传统的饮用水处理需要通过混凝、絮凝和沉淀去除悬浮固体。使用凝结剂处理水成本高且复杂，目前，水处理公司必须依靠水处理专家的技能和直觉来获得良好的效果，通常与罐式测试相结合。数学模型将使这一过程得到优化，并可能实现自动化。我们的研究重点是优化硫酸铁和Clar+离子的剂量，以在常规处理厂达到给定的过滤器顶部浊度水平。我们从肯塔基州北部一家当地公用事业公司获得了四年的水库水处理数据，这两种混凝剂在那里经常使用。 数据包括进水浊度、混凝剂剂量、温度、pH值、碱度、硬度、处理水量和过滤器顶部（输出）浊度的日常记录。我们的目标是根据这些其他变量的值预测过滤器顶部的浊度。我们开发了从两个方面解决这个问题的模型。一个基本线性回归模型（带有一个非线性项）很好地拟合了数据，并显示进水浊度、温度、硫酸铁和Clar+离子剂量是显著变量，而pH、碱度、硬度和处理水的量则不显著。接下来，我们构建了一个前馈非线性神经网络模型，该模型使用反向传播算法来发现这四个输入变量与输出浊度之间的关系。 研究了各种网络；典型的一个包含84个输入单元、10个单元的隐藏层和10个单元的输出层。所有实数都被数字化了。该网络是根据1997年至1999年的一部分数据进行训练的，训练后，它成功地预测了2000年前六个月的输出浊度，但早春的几天除外。与训练数据相比，2000年春季的浊度要高得多；因此，2000年春季的情况对模型来说是新的，这让我们的拟合度很差。随着联邦和州法规的日益严格，如果饮用水公用事业公司要以最低的价格保证客户的饮用水质量，就必须提高其效率、教育程度和自动化程度。 我们由来自三个领域（计算机科学、生物学和化学）的学生和教师研究人员组成的多学科团队，使我们能够构建一个讨论支持工具，用于选择最佳混凝剂剂量。包括数字。

Conventional drinking water treatment requires the removal of suspended solids by means of coagulation, flocculation and sedimentation. Treating water with coagulating chemicals is costly and complex, and at present water companies must rely on the skills and intuitions of water treatment experts to achieve good results, usually in conjunction with jar testing. A mathematical model would allow this process to be optimized and, potentially, automated. Our research focused on optimizing ferric sulfate and Clar+Ion dosages to achieve a given level of top-of-filter turbidity at a conventional treatment plant. We obtained four years of data on the treatment of reservoir water from a local northern Kentucky utility, where these two coagulants are regularly used. The data consisted of daily records of influent turbidity, coagulant dosages, temperature, pH, alkalinity, hardness, amount treated water, and top-of-filter (output) turbidity. Our goal was to predict top-of-filter turbidity based on the values of these other variables. We developed models to address this problem from two sides. An essentially linear regression model (with one nonlinear term) provided a good fit for the data, and revealed that influent turbidity, temperature, ferric sulfate and Clar+Ion dosages were significant variables while pH, alkalinity, hardness and amount of treated water were not significant. We next constructed a feed-forward nonlinear neural network model that used the back-propagation algorithm to discover relationships between these four input variables and output turbidity. A variety of networks were studied; a typical one contained 84 input units, a hidden layer of ten units, and an output layer of ten units. All real numbers were digitized. The network was trained on a subset of data from the years 1997 through 1999, and, once trained, it successfully predicted output turbidity for the first six months of 2000, with the exception of a period of days in early spring. The spring of 2000 had much higher turbidities relative to the training data; hence the conditions of the spring 2000 were new to the model, giving us a poor fit. With increasingly strict federal and state regulations, it is essential that drinking water utilities become more efficient, educated, and automated if they are to guarantee their customers quality drinking water at the lowest price. Our multidisciplinary team of student and faculty researchers drawn from three fields (computer science, biology, and chemistry) allowed us to construct a discussion support tool for choosing optimal coagulant dosages. Includes figures.