دانشگاه تهرانمحیط شناسی1025-862040420141222Efficiency Consideration of Wastewater Treatment Plant of Tabriz using Artificial Intelligence Modelsبررسی کارایی تصفیهخانۀ فاضلاب تبریز با استفاده از مدلهای هوش مصنوعی8278445299910.22059/jes.2014.52999FAسیماشکریدانشجوی کارشناسی ارشد زمینشناسی زیستمحیطی دانشکدۀ علوم طبیعی، دانشگاه تبریزعطا اللهندیریاستادیار هیدروژئولوژی گروه علوم زمین دانشکدۀ علوم طبیعی، دانشگاه تبریزاصغراصغری مقدماستاد هیدروژئولوژی گروه علوم زمین دانشکدۀ علوم طبیعی، دانشگاه تبریزJournal Article20140104Introduction <br />According to the shortage of water resources in the world, it seems necessary the use of refined water, particularly in arid and semi-arid areas such as Iran. The correct treatment, management and the control of refining process needs to investigate effective parameters in this process precisely. Therefore, because of the uncertainty and complexity in refining qualitative parameters and their relationship, artificial intelligence such as a fuzzy model (FL) and artificial neural networks (ANNs) were used in this study for modeling the behavior of Tabriz plant wastewater treatment. <br />Tabriz city as the capital of the East Azarbaijan province is the most industrialized and urbanized cities of North West of Iran (Figure 1).The sewage of Tabriz city, including industrial and domestic wastewaters, collected gravitationally and enter into the wastewater treatment plant in Tabriz city. It is located in Qaramalek District, four kilometers away in west of downtown on the south side of the Ajichay river and on the lowest part of the city at 1334 (amsl) level. The wastewater treatment plant is designed in three stages. The first phase of Tabriz wastewater treatment plant with a capacity of 612 thousand people has been established in July of (2001). Currently, due to the being incomplete of sewerage network in Tabriz city, only the first phase of this plant is exploited with around 30% of total capacity. The second and third phases are under study. Tabriz wastewater refinery with an annual average rate 1.5 cubic meters per second and peak flow in the rainy and non-rainy days are 3.8 and 2.5 cubic meters per second, respectively. Refinery process including both primary and secondary treatment stages, the physical refining is done in first step then biological treatment and finally disinfection. <br /> <br />Figure 1. Location of wastewater treatment plant of Tabriz city in Tabriz plain. <br />Material and method <br />The data used in this study to develop an artificial neural network and two fuzzy logic model (Sugeno and Mamdani FL models for evaluating the performance of wastewater treatment plant of Tabriz city is influent and effluent of refinery system. In this study, the system influent applied as input model to estimate the qualitative factors of effluent. To quality assessment in wastewater and sewage treatment plant, the parameters that are usually measured and recorded are the Temperature (T), Biochemical Oxygen Demand (BOD), Chemical Oxygen Demand (COD), Total Suspended Solids (TSS) and pH compared in the input and Output of wastewater treatment. In this regard, the monthly values of BOD, COD, TSS, T, and pH of effluent raw wastewater and effluent treated water of treatment plant were analyzed for the 11 years 2002 to 2012 <br />At first the data set that contains 660 data from Tabriz refinery, were divided into two parts include testing (130 data) and training (530 data). A FL method consists of three main parts, fuzziﬁcation, inference engine (fuzzy rule base), and defuzziﬁcation. In the fuzzification step, the four crisp inputs change to fuzzy set for constructing the inference engine. The inference engine consists of rules. Each rule, in turn, is formed from multiple inputs and a single output. When the antecedents of fuzzy rules include more than one rule, then fuzzy operators are used to connect them. The most common fuzzy operators are AND which supported min (minimum) and prod (product), OR (maximum) and NOT. The consequences of a fuzzy rule assign an entire fuzzy set to the output through the process which is called implication. The input to the implication process is a single number given by the antecedent, and the output is a fuzzy set. Since decisions are based on the testing of all of the rules in an FIS, the rules must be combined via aggregation processes in order to make a decision. The process of transforming the aggregation result into a crisp output is termed defuzziﬁcation. The most common defuzziﬁcation methods are centroid, bisector, middle of maximum (the average of the maximum value of the output set), largest of maximum, and smallest of maximum. <br />Discussion of Results <br />The Sugeno fuzzy logic model is constructed by subtractive clustering method. The optimal cluster radius is assigned based on the minimum Root Mean Square Errors (RMSE) and MAPE are 2.71 and 8.08. Based on the optimal clustering radius, five clusters and five If-Then rules were determined using Gaussian membership functions. The average correlation coefficient of this model is 0.82. <br />Also, for constructing Mamdani fuzzy model, FCM clustering method is applied. By using minimum RMSE and MAPE that are 2.83 and 8.88 respectively, the optimal numbers of input and output clusters are assign 17. The Membership functions are Gaussian membership functions and average correlation coefficient of this model is 0.8. <br /> The perceptron network, three layers artificial neural network, includes input, hidden and output layers. Data division in train and test steps is the same as the FL models. Input layer have 5 nodes include BOD, COD, TSS, T and pH that is used as input data for predicting 3 output parameters (i.e., BOD, COD and TSS). To select the appropriate number of neurons in the hidden layer has been proposed the various methods such as trial and error and also mathematical rules. In this study, the number of neurons in the hidden layer determined 3 neurons by the trial and error method. At the first step, the network must be trained. The purpose of the learning network is achieving Minimum Absolute Percentage Errors and optimal weights of the network. Due to the high R² and lowest RMSE in the optimal structure is 5-3-3. The obtained results for test step (average RMSE= 3.63) confirm of high ability of artificial neural network in estimating parameters of wastewater treatment plant. <br />Finally, to evaluate the output of wastewater treatment plant and the results of artificial intelligence models in removing and reducing studying parameters, percentage of removal efficiency was used according to equation 1. <br /> <br /> (1) <br />This index shows indicates the isolation of each of the quality parameters of the wastewater after treatment. In this equation, <em>RE<sub>X</sub></em> is the index of removal efficiency percent of x, amount of input mass of parameters x to treatment plant and artificial intelligence, amount of output mass of the parameters x from the treatment plant which was estimated by artificial intelligence and n is the number of data for each pollutant. This equation was performed in two cases, one based on the measured data in the input and output of treatment plant and the other based on the measured data in the input and the estimates of the artificial intelligence. At the end calculated and compared the efficiency of reducing emissions separately for each pollutant for both cases. <br /> The removal efficiency of each three contaminant is close together and the TSS pollutant has maximum removal efficiency (93.74%) for three artificial intelligence models. The results represent a good performance of applied models. Consequently, the neural network and Mamdani and Sugeno fuzzy models have good accuracy in evaluating the performance of treatment. <br />Conclusions <br />Although three artificial intelligence models have acceptable results, but based on the correlation coefficient and RMSE for each of the parameters in the models, the Sugeno fuzzy model is more superior than other models. The superiority of the Sugeno fuzzy model over an artificial neural network is due to high uncertainty of wastewater treatment plant parameters. Also, calculate the percentage of contaminant removal efficiency was determined in the output treatment. The maximum removal efficiency was related to TSS pollutant that is equal to 93%. TSS values were also very close to other pollutants. Similarly, the removal efficiency of pollutants from the estimated values by fuzzy model and neural network is the same and close to the observed value due to the good performance of used models. با توجه به محدودبودن منابع آب شیرین در جهان، استفادۀ مجدد از پسابهای شهری تصفیهشده به خصوص در مناطق خشک و نیمهخشک مانند ایران ضرروی به نظر میرسد. از این رو تصفیۀ صحیح، مدیریت و کنترل فرایند تصفیه نیازمند بررسی دقیق پارامترهای مـؤثر در این فرایند است. لذا با توجه به عدم قطعیت در پارامترهای کیفی فرایند تصفیه و پیچیدگی ارتباط آنها، در این مطالعه از مدلهای هوش مصنوعی مانند فازی (FL) و شبکههای عصبی مصنوعی (ANN) برای مدلسازی رفتار تصفیهخانۀ فاضلاب تبریز استفاده شده است. بدین منظور با مبنا قراردادن مشخصههای کیفی اندازهگیریشده در ورودی تصفیهخانه، مقدار متناظر این مشخصهها در خروجی تصفیهخانه پیشبینی و از مدلهای فازی ممدانی و ساجنو و شبکههای عصبی مصنوعی استفاده شد. دادههای ورودی این مدلهای هوش مصنوعی شامل پارامترهای دما، اکسیژن مورد نیاز بیوشیمیایی (BOD)، اکسیژن مورد نیاز شیمیایی (COD)، کل جامدات معلق (TSS) و pH فاضلاب و دادههای خروجی شامل مشخصۀ BOD³، COD و TSS است. بر اساس نتایج، هر سه مدل شبکۀ عصبی مصنوعی، فازی ممدانی و ساجنو به ترتیب متوسط RMSE کمتر از 1/3 برای مراحل آموزش و آزمایش داشتند و نتایج قابل قبولی ارائه دادند. مدل فازی ساجنو با متوسط RMSE برابر 83/2 بهترین مدل شناخته شد. همچنین، با محاسبۀ درصد بازده حذف آلایندهها در خروجی تصفیهخانه مشخص شد حداکثر بهرهوری حذف در تصفیهخانه مربوط به آلایندۀ TSS بوده و معادل 93 درصد است. سایر آلایندهها نیز مقادیری بسیار نزدیک به TSS داشتند. به همین ترتیب بازده حذف این آلایندهها از طریق مقادیر برآوردی با مدلهای فازی و شبکۀ عصبی نیز به همین منوال است که به واسطة نزدیکی با مقادیر مشاهداتی مبین کارایی خوب مدلهای به کار برده شده است.https://jes.ut.ac.ir/article_52999_bfb69dd794554ad27878276e252ccbf6.pdf