A Deep Learning-Based Cloud Groundwater Level Prediction System

Abstract

In the context of global change, understanding changes in water resources requires close monitoring of groundwater levels. A mismatch between water supply and demand could lead to severe consequences such as land subsidence. To ensure a sustainable water supply and to minimize the environmental effects of land subsidence, groundwater must be effectively monitored and managed. Despite significant global progress in groundwater management, the swift advancements in technology and artificial intelligence (AI) have spurred extensive studies aimed at enhancing the accuracy of groundwater predictions. This study proposes an AI-based method that combines deep learning with a cloud-supported data processing workflow. The method utilizes river level data from the Zhuoshui River alluvial fan area in Taiwan to forecast groundwater level fluctuations. A hybrid imputation scheme is applied to reduce data errors and improve input continuity, including Z-score anomaly detection, sliding window segmentation, and STL-SARIMA-based imputation. The prediction model employs the BiLSTM model combined with the Bayesian optimization algorithm, achieving an R2 of 0.9932 and consistently lower MSE values than those of the LSTM and RNN models across all experiments. Specifically, BiLSTM reduces MSE by 62.9% compared to LSTM and 72.6% compared to RNN, while also achieving the lowest MAE and MAPE scores, demonstrating its superior accuracy and robustness in groundwater level forecasting. This predictive advantage stems from the integration of a hybrid statistical imputation process with a BiLSTM model optimized through Bayesian search. These components collectively enable a reliable and integrated forecasting system that effectively models groundwater level variations, thereby providing a practical solution for groundwater monitoring and sustainable water resource management.