An AI Modelling Framework for Preconstruction Cost Estimation in Power-Sector Projects Using Expert-Based Predictor Selection

Reliable cost estimation during the preconstruction phase remains challenging because decisions must be made with limited project information, substantial uncertainty, and project-specific cost drivers. This study proposes and empirically evaluates an AI modelling framework for preconstruction cost estimation (PCE) in power-sector projects that incorporates expert-based predictor selection with predictive and probabilistic modelling. A Modified Delphi process involving 20 experts validated cost predictors for three power-sector project domains: Power Generation, Transmission and Distribution (T&D), and Associated Infrastructure. The resulting domain-specific predictor sets were mapped to a historical database of 300 completed projects and evaluated using Bayesian Network (BN), Multiple Linear Regression (MLR), and Random Forest (RF) models under a common validation framework. The final hold-out evaluation showed that MLR achieved the lowest root mean square error (RMSE) and the highest coefficient of determination (R2) across all three domains. In contrast, RF achieved the lowest observed mean absolute percentage error (MAPE) in all three hold-out samples. Although less accurate for point prediction, the BN provided posterior probability information and interpretable probabilistic dependency structures. Cross-method analysis further showed consistency within the present dataset in identifying Installed Capacity, Transformer Capacity, and Project Size as the principal cost predictors for Power Generation, T&D, and Associated Infrastructure, respectively. Collectively, the results demonstrate a clear trade-off among predictive accuracy, probabilistic information, and interpretability, rather than identifying a universally superior modelling approach. The resulting framework, therefore, moves beyond algorithm comparison by linking expert-based predictor selection with empirical model evaluation to support the assessment of complementary modelling capabilities across project types and decision requirements. This integration provides a transparent and context-sensitive basis for supporting preconstruction cost-estimation decisions under uncertainty.