Revolutionizing Groundwater Level Prediction in Taiwan: Unleashing the Power of Transformer Neural Networks

Abstract

    In the face of evolving global weather patterns attributed to climate change, precise prediction of groundwater levels is increasingly essential for effective water resource management. This significance is particularly pronounced in regions like Taiwan, where groundwater is a pivotal water source. This study focuses on the Zhuoshui River basin in central Taiwan and explores a Transformer Neural Network (TNN) based on a 20-year hydrometeorological dataset at a 10-day scale to predict groundwater levels. Our investigation reveals that the innovative TNN model outperforms conventional models, such as the Convolutional Neural Network (CNN) and the Long Short-Term Memory neural network (LSTM). The TNN model's superiority is evidenced by its enhanced predictive capabilities, as measured by metrics like R2 and MAE. Notably, the TNN model excels in providing precise forecasts (MAE < 1 m) for the majority of groundwater monitoring stations, notwithstanding challenges in areas facing overexploitation.     This groundbreaking study marks the first attempt of the TNN model to predict groundwater levels, showcasing its robust performance and broad applicability. The TNN model emerges as a valuable tool for groundwater level prediction, contributing to sustainable groundwater management and effective resource utilization amid the backdrop of climate change. With the potential to address climate-related challenges, the TNN model stands as a pivotal asset for optimizing strategies in groundwater resource management.