基于BP神经网络的城镇污水厂活性炭自动投加系统的研究

发布时间：2018-10-22 10:30

【摘要】：近年来,粉末活性炭凭借吸附效果好和成本低的优势,不再局限于应对突发性污染应急处理,正逐步发展成为深度处理的主流技术。但是粉末活性炭投加系统仍处于人工控制阶段,现场需要技术人员依靠经验确定活性炭投加量,存在出水水质不稳定,活性炭药耗大等问题。因此,如何根据进水水质自动调节活性炭投药量,是粉末活性炭投加系统推广应用中亟需解决的问题,对中小型城镇污水处理厂实现稳定达标和节能减耗具有重要意义。本论文以浙江省嘉善县某城镇污水处理厂的深度处理工艺作为研究背景,以粉末活性炭投加系统作为研究对象,经分析确定了影响活性炭投药量的主要因素是原水COD、pH和流量。针对活性炭投加控制系统大滞后、非线性和复杂性等问题,创新性地提出基于BP神经网络的前馈预测投药控制器,并确立了 BP神经网络前馈预测-PID反馈的复合控制系统方案。主要研究结果如下:1、通过烧杯试验获取46组成功的样本数据,分别选用BP神经网络和多元线性回归构建活性炭投加系统的前馈控制模型。BP神经网络采用三层结构,其中输入层有2个节点,输出层有1个节点,隐含层有11个节点,利用离线训练后的模型对未学习的样本进行仿真,得到与实际投加量的拟合度R2=0.968,均方根误差RMSE=0.0091。多元线性回归模型的表达式为U=0.0170X1+0.0020X2-0.2078,按相同方法仿真后得到与实际投加量的拟合度R2=0.909,均方根误差RMSE=0.0145。对比仿真结果可知,活性炭投加系统的前馈控制器采用BP神经网络模型具有明显的优越性,不仅预测精度高,而且学习能力强,可以适应不同水质的变化。2、通过对被控对象建模的理论分析得到传递函数表达式。利用粉末活性炭投加飞升实验得到活性炭投加飞升曲线,找到阶跃响应曲线拐点(50,49.78),确定被控对象的传递函数为:G0(s)=5.4/(1+33.75s)(1+16.875s)e-26.6s。选用临界比例度法确定PID控制器的参数,分别为:Kp=0.36;KI=0.006;KD=5.4。阶跃响应曲线表明,反馈控制下的系统出水COD超调量大于40%,达到稳定所需的时间为300 min 左右。3、在Matlab的Simulink环境下进行前馈BP神经网络预测-PID反馈复合控制系统的仿真,结果表明复合控制系统下出水COD的超调量小于20%,调节时间在70 min左右。示范工程中的上位机采用装有WinCC 7.0的PC,通过以太网与各PLC站点进行数据传输,通过OPC通讯工具来完成PC Matlab与PLC之间的通讯,活性炭投加计量泵流量q根据公式q=x/10ω自动计算而得。该投加系统经过一个月自动控制模式下的试运行,出水COD达标率升至90.63%,相较于人工控制提高了 8.88个百分点,同时削减活性炭日均消耗量16.61%,每月可减少活性炭使用成本13.5万元。
[Abstract]:In recent years, with the advantages of good adsorption effect and low cost, powdered activated carbon is no longer limited to the emergency treatment of sudden pollution, and is gradually developing into the mainstream technology of advanced treatment. However, the powdered activated carbon feeding system is still in the manual control stage. It is necessary for technicians to determine the dosage of activated carbon based on experience, which leads to unstable effluent quality and large consumption of activated carbon. Therefore, how to automatically adjust the dosage of activated carbon according to the influent water quality is an urgent problem to be solved in the popularization and application of the powder activated carbon dosing system. It is of great significance for the small and medium-sized town wastewater treatment plants to achieve stable standards and save energy and consumption. In this paper, the advanced treatment process of a town sewage treatment plant in Jiashan County, Zhejiang Province is taken as the research background, and the powder activated carbon dosing system is taken as the research object. The main factors affecting the dosage of activated carbon are COD,pH and flow rate of raw water. Aiming at the problems of large lag, nonlinearity and complexity of active carbon feeding control system, a novel feedforward predictive dosing controller based on BP neural network is proposed, and a compound control system scheme based on BP neural network feedforward predictive and PID feedback is established. The main results are as follows: 1. 46 groups of successful sample data were obtained by beaker test. BP neural network and multivariate linear regression were used to construct feedforward control model of activated carbon feeding system. BP neural network adopted three-layer structure. There are 2 nodes in the input layer, 1 node in the output layer and 11 nodes in the hidden layer. Using the model after off-line training, the unlearned samples are simulated, and the fitting degree R2C 0.368 and the root mean square error (RMSE=0.0091.) are obtained. The expression of multivariate linear regression model is U=0.0170X1 0.0020X2-0.2078. According to the same method, the fitting degree R _ 2o _ (0.909) and root mean square error (RMSE=0.0145.) are obtained. Compared with the simulation results, the feedforward controller of the active carbon feeding system has obvious advantages by using the BP neural network model, which not only has high prediction accuracy, but also has a strong learning ability. It can adapt to the change of different water quality. 2. The expression of transfer function is obtained by theoretical analysis of the model of controlled object. The inflection point of the step response curve (50 ~ 49.78) was obtained by using the experiment of adding powder activated carbon. The transfer function of the controlled object was determined as G0 (s) = 5.4 / (1.33.75s) (116.875s) e-26.6s. The critical proportion method is used to determine the parameters of PID controller, which are: Kp=0.36;KI=0.006;KD=5.4. The step response curve shows that the output COD overshoot of the system under feedback control is greater than 40 and the time required to achieve stability is about 300 min. The feedforward BP neural network predictive control system and PID feedback compound control system are simulated under the Simulink environment of Matlab. The results show that the overshoot of effluent COD is less than 20 and the adjusting time is about 70 min. In the demonstration project, the upper computer uses PC, with WinCC 7.0 to transmit data to each PLC station through Ethernet. The communication between PC Matlab and PLC is accomplished by OPC communication tool. The flow rate of activated carbon dosing metering pump Q is automatically calculated according to the formula q=x/10 蠅. After one month's trial operation under automatic control mode, the effluent COD reaches the standard rate of 90.63%, which is 8.88% higher than that of manual control. At the same time, the average daily consumption of activated carbon is reduced by 16.61%, and the cost of using activated carbon is reduced by 135000 yuan per month.
【学位授予单位】：浙江大学
【学位级别】：硕士
【学位授予年份】：2017
【分类号】：TP183;TU992.3

【相似文献】