Early warning systems (EWSs) have proven to be useful in identifying students at risk of failing both online and conventional courses. Although some general systems have reported acceptable ability to work in modules with different characteristics, those designed from a course-specific perspective have recently provided better outcomes. Hence, the main goal of this article is to design a tailored EWS for a conventional course in power electronic circuits. For that purpose, effectiveness of some common classifiers in predicting at-risk students has been analyzed. Although slight differences in their performance have only been noticed, an ensemble classifier combining outputs from several of them has provided to be the best performer. As a major contribution, a novel weighted voting combination strategy has been proposed to exploit global information about how basic prediction algorithms perform on several time points during the semester and diverse subsets of student-related features. Predictions at five critical points have been analyzed, revealing that the end of the fourth week is the optimal time to identify students at risk of failing the course. At that moment, accuracies about 85%-90% have been reached. Moreover, several scenarios with different subsets of student-related attributes have been considered in every time point. Besides common parameters from student's background and continuous assessment, novel features estimating student's performance progression on weekly assignments have been introduced. The proposal of this set of new input variables is another key contribution, because they have allowed to improve more than 5% predictions of at-risk students at every time point.