基于动态模型与神经网络的微燃机在混合燃料

Abstract Micro gas turbines (MGTs) are emerging as a versatile off-grid power solution for remote regions and islands, and as a reliable power source for micro smart grids, offering a viable alternative to conventional centralized generation and transmission. This study introduced a novel hybrid performance prediction framework that integrates a physics-based white box model (WBM) with a data-driven black box model (BBM), enabling high-precision and computationally efficient evaluation of MGT performance. Taking a 100-kW MGT as the reference system, both regenerative cycle (RC) and simple cycle (SC) configurations were analyzed across ambient temperatures of 258.15 K, 273.15 K and 288.15 K, while systematically varying the hydrogen volume fraction (HVF) in natural gas (hydrogen from 0 % to 100 %). To further enhance predictive capability, advanced machine learning algorithms (KNNR, RFR, Extra Trees and BR) were trained for key performance metrics, and multi-objective optimization was performed using NSGA-III and MOEA/D to achieve simultaneous improvement of efficiency and operational flexibility. The results revealed that RC configuration significantly outperform SC configuration in fuel efficiency and waste heat recovery, particularly at high HVFs, with the RFR model delivering the highest predictive accuracy. The proposed fast-start MGT energy storage coupling mechanism demonstrated strong potential for autonomous, resilient power supply in isolated microgrids. By combining thermodynamic modeling, machine learning and multi-objective optimization, this research provides a reproducible methodology for the scientific community to accelerate the design, operation and control of hydrogen-based hybrid-fuel MGT systems. The framework is readily extensible to other distributed energy technologies, supporting their future integration with renewable generation, decarbonization strategies and intelligent energy management systems.