Intelligent control and optimization by reinforcement learning (RL) agents have emerged as a promising framework for biological nutrient removal (BNR) processes. However, most existing studies remain confined to simulation environments, limiting their credibility and engineering relevance. In addition, the black-box nature of RL agents hinders operator trust. This study establishes a reactor-agent integration that enables direct interaction between physical bioreactors and an RL-based agent, demonstrating that agent-based control outperforms conventional strategies in managing influent disturbances. Experimental results show that, under short-term perturbations, RL-based control achieved approximately 30% reduction in nitrogen species exceedance compared with knowledge-based control, while simultaneously realizing a 36.5% decrease in operational cost via coordinated adjustment of dissolved oxygen (DO) set points and internal mixed liquor recirculation. To enhance interpretability, this study proposes an analysis framework that bridges algorithmic intelligence with process engineering transparency. It combines controller-action visualization, surrogate decision trees, Sobol sensitivity analysis, and decision-trajectory analysis to elucidate the RL agent's decision logic and relate it to process kinetics and control principles. These insights transform the RL policy from a black box into an interpretable, process-consistent, and auditable control strategy. Overall, the results demonstrate the feasibility, robustness, and transparency of agent-based intelligent control under real-world conditions, paving the way for its reliable deployment in full-scale WWTPs.
Zhu et al. (Tue,) studied this question.