Machine Learning-Driven Cost Estimation and Resource Allocation Optimization in Industrial Engineering Projects: Evidence from Albania

Abstract

Cost overruns and inefficient resource allocation remain persistent problems in industrial engineering projects, particularly in transition economies where historical data are scarce and price volatility is high. This study develops an integrated predict-then-optimize framework that couples machine learning (ML) cost estimation with risk-aware resource allocation, and demonstrates it on a calibrated case study of 420 industrial projects (manufacturing plants, logistics facilities, agro-food processing, energy, water treatment and mineral processing) in eight Albanian regions over 2015–2025. Five learners’ multiple linear regression, support vector regression, random forest, artificial neural network and gradient boosting were trained in a direct formulation and in a hybrid formulation that learns the deviation from a conventional unit-rate estimate indexed to the national construction cost index. Under a strict temporal hold-out (training 2015–2022, testing 2023–2025), the hybrid gradient boosting model reduced the mean absolute percentage error from 19.1% for the conventional estimate to 8.2% (R² = 0.974), whereas direct tree-based models degraded under post-2021 price escalation. Design completeness at tender, site difficulty and region were the dominant cost-risk drivers. Conformalised quantile regression provided P10–P90 cost intervals that were fed into a mixed-integer linear programme selecting execution modes and start times for a 14-project portfolio under three resource capacities. In Monte-Carlo evaluation the risk-aware schedule lowered expected portfolio cost by 3.1% (≈ EUR 2.4 million), reduced the expected number of late projects by 26% and cut expected delay-months by 67% relative to earliest-due-date practice, with the largest gains when resources were scarce. The framework offers project-based firms and public clients in Albania a transparent, reproducible decision tool.