Rainfall-runoff modeling is a long-standing challenge in hydrological sciences, with various approaches attempted over time. However, the computational costs and reliability of existing models remain issues, leading to the use of less complex but less reliable models.
In this project, a Long Short-Term Memory (LSTM) network and an Autoencoder have been used to address this challenge. The goal is to assess whether the encoder, capturing the key characteristics of different hydrological basins as summary statistics, can improve predictions for basins with limited information.
For this analysis, the CAMELS dataset (Catchment Attributes for Large-Sample Studies) was chosen, consisting of daily meteorological data and measured discharge for 671 minimally disturbed catchments across the United States. The dataset spans a 30-year period from 1980, encompassing factors such as precipitation, temperature, and discharge. The basins are categorized into 18 hydrological units, reflecting variations in aridity and humidity.
Our findings indicate that the encoder improves the model's hydrological prediction performance, reflected in lower loss values and higher Nash-Sutcliffe efficiency (NSE) coefficients. The improvements are particularly evident in arid basins, although better performance is observed across all basin types.
This project has been developed in May/June 2022 by Group 1 for Laboratory of Computational Physics MOD B course (academic year: 2021-2022) at the University of Padova. The team was composed by:
- Cano Amaro German: https://github.com/german-cano
- Monti Sebastiano: https://github.com/MontiSebastiano
- Tonazzo Valentina: https://github.com/ValentinaTonazzo
- Zoppellari Elena: https://github.com/zoppellarielena