Machine learning for uncertainty quantification in climate precipitation prediction over South America

Abstract

In recent times, machine learning algorithms have found applications in various fields. Nice results have been obtained by employing these algorithms, including precipitation climate prediction. Precipitation is a key meteorological variable in South America, related to the intense rainfall under different climate seasons. For Brazil, the importance is beyond agriculture, civil defense, and tourism, because most part of the electric energy is produced by hydroelectric power plants. The inputs are some meteorological variables (wind fields, temperature, and precipitation from the previous month). Decision tree is the approach applied for monthly precipitation prediction, as implemented in the Light Gradient Boosting Machine (LightGBM) framework, after the prediction, another important issue is to quantify how good is the forecasting, in other words, it is important to estimate the predictability -- or uncertainty quantification to the prediction. Therefore, two decision trees were designed to estimate the precipitation and the prediction uncertainty over South America. Optimal hyperparameters for the machine learning tools were determined by Optuna optimizer. Data from the years January-1980 up to December-2017 were addressed for algorithms training, while the years 2018 and 2019 were used for testing. Very good results were obtained for precipitation prediction, with uncertainty maps calculated for all South America territory.