Comparing MLP and LSTM neural networks for short-term water level forecasting: insights from the 2024 major flood in Southern Brazil

Abstract

ABSTRACT This study develops and compares two data-driven models for short-term water level forecasting based on a MLP and a LSTM network. Both models were trained for a six-hour forecasting horizon using hourly water level and rainfall data from seven stations in the Guaíba Lake watershed, southern Brazil, from November 2023 to May 2024, a period that includes the region’s most severe flood on record. A sliding-window learning scheme with incremental weight updates was employed to improve temporal consistency and computational efficiency. Four experiments assessed the effects of different input configurations, recursive forecasts, and model architectures. Results indicate that the MLP performed comparably to, and slightly better than, the LSTM, with MAE and NSE values supporting its suitability for near-real-time forecasting. Overall, the findings show that relatively simple neural architectures, when trained with adaptive schemes, can deliver robust and efficient predictions for operational early-warning systems.