Seasonal forecasts for hydropower: downscaling of precipitation in South American basins

Hydropower is one of the industrial sectors more strongly affected by the timing and intensity of extreme climatic conditions, especially related to precipitation. Particularly, the recent expansion of the hydropower capacity in South American regions has raised the interest of this sector in seasonal forecasts that can be used to anticipate persistent precipitation anomalies (e.g. meteorological droughts) over specific basins. Some of the limitations for the generation of seasonal forecast products that can be integrated in hydropower decision-making processes are the coarse spatial resolution and the limited forecast quality (i.e. systematic errors, low skill) of the current operational seasonal forecast systems. To overcome this problem, we propose a methodology based on machine learning-informed analogs. Information-theoretic preprocessing has been used to identify the large-scale drivers of precipitation in a drainage basin located in southern Brazil, where hydroelectric energy is produced. This information allows us to exploit the ability of the dynamical seasonal forecast systems to predict these large-scale drivers in combination with the statistical link between these drivers and precipitation. We have employed the global CMCC-SPS35 seasonal forecasts at 1° spatial resolution and the CHIRPS-V2 precipitation dataset at 0.05° to produce dynamical-statistical seasonal forecasts of precipitation. The results show that downscaled forecasts exhibit higher skill and are affected by smaller biases than those obtained directly from the operational dynamical systems. This suggests that there is potential for the use of hybrid forecasts in the optimal management of South Brazilian hydropower production.