Hybrid Deep Learning with Multi-Source Data Fusion for Deep Groundwater-Level Prediction and Anomaly Detection

Deep groundwater dynamics, influenced by multiple factors, exhibit multi-scale periodicity, nonlinearity, and non-stationarity, posing significant challenges for reliable groundwater level prediction and anomaly detection. To address this issue, this study develops a hybrid deep learning model for simulating and predicting normal groundwater dynamics, which integrates multi-source observations—including barometric pressure, solid tide, and historical groundwater levels—with convolutional neural networks (CNNs), Long Short-Term Memory (LSTM), and Attention mechanisms. An Exponentially Weighted Moving Average (EWMA) control chart is then applied to the prediction residuals for statistical anomaly detection. The results show that CNN kernels capture multi-period local features induced by barometric and tidal effects; LSTM gating mechanisms learn long-term temporal dependencies; and the Attention mechanism dynamically assigns weights to both temporal steps and input features. Compared with single-source inputs or standalone LSTM/CNN models, the hybrid model enhances prediction accuracy (MAE < 0.02, RMSE < 0.03, R2 > 0.99), while integration with the EWMA control chart helps to identify the possible anomalies. Collectively, these results provide a reliable means for predicting deep groundwater levels and detecting hydrological anomalies under complex multi-scale periodic forcings, holding significant practical value in groundwater resource management and disaster warning.