Flight delays in commercial aviation have become a common problem worldwide due to increasing air transportation demands, resulting in higher costs and continuous adjustments in flight management. Efficiently predicting these delays is a recurring research topic in air traffic operations. Machine learning has emerged as a robust alternative to classical statistical methods, particularly Artificial Neural Networks (ANNs), known for their modular adaptability. However, varying study scenarios yield different accuracy outcomes. This study aims to define an appropriate prediction and classification scenario for regular departures at an airport, considering influencing variables, acceptable accuracy levels, and the most effective prediction model. It proposes a class-based predictive system applicable to São Paulo International Airport (SBGR) and the national airline Gol Linhas Aéreas (GLO). Bibliometric review and meta-analysis were employed to establish a baseline reference for method accuracy evaluation. Furthermore, studies were categorized using Data Envelopment Analysis by metafrontier to identify ideal prediction scenarios. The analyses reveal that the MultiLayer Perceptron (MLP) Neural Network demonstrates better predictive efficacy for route or airline-specific analyses, irrespective of delay reasons. A comparative case study was conducted for the defined prediction scenario, confirming that classifying delay time yielded the highest accuracy regardless of delay reasons. It was found that the specificity of scenarios and delay reasons within the chosen analysis area had a stronger relationship with predictive accuracy than the volume of data obtained.