مدلسازی فرآیند انعقاد و لخته سازی توسط روش های استنتاج عصبی- فازی تطبیقی، شبکه های عصبی مصنوعی و رگرسیون فازی

Surface water contains various type of suspended impurities that cause turbidity and color. Coagulation is the main process of integrating fine particles and turn them into larger particles. In this study, replacement of the modeling methods by time-consuming and expensive experimental techniques such as JAR test has been discussed. For this purpose, two models of Feedforward and radial basis of artificial neural networks and Adaptive network-based fuzzy inference system and the various kinds of fuzzy regression analysis to predict the ultimate extent of turbidity after coagulation and flocculation process in 3 and 4 Tehran water treatment plants, were studied. The coagulant used in the treatment plant was poly-aluminum chloride (PAC) and the type and concentration of coagulant, pH and turbidity of the raw water, was opted from the basic information. Radial basis model due to the possibility of automatic raising of hidden layer’s neurons to achieve performance function with minimum error, is highly capable in simulating the process of coagulating. Unlike Feedforward networks, radial basis networks required a smaller number of neurons, and also had the ability to change parameters to achieve the desired results. Increasing the number of hidden layer’s neurons and normalizing the input data to the network enhanced the predictability of artificial neural networks. The study also generalize Feedforward networks to predict data validation and correction of the increasing of performance function. Due to the uncertainty which caused by human error in the laboratory, adaptive network-based fuzzy inference system and fuzzy regression, in which the data sets in the form of fuzzy, were used. The results showed that artificial neural networks and fuzzy regression analysis had more ability in simulating the coagulation process and turbidity removal in different experimental conditions rather than adaptive network-based fuzzy inference system and had the ability to replace the JAR test with time-consuming and expensive methods. The best network built to predict the filtered water turbidity in this study was feed forward network with two hidden layers and 6 and 8 neurons and Tansig and Purelin transfer functions respectively in the first and second layers, using normalized data with performance function. This network is able to predict the coagulation process with a Correlation Coefficient of 0.96 and 0.99 Agreement Index and root mean square error 0.0106. Best predicting done by regression analysis using fuzzy quadratic function. This function was able to predict the data validation with a correlation coefficient, and Agreement Index and root mean square error, respectively, 0.94, 0.96 and 0.75. adaptive network-based fuzzy inference system with the use of Gaussmf membership functions for raw water turbidity and pH input ,and type and Trimf had best efficiency to apply coagulant concentration data into network and estimated the filtered water turbidity with correlation coefficient of 0.89, Agreement Index of 0.91, and squares error of 1.02. This system showed that increasing initial turbidity caused removal efficiency increased and the best impaction of coagulation process for the removal of turbidity would be occurred in the range of pH, 7.6 to 8. The best efficiency in operation condition was determined 99.5% in initial turbidity of 160 NTU, pH=8 and 19 mg/L dosage of PAC coagulant type I.