The classification of voltage sags source location as downstream (DS) or upstream (US) of a monitor is a significant issue that should be taken into account when establishing mitigation strategies. Given the weak accuracy -bellow 90%- of single or two criteria analytical methods that are usually applied to locate sags, the application of intelligent methods is highly desirable. Therefore, this paper presents two classifiers of the Support Vector Machine (SVM) (three kernels) and Ensemble (three learners and nine methods) using genetic algorithm (GA) and a 10-fold cross-validation (CV). These methods throught extensive simulations on a Brazilian regional utility showed a 96.28% classification performance with Polynomial-SVM and a 99.11% performance for Decision Tree (DT)-Ensemble with the Totally Corrective Boosting (TotalBoost) method. Also, a comprehensive strategy to enhance the SVM accuracy and to keep the Ensemble performance by fewer appropriate features (which determine relative location of voltage sags sources) is presented. After testing three different feature selectors, an effective forward selection applied to the Polynomial-SVM concluded five appropriate optimum features and improved the accuracy of SVM up to 98.6%. The obtained optimum features applied to Ensemble showed a 99.2% performance in the DT-Ensemble-TotalBoost. Using the minimum obtained optimum features, a novel analytical rule based on maximum wins strategy has been proposed as well.