Exploring continuity information in sparse Data: Mechanism-Informed data augmentation strategy to enhance prediction accuracy of Anammox process models.

The stability of the anaerobic ammonium oxidation (Anammox) process, crucial for low-carbon wastewater treatment, relies on effective control of nitrite levels and other conditions. Accurate prediction of effluent nitrite and related indicators is essential for process control, yet sparse data make it challenging to train a machine learning (ML) model. This study introduces a mechanism informed data augmentation (MIDA) framework to address the issue of data scarcity. MIDA assesses sampling adequacy by considering process time constants and the Nyquist criterion, and then employs cubic spline interpolation to expand datasets. When applied to our Anammox reactor data, MIDA reduced the mean squared error of six ML models to 5% of its original value and improved the R-square from 0.45 to 0.97. Transferability was validated through enhancements on four independent datasets from varied reactors. Noise injection analysis indicated that gains stem from continuous trend information in the augmented data, and not merely from increased sample size. MIDA can provide a solution to small-sample challenges, facilitating real-time control of Anammox process and other biological processes.