Development of Physics-informed Recurrent Neural Network to Predict Actual Groundwater Level Fluctuation according to Precipitation Time-series Event

A physics-informed neural network (PINN) is developed to predict groundwater level (GL) fluctuations based on precipitation time-series data, integrating both physics-based principles and data-driven learning to improve the prediction accuracy and robustness. The proposed PINN model embeds the governing equations of groundwater flow dynamics within a gated recurrent unit (GRU), ensuring that predictions adhere to physical laws while leveraging historical data patterns. The model’s performance is evaluated against two benchmark models: (i) a purely physics-based linear reservoir model and (ii) a data-driven GRU model. The results demonstrate that the PINN model outperforms both benchmarks, particularly under reduced time resolution, maintaining stable accuracy through its integration of physics-based information. Quantitative metrics, including the root mean squared error (RMSE) and correlation coefficient (CC), confirm the superior predictive capability of the PINN model, indicating its resilience to data limitations and noise in real-world monitoring data. As such, this study underscores the advantages of incorporating physics information into neural networks, and demonstrates that the PINN approach provides robust predictions even with limited data, which makes it ideal for complex aquifer systems and endows it with significant potential for supporting real-world groundwater management.