Robust LSTM-Based River Stage Prediction Under Rainfall Data Incompleteness

Accurate and timely prediction of river stage levels during typhoon events is critical for effective flood risk management and emergency response. This study develops a data-driven forecasting model utilizing a Long Short-Term Memory (LSTM) network to simulate the rainfall-to-water-level relationship at the Lanyang Bridge in Yilan County, Taiwan. The model leverages historical hydrometeorological data acquired from a network of rainfall observation stations and a water-level gauge, encompassing ten major typhoon events recorded between 2014 and 2022. A cross-correlation analysis was conducted to identify optimal temporal lags between rainfall and river stage response, which informed data preprocessing for the LSTM model. Notably, the developed forecasting framework does not require rainfall forecasts, relying instead on recent observed rainfall and current water level, thereby mitigating errors introduced by precipitation prediction. To address situations of missing data frequently encountered during extreme events, a masking technique was incorporated into the model, allowing the model to handle incomplete input sequences robustly. The resulting model demonstrates strong potential for short-term river stage level forecasting in data-constrained and rapidly changing conditions.