The rapid growth of the COVID-19 epidemic put the education system in a problematic condition for examination analysis of students, particularly when the virtual engagement was challenging for learners. With the shift to mandatory virtual education during the outbreak, it has become especially more crucial to predict these students and provide educational programs to prevent dropping them ahead. This prediction is done by the different Data Mining (DM) and Artificial Intelligence (AI) algorithms in the previous research. However, only a limited study was available to estimate learners who fail in learning during the epidemic. Additionally, there is a need for students’ performance data while participating in virtual learning during the outbreak. From this perspective, this article presents an automated student performance prediction framework, which implements partially available student learning records in the online education system. This framework encompasses four different phases: data gathering, attribute selection, prediction and performance analysis. At first, the student’s database including student’s sex, age, overall time spent in virtual classrooms, their emotions towards Covid-19, etc., is collected. After that, a new metaheuristic optimization algorithm called Tuna-Swarm Optimization Algorithm (TSOA) is introduced to choose the most relevant attributes and minimize the data dimension. Once all the relevant attributes are selected, an Improved Conditional Generative Adversarial Network-based Ensemble Deep Long Short-Term Memory (ICGAN-EDLSTM) classification system is developed to learn those attributes and predict the student’s outcome during the Covid-19 outbreak. Finally, these algorithms are applied to the 3 different benchmark databases to evaluate their prediction efficiency compared to the other classical algorithms.