摘要:
厌氧膜生物反应器(AnMBR)是一种高效的废水资源化处理技术,膜污染会导致跨膜压差(TMP)增大,进而影响反应器运行性能以及膜寿命。为实现膜污染的精准预测,构建了一种基于注意力机制(AM)的卷积神经网络(CNN)-门控循环单元网络(GRU)的混合机器学习模型(CNN-GRU-AM)。首先,在模型构建过程中,通过参数优化确定RMSprop为最佳优化器,卷积层神经元数优选[64,128]。然后,采用皮尔逊(Pearson)相关性分析对膜污染影响因素进行特征选择,发现活性污泥浓度(MLSS)与胞外聚合物中多糖(EPS-PS)贡献度最大,分别为58.34%和52.47%;经特征选择后,利用对模型拟合影响较大的5个输入变量进行预测,决定系数R 2从0.948 1提升至0.981 0,预测精度提升了3.5%。最后,评估了基于CNN、长短期记忆人工神经网络(LSTM)和GRU建立的不同模型对TMP的预测精度,CNN-GRU-AM混合模型的决定系数(R 2=0.981 0)、平均绝对误差(MAE=0.280 9)、均方误差(MSE=0.250 3)均优于其他模型,证实了CNN-GRU-AM混合模型预测AnMBR膜污染的合理性和实用性。
关键词:
厌氧膜生物反应器,
机器学习,
特征选择,
注意力机制,
膜污染
Abstract:
Anaerobic membrane bioreactor(AnMBR) is a very effective wastewater resource treatment process, while membrane fouling can lead to the increase of transmembrane pressure(TMP), affecting operational performance and membrane life. A machine learning model (CNN-GRU-AM) combined with convolutional neural network(CNN) and gated recurrent unit network (GRU) based on the attention mechanism(AM) was constructed to predict the membrane fouling. Firstly, during the model construction process, RMSprop was identified as the optimal optimizer, and the preferred number of neurons in the convolutional layers was determined to be [64, 128]. Then Pearson correlation analysis was applied to select the features of the membrane fouling influence factors, with MLSS and EPS-PS contributing the most with 58.34% and 52.47%, respectively. After feature selection, five key input variables with significant influence on model fitting were used for prediction. The coefficient of determination(R 2) increased from 0.948 1 to 0.981 0, improving the prediction accuracy by 3.5%. Finally, the prediction accuracy of different models based on CNN, long and short-term memory artificial neural network(LSTM) and GRU for TMP was evaluated. The results showed that the CNN-GRU-AM model was more accurate in terms of the coefficient of determination (R 2=0.981 0), the mean absolute error (MAE=0.280 9), and the mean square error (MSE=0.250 3). The results indicated that the rationality and practicality of CNN-GRU-AM for predicting membrane fouling in AnMBR.
Key words:
anaerobic membrane bioreactor,
machine learning,
feature selection,
attention mechanism,
membrane fouling
中图分类号:
谢伯逊, 赵海岑, 陈杰, 张梦, 张新波. 注意力机制驱动的AnMBR膜污染预测机器学习模型构建[J]. 工业水处理, 2025, 45(9): 78-86.
Boxun XIE, Haicen ZHAO, Jie CHEN, Meng ZHANG, Xinbo ZHANG. Construction of machine learning model driven by attention mechanism for AnMBR membrane fouling prediction[J]. Industrial Water Treatment, 2025, 45(9): 78-86.