LSTM-based prediction and early warning of nitrogen removal in a partial denitrification-anammox bioreactor for municipal wastewater.

Intelligent operational optimization of the partial denitrification and anammox process is crucial for achieving efficient and stable nitrogen removal, yet it presents a challenge due to the system complexity and sensitivity. This study proposes and experimentally validates an integrated Long Short-Term Memory (LSTM)-based framework, designed for both predictive performance assessment and proactive risk management. The framework comprises a high-accuracy predictive model for estimating total nitrogen (TN) removal efficiency, achieving a root men square error of 5.89% using operational data sampled at 2-h intervals. Concurrently, an auxiliary real-time early-warning model was developed to proactively forecast potential violations of effluent discharge standards for ammonia nitrogen, nitrate nitrogen, and TN, demonstrating an overall classification accuracy of 97.96% Feature importance analysis identified the influent nitrogen concentration (specifically nitrate and TN) and the mixing ratio as the primary determinants governing system performance. By integrating these analytical insights with real-time early-warning signals, an optimal control strategy was formulated and successfully implemented. Validation results confirmed the efficacy of strategy in regulating operational parameters and its robustness in mitigating the risk of effluent nitrogen exceedance, thereby ensuring consistent compliance with discharge standards. This work presents a powerful and intelligent control tool to enhance the operational stability and ensure the effluent compliance of PDA systems, thereby contributing to advancing the practical application of PDA technology for municipal wastewater treatment.