Loading...

An artificial neural network model for the prediction of pressure filters performance and determination of optimum turbidity for coli-form and total bacteria removal

Badalians Gholikandi, G ; Sharif University of Technology | 2012

444 Viewed
  1. Type of Document: Article
  2. Publisher: 2012
  3. Abstract:
  4. In water treatment processes, because of complicated and nonlinear relationships between a number of physical, chemical and operational parameters, using analytical models with the ability to capture underlying relationships using examples of the desired input-output mapping is quite suitable. Artificial Neural Networks (ANN) has been increasingly applied in the area of environmental and water resources engineering. The main advantage of Artificial Neural Networks over physical-based models is that they are data-driven. The purpose of this research is to study the performance of pressure filters on turbidity removal from water according to several parameters such as turbidity, filtration rate, and pressure, and to ustilize these results for introducing the artificial neural network model as a powerful tool in predicting the performance of full scale plants. The reason of using pressure filters in this research is lack of enough investigations on this kind of filters in the past, and also the need for studying the performance of these filters according to several parameters, simultaneously. Materials and methods Pilot plant A pressurized pilot filter with a circular metal area of 4 mm thickness was used in this research. Prior to use, the interior surface of the filter was painted with two layers of epoxy. Thereafter, all mechanical parts such as influent and effluent pipes, valves, barometer, nozzles etc., were installed in appropriate places. In Table 1, the characteristics of the filter are shown. The main variables in this research are influent turbidity, filtration rate, and influent pressure. The influent water with different turbidities and flows and operational pressures were studied. The characteristics of the variables are provided in Table 2. Artificial Neural Network Neural networks are composed of simple elements operating in parallel. These elements are inspired by biological nervous systems. As in nature, the network function is determined largely by the connections between elements. Commonly, neural networks are adjusted or trained so that a particular input leads to a specific target output. The back-propagation neural network (BPN) is the most representative learning model for the ANN. Among several algorithms used in artificial neural networks, the "Multilayer Perseptron Algorithm" with "Back Propagation Training Algorithm" is prevalent in engineering fields. In the present research, this algorithm is used to update the parameters used in artificial neural network. In this method, errors are minimized while network parameters are adjusted. Mean squares of errors are used as a scale to measure teaching data, and the parameters minimizing error, are then measured. Results and Discussion After setting up the filter successfully and sampling the effluent from the filter for different input conditions such as different turbidities, influent filtration rate, and the filter pressure, the output results were statistically analyzed. These results, which are determined according to 1300 samples from different situations, have been used randomly as the artificial neural network inputs, after being normalized. The goal of using artificial neural network in this study is to provide a pattern for predicting the minimum and maximum values of probable turbidity in outputs of pressure filter systems. In this study, MATLAB software was used for training and testing the data in the artificial neural network (Kia, 2005). After investigating and controlling different neural network conditions, such as number of hidden layers and number of neurons in each, 3-11-11-2 was determined as the best network structure, in which 3 and 2 are referred to input and output layers numbers, respectively, and 11-11 is the number of neurons in two hidden layers. After investigating different conditions, the best values for momentum coefficient optimum range and training rate coefficient were determined to be 0.5 and 0.2, respectively. These values were used in different structures, such as the final structure. The results of investigating some of different structures of artificial neural network are provided. The results for the best conditions are also provided in Table 4. Diagrams plotted according to these results are shown in Fig. 1 to Fig. 3. Also in this study, the optimum turbidity for total bacteria and coli-forms removal was 82m/d (see Fig. 4.). Conclusion The following results were obtained from this study, in which the efficiency of pressure filters and providing a pattern to show this efficiency due to artificial neural network was investigated. Different network architectural parameters such as momentum coefficient and training rate were investigated in different conditions. The best conditions were 0.5 and 0.2, respectively, and several neural network structures with different number of neurons were investigated to determine the optimum condition. The 3- 11-11-2 structure for the artificial neural network was finally obtained as the optimum pattern
  5. Keywords:
  6. Artificial neural networks ; Bacteria removal ; Pressure filters modeling ; Turbidity ; Bacteria (microorganisms)
  7. Source: Journal of Environmental Studies ; Volume 37, Issue 60 , 2012 , Pages 129-136 ; 10258620 (ISSN)
  8. URL: http://en.journals.sid.ir/ViewPaper.aspx?ID=242532