基于反向传播神经网络模型优化冶金废水单级脱氮试验研究

Optimization of Single-stage Denitrification of Metallurgical Wastewater Based on Back Propagation Neural Network

研究了基于前馈反向传播人工神经网络模型,结合Box-Behnken设计方法,采用厌氧氨氧化和部分硝化(SNAP)工艺、以上流式污泥床(upflow-sludge-bed,USB)反应器进行单级脱氮处理从含氮废水中去除NH+4和总氮。通过优化网络结构参数,开发神经网络;基于拟合优度标准,用Levenberg-Marquardt算法训练3层NH+4和总氮的去除。神经网络模型偏差较小(±2.1%),测定系数和分数方差较理想,协议指数(IA)分别为0.989~0.997、0.003~0.031和0.993~0.998。计算结果表明,优化后的神经网络结构可提高神经网络模型效率;利用人工神经网络模型对复杂的生物系统进行建模,可以提高去除效率、实施过程控制策略和实现优化性能;微生物群落的16S rRNA高通量法分析结果表明,库尼尼假丝酵母属的作用最显著(13.11%),其次是亚硝基单胞菌属(6.23%)和蛋白链球菌属(3.1%),这进一步说明USB的脱氮途径主要是部分硝化/氨氧化过程。

Based on the feedforward back propagation artificial neural network model,combined with Box-Behnken design,the single-stage denitrification treatment was carried out using anammox,partial nitrification(SNAP)process and upflow sludge bed(USB)reactor to remove NH+4 and total nitrogen in nitrogen-containing wastewater.Based on the goodness of fit criterion,Levenberg Marquardt algorithm is used to train the removal of NH+4 and total nitrogen in three layers.The deviation of the neural network model is small(±2.1%),and the measurement coefficient and fractional variance are ideal.The agreement index(IA)is 0.989~0.997,0.003~0.031 and 0.993~0.998,respectively.The results show that the optimized neural network structure can improve the efficiency of the neural network model;the artificial neural network model can improve the removal efficiency,implement the process control strategy and realize the optimization performance;the 16S of microbial community can improve the removal efficiency,implement the process control strategy and achieve the optimization performance The results of rRNA high-throughput analysis show that Candida cunini is the most effective(13.11%),followed by nitrosomonas(6.23%)and streptococcus albuminus(3.1%).This further indicates that the denitrification pathway of USB is mainly partial nitrification/ammoxidation.