Comparative Study of Neural Networks and Support Vector Regression for Final Project Duration Prediction using Simulated Scheduling Scenarios

The prediction of the duration of a construction project is a highly complex task influenced by multiple factors. To tackle the heterogeneity and noise of historical datasets, this paper proposes the use of a highly controlled, simulation-generated dataset comprising 157 realistic schedules of school construction projects. The schedules were generated using six core planning variables. A distinctive contribution of this research is the systematic evaluation of the combinatorial performance between data normalization techniques (Min-Max and Z-score) and learning algorithms (Support Vector Regression and Artificial Neural Networks). The models were globally optimized using cross-validation to prevent data leakage. Comparison between the results of the SVR model with an RBF kernel under Min-Max normalization, and a compact ANN with one hidden layer using the Levenberg-Marquardt training function, demonstrated high theoretical accuracy within the simulated boundaries. The results indicate that SVR yields superior generalization accuracy across the analyzed scenarios. Furthermore, a conceptual reliability-aware framework, ScheduleRelia, is introduced. ScheduleRelia serves as a proof-of-concept application designed to estimate project duration and assess schedule health using critical path metrics. While broad generalization requires future external validation with real-world inputs, this study establishes a rigorous baseline for algorithm selection and data formatting in schedule forecasting.