Deep reinforcement learning-based stability control for biomass thermal-electric conversion system: Experimental analysis and multi-physics coupled modeling

Remote mountainous regions and temporary installations face dual energy challenges: grid infrastructure deficiencies cause unstable power supply, further exacerbated by the inherent limitations of conventional biomass power systems in dynamic control of energy conversion processes. This study experimentally investigated the impact of airflow on the performance of the biomass thermal-electric conversion system (BTECS) and developed an intelligent control framework using deep reinforcement learning for BTECS. The framework established multi-physics-coupled differential equations for gasification-heat transfer-power generation processes. Comparative evaluation was performed for four algorithms: Double Deep Q-Network (DDQN), Proximal Policy Optimization (PPO), Soft Actor-Critic (SAC), and Asynchronous Advantage Actor-Critic (A3C). Key findings revealed that an optimal airflow rate of 9.2 m3/h reduced the target power achievement time to 630 s (a 79 % reduction from 2980 s at 6.5 m3/h). Moreover, a physically interpretable simulation model was established via Newton's law of cooling, polynomial fitting, and multiple regression based on experimental data. The DDQN algorithm demonstrated superior performance compared to other algorithms and conventional controllers, achieving 98.32 % of the target power output (786.56 ± 30.85 W), with 2.4 times faster convergence and 65.5 % lower computational cost than SAC. This study demonstrates the technical superiority of deep reinforcement learning in optimizing the BTECS while offering implementable energy solutions for off-grid applications.