Modelos hidrológicos híbridos para a bacia do rio Paraíba do Sul: acoplando redes neurais artificiais com transformada wavelet para previsão de vazão em curto prazo com ênfase na previsão de vazões extremas

Abstract

The Paraíba do Sul River flows through the most important industrial region in Brazil, between the cities of Rio de Janeiro and São Paulo. According to the National Water Agency, the Paraíba do Sul river basin is characterized by conflicts of multiple uses of water resources (urban supply, sewage dilution, irrigation and hydroelectric power generation). Due to its importance, the development of accurate flow prediction models can assume strategic value for the management of the quantity and quality of water in this basin. These models can support decision-making by public managers on alerts for extreme flood or drought conditions and can become a key element to protect society and favor a timely reaction, effectively reducing socio-economic damage. Models based on artificial intelligence have been successfully applied to solve nonlinear problems in hydrology. In the last two decades, the wavelet transform, a data pre-processing technique, has been studied for use in time series analysis and has been shown to be very effective in non-stationary data. The wavelet transform divides the original time series into sub-components that serve as inputs to artificial intelligence models. Thus, they provide a time-frequency representation of a signal at different periods in the time domain, as well as information about the physical structure of the data. This research aims to contribute to the provision of a methodology for the development of two types of hybrid models called WANNone and WANNmulti coupling artificial neural networks with a data pre-processing technique based on a wavelet decomposition algorithm À T rous. The developed models are applied to forecast the average daily natural flow of 1, 3, 5 and 7 days ahead in 8 fluviometric stations located in the Paraíba do Sul river basin. The performance of models based on wavelets and artificial neural networks is compared with that conventional models of artificial neural networks (ANN). The results reveal that the performance of the WANNone and WANNmulti models are significantly superior to the ANN models in terms of MAPE, RMSE and R2 forecast efficiency measures at all horizons for all modeled stations. In terms of applicability to certain scenarios, the WANNmulti model is able to better represent low flow cycles, in drought scenarios, while WANNone can better follow the magnitude of high flows extremes, in situations of risk of flooding.