Description of optimal model parameters.

The measurement of chemical oxygen demand (COD) is very important in the process of sewage treatment. The value of COD reflects the effectiveness and trend of sewage treatment to a certain extent, but obtaining accurate data requires high cost and labor intensity. To1 solve this problem, this paper proposes an online soft measurement method for COD based on Convolutional Neural Network-Bidirectional Long Short-Term Memory Network-Attention Mechanism (CNN-BiLSTM-Attention) algorithm. Firstly, by analyzing the mechanism of the aerobic tank stage in the Anaerobic-Anoxic-Oxic (A2O) wastewater treatment process, the selection range of input variables was preliminarily determined, and the collected sample dataset was subjected to correlation analysis. Finally, pH, dissolved oxygen (DO), electrical conductivity (EC), and water temperature (T) were determined as input variables for soft measurement prediction of COD.Then, based on the feature extraction ability of CNN and the advantage that BiLSTM is able to capture the backward and forward dependencies in time series data, combined with the attention mechanism that can assign higher weights to the key data, a CNN-BiLSTM-Attention algorithm model was established to soft measure COD in the effluent from the aerobic zone of the A2O wastewater treatment process. At the same time, root mean square error (RMSE), mean absolute error (MAE), mean absolute percentage error (MAPE) and coefficient of determination (R2) were utilized Three indicators were used to evaluate the model, and the results showed that the model can accurately predict the value of COD and has a high accuracy. At the same time, compared with models such as CNN-LSTM-Attention, CNN-BiLSTM, CNN-LSTM, LSTM, RNN, BP, SVM, XGBoost, and RF etc., the results showed that the CNN-BiLSTM Attention model performed the best, proving the superiority of the algorithm model.The Wilcoxon signed-rank test indicates significant differences between the CNN-BiLSTM-Attention model and other models.化学需氧量（chemical oxygen demand, COD）的测量在污水处理流程中至关重要。COD数值可在一定程度上反映污水处理的效果与趋势，但获取精准数据往往需要高昂的成本与人力投入。为解决该问题，本文提出了一种基于卷积神经网络-双向长短期记忆网络-注意力机制（Convolutional Neural Network-Bidirectional Long Short-Term Memory Network-Attention Mechanism, CNN-BiLSTM-Attention）算法的COD在线软测量方法。首先，通过分析厌氧-缺氧-好氧（Anaerobic-Anoxic-Oxic, A2O）污水处理工艺中好氧池阶段的作用机制，初步确定输入变量的选择范围，并对采集得到的样本数据集开展相关性分析，最终选定pH、溶解氧（dissolved oxygen, DO）、电导率（electrical conductivity, EC）以及水温（water temperature, T）作为COD软测量预测的输入变量。随后，基于卷积神经网络（CNN）的特征提取能力、双向长短期记忆网络（BiLSTM）能够捕捉时序数据前后向依赖关系的优势，结合可对关键数据赋予更高权重的注意力机制，构建了CNN-BiLSTM-Attention算法模型，用于对A2O污水处理工艺好氧区出水的COD进行软测量。同时，采用均方根误差（root mean square error, RMSE）、平均绝对误差（mean absolute error, MAE）、平均绝对百分比误差（mean absolute percentage error, MAPE）与决定系数（coefficient of determination, R²）四项指标对模型进行评估，结果显示该模型可精准预测COD数值，具备较高的准确度。此外，与CNN-LSTM-Attention、CNN-BiLSTM、CNN-LSTM、LSTM、RNN、BP、SVM、XGBoost以及随机森林（RF）等模型相比，CNN-BiLSTM-Attention模型表现最优，验证了该算法模型的优越性。威尔科克森符号秩检验（Wilcoxon signed-rank test）结果表明，CNN-BiLSTM-Attention模型与其他模型之间存在显著差异。