Artificial Intelligence and Soft Computing-Based Multi-Objective Optimization for Risk and Efficiency Management in Thermal Power

Running a thermal power plant means constantly balancing efficiency, reliability, and emissions—three goals that are deeply connected in practice, yet often treated separately in conventional optimization. This study introduces an integrated AI framework designed to handle them together, not one at a time. The idea is to find solutions where all three improve simultaneously, rather than advancing one at the expense of others. We tested this using three years of operational data (2022–2024) from three Iranian power plants with different technologies: Neka (combined-cycle), Shazand (steam turbine), and Montazeri (gas turbine). The dataset includes more than 36,000 daily records and 1,100 monthly reports, covering heat rate, fuel use, equipment failures, and continuous CO₂ measurements. A deep neural network learned to predict plant performance under different conditions, achieving R² values above 0.90 across all indicators. A fuzzy logic system assessed operational risk by considering equipment health, maintenance backlog, and operating severity. At the core, a hybrid optimizer combining NSGA-II and MOPSO searched for the best trade-offs among efficiency, reliability, and emissions. The results were consistent and meaningful. Across the three plants, heat rate dropped by 2.8%, saving about 25,000 tons of fuel annually. Mean time between failures increased by 18%, and critical equipment failures fell by 15%. CO₂ intensity decreased by 4.2%, avoiding roughly 45,000 tons of emissions per year. Notably, one plant achieved its largest emission reduction despite modest efficiency gains—a reminder that under certain fuel and combustion conditions, emissions can be partially uncoupled from efficiency. Each plant responded differently: the combined-cycle unit led in efficiency, the steam turbine plant excelled in emission cuts, and the aging gas turbine showed the biggest reliability gains. Rather than inconsistency, this reflects the framework's ability to adapt to each plant's unique circumstances.