Recognition and analysis of voltage sags (dips) allow network operators to predict and prevent problems in real-life applications. Clearing the voltage sag source by direction detection methods is the most effective way to solve and improve the voltage sags and their related problems. However, the existing analytical methods use single or two input features as phasor-based (PB) or instantaneous-based (IB) values. Hence, their limited maximum accuracy is given at 93% and 84% when using PB features for noiseless and high-level noise signals, respectively. To increase the detection accuracy, the main contributions of this research by proposing machine learning (ML) methods include: (a) Developing nine supervised methods including support vector machine (SVM)-based, tree-based, others, and an ensemble learning of said methods, and providing a comparative analysis (b) Employing a set of PB, IB, and both PB and IB input features as noiseless and noisy; (c) Finding the best developed supervised methods by highest possible accuracy under subsets said in (b); (d) Proposing a new unsupervised method fed by both PB and IB features using a sparse principal component analysis (SPCA) applied to a k-means clustering with an internal SPCA approach. The proposed unsupervised schema does not use the upstream/downstream labels in developed supervised methods. Extensive simulations of voltage sags due to fault and transformer energizing on a Brazilian regional network show that regardless of the sag sources, input feature subset, and noise levels, the random forest (RF) models yield the best performance so that noiseless-RF (99.84%) using both PB and IB features is the most effective one. The proposed unsupervised method outcomes an overall accuracy of 99.17%-noiseless and about 90% for high-level noises. This performance is higher than analytical methods, very close to SVM-based supervised methods, and uses no predefined labels. Moreover, the results of Slovenian field measurements confirm the effectiveness of the best-developed supervised methods and the proposed unsupervised learning.