A physics-informed neural network framework for plant-wide prediction and dynamic control of wastewater treatment

Mechanistic models and machine learning methods provide powerful capabilities for simulating and controlling wastewater treatment processes; however, their application to real-time control faces major challenges due to the high computational cost of the former and the limited physical consistency of the latter. Accordingly, this study develops a physics-informed neural network (PINN) framework that integrates a physics-constrained surrogate model with a rolling optimization loop for pollutant prediction and dynamic control. Evaluation against ten critical water quality indicators demonstrates high predictive accuracy, achieving an average R² of 0.8860. Comparative analysis with the ASM2D model reveals that the proposed PINN not only enhances prediction accuracy but also reduces computational costs by approximately 786-fold, underscoring its strong potential for real-time deployment. Steady-state multi-objective optimization further validates the computational efficiency of the PINN, yielding a distinct Pareto front. Subsequently, the integration of dynamic optimization with the rolling prediction mechanism facilitates rapid control parameter updates (0.83 s). This strategy reduces energy consumption by 40.3% while ensuring compliance with discharge standards, demonstrating significant advantages in energy efficiency. Biochemical kinetic analysis confirms that the superior performance stems from precise regulation of core reaction rates, facilitated by the hybrid structure of the PINN that combines the computational efficiency of machine learning with the interpretability of biochemical kinetics. Overall, this study provides a reliable intelligent decision-support tool for WWTP operation under varying influent conditions within the same plant configuration.