An Explainable Bayesian TimesNet for Probabilistic Groundwater Level Prediction
Zhen Peng, Shaoxing Mo, Alexander Y. Sun, Jichun Wu, Xiankui Zeng, Miao Lu, Xiaoqing Shi
Abstract
Abstract Reliable groundwater level (GWL) prediction is essential for sustainable water resources management. Despite recent advances in machine learning (ML) methods for GWL prediction, further improvements may be made in uncertainty quantification and model interpretability. This study proposes Bayesian TimesNet (BTimesNet), a novel deep learning model for probabilistic and explainable GWL prediction. BTimesNet transforms 1D time series data into 2D matrices based on periodicity, enhancing the capture of temporal patterns through convolutional filters. A Bayesian framework using Stein Variational Gradient Descent is implemented to quantify predictive uncertainties. For model interpretability, SHapely Additive exPlanations (SHAP) is utilized to quantify predictor contributions. The efficacy of BTimesNet for multi‐step‐ahead GWL prediction is evaluated using monthly data collected from 19 monitoring wells across three hydroclimatic regions in the U.S., and compared against the widely used long short‐term memory (LSTM) and Autoformer models. Results show that BTimesNet consistently outperforms LSTM and Autoformer, providing more accurate predictions up to 4 months ahead. SHAP analysis reveals that historical GWLs are the most informative features, with meteorological predictors making secondary contributions. BTimesNet's superior performance stems from its ability to extract both short‐ and long‐term temporal features. This approach represents a valuable advancement for accurate GWL prediction and risk‐informed decision‐making, providing critical lead time for proactive groundwater and ecosystem management and agricultural irrigation planning. Its data‐driven nature also facilitates broader applications across hydrological and environmental prediction domains.