Deep Learning-Based Downstream Water Level Prediction Enhanced by Upstream Predict Information

Climate change and urbanization have increased the precipitation variability and extreme hydrological events, highlighting the need for accurate river water level prediction. This study proposes a two-step sequential prediction framework based on a Long Short-Term Memory (LSTM) model and evaluates the impact of hydrological connectivity among observation stations on predictive performance. In Step 1, water levels at upstream and downstream stations are predicted. In Step 2, these predictions are incorporated as additional inputs for forecasting water levels at a target station. Input variables are selected using information gain (IG), and multicollinearity is assessed with the variance inflation factor (VIF). Results show that at Pojin Bridge, where short-term fluctuations are significant, incorporating predicted upstream and downstream water levels improves the coefficient of determination (R2) by approximately 3.9% to 9.24% as lead time increases. In contrast, at Andong Bridge, where hydrological responses are relatively stable, the additional inputs reduce model performance. These findings indicate that the effectiveness of incorporating hydrological connectivity depends on station-specific characteristics. The study provides practical guidance for designing data-driven river forecasting models under varying hydrological conditions.