Artificial Neural Networks for Flood Prediction in Current and CMIP6 Climate Change Scenarios

Abstract

Researchers have widely applied discharge simulation using artificial neural networks (ANNs) and have gained prominence in water resources. Morphological features, watershed urbanization, and climate change influence hydrological variables. Thus, data-driven models need to be able to identify the hydrological relationships without explicitly stating the physical processes. The main objectives of this work were (i) to evaluate an ANN Multilayer Perceptron for flood forecasting in an urban basin and its efficiency for several lead times; (ii) to evaluate discharge variation considering climate change scenarios. The study applied the methodology in a basin occupied by the Cerrado biome, with its intermediate outlet in an urban area that suffers from recurrent floods. The selection of climate change models followed from the Coupled Model Intercomparison Project Phase 6 scenarios Shared Socioeconomic Pathway (SSP)2-4.5 and SSP5-8.5 for two futures: 2021–2050 and 2071–2100, with the period of 1976–2019 as reference. The model obtained satisfactory results for the discharge prediction at the current time and for a horizon of up to 4 days. However, forecasts for longer lead times led to metrics deterioration. Furthermore, future projections suggest decreased discharges, more extreme events, and increased short return-period floods. The developed model is valuable for short-term forecasting and water resources management in the face of changing climates.