Related Experiment Videos
From black-box prediction to transparent process control: A physics-informed interpretable AI framework for total
Pengfei Li1, Jiaqi Li1, Ning Deng2
1College of Information, Shanghai Ocean University, Shanghai, 201306, PR China.
Water Research
|July 28, 2026
Summary
This study introduces a novel physics-informed dual-attention BiLSTM-Kolmogorov-Arnold network (PI-DABiLSTM-KAN) for accurate effluent total nitrogen (TN) forecasting in wastewater treatment plants (WWTPs). The model enhances prediction reliability by integrating physical laws and attention mechanisms for better operational management.
Area of Science:
- Environmental Engineering
- Wastewater Treatment
- Artificial Intelligence in Environmental Science
Background:
- Accurate effluent total nitrogen (TN) forecasting is critical for wastewater treatment plant (WWTP) operational stability and compliance.
- Existing data-driven models struggle with nonlinearities, temporal dependencies, and physical coherence, limiting prediction reliability.
- High-dimensional monitoring data in WWTPs present challenges for effective variable selection and relationship modeling.
Purpose of the Study:
- To develop a physics-informed deep learning framework (PI-DABiLSTM-KAN) for accurate and interpretable effluent TN prediction.
- To enhance model physical plausibility by incorporating soft constraints and process-informed descriptors.
- To improve the identification of key variables and long-range temporal dependencies for reliable forecasting.
Main Methods:
- Proposed a physics-informed dual-attention BiLSTM-Kolmogorov-Arnold network (PI-DABiLSTM-KAN) framework.
- Utilized LightGBM-SHAP for feature selection and constructed ASM1/ASM3-informed descriptors for nitrogen transformation.
- Implemented feature and temporal attention mechanisms, a KAN decoder, and a two-stage adaptive training strategy with soft physics constraints.
Main Results:
- Achieved high prediction accuracy on test and external validation datasets (R² > 0.96, RMSE < 0.36 mg/L, MAE < 0.10 mg/L).
- Demonstrated robustness to noise perturbation (<0.8% R² degradation) and stable seasonal performance.
- Physics-informed training reduced physical violations from 4.21% to 1.87%, and identified key control variables for risk reduction.
Conclusions:
- The PI-DABiLSTM-KAN framework offers a reliable and physically plausible approach for effluent TN prediction in WWTPs.
- The integration of physics-informed learning and attention mechanisms significantly enhances forecasting accuracy and interpretability.
- This decision-support framework aids in operational screening and risk management for enhanced WWTP performance.
Related Concept Videos
01:28Bioreactor Controls-I
Maintaining optimal conditions within fermenters is essential for maximizing microbial productivity and ensuring process efficiency. This lesson focuses on key parameters—temperature, foam, pH, carbon dioxide, oxygen, and pressure—and their precise measurement and control strategies in fermentation systems.Temperature ControlTemperature regulation is critical due to the exothermic nature of many fermentation processes. In small laboratory fermenters, temperature is commonly monitored using...
01:27Upstream Processing
Upstream processing represents a critical phase in biomanufacturing, wherein biological systems such as microorganisms, mammalian cells, or insect cells are cultivated to produce therapeutic proteins, vaccines, enzymes, or other biologically derived products. This phase encompasses all steps from the selection and genetic manipulation of the production organism to the cultivation of cells in bioreactors under tightly controlled environmental conditions.Host Selection and Genetic OptimizationThe...