Water Online the Magazine gives Water & Wastewater Engineers and end-users a venue to find project solutions and source valuable product information. We aim to educate the engineering and operations community on important issues and trends.
Issue link: http://wateronline.epubxp.com/i/303843
wateronline.com ■ Water Online The Magazine control based on operator's knowledge and prediction of the settled water turbidity from the entries of the plant) are the construction of linear or nonlinear models. Adopting a linear modeling technique would demand a multivariable and high complexity model. In addition, the representation of these systems from the impulse response is not possible due to the natural coupling of input variables and also due to the fact that these impulses interfere in the operation of the water treatment station that cannot have its productive process disturbed or interrupted. Therefore, we opted for an identification approach based on the techniques of artificial intelligence. In this way we create a tool to aid the operator. As the operator changes the dosage, information of the estimated turbidity of settled water is obtained, which will allow the operator to perform interventions in the process in order to improve the quality of the settled water. Once the behavior of the operator controlling the coagulant dosage is identified, it is possible to provide references for dosage adjustment for the different conditions of the process. Such models can also be used in a strategy to perform the automatic control of the coagulant dosage. Correlation Of Variables For Determining The Parameters Analyzing the correlations of the variables of the coagulation process, one can identify in an appropriate way which of them may compose the inputs of the neural networks that will identify the operator's behavior and predict the turbidity of the settled water. We used the MATLAB routine called crosscorr.m to generate graphs of cross-correlation. Using the graphs of cross-correlation, we made an analysis to check whether there is a relationship between the variables, the direct or inverse way that they relate, and the dead time of the process. Modeling Prediction Of The Settled Water Turbidity Several models were developed performing variations in parameters and in the topology of the ANN. See Table 1. Identification Of The Operator's Behavior In order to model the system to identify the operator's behavior, when performing the coagulant dosage control activity, we used the same database used for predicting the settled water turbidity and cross-correlation of the graphics obtained. The prediction of the settled water turbidity was not initially inserted as neuro-fuzzy model input because the operator does not currently use this information to control the dosage. The prediction will be inserted as information for the operator to improve the performance of the coagulation process and consequently generate a new database useful in remodeling the identification of the operator's behavior system. The prediction is also a feature of great value when there is a failure in any instrument, working as an estimator of the settled water turbidity. ANFIS network was used to model the operator's behavior. Gaussian membership functions were used to perform the training. The type of equation used in defuzzification was a linear equation in the parameters ΣPij. Three membership functions were used. Presentation Of Results In order to obtain results in the field, automation resources available in the WTP were used. The existing 25 Tutorial IMPLEMENTED MODEL STRUCTURE INPUTS Preliminary model based on a MLP (multilayer perceptron). 1 ANN with 10 inputs, 15 nodes in the hidden layer. Turbidity of the raw water, conductivity, fl ow, pH, dosage, and their respective values 30 min before and the type of coagulant used. Remodeling of the previous attempt. 1 ANN with 10 inputs, 25 nodes in the hidden layer. Set of models with simple selector. 2 ANN with 10 inputs, 25 nodes in the hidden layer using the turbidity of the raw water as parameter of model selection. Set of models with selector based on LVQ algorithm. 2 ANN with 5 inputs, 25 nodes in the hidden layer using the LVQ algorithm with the raw water turbidity and the true color as parameters for the model selection. Neuro-fuzzy structure with online training. Yamakawa Network considering 2 membership functions. Turbidity of the raw water, conductivity, fl ow, pH, dosage, settled water turbidity, and their respective values 1.5 h before. Table 1: History of attempts to obtain the model for prediction of the settled water turbidity Figure 3: Settled Water Turbidity Turbidity Turbidity 2 4 _ V E R T _ 0 5 1 4 C l e a n w a t e r _ C o p a s a _ D G . i n d d 2 24_VERT_0514 Cleanwater_Copasa_DG.indd 2 4 / 2 2 / 2 0 1 4 2 : 1 2 : 2 4 P M 4/22/2014 2:12:24 PM