Advancing AI‐powered hybrid plants: reinforcement learning control of small‐scale solar Brayton turbines via surrogate modelling and fuel‐air ratio insensitivity discovery

Recently, sustainable energy sources have attracted significant attention from researchers, developers, companies, and investors. One of these sources is solar energy, particularly concentrated solar power (CSP) which is integrated with other conventional sources of energy. Small-scale solar Brayton turbines suffer from severe inlet temperature fluctuations due to varying solar radiation, demanding fast-adaptive control systems for higher station performance. This paper presents a Reinforcement Learning (RL) agent capable of controlling Fuel-to-Air Ratio (FAR) and mass flow rate (ṁ) to maximize electrical power output. An accurate surrogate model was built using Random Forest based on 80 operating points from ANSYS VISTA, achieving exceptional accuracy (R² = 0.999). The agent was trained using the Proximal Policy Optimization (PPO) algorithm for 30,000 time steps. Results show that the agent automatically learned to fix FAR at its minimum value (0.02) and raise ṁ to its maximum (0.07 kg/s). Sensitivity analysis revealed that changing FAR from 0.02 to 0.08 changes power by less than 1.11 %, and feature importance ranking confirmed that FAR has a negligible impact (importance < 0.1 %) on power output within the VISTA model and the analysed operating range. Compared to a well-tuned PID controller, the RL agent achieves a 36.4 % reduction in fuel consumption and 13.5 % improvement in the turbine's output power. Moreover, it outperformed random and fixed policies by 18 % and 12 % in power improvement and 46 % and 49 % in fuel savings, respectively. These findings open the door to radically simplifying control systems for small-scale solar Brayton turbines and achieving higher performance for the power station system.