BoilerNet: Deep reinforcement learning-based combustion optimization network for pulverized coal boiler

Reinforcement learning is considered a potential technology for the next phase of intelligent control. However, its reliance on trial-and-error learning prevents direct interaction between the agent and the physical boiler, driving the development of the digital twin boiler. To address the challenges of low prediction accuracy under transient loads and in time-consuming decision-making, we propose an efficient deep reinforcement learning combustion optimization network (BoilerNet), coupled with a digital twin boiler. The digital twin boiler integrates a multi-objective model employing advanced triangular convolutional neural networks (TR-CNN), which reduces model complexity by adjusting the width factor. To enhance decision-making efficiency, a combustion optimization agent based on soft actor-critic (SAC) was designed, with policy and value functions developed for the combustion state and manipulated variables. Simulation experiments using historical boiler data demonstrate that with a TR-CNN width factor of W = 0.25, the inference time was 11.894 μs, a 28.92 % reduction compared to the pre-improved model. Compared with the traditional deep deterministic policy gradient (DDPG), the SAC-based combustion optimized a greater portion of samples, achieving 99.36 % optimization, while DDPG achieved 89.98 %. Additionally, SAC increased thermal efficiency by 0.357 % and reduced NOx emissions by 20.244 mg/m3.