable improvements in model performance, and to highlight the potential advantages of synthesizing cross-data information for model training when suffering from hydrogeological heterogeneity in fields.