Comparative Performance of Deep Learning and Statistical Models for River Stage Forecasting

Accurate river water level prediction is vital for early flood warning and effective water resource management. The traditional methods of river water level prediction based on Autoregressive Integrated Moving Average (ARIMA) and linear regression struggled to capture the nonlinear and dynamic nature of hydrological systems. Therefore, this study evaluates the performance of deep learning models, namely Long Short-Term Memory (LSTM), Basic LSTM, Stacked LSTM, Bidirectional LSTM, and LSTM with Dropout, alongside classical models such as Linear Regression, Moving Average, and ARIMA across five river monitoring stations. Model performance was assessed using Root Mean Square Error (RMSE), Mean Absolute Error (MAE), Mean Absolute Percentage Error (MAPE), Mean Squared Error (MSE), and R-squared (R2) metrics on test data. The results consistently showed that the LSTM with Dropout model outperformed all other approaches in terms of predictive accuracy and generalization, achieving the lowest RMSE and MAE values and the highest R2 scores at four out of five stations. The standard LSTM model achieved the best performance at one station but was generally outperformed by its dropout-regularized counterpart. In contrast, classical models exhibited significantly lower performance, particularly in capturing temporal dependencies, with R2 values often below zero. These findings validated the effectiveness and robustness of LSTM-based deep learning, particularly with dropout regularization, in modeling and forecasting river water level dynamics across diverse hydrological conditions. The study supports the integration of advanced machine learning models in river monitoring systems to enhance flood preparedness and mitigation strategies.