Climate Indices Impact in Monthly Streamflow Series Forecasting

Abstract

Hydroelectricity has been widely deployed in many countries for decades, and remains the main source of electricity in Canada, Norway, and Brazil. Despite the recent diversification, hydroelectricity accounts for approximately 65% of the electricity generated in Brazil. It also represents the largest non-polluting source in the country, and is important for complying with the global goals of reducing carbon emissions. Meeting energy and environmental sustainability requirements impose challenges on the hydroelectric sector in terms of resilience to climate change observed around the planet. These changes affect electricity generation by altering seasonality and increasing the variability of streamflow and evaporation losses in reservoirs. Therefore, in this study, among a set of 27 climate indices, we identified the most relevant for improving the performance of the models applied to monthly seasonal streamflow series forecasting. A database provided by the NOAA Physical Sciences Laboratory was used as exogenous variables for three machine learning models (support vector regression, extreme learning machine, and kernel ridge regression) and one linear model (seasonal autoregressive integrated moving average with exogenous factors, SARIMAX). Random forest with recursive feature elimination was used as the feature-selection technique. The results obtained allowed for the identification of the most relevant set of indices for the analyzed plants, thereby improving streamflow predictions.

Keywords

• Climate indices
• machine learning
• random forest
• streamflow forecasting