Comparative Analysis of Statistical and Recurrent Neural Network Models for Short-Term River Level Forecasting in the Paraguay River

The accurate short-term forecast of river levels is crucial in mitigating flood and drought risks and supporting the ecosystem, transportation, and national production. However, hydrological models face significant challenges due to seasonal variations in river levels influenced by rainfall patterns and land use changes. In this context, this study applies statistical and deep learning models to time series data to predict water levels in the Paraguay River, at the port of Asunción. This work studies the performance of Autoregressive Integrated Moving Average and three recurrent-based deep learning approaches - Recurrent Neural Network, Long Short Term Memory, and Gated Recurrent Unit models-to determine which one is most suitable for predicting water levels. The experimental results indicate the superiority of recurrent-based models, e.g., Long Short Term Memory and Gated Recurrent Unit, over the traditional Autoregressive Integrated Moving Average statistical model, especially when incorporating other station levels as past covariates and forecasting with various horizons ranging from 7, 14, 21, and 28 days. The recurrent-based deep learning models achieved lower Root Mean Squared Error values of 16.7 cm, 33.8 cm, 49.36 cm, and 60.9 cm for the respective forecasting horizons. These findings emphasize the critical importance of including additional station levels as past covariates to enhance the accuracy and reliability of the forecasting models.