Enhanced prediction of partial nitrification-anammox process in wastewater treatment by developing an attention-based deep learning network

In the process of partial nitrification and anaerobic ammonia oxidation (anammox) for nitrogen removal, the process offers simple metabolic pathways, low operating costs, and high nitrogenous loading rates. However, since the partial nitrification-anammox (PN-anammox) process combines partial nitrification and anammox reactions within the same reactor, strict control of dissolved oxygen (DO) is essential. Additionally, assessing treatment performance through chemical measurement involves time lag, making it challenging to recover the biological process when issue arise, especially in the PN-anammox process, where strict DO control and the sensitivity of anammox bacteria to conditions and substrates demand timely intervention. Therefore, modelling for PN-anammox process is of great significance. Because of traditional modelling methods have limitations in this process, in this study, deep learning networks were applied to model and predict the process by constructing a dataset based on long-term experiments. Specifically, Long Short-Term Memory Network (LSTM) and Densely Connected Convolutional Network (DenseNet) were employed, and an enhanced Attention-based DenseNet (AttentionNet) was developed to further improve the prediction performance. These networks were utilized for long-term continuous PN-anammox reactors for nitrogen removal in low-strength wastewater. The results demonstrated that both DenseNet and the proposed AttentionNet effectively modeled the PN-anammox processes, even under conditions of unstable influent quality and relatively poor treatment performance. The mean relative error (MRE) for DenseNet was under 15%, while AttentionNet achieved an MRE of approximately 10% or less, highlighting the superior performance of the Attention layer. These findings were further validated by Bland-Altman analysis. Additionally, further analysis of AttentionNet showed that predictions for pH, nitrite nitrogen, and total nitrogen in the effluent were more accurate than those for other output parameters.