Abstract：
A continuous stirred tank bioreactor (CSTB) with cell recycling combined with ceramic membrane technology and inoculated with Rhodococcus opacus PD630 was employed to treat petroleum refinery wastewater for simultaneous chemical oxygen demand (COD) removal and lipid production from the retentate obtained during wastewater treatment. In the present study, the COD removal efficiency (CODRE) (%) and lipid concentration (g/L) were predicted using two artificial intelligence models, i.e., an artificial neural network (ANN) and a neuro-fuzzy neural network (NF-NN) with a network topology of 6-25-2 being the best for NF-NN. The results revealed the superiority of NF-NN over ANN in terms of determination coefficient (R2), root mean square error (RMSE), and mean absolute percentage error (MAPE). Three learning algorithms were tested with NF-NN; among them, the Bayesian regularization backpropagation (BR-BP) outperformed others. The sensitivity analysis revealed that, if solid retention time and biomass concentrations were maintained between 35 and 75 h and 3.0 g/L and 3.5 g/L, respectively, high CODRE (93%) and lipid concentration (2.8 g/L) could be obtained consistently.
摘要：
采用连续搅拌槽生物反应器（CSTB），将细胞循环与陶瓷膜技术结合使用，并接种了不透明红球菌PD630，用于处理炼油厂废水，以同时去除化学需氧量（COD）并从废水处理过程中获得的滞留物中产生脂质。在本研究中,利用两个人工智能模型预测COD去除效率(CODRE)(%)和脂质浓度(g/L),即，网络拓扑为6-25-2的人工神经网络（ANN）和神经模糊神经网络（NF-NN）是NF-NN的最佳选择。结果表明，NF-NN在决定系数（R2）方面优于ANN，均方根误差(RMSE)，和平均绝对百分比误差(MAPE)。用NF-NN测试了三种学习算法；其中，贝叶斯正则化反向传播(BR-BP)优于其他算法。敏感性分析表明,如果固体保留时间和生物量浓度保持在35和75小时之间，3.0g/L和3.5g/L，分别,可以一致获得高CODRE（93％）和脂质浓度（2.8g/L）。
