Modelling Temporal Dynamics and Uncertainty in River Water Levels Using <scp>TP</scp> ‐Transformer: Advancing Hydrological Process Understanding

ABSTRACT Reliable river water level forecasting is essential for hydrological risk management and process understanding. This study presents TP‐Transformer, a transformer‐based model integrating a time perception attention mechanism and Monte Carlo dropout to capture temporal dependencies and quantify predictive uncertainty. Unlike standard sequence models, TP‐Transformer adaptively weights historical inputs and provides confidence intervals, enabling both accurate forecasts and interoperability. Applied to over 60 years of daily water level data from Yichang Station on the Yangtze River, the model outperforms state‐of‐the‐art baselines across short‐, medium‐, and long‐term horizons. Uncertainty analysis reveals periods of high predictive variance, offering insights into hydrological dynamics and model reliability. This framework contributes to advancing data‐driven hydrological modelling by combining temporal learning with uncertainty quantification and is transferable to other basins and variables.