Daily streamflow forecasting in Sobradinho Reservoir using machine learning models coupled with wavelet transform and bootstrapping

Improving forecasting techniques for streamflow time series is of extreme importance for water resource planning. Among the available techniques, those based on machine learning models have sparked the interest of the scientific community in recent decades due to their good adaptability and forecasting accuracy, particularly when coupled with data pretreatment methods. These algorithms include artificial neural networks (ANNs) and support vector machines (SVMs). In the present study, a comparative analysis of a set of machine learning models, namely, an ANN and an SVM coupled with wavelet transform and data resampling with the bootstrap method, was performed. The analysis of these models was performed with daily streamflow time series for Sobradinho Reservoir, which is located in northeastern Brazil, corresponding to the period between 1931 and 2015. The results show that higher accuracy is achieved by the ANN than by the SVM. Furthermore, the best combination for predicting the streamflow into the Sobradinho Reservoir was the bootstrap, wavelet and neural network (BWNN) method. For a forecast 3 to 15 days ahead, this method yielded a mean square error (MSE) 15 to 25 times lower than that of the purely neural model (ANN), and the R-2 and mean absolute error (MAE) coefficients increased by more than 7 to 28% and 13 to 50%, respectively. Therefore, the BWNN is an alternative approach for future time series forecasting studies. (C) 2021 Elsevier B.V. All rights reserved.

Automated summary

This study conducted a comparative analysis of a set of machine learning models, including an ANN and an SVM coupled with wavelet transform and data resampling. Results showed that the ANN outperformed the SVM in terms of accuracy, with the best performing combination being the BWNN method. The BWNN method yielded lower mean square error and higher R-2 and MAE coefficients for streamflow forecasting 3 to 15 days ahead.