-
厌氧氨氧化反应是厌氧氨氧化菌在厌氧或缺氧条件下,利用CO2或H2CO3等无机碳源,以
${\rm{NO}}_2^ - $ 为电子受体,直接将${\rm{NH}}_4^ + $ 还原为N2排放到大气中的新型脱氮工艺[1],具有工艺流程短、脱氮效率高、需氧量少、几乎不产生污泥、不产生二次污染等优点[2-3]。在当今许多污水处理系统工艺复杂、能耗高、产泥量大的背景下,实现厌氧氨氧化工艺的工程应用显得尤为重要。然而,由于厌氧氨氧化过程苛刻的工艺条件和复杂的影响因素[4],使其对工艺过程控制的要求高,且难以确定关键运行控制参数。通过建立厌氧氨氧化工艺系统水质预测模型,可以掌握该工艺系统的运行规律,反映其运行状态,对明确厌氧氨氧化工艺过程的关键控制参数,实现工艺过程的工程应用具有重要意义。与所有污水处理工艺一样,厌氧氨氧化属于复杂的生化反应过程,具有非线性、时变性和大时滞的特点[5],难以建立精确的数学模型。目前,用于水质预测的数学模型主要有多元线性回归模型、灰色预测模型、支持向量机、神经网络模型等。多元线性回归模型因其算法精度低、智能化程度弱,达不到很好的预测效果[6];灰色预测模型因其不考虑系统的内在机理,在作较长时间预测时往往精度不高[7];虽然支持向量机模型算法在处理小样本预测方面较有优势,但基于小样本的学习随机性较强,易忽略数据样本的关键信息,且模型运行时间较神经网络长[8]。有研究[9-11]表明,BP神经网络是解决非线性、大时变系统的分析、预测及建模问题的有效工具,具有智能化程度高、仿真能力强的优势,但也存在易收敛至局部极值点的缺陷,须辅助其他优化模型以改善其性能。常用的解决方法是对神经网络模型各层之间的权重和阈值进行优化,如BAGHERI等[12]利用遗传算法(genetic algorithm,GA)优化人工神经网络,对神经网络的权值、阈值所组成的种群随机搜索并进行选择、交叉、变异等数值计算,成功实现了对活性污泥膨胀的预测。但GA计算复杂,容易陷入“早熟”,近年来,粒子群算法(particle swarm optimization,PSO)因其算法简单、参数少、求解速度快、通用性强等优点而备受人们关注。粒子群中的每一个粒子都代表一个问题的可能解,通过粒子个体的简单行为、群体内的信息交互可实现问题的求解,常用于智能系统模型的优化。但PSO也存在着搜索精度较低、后期迭代效率不高等不足[13]。
厌氧氨氧化工艺的优势在于其较高的脱氮效率,总氮去除率是评判工艺水平和运行水平的重要指标之一[14]。为确定厌氧氨氧化工艺工程应用的关键运行控制参数,本研究以SBBR单级自养脱氮厌氧氨氧化工艺系统为研究对象,以出水总氮去除率为目标,建立了基于BP神经网络的多级预测模型。其中:一级模型通过灰色关联度分析,对影响出水总氮去除率的关键性指标进行了预测;二级模型在基于一级模型上增加数据维度,并通过改进粒子群算法优化人工神经网络、借鉴遗传算法变异的思想扩大搜索范围,提高出水总氮去除率的预测精度。通过模型预测和结果分析,明确了厌氧氨化工艺系统运行的关键控制参数,为工艺过程的优化与进一步工程应用提供参考。
基于多级神经网络模型的厌氧氨氧化系统参数预测
Parameter prediction of anaerobic ammonium oxidation system based on multi-level neural network model
-
摘要: 厌氧氨氧化菌生长条件复杂、影响因素多,其工艺系统运行控制复杂,为解决上述问题,研究构建了1个多级神经网络预测模型,以提高SBBR单级自养脱氮厌氧氨氧化系统出水总氮去除率预测精度,并确定了系统工程应用的关键控制参数。一级神经网络模型通过灰色关联度分析,对影响出水总氮去除率的关键性指标进行预测;二级神经网络模型基于一级模型增加数据维度,并通过改进粒子群算法优化网络、借鉴遗传算法变异的思想扩大搜索范围,提高了出水总氮去除率的预测精度。多级神经网络模型预测结果表明,其总氮去除率平均相对误差为0.54%,相对误差为5.76%,均方根误差为1.132 1,预测数据基本上与实际值相符;与其他预测模型相比较,该模型表现出较优的预测精度。进一步分析发现,通过控制工艺系统的曝气量调节出水亚氮浓度,是保证工艺反应的稳定和实现厌氧氨氧化工艺工程应用的有效控制方式。Abstract: Due to the complex growth conditions and multiple factors of anaerobic ammonium oxidizing (anammox) bacteria, the operation and control of the anammox process is very complex. In this study, a multi-level neural network prediction model was developed to improve the prediction accuracy of total nitrogen removal rate in the effluent of a single-stage SBBR autotrophic anammox system, and to determine the key control parameters for engineering applications of the system. The first-level artificial neural network model predicted the key indicators affecting the total nitrogen removal rate in the effluent through gray correlation analysis. The second-level artificial neural network model added the data dimension based on the first-level model, its artificial neural network was optimized with improving particle swarm optimization algorithm and its search range of particles was expanded by using the idea of genetic algorithm variation, then its prediction accuracy of total nitrogen removal rate was improved. The results showed that the predicted data basically matched with the actual value, the average relative error of total nitrogen removal rate was 0.54%, the relative error was 5.76%, and the root mean square error was 1.132 1. Compared with other predictive models, this model showed better prediction accuracy. Further analysis showed that it would be an effective control method to adjust the nitrous concentration in the effluent by controlling the aeration rate of the process system, which can ensure the stability of the process reaction and realize the engineering application of anammox process.
-
表 1 部分实验数据
Table 1. Partial experimental data
序号 进水氨氮/(mg·L−1) 溶解氧/(mg·L−1) 温度/℃ 出水氨氮/(mg·L−1) 出水硝氮/(mg·L−1) 出水亚氮/(mg·L−1) 总氮去除率/% 1 200 0.98 31 18.83 26.75 0.24 77.09 2 200 0.48 30.9 21.25 26.00 0.31 76.22 3 200 0.77 31.2 1.61 43.66 0.24 77.24 4 200 0.89 30.8 34.70 14.87 0.29 75.07 5 200 0.54 31.1 41.96 15.62 0.29 71.06 6 200 1.11 29.2 23.13 16.91 0.18 79.89 7 185 0.7 30.5 33.89 11.88 0.26 75.12 8 173 0.56 32.1 18.56 32.64 0.21 70.29 9 173 0.88 32.3 6.19 33.28 0.23 77.06 10 100 1.4 30.8 4.84 22.79 0.16 72.21 11 100 1.05 30.3 9.15 20.22 0.16 70.55 12 185 1.09 31.1 25.29 8.35 0.28 81.67 13 173 0.77 32.3 9.15 35.10 0.26 74.28 14 100 1.6 29.5 0.00 22.47 0.11 77.42 15 86 1.35 31.7 12.37 4.28 0.16 80.45 表 2 各模型性能误差
Table 2. Performance error of each model
模型 Emape/% Ermse Ere/% BP 6.87 5.970 1 14.94 GA-BP 4.60 3.991 4 10.37 PSO-BP 4.43 3.920 7 10.67 多级预测模型 0.54 1.132 1 5.76 -
[1] JETTEN M S M, STROUS M, VAN DE PAS-SCHOONEN K T, et al. The anaerobic oxidation of ammonium[J]. FEMS Microbiology Reviews, 1998, 22(5): 421-437. doi: 10.1111/j.1574-6976.1998.tb00379.x [2] MA B, WANG S, CAO S, et al. Biological nitrogen removal from sewage via anammox: Recent advances[J]. Bioresource Technology, 2016, 200: 981-990. doi: 10.1016/j.biortech.2015.10.074 [3] CAO Y S, VAN LOOSDRECHT M C M, DAIGGER G T. Mainstream partial nitritation-anammox in municipal wastewater treatment: Status, bottlenecks, and further studies[J]. Applied Microbiology and Biotechnology, 2017, 101(4): 1365-1383. doi: 10.1007/s00253-016-8058-7 [4] 唐崇俭. 厌氧氨氧化工艺特性与控制技术的研究[D]. 杭州: 浙江大学, 2011. [5] PICOS-BENITEZ A R, LOPEZ-HINCAPIE J D, CHAVEZ-RAMIREZ A U, et al. Artificial intelligence based model for optimization of COD removal efficiency of an up-flow anaerobic sludge blanket reactor in the saline wastewater treatment[J]. Water Science and Technology, 2017, 75(6): 1351-1361. doi: 10.2166/wst.2017.005 [6] SEEBOONRUANG U. A statistical assessment of the impact of land uses on surface water quality indexes[J]. Journal of Environmental Management, 2012, 101: 134-142. [7] 蒋白懿, 牟天蔚, 王玲萍. 灰色遗传神经网络模型对居民年需水量预测[J]. 给水排水, 2018, 54(1): 137-142. doi: 10.3969/j.issn.1002-8471.2018.01.028 [8] 黄辉, 马思佳, 王庆, 等. 多参数影响下污水总氮浓度预测最优方法研究[J]. 南京大学学报(自然科学), 2017, 53(6): 1194-1202. [9] ZHANG H G, CUI L L, ZHANG X, et al. Data-driven robust approximate optimal tracking control for unknown general nonlinear systems using adaptive dynamic programming Method[J]. IEEE Transactions on Neural Networks, 2011, 22(12): 2226-2236. doi: 10.1109/TNN.2011.2168538 [10] ZHAO Y, GUO L, LIANG J B, et al. Seasonal artificial neural network model for water quality prediction via a clustering analysis method in a wastewater treatment plant of China[J]. Desalination and Water Treatment, 2016, 57(8): 3452-3465. doi: 10.1080/19443994.2014.986202 [11] YE J, ZHANG P, HOFFMANN E, et al. Comparison of response surface methodology and artificial neural network in optimization and prediction of acid activation of bauxsol for phosphorus adsorption[J]. Water, Air & Soil Pollution, 2014, 225(12): 2225. [12] BAGHERI M, MIRBAGHERI S A, BAGHERI Z, et al. Modeling and optimization of activated sludge bulking for a real wastewater treatment plant using hybrid artificial neural networks-genetic algorithm approach[J]. Process Safety and Environmental Protection, 2015, 95: 12-25. doi: 10.1016/j.psep.2015.02.008 [13] 乔俊飞, 逄泽芳, 韩红桂. 基于改进粒子群算法的污水处理过程神经网络优化控制[J]. 智能系统学报, 2012, 7(5): 429-436. doi: 10.3969/j.issn.1673-4785.201205034 [14] 李佟, 李军. 基于BP神经网络与马尔可夫链的污水处理厂脱氮效果模拟预测[J]. 环境科学学报, 2016, 36(2): 576-581. [15] HU K, WAN J Q, MA Y W, et al. A fuzzy neural network model for monitoring A2/O process using on-line monitoring parameters[J]. Journal of Environmental Science and Health, 2012, 47(5): 744-754. [16] PLAZAS-NOSSA L, HOFER T, GRUBER G, et al. Forecasting of UV-Vis absorbance time series using artificial neural networks combined with principal component analysis[J]. Water Science and Technology, 2017, 75(4): 765-774. doi: 10.2166/wst.2016.524 [17] 乔俊飞, 安茹, 韩红桂. 基于RBF神经网络的出水氨氮预测研究[J]. 控制工程, 2016, 23(9): 1301-1305. [18] 刘思峰, 杨英杰, 吴利丰, 等. 灰色系统理论及其应用[M]. 7版. 北京: 科学出版社, 2014. [19] DU R, CAO S B, LI B K, et al. Performance and microbial community analysis of a novel DEAMOX based on partial-denitrification and anammox treating ammonia and nitrate wastewaters[J]. Water Research, 2017, 108: 46-56. doi: 10.1016/j.watres.2016.10.051 [20] DECANETE J F, DEL SAZ-OROZCO P, BARATTI R, et al. Soft-sensing estimation of plant effluent concentrations in a biological wastewater treatment plant using an optimal neural network[J]. Expert Systems with Applications, 2016, 63: 8-19. doi: 10.1016/j.eswa.2016.06.028 [21] LACKNER S, GILBERT E M, VLAEMINCK S E, et al. Full-scale partial nitritation/anammox experiences: An application survey[J]. Water Research, 2014, 55: 292-303. doi: 10.1016/j.watres.2014.02.032 [22] 严聪聪, 徐凯琳, 张建昆, 等. 基于改进的粒子群算法的活性污泥神经网络控制模型[J]. 给水排水, 2018, 54(1): 130-136. doi: 10.3969/j.issn.1002-8471.2018.01.027 [23] 王元月, 魏源送. 新型控制模式下短程硝化/厌氧氨氧化法处理污泥水[J]. 中国给水排水, 2013, 29(17): 24-27. doi: 10.3969/j.issn.1000-4602.2013.17.006