Lightweight Deep Neural Network Design for Edge-Based Gas Turbine Efficiency Monitoring Using Multi-Objective Secretary Birds Optimization Algorithm

Abstract

This research presents a comprehensive framework for optimizing artificial Deep neural networks to predict MS5002B gas turbine efficiency using advanced multi-objective metaheuristic optimization techniques. The study systematically compares three nature-inspired algorithms to determine the optimal Deep neural architecture that balances predictive accuracy against computational efficiency. The Secretary Bird Optimization Algorithm (SBOA), inspired by the unique hunting behavior of secretary birds, which combines strategic walking patterns with precise strikes, demonstrated exceptional performance in navigating the complex search space of Deep neural architectures. Through rigorous experimentation, SBOA yielded an optimal network configuration of layers with a learning rate of 0.1, achieving near-perfect prediction accuracy (R² = 0.999998) while maintaining the fastest training time of 4.475 seconds among all evaluated algorithms. The research incorporates critical physical constraints, particularly the zero-power-to-zero-efficiency relationship, to ensure thermodynamic validity in all model predictions. The resulting optimized Deep neural network provides a powerful tool for real-time performance monitoring, operational optimization, and predictive maintenance in gas turbine power generation systems.