第三届挑战赛C1-城市供水处理混凝投药过程的建模与控制

全国大学生数据挖掘竞赛 优 作品名称：城市供水处理混凝投药过程的建模与控制 荣获奖项：一等奖 作品单位：东北大学秦皇岛分校 作品成员：李起旺唐鑫桂 基于非线性预测建模的混凝投药过程控制 摘要： 针对水厂混凝投药过程，本文根据广东省某水厂历史数据，建立了进水流量、浊度、PH值、加药量和沉淀池出水浊度之间的四个阶段性数学模型，并对模型进行优化最终建立了最佳投药量的闭环预测系统，实现对投药量的实时控制。 第一阶段：忽略PH对投药量的影响，综合已有文献[3]中的投药量与各因素间的指数关系，建立了基于非线性回归辨识指数模型，并采用取对数方法将其转换为线性回归问题，最终求得出水浊度与投药量、原水浊度及取水量之间的函数表达式（可信度为99%），并通过式（2-5）计算反应到沉淀结束的时间为90min。 第二阶段：考虑时间的滞后性，本文建立了基于改进的BP神经网络的非线性黑箱投药控制模型，将原水PH、原水浊度、取水量以及投药量作为输入参数，出水浊度作为输出参数。通过设置学习参数和动量系数，并控制出水浊度达到标准值，最终确定了4-12-1最佳的神经网络结构，其泛化误差为0.0014。运用此模型，计算了在不同原水条件下的最佳投药量，具体可见表2-6；并分析了原水PH、原水浊度及取水量对最佳投药量的影响，可见图2-7、2-8、2-9。 第三阶段：鉴于混凝投药过程是一个反馈控制过程，在投药时混凝剂量不仅与控制本次出水浊度有关，还与上次的出水浊度相关，模型二只是通过控制出水浊度来寻找最佳的投药量的正向反馈。为使模型更有效，与实际情况更接近，本文在模型二基础上，增加出水浊度作为输入量，投药量作为输出量建立新的BP神经网络模型，并通过样本数据的训练和误差控制确定了最佳网络结构。出水浊度作为反馈参数实现对最佳投药量的反馈控制：以1.0NTU为临界点，大于1.0NTU会使投加量增加，且越大增加越多，小于1.0NTU会使投加量减少，且越小减少越多，具体可见表2-7和图2-12. 第四阶段：由于温度会影响分子的布朗运动和水的动力粘度，最终会影响到最佳PAC量的投加，故需要将温度也作为输入参数，研究温度对最佳PAC量投加影响，其模型是在模型三基础上增加温度作为输入量构建新的神经网络模型，从气象局搜集到2013年8月22日到2014年9月5日的气温数据，并利用文献[11]气温与水温函数关系计算得日均水温数据，同时将附件数据集预处理后的数据进行日均化后与温度数据进行标准化处理后作为输入参数进行网络训练，最终确定较优的神经网络结构，在10-250C时最佳投药量几乎不变，而在温度升高投药量降低；反之升高，具体可见图2-17. 最后，本文根据BP神经网络的缺陷，设计了基于RBF神经网络结构的投药控制模型，并基于之前的模型最终设计出最佳投药量的闭环预测系统（图2-20），实现对投药量在水发生变化情况下的实时控制。 关键词：混凝投药，非线性预测，回归模型，BP神经网络，RBF神经网络 Process Control of Coagulant Dosage Based on theNonlinear Predictive Modeling Abstract: As for coagulant dosing process In a water industry, this paper applies the historical data of thewaterworks in Guangdong Province to establish the four stages of mathematical model about influent flowrate, turbidity, pH, dosage and effluent turbidity, and further optimizes the model and finally establish theoptimal dosage of closed-loop prediction system to realize the real-time control of coagulant dosage. The first stage: Neglecting the influence of pH on the dosage and mirroring dosage among variousfactors and the exponential relationship in literatures[3], this paper establishes the identification indexmodel basedon nonlinear regression,and transforms it into a linear regression problem with thelogarithmic method, and get the final function between these variables, and reaction precipitation over timeis obtained 90min by formula 2-5 . The second stage: Considering the time lag, this paper establishes the nonlinear black box dosingcontrol model based on Improved BP neural network , where the raw water pH, turbidity, water and dosageare input parameters, the effluent turbidity is output parameter. By setting the learning parameters andmomentum coefficient, and controlling the turbidity of the effluent to reach the standard value, the optimalneural network structure of 4-12-1 is determined, and its generalization error is 0.0014. Using this model,the optimal dosage in different raw water conditions were calculated, Tab2-6 and Fig 2-7/8/9 can showyou . The third stage: Considering the process of coagulant dosage is a feedback control process, thedosage of PAC is not only controlled by the effluent turbidity but also the effluent turbidity. Model 2 findsthe best dosage just form positive feedback by controlling effluent turbidity. In order to make the modelmore effective and more close to the actual situation, this paper increases turbidity as another input basedon the second model, the dosage as output, to establish new model based on BP neural network, anddetermined the optimal network structure through training and error control sample data. The detailedoptimal dosage changes under different raw water conditions can be seen Tab 2-7 and Fig 2-12. Thefourth stage:Because the dynamic viscosity temperature will affect the molecular Brownmovement and water, eventually affect the best amount of PAC. There is a need to consider the temperatureas the input parameter, and study the effects of temperature on the optimum PAC. This model is based onthe third model by increasing temperature as input and constructs a new neural network model. Thetemperature data from August 22, 2013 to September 5, 2014 are collected from the Meteorological Bureau,andgets the daily water temperature through the relationship between air temperature and watertemperature in literature[11], and normalizes the attachment data set after pretreatment and the temperaturedata as the input parameters of the network training, and ultimately determines the neural network anoptimal structure. The optimal dosage doesn’t change between 100C and 250C,which can be shown in Fig2-17. In the end, according to the defects of BP neural network, this paper builds a control model based onthe administration of the structure of RBF neural network, and design the optimum dosage of closed-loopprediction system (Fig 2-20), implementation of coagulant dosage in water with changes in the real time . Key words:Coagulant Dosage, Nonlinear Prediction, Regression Model, BP Neural Network, RBF NeuralNetwork 目录 1.研究目标...............................................................................................12.分析方法与过程...........................