Modelos de aprendizado de máquinas híbridos aplicados à previsão de curto prazo da vazão do Rio Zambeze afluente à barragem hidroelétrica de Cahora-Bassa, em Moçambique

Abstract

Rivers are still one of the greatest and most important resources for the survival of humanity. They provide much of the water we consume and that we use to produce our food and for our hygiene. In addition, we use it to irrigate the soil in agricultural areas, for navigation, fishing, electricity production, in addition to maintaining biodiversity. Due to water scarcity and concern about its existence in the future, it becomes essential to develop studies that can help to understand its dynamics, in order to manage it. The Zambezi River is the fourth largest watercourse on the African continent and is of vital importance in the region, as it is the main source of electricity production in Mozambique, contributes to the development of the economy of the delta through the viability of economic activities such as agriculture, pastoralism, fishing, construction of access roads and, in reducing the risk of occurrence of natural disasters, such as drought and floods. Despite this need, studies of water resources for this region are still scarce. The present work presents the development of hybrid machine learning models for predicting natural flows of water bodies of the Zambezi River at the Cahora-Bassa Hydroelectric dam, using antecedent values of influent flows, precipitation, evaporation and relative humidity as input variables. Five models were considered in the analysis: Extreme Gradient Boosting (XGB), Extreme Learning Machine (ELM), Support Vector Regression (SVR), Elastic Net linear (EN) and Multivariate Adaptive Regression Splines (MARS) . Evolutionary/bioinspired optimization algorithms: Grey wolf optimization (GWO), Genetic Algorithms (GA), Differential Evolution (DE) and Particle Swarm Optimization (PSO) were used as selection strategies for all parameters internals of the models. And, in addition to evaluating the potential of the five hybrid machine learning models for flow simulation, this work also focused on two input variable selection methods, Least Absolute Shrinkage and Selection Operator (LASSO) and Partial Mutual Information variable selection (PMI), used to choose the subset that allows the best prediction of future values, and that provide comprehensive information about the influence of variables on the simulated flow. The developed hybrid models performed predictions for 1, 3, 5 and 7 days ahead on data collected at the Cahora-Bassa Hydroelectric dam in Mozambique. Combined performance measures and statistical tests were applied to evaluate the performance of the models. Uncertainty analysis techniques based on confidence intervals and Montecarlo simulation were also used to analyze the reliability of forecasts provided by the models. The results show that all models performed well to predict river discharge several steps ahead, indicating that the integration of evolutionary/bioinspired algorithms is a successful alternative to produce accurate discharge predictions. The hybrid model XGB with integrated GWO and its respective variations with selection of input variables, XGB-LASSO and XGB-PMI outperformed the hybrid models EN, ELM, MARS and SVR, having presented the best performance, respectively, for 1, 3, 5, and 7 steps ahead. These models provide reliable forecasts, and can be used as an alternative to aid in flow forecasting, which is crucial for sustainable or optimized management and planning of water resources in hydroelectric power plants.