A gradient boosting surrogate-assisted multi-objective optimization framework for steam thermal power plants

Steam thermal power plants, as long-standing pillars of electricity generation, are increasingly under pressure to adapt to a landscape shaped by environmental imperatives, fluctuating fuel economics, and growing energy demands. Their inherent operational complexity—stemming from a multitude of thermodynamic, economic, and environmental interactions—necessitates sophisticated, multi-faceted optimization approaches that transcend conventional methods. In response to this challenge, the present study introduces a hybrid, data-driven optimization framework tailored to the real-world operational intricacies of a Steam Power Plant in Iran. A high-resolution simulation model was first constructed using Steam Pro software, achieving close conformity with empirical plant data (RMS errors < 0.08 %), thereby establishing a reliable foundation for subsequent modeling phases. Drawing upon this validated platform, a comprehensive dataset was generated to train and evaluate four machine learning algorithms—namely, Linear Regression, Ridge Regression, Random Forest, and Gradient Boosting—across six core performance indicators: net electricity output, thermal efficiency (LHV), levelized cost of electricity (LCOE), CO₂ emissions, total exergy destruction, and cumulative exergy loss. Among the tested models, Gradient Boosting emerged as the most robust, exhibiting superior generalization with determination coefficients up to R² = 0.9912. These surrogate models were subsequently embedded within a multi-objective optimization framework incorporating three state-of-the-art metaheuristic algorithms: NSGA-III, Multi-Objective Multi-Verse Optimizer (MOMVO), and Multi-Objective Grasshopper Optimization Algorithm (MOGOA). Optimization scenarios ranged from tri-objective to six-objective formulations, facilitating nuanced trade-off analyses between energetic, economic, and environmental metrics. Results underscore the exceptional performance of NSGA-III, which consistently identified optimal operational regimes characterized by maximum power output (1205,060 kW), minimal CO₂ discharge (171.64 t/h), and competitive electricity costs, closely approaching the theoretical ideal point. In contrast, MOMVO prioritized cost minimization (0.04311 $/kWh) at the expense of marginally higher emissions. Feature sensitivity analyses revealed that turbine inlet conditions (x3, x5, x13), excess air ratio (x17), and pump isentropic efficiencies (x1, x6, x9) exerted the most pronounced influence across objectives. The proposed methodology by harmonizing accurate machine-learned surrogates with advanced evolutionary optimization offers a scalable, interpretable, and computationally efficient pathway for real-time performance enhancement of large-scale thermal systems. It paves the way for integrating smart control strategies, digital twins, and predictive maintenance mechanisms in modern power infrastructure, and stands as a replicable blueprint for data-driven optimization in other energy-intensive domains.