Bayesian optimized ensemble learning system for predicting conceptual cost and construction duration of irrigation improvement systems

Abstract

Linear construction projects, such as pipeline irrigation projects, are prone to delays and cost overruns owing to inaccurate cost and duration estimates. The research gap pertains to studies that concentrated exclusively on predicting costs or durations using backbox artificial intelligence models. Consequently, this study introduces an innovative approach that utilizes explainable artificial intelligence to forecast the conceptual cost and duration of irrigation projects simultaneously. This study analyzed data from 1,277 historical cases using factor analysis and stepwise regression to distill 25 parameters down to five key drivers. It evaluates 12 machine learning models, including multiple linear regression, artificial neural networks, and decision tree-based ensemble methods. Bayesian optimization was employed to fine-tune the performance of each algorithm. The light gradient boosting machine is identified as the most effective algorithm for cost prediction, with a Mean Absolute Percentage Error (MAPE) of 2.989 % and an Adjusted Determination Coefficient (R ⁎2 ) of 0.931. For duration prediction, the extremely randomized tree model stands out, achieving a MAPE of 2.533 % and an R ⁎2 of 0.961. The study further employs the Shapley additive explanation technique to improve the interpretability of the key drivers used for predicting both the budget and the timeline.