Hydropower Generation Forecast Using a Bayesian Hierarchical Model

Abstract

The ability to properly forecast the power generated by hydroelectric power units have been gaining increasing interest in the literature, especially in Brazil, where hydropower represents 58% of the power generation matrice. Several works in the literature have tried to forecast the generation of hydropower, both in large and small hydroelectric power plants. The atmospheric variables, specially the ones related to rainfall, are commonly used to forecast hydropower. Additionally, there is a need to consider the uncertainties related to the covariates and the final generation forecast. With this in mind, this work proposes a Bayesian Hierarchical Model to forecast the Hydropower Generation of both larger and small plants. A Bayesian approach is used given its ability to deal with uncertainties and missing data in the learning process, as well as the forecast process. Several weather conditions are used as covariates to feed the model. With data from 2020 and 2021, the generation from 2022 is predicted. Results shows the effectiveness of the proposed work, especially in face of disturbance in the input variables, since the forecast achieved a normalized mean squared error (NMSE) of less than 0.39 even in an extremely disturbed scenario.