The research field of weather forecasting is dominated by numerical models. Despite presenting mathematical modeling of atmospheric dynamics very close to reality, these models have their limitations, as the modeling is based on approximations and a large volume of data is necessary to generate better results. This work proposes the minimization of the prediction error of numerical models through the application of machine learning (ML) algorithms: Random Forest (RF), Support Vector Regression (SVR), and Multilayer Perceptron (MLP) considering an ensemble forecasting methodology. The proposed strategy differs from the classical approaches by weighting the final results of the ML models to favor the best predictor. The results showed that this methodology reduces the variance of the results in the application of a cross-validation analysis and that the combination of a numerical model with hML algorithms can generate an improvement of up to 23% in the coefficient of determination and 29% in the root mean squared error of the predictions when compared with a purely numerical model.