Sei sulla pagina 1di 10

Computers and Chemical Engineering 30 (2006) 490499

Dynamics of an activated sludge process with nitrication and denitrication: Start-up simulation and optimization using evolutionary algorithm
Saziye Balku, Ridvan Berber
Department of Chemical Engineering, Faculty of Engineering, Ankara University Tandogan, 06100 Ankara, Turkey Received 3 June 2004; received in revised form 3 October 2005; accepted 6 October 2005 Available online 28 November 2005

Abstract An alternative but benecial way of achieving nitrication and denitrication in activated sludge systems is to apply consecutive aeration and non-aeration periods in a single reactor. In this paper, dynamic modeling and start-up simulation in order to establish the continuous operation mode in such systems are rst presented. Newly developed ASM3 is applied to the modeling of microbiological processes in the aeration tank and a 10 layer settling model is adapted to the settling tank. Then, the constrained optimization problem is tackled by an evolutionary algorithm, and the optimum aeration schedule for minimum energy consumption is identied. The constraint handling is achieved by three comparative methods using rejection of infeasibles, penalizing infeasibles and tournament selection without using any penalty parameters. The results indicate that the suggested optimum operation strategy may be economically benecial in all methods but the third one, which is proposed by Deb [Deb, K. (2000). An efcient constraint handling method for generic algorithms. Computer Methods in Applied Mechanics and Engineering, 186, 311338] gives better solutions. 2005 Elsevier Ltd. All rights reserved.
Keywords: Wastewater treatment; Dynamic optimization; Evolutionary algorithm

1. Introduction Activated sludge process, one of the ways of biological treatment, is commonly used in domestic wastewater treatment, or in the secondary treatment of the industrial wastewater. In the classical biological treatment systems anoxic denitrication is maintained in a separate zone, or carrousel type ditches are used for this purpose. In recent years alternating activated sludge processes have been developed in a way through which the aerobic and anoxic zones are periodically exchanged (Zhao, Isaacs, Soeberg, & K ummel, 1995). In the present study a single aeration tank has been used for both aerobic and anoxic operation by an on/off strategy applied to the aeration devices determined by dynamic optimization. Such a system using aeration and non-aeration modes periodically in the aeration tank will ensure carbon removal, nitrication
Corresponding author. Tel.: +90 312 212 2130; fax: +90 312 212 7464. E-mail addresses: balku@atilim.edu.tr (S. Balku), berber@eng.ankara.edu.tr (R. Berber). 0098-1354/$ see front matter 2005 Elsevier Ltd. All rights reserved. doi:10.1016/j.compchemeng.2005.10.014

and denitrication. The system, which can also be named as completely stirred alternating aerobicanoxic (AAA) process, consists of basically an aeration tank and a settler as shown in Fig. 1. In the aeration tank, the microorganisms are used to convert the colloidal and dissolved carbonaceous organic matter into various gases and into cell tissue. Because cell tissue has a specic gravity slightly greater than that of water, the resulting cells can be removed from the treated liquid by gravity settling in the settler. Most of the cells are recycled and mixed with incoming wastewater to maintain convenient sludge age characteristics and such high degradation rates. In activated sludge systems, the main operational cost is due to the electrical energy needed to operate the aerating devices in order to obtain necessary aeration and mixing. Determining the optimum durations of consecutive aeration and non-aeration periods in order to minimize the energy consumption is, therefore, a non-trivial dynamic optimization problem. The dynamical character comes from the fact that the complicated dynamic model, which forms the constraints, needs to be satised at every step. Furthermore, the efuent from the plant has to comply with

S. Balku, R. Berber / Computers and Chemical Engineering 30 (2006) 490499

491

Nomenclature abs at ave error in COD DO eff fns h J kL a Q R req rh rp absolute aeration average deviation from discharge criteria inuent chemical oxygen demand dissolved oxygen efuent non-settleable fraction of the inuent suspended solids height gravity ux oxygen transfer coefcient ow rate appearance/disappearance rates of components required settling parameter associated with the hindered settling component settling parameter associated with the low concentration and slowly settling component of the suspension returning sludge saturation settler total suspended solids total operation time total nitrogen tank volume maximum theoretical settling velocity maximum practical settling velocity settling velocity waste sludge

Fig. 1. Schematic diagram of the activated sludge process.

rs sat set SS T TN V v0 v0 VS w

ASM3 components SO dissolved oxygen concentration SI inert soluble organic material SS readily biodegradable organic substrate SNH ammonium plus ammonia nitrogen S N2 dinitrogen SNO nitrate plus nitrite nitrogen SHCO alkalinity XI inert particulate organics slowly biodegradable substrates XS XH heterotrophic biomass XSTO organics stored by heterotrophs XA autotrophic, nitrifying biomass XSS total suspended solids

strict discharge regulations, and this makes the hard constraints. The fact that dynamic optimization problems are inherently difcult to solve in the rst place makes the problem even more challenging. In this perspective, there seems to be very limited work, in fact only one, reported in the current literature

(Chachuat, Roche, & Lati, 2001), which was based upon ASM1 (Henze, Grady, Gujer, Marais, & Matsuo, 1987). Similar activated sludge systems (i.e. alternating aerobicanoxic process) have been studied from the perspective of process evaluation (Hao & Huang, 1996), and dynamic modeling (Huang & Hao, 1996). Their dynamic model was constructed on ASM1 with some revisions including pH and alkalinity. It was pointed out that the energy saving could be possible by the reducing aeration time in the steady state process evaluation. However, no quantitative assessment was made because they looked at the case primarily from the perspective of parameter estimation and process control. Therefore, the question remains open except the previous attempt by Chachuat et al. (2001). The present work uses the newly developed ASM3 (Gujer, Henze, Mino, & Loosdrecht, 1999), which corrects some defects in ASM1 that have been noticed over 10 years of experience. Although ASM3 is based on the same dominating phenomena in terms of the oxygen consumption, sludge production, nitrication and denitrication, it has become a new standard for future modeling. Some important features that ASM3 covers, compared to ASM1, are: (i) the storage of poly-hydroxy-alkanoates observed under aerobic and anoxic conditions is included in ASM3, (ii) the decay rates of nitriers and heterotrophs under aerobic and anoxic conditions are separated and (iii) all the conversion processes of the two groups of organisms are clearly separated. ASM3 is considered as the best choice in the modeling of the biological processes involved in the activated sludge tank for the time being especially in dealing with the aerobic and anoxic processes in the same tank. Furthermore, the dissolved oxygen (DO) constraints are added to the optimization problem in order to avoid obtaining misleading results. Since the growth of lamentous bacteria may be possible and also the growth rate and the activity of the nitrifying organisms decrease at low DO concentrations. Thus, the work tackles not only an academic but also an industrially important optimization problem, and attempts to dene an effective and practical method to reach a solution by using more realistic means. Initially, this paper describes the dynamic model, and provides the start-up simulation in order to establish the continuous operation mode. Finally, the problem of constrained optimization is tackled to nd the optimum aeration schedule for minimum energy consumption. In the solution of the optimization

492

S. Balku, R. Berber / Computers and Chemical Engineering 30 (2006) 490499

problem, it is proposed to use the evolutionary algorithm, which has become popular in recent years. One of the main advantages of evolutionary computational techniques is that they do not have much mathematical requirements about the optimization problem; all they need is an evaluation of the objective function. The presence of the constraints makes the solution of the problem rather difcult and three methods of dealing with constraints are proposed here; the rejection of infeasible individuals, penalizing infeasibles and Debs method (Deb, 2000). The results show that carbon removal, nitrication and denitrication can be well achieved by the optimal aeration prole determined by using evolutionary algorithm and 18.98% energy saving is possible in comparison to a constant arbitrary aeration prole and 5.81% in comparison to an optimized constant prole. The methods applied and results obtained by these methods are also compared in this study. 2. Treatment plant model Treatment plant model consists of an aeration tank model in which the microbiological processes take place and a settling tank model where the settling velocity of activated sludge is considered. 2.1. Aeration tank Activated Sludge Model No. 3 was used for the biological processes involved in the aeration tank. The dynamical work previously reported in the literature was based on ASM1; but the problem has been dened here by making use of a more realistic model, and therefore reports an improvement in this context. The components of the system, i.e. state variables, are divided into two categories; soluble components, concentrations of which are denoted by S, are assumed to be transported by water, whereas particulate components, whose concentrations are denoted by X, are assumed to be associated with the activated sludge concentrated in the settling tank. Mass balances in the aeration tank result in:
at in + Q Xrs (Q + Q )Xat dXi Qin Xi rs i rs in i + Ri = dt Vat

concentrations in each layer. In order to calculate the gravity uxes, the settling velocities should be calculated for each layer of the settler. The soluble components of ASM3 are assumed to follow the water ow, and their gravity uxes are set to zero. Hence, the concentrations of soluble components of ASM3 are equal in the inlet and outlet of the settling tank. For the particulate components of ASM3, the settling velocity is derived from Tak acs expression as follows:
set ) = max(0, min(v0 , v0 (erh (Xt VS (Xt
set f Xset ) ns in

erp (Xt

set f Xset ) ns in

)))

(2)

where v0 , v0 , rh , rp and fns are model parameters dened in set the suspended solids concentrations Tak acs et al. (1991), Xin set is the concentration of the suspended entering the settler and Xt solids calculated for each layer of the settler based upon ASM3 as follows:
set set set set set set = 0.75(XI (3) + XS Xt ) + 0.90(XH + XA ) + 0.60XSTO

Thus, the gravity ux for the particulate components at each layer j is:
set (j ) J (j ) = VS (j )Xt

(4)

The minimum uxes and the threshold value for solid concentration are considered in the mass balances around each layer of the settler. Differential equations related to the mass balances dened by the ux theory for each 10 layer and each particulate component then becomes: dXset = fset (Xat , Xset ) dt where Xset is a 60-dimensional vector. 2.3. General model The general model then consists of two sets of differential equations as follows: dX = f (1) (X) dt dX = f (2) (X) dt non-aeration periods aeration periods (6) (7) (5)

(1)

where Xat , Xrs and Xin are 13-dimensional vectors consisting of ASM3 components. In the aeration period, the mass balance related to dissolved oxygen includes an additional term on sat S at )) representing the oxygen the right hand size (+kL a(SO O transfer. 2.2. Settling tank The settler is modeled as a cylindrical tank with 10 horizontal layers and settling velocity of Tak acs is used (Tak acs, Patry, & Nolasco, 1991). In the settling tank, each layer is assumed to be perfectly stirred. No reaction takes place and no radial concentration changes. According to the ux theory; total ux is equal to the sum of the bulk and gravity uxes. The bulk uxes are the functions of the linear ow rate and the solid

where X is a 73-dimensional vector, 13 aeration tank variables and 60 settling tank variables. Both sets, despite discrimination for the aerobic and anoxic cycles in the aeration tank, include the equations for the settling tank since there is a process of recycle from the settling tank to the aeration tank. 3. Dynamic simulation Simulation of the start-up period of the plant was accomplished by starting with an aeration tank and a settler lled in by the incoming wastewater. The inlet wastewater composition, stoichiometric matrix and kinetic parameters (at 20 C) were taken from ASM3 as they were. Parameters involved in settling

S. Balku, R. Berber / Computers and Chemical Engineering 30 (2006) 490499

493

velocity model were used as given by Tak acs for low load feeding. The feed to the settler was into the 7th layer from the top. In the recycling stream the concentrations of SO and SN2 were assumed to be equal to zero. The threshold concentration Xt was equal to 3000 g m3 . The design parameters related to the activated sludge plant were chosen according to the basic principles of wastewater treatment plant design (Tchobanoglous & Burton, 1991) as follows: Qin : 1000 m3 day1 Qrs : 800 m3 day1 Vat : 450 m3 Vset : 400 m3 hset : 3.5 m kL a: 4.5 h1 Simulation program was implemented in MATLAB by using ode15s integration routine for stiff differential equations with adjustable step size. To maintain a balance between computing accuracy and time, the maximum step size was 0.01 h for start-up simulation. A constant arbitrary aeration sequence of 0.9 h nonaerated plus 1.8 h aerated period was employed for 20 days. The aim was to maintain the required microorganism concentration for convenient treatment and proper settling characteristics, and also to achieve the initial concentrations for the system before the continuous operation mode for which the optimum aeration sequence was to be obtained. The results of the simulation for 20 days of start-up operation for the system are shown in Table 1. The changes in the concentrations of ASM3 components in the aeration tank as a function of time can be followed in the gures. The heterotrophic bacteria concentration increases from 30 to 1414.8 g m3 in 20 days (Fig. 2) and the autotrophic bacteria concentration increases from 0.1 to 68.7 g m3 (Fig. 3). The changes in concentrations of dissolved oxygen, ammonium nitrogen and nitrate nitrogen and also the alkalinity are shown in Fig. 4. Dissolved oxygen concentration and nitrate nitrogen concentration vary continuously due to the on/off strategy applied to the aeration devices. Alkalinity shows a slight and continuous decrease in the start-up period. During the start-up period
Table 1 The simulated concentrations of ASM3 components ASM3 components (g m3 ) SO SI SS SNH SN2 SNO SHCO (moles m3 ) XI XS XH XSTO XA XSS Initial values of aeration tank 0.0 30.0 100.0 16.0 0.0 0.0 5.0 25.0 75.0 30.0 0.0 0.1 125.0 Aeration tank after 20 days 2.6 30.0 0.1 0.3 6.3 3.4 1390.7 57.7 1414.8 124.7 68.7 3312.9 Efuent from settler 30.0 0.1 0.3 6.3 3.4 3.4 0.1 3.5 0.3 0.2 8.1 Sludge from settler bottom and recycle 30.0 0.1 0.3 6.3 3.4 3114.5 129.7 3178.6 282.7 154 7431.9 Fig. 2. Twenty days simulation proles for suspended solids (XSS ), heterotrophic bacteria (XH ) and inert particulate (XI ).

Fig. 3. Twenty days simulation proles for organics stored (XSTO ), autotrophics (XA ), slowly biodegradables (XS ), inerts soluble (SI ) and readily biodegradables (SS ).

494

S. Balku, R. Berber / Computers and Chemical Engineering 30 (2006) 490499

Fig. 4. Twenty days simulation proles for dissolved oxygen (SO ), ammonium (SNH ), nitrate nitrogen (SNO ) and alkalinity (SHCO ).

no waste sludge was removed from the system. The concentrations of the particulate components of ASM3 in the layers of the settler at the end of 20 days start-up operation are given in Table 2. After the operation period of 20 days the efuent water (from the top layer of the settler) has the following characteristics: CODeff : 37.6 g m3 TNeff : 7.3 g m3 SSeff : 8.1 g m3 These results show that the above-mentioned system operates in conformity with the European Environment Commission (EEC) discharge criteria (EU, 1998). 4. Evolutionary algorithm After the continuous operation dynamics were identied, the next attempt was devoted to determining the optimal aeration prole using the evolutionary algorithm. 4.1. General Evolutionary algorithms (EA) are based on Darwins natural selection principle. Traditional deterministic search methods

nd the next sampling point using the problem features (such as gradients, Hessians, linearity and continuity). However, stochastic search approaches do not use these features, but randomly sample the search space. The evolutionary algorithm is one of the stochastic search methods. The state variables scope is scanned by means of evolutionary resembling operations (selection, crossover and mutation), which are applied on individuals in a population (Ronen, Stabtai, & Guterman, 2002). A chromosome represents each individual in the relevant population of solutions. Each chromosome is composed of successive genes of a certain alphabet. This alphabet could be binary digits, real numbers, integers, symbols, matrices, etc. The original genetic algorithm uses binary digits. It is shown that the more natural representation to a real number problem, i.e. real numbers, is more efcient and produces better solutions space while combinatory problems are better solved using binary representation. The initial population provided for the EA is usually a random population. A positive value, called the tness value, reects the goodness of the chromosome and is tightly related to the value of the objective function. During the course of the evaluation, the surviving chromosome is considered as the best solution. In a practical EA problem, a pool of chromosomes is set. Each upcoming generation is created from the chromosomes of the current pool. Parent chromosomes are selected by a certain procedure (selection). The selection function denes the search space of every generation. Some selection functions are: the roulette wheel, ranking methods, elitist models, scaling methods and tournament methods. Then the genes of the parents are mixed and reassembled to create their offspring. The genetic operators are the basis of the EA search mechanism. They create new solutions based upon the current population. The main operators are mutation and crossover. The mutation changes an individual to create a new one, while the crossover mixes two individuals and creates two new ones. It is expected that the better chromosome will create more offspring and hence, will gain a survival capability through the next generations. The genetic operators used in crossovering may be arithmetical or heuristic, and in mutation they may be uniform, non-uniform, boundary and Gaussian (Michalewiczs real-code genetic operators, Roubus, van Straten, & van Boxtel, 1999). The binary representation used in classical genetic algorithms (GA) has some drawbacks when applied to multidimensional, high precision numerical problems. The bit-strings become very long and the search-space is blown up. Furthermore, much computational time is lost in conversion of the bit strings to real

Table 2 The simulated concentrations of the particulate components in the layers of settler Settler layers/particulates (g m3 ) XI XS XH XSTO XA XSS 1 3.4 0.1 3.5 0.3 0.2 8.1 2 4.3 0.2 4.4 0.4 0.2 10.2 3 5.4 0.2 5.5 0.5 0.3 12.9 4 7.3 0.3 7.4 0.7 0.4 17.3 5 11.6 0.5 11.9 1.1 0.6 27.8 6 27.6 1.1 28.1 2.5 1.4 65.7 7 144.9 6.0 147.5 13.0 7.2 345.3 8 145.2 6.0 147.5 13.0 7.2 345.7 9 144.9 6.0 147.5 13.0 7.2 345.3 10 3114.5 129.7 3178.6 282.7 154.0 7431.9

S. Balku, R. Berber / Computers and Chemical Engineering 30 (2006) 490499

495

values and vice versa. The population size is a parameter to control the optimization progress of GAs. There are no xed rules for the population size when using a GA. Generally, it must not be too low (lower than 1020) because too much information will be lost in every generation and as a consequence the algorithm may converge too fast and end. 4.2. Presence of the constraints From the optimization viewpoint, one of the main advantages of evolutionary computational techniques is that they do not have much mathematical requirements about the optimization problem; all they need is an evaluation of the objective function. But the presence of constraints signicantly affects the performance of the evolutionary algorithm like any other algorithm. One general way of dealing with constraints, whatever the optimization method is, is penalizing infeasible solutions. However, there are no guidelines on designing penalty functions. Michalewicz, Dasgupta, Le Riche, & Schoenauer (1996) focused on and classied the constraints handling methods in evolutionary computation techniques. An efcient constraint handling method is proposed by Deb (2000). In his paper, GAs population-based approach and ability to make pair-wise comparison in tournament selection operator are exploited to devise a penalty function approach that does not require any penalty parameter. The other main efcient features proposed by Deb are the usage of the simulated binary crossover and the parameter-based mutation. In the present paper rejection of infeasible individuals, penalizing infeasible individuals and constraint handling method proposed by Deb are considered: (I) Rejection of infeasible individuals: In this method, the individuals that do not t the constraints are rejected from the population. This offers a few simplication of the algorithm, for example, there is no need to evaluate the objective functions of the infeasible solutions and to compare them with feasible ones. Since this method does not require any penalty function approach, it is easy to apply any kind of constrained problems. Such a method may work reasonably well when the feasible search space is convex and constitutes a reasonable part of the whole search space (e.g. evolution strategies do not allow equality constraints since the ratio between the sizes of feasible and infeasible search spaces is zero with such constraints). Otherwise the method has serious limitations. For example, for many search problems where the initial population consists of infeasible individuals only, no solution is obtained. (II) Penalizing infeasible individuals: This method is the most common approach in the genetic algorithms community. The domain of objective function f is extended as follows: eval(p) = f (p) Q(p), where Q(p) represents either a penalty for infeasible individual p or a cost for repairing such an individual. The

major difculty in this method is the estimation of the penalty parameters. (III) Debs method: The third method proposes to use a tournament selection operator, where two solutions are compared at a time always enforcing the following criteria: 1. Any feasible solution is preferred to any infeasible solution. 2. Among two feasible solutions, the one having better objective function value is preferred. 3. Among two infeasible solutions the one having smaller constraint violation is preferred. In the proposed method, penalty parameters are not needed because in any of the above three scenarios, solutions are never compared in terms of both objective function and constraint violation information. Of the three tournament cases mentioned above, in the rst case neither the objective function value nor the constraint violation information is used, simply the feasible solution is preferred. In the second case, solutions are compared in terms of objective function values alone and in the third case, solutions are compared in terms of constraint violation information alone. Instead of calculating the objective functions of the infeasible solutions the author chooses to place the maximum objective function value among the feasible solutions in the current population. For the real-coded GA, Debs method uses simulated binary crossover (SBX) and a parameter-based mutation. Once sufcient feasible solutions are found, a niching method is used to maintain diversity among the feasible solutions. The advantages of the Debs method are: 1. No need for the calculation of the penalty parameters in the selection procedure. 2. Using SBX and parameter-based mutation ensures the new generations remaining in the boundaries. However, weighing the constraints continues to be a question, and the evaluation of the maximum objective value for the feasible solutions is preferred only when the evaluation of the objective functions spends an important CPU time in the solutions and only if any other selection method depending upon the tness values, such as roulette selection, etc. is to be used together with the tournament selection proposed by Deb. 5. Optimization problem Energy optimization is achieved by minimizing the aerated fraction in the total operating time. The objective function is then dened as the ratio of two quantities: total aeration time divided by total operation time, the sum of total aerated time and total non-aerated time. min J =
M k k=1 b M k k k=1 (a + b )

(8)

496

S. Balku, R. Berber / Computers and Chemical Engineering 30 (2006) 490499

The time scale during which the optimum sequence of nonaeration/aeration periods to be found is discretized with M intervals such that k = 1, 2, . . ., M. Each interval (k) was composed of a pair with one non-aeration period and one aeration period, whose durations in time scale were ak and bk , respectively. The objective function is subject to the mass balances around the aeration tank and the settling tank such as represented in Eqs. (6) and (7). The boundary conditions for the aeration and nonaeration sequences are:
on on tmin bk tmax off off tmin ak tmax

(Arceivala, 1986). For above-mentioned reasons such a limitation is added to the optimization algorithm. In the simulation of continuous operation, a waste sludge stream of 2% of inlet ow rate was introduced to the settler model. 6. Solution algorithm The optimization problem is subject to a set of differential equations, which represent the process behavior and has some hard constraints on the system. Solution algorithms for the dened optimization problem by three different methods of handling constraints are given in this section. All of the methods use real numbers and the same solution approach. Running the global model for each individual, the objective function, the deviation from discharge standards and average DO at the end of the aeration periods are calculated. The application of the GA operators is the same in the rst and second methods but differs in the third method. The algorithm in the application of the GA operators for the rst and second methods is as follows: 1. Selection: Arrange objective functions in ascending order. Take their inverse, and nd percentages in order to nd the tness values. By roulette wheeling rearrange the individuals. A certain percentage of the individuals is chosen for the future operations where the chance to be chosen is proportional to the goodness of the individual which depends upon the objective function. 2. Crossover: The chromosomes are paired randomly and then starting from a random position, the genes of the chromosomes are replaced by simple arithmetical crossover. 3. Mutation: A randomly selected gene of a randomly selected chromosome is replaced by a value randomly given but in compliance with the upper and lower boundaries. Mutation is applied in order to prevent the chromosomes repeating themselves (whole uniform mutation). 4. Elitism: The individual having the best tness function value in the selection is added to the new generation without crossover and mutation. In the third method, tournament selection (Section 4.2), simulated binary crossover and parameter-based mutation are applied (Deb, 2000). 6.1. Rejection of infeasible individuals The rst solution algorithm is based upon the general evolutionary algorithm with the known GA operators, only the infeasible solutions are rejected during the computation. The ow chart of the method for the rejection of the infeasible individuals is shown in Fig. 5. 6.2. Penalizing infeasible individuals In this solution algorithm, instead of rejection of the infeasible individuals, a penalty function is added to the objective function in order to calculate the tness value before roulette

The minimum and maximum lengths of non-aeration and aeration sequences are, 15 min for both sequences as minimum, and 60 min for non-aeration and 120 min for aeration periods as maximum. The constraints on the system are the efuent discharge criteria and the dissolved oxygen concentration. The deviation from the efuent discharge criteria was formulated as: Error = max(0, TNeff TNmax ) + max(0, CODeff CODmax ) + max(0, SSeff SSmax ) (9)

where TNeff , CODeff and SSeff are calculated by their denitions in ASM3. The required efuent criteria are taken as: TNmax : 10 g m3 CODmax : 125 g m3 SSmax : 30 g m3 According to our numerical experiments the most important part of the optimization is to control the dissolved oxygen concentration in the aeration tank. Since the aerobic and anoxic modes are both considered in the same tank, SO varies between 0 and 2.6 g m3 . ASM3 works at low dissolved oxygen concentrations. Therefore, it seems that a successful treatment may be achieved also at low SO values during aerating periods, in contradiction to the practice. In order to overcome this difculty, an additional constraint is added to the algorithm and the DO concentrations at the end of the aeration periods are expected to be controlled by holding the average DO concentration at the end of the aeration periods to be above 2 g m3 . Such a constraint is not involved in the current literature (Chachuat et al., 2001). Our simulation studies show that if DO concentration is not constrained, since the overall economical objective is to minimize the aeration period, the DO levels may decrease under the limits dened in literature in order to meet this criterion (Balku, 2004). Obviously, such low levels of DO concentrations are not suitable for the activated sludge systems including nitrication because the activities of nitrifying bacteria decrease at low levels of dissolved oxygen. Therefore, it is known that the DO concentrations are generally maintained, around (though not strictly) 2 g/m3 in practice. Arceivala denes a DO concentration range between 1.5 and 2.0 mg/l for proper nitrication

S. Balku, R. Berber / Computers and Chemical Engineering 30 (2006) 490499 Table 3 The results obtained by the application of the proposed methods Method applied Initial number of population Number of variables Max. no of iteration Objective function Constraint violation Total computational time (h) Rejection of feasibles 1000 30 50 55.04 68.00 Penalizing feasibles 21 30 100 58.07 65.36

497

Debs method 20 30 50 54.01 34.65

Fig. 5. Flow-chart for rejection of infeasible individuals handling algorithm.

selection. The penalty function consists of some linear equations dening violations from the constraints. Therefore, the values of the objective functions increase while the solution individuals are moving far from the constraints. The application of the penalty function is as follows: Q(p): the penalty function; a function of the violations from the constraints and of the parameters dened according to the degree of violation. Q(p) = [1 abs(Tcalc Treq ) + 2 abs(error) + 3 abs(DOcalc,ave DOreq )] where is the weighing factor between objective function and penalty function which was chosen as 9. Alphas () are the penalty parameters. If the constraints are satised to an acceptable extent, zero is given to penalty parameters. Otherwise they are chosen as 10. 6.3. Debs method In the third solution algorithm, instead of roulette selection, tournament selection is applied based upon the method proposed by Deb. The suggested algorithm: 1. Choose the number of population and the number of genes (an even number, 2 M) in a chromosome. A chromosome represents an aeration prole consisting of non-aerated and aerated periods. 2. Give real, random values for each gene in all chromosomes in the population. 3. Control whether upper and lower limits are satised, if not, replace the disconformities with the upper and lower boundaries. 4. Assign the initial values (X0 ) calculated by the start-up simulation for the state variables.

5. Integrate the non-aerated model (Eq. (6)) forward in time starting from X (t0 ) = X0 , nd the values of Xk at the end of the time period of ak , and integrate the aerated model (Eq. (7)) forward in time starting from Xk for the time period of bk and so on until it is applied to the each gene in one chromosome and for all the chromosomes in the population. 6. Evaluate the objective functions (this evaluation takes nearly no CPU time) and the constraints (can only be evaluated by running the model and most of the CPU time is spent for this evaluation because of the differential equations). 7. Tournament selection is performed by Debs method (dened in Section 4.2, the only difference is the placement of real objective function values of the infeasibles in the tness function). 8. SBX method is used in crossover. 9. Mutation is applied according to the parameter-based mutation. 10. Go to the 5th step in order to apply the GA operators to the new generation. 11. Continue the iteration until a predened maximum iteration number is met. 7. Results and discussion The results obtained by the three different optimization methods dealing with constraint handling are shown in Table 3. The objective functions evaluated are 55.04, 58.07 and 54.01 by rejection, penalizing and Debs method, respectively. The best result is obtained by Debs method in terms of objective function in a predened maximum iteration number as it can be seen in Table 3. The changes in objective function by iteration numbers for the rejection of the infeasible are shown in Fig. 6. The values are

Fig. 6. Changes in objective function (rejection).

498

S. Balku, R. Berber / Computers and Chemical Engineering 30 (2006) 490499

Fig. 7. Changes in objective function (penalizing).

up to the 45th iteration since the number of individuals decreases to one at that iteration and the program is automatically stopped. Keeping the elite chromosome can maintain the objective function constant for a number of proceeding iterations, until another elite chromosome directing towards a better objective function is found, as reected in Fig. 6. This algorithm eliminates the chromosomes (individuals), which do not satisfy the constraints. For this reason, the parent population should be chosen as much as large. If it is a small number, a feasible population may not be achieved, on the other hand, if a high population number is selected; the solution requires high CPU times. How big a maximum number of iterations is selected is not so important in this algorithm in comparison to the population size, since the GA operators acting on remaining limited ones cannot considerably change the results in proceeding iterations. The suggested second algorithm contains an augmented penalty term added to the objective function, and thus the values of the tness function may become occasionally large for the infeasible solutions. The maximum number of iterations is very important in this case, because if it is not big enough, the solution population may consist of only infeasible individuals. On the other hand, when a large maximum number of iterations is chosen, the drawback is the increased computational time. As it is seen from Table 3 using the penalizing feasibles method the penalty value has decreased to zero, which is the desired value. The changes in the tness function with respect to the iteration numbers are shown in Fig. 7. The third method that is proposed by Deb has some advantages in comparison to the previous methods. The method does not require the evaluation of any penalty parameters which is

Fig. 9. Changes in dissolved oxygen (SO ), ammonium plus ammonia nitrogen (SNH ), nitrate plus nitrite nitrogen (SNO ) concentrations as a function of time by the variable optimized prole.

the most important difculty in penalty function approach and further using SBX and parameter-based mutation results in better solutions in the next generations. In the solution generation 70% of the population is in the feasible region. The variable optimized prole obtained by Debs method is shown in Fig. 8. The black bars indicate durations of the non-aerated periods, whereas the white bars indicate those of the aerated periods. The same problem related to the high CPU times is also valid in this method anyhow due to the features of the problem dened in this paper. The values obtained from the suggested optimization algorithm have satised the constraints, and 18.98% energy is saved by the optimal prole in comparison to a constant arbitrary prole (0.9 h non-aerated and 1.8 h aerated). The changes in nitrogen compounds and the dissolved oxygen concentrations in time are shown in Fig. 9. In aerated periods SNO concentration increases, and contrarily SNH concentration decreases

Fig. 8. Optimal aeration prole using Debs method.

Fig. 10. Changes in chemical oxygen demand (CODeff ), total suspended solid (SSeff ) and total nitrogen (TNeff ) as a function of time by the variable optimized prole.

S. Balku, R. Berber / Computers and Chemical Engineering 30 (2006) 490499

499

depending upon increasing SO . The operation by optimal aeration prole results the CODeff , SSeff and TNeff changes given in Fig. 10 showing that at any time the discharge criteria are met. The average DO concentration at the end of the aeration periods is 2.01 g m3 and the deviation from the discharge criteria is zero. Additionally, one run is executed by Debs method in order to nd an optimized constant prole. The results show that the value of objective function is 57.34 and the optimized constant prole consists of 1.00 h for non-aerated periods and 1.34 h for aerated periods. The energy savings achieved by the variable optimized prole obtained by Debs method in comparison to the constant optimized prole will be 5.81%. Solution of the optimization problem has taken quite substantial computation time probably because of the complicated dynamic model and the maximum step size of 0.0005 h in ode15s integration routine in all three methods. MATLAB 6.5 software package and a 2.4 GHz Pentium 4 processor having 512 MB memory were used for all computations. 8. Conclusion In the present study, the dynamic start-up simulation of an activated sludge system including nitrication and denitrication has been pursued in order to evaluate the normal operating conditions. Then the determination of the optimal aeration prole for the system has been accomplished by the evolutionary algorithm using three different constraint handling methods. It is seen that all the methods give satisfactory results in terms of the required treatment despite none of them being global optima and energy savings to the extent of 17.44, 12.90 and 18.98% for rejection, penalizing and Debs methods, respectively, are possible in comparison to a constant arbitrary prole, but all of them have suffered high computational times due to the integration of the large set of differential equations at every stage of the solution algorithm. The other features that may have contributed to the large computation time requirement are constraints, such as the deviations from discharge standards and DO concentration that must be controlled. High computational times cause hesitation to search global minima and the attempts to nd the global optima are left for the future studies. Consequently, the proposed evolutionary solution methods to the constrained dynamic optimization problem of an activated sludge system show that denitrication can also be achieved

in the aeration tank by an on/off strategy applied to aeration devices determined by optimal aeration prole. Energy savings could be possible to the extent of 18.98% in comparison to a constant arbitrary aeration prole and 5.81% in comparison to the constant optimized prole. It is to be noted that the best objective function can be achieved is a variable period optimum prole by Debs method of constraint handling. References
Arceivala, S. J. (1986). Wastewater treatment for pollution control (p. 187). New Delhi: Tata McGraw-Hill Publishing. Balku, S. (2004). Energy optimization in activated sludge systems with nitrogen removal. Ph.D. thesis (in Turkish). Ankara University, Department of Chemical Engineering. Chachuat, B., Roche, N., & Lati, M. A. (2001). Dynamic optimization of small size wastewater treatment plants including nitrication and denitrication processes. Computers and Chemical Engineering, 25, 585593. Deb, K. (2000). An efcient constraint handling method for generic algorithms. Computer Methods in Applied Mechanics and Engineering, 186, 311338. EU-Directive 98/15/EEC amending directive 91/271/EEC. http://europa.eu. int/. Gujer, W., Henze, M., Mino, T., & Loosdrecht, M. (1999). Activated sludge model no. 3. Water Science Technology, 39(1), 183193. Hao, O. J., & Huang, J. (1996). Alternating aerobicanoxic process for nitrogen removal: Process evaluation. Water Environment Research, 68(1), 8393. Henze, M., Grady, C. P., Gujer, W., Marais, G., & Matsuo, T. (1987). Activated sludge model no. 1. Scientic and Technical Reports No. 1. London: IAWQ. Huang, J., & Hao, O. J. (1996). Alternating aerobicanoxic process for nitrogen removal: Dynamic modeling. Water Environment Research, 68(1), 94104. Michalewicz, Z., Dasgupta, D., Le Riche, R. G., & Schoenauer, M. (1996). Evolutionary algorithms for constrained engineering problems. Computers & Industrial Engineering, 30(4), 851870. Ronen, M., Stabtai, Y., & Guterman, H. (2002). Optimization of feeding prole for a fed-batch bioreactor by an evolutionary algorithm. Journal of Biotechnology, 97, 253263. Roubus, J. A., van Straten, G., & van Boxtel, A. J. B. (1999). An evolutionary strategy for fed-batch bioreactor optimization; concepts and performance. Journal of Biotechnology, 67, 173187. Tak acs, I., Patry, G. G., & Nolasco, D. (1991). A dynamic model of the claricationThickening process. Water Research, 25(10), 1263 1271. Tchobanoglous, G., Burton, F. L. (1991). Wastewater engineering: Treatment, disposal and reuse (3rd ed., p. 1334). McGraw-Hill. Zhao, H., Isaacs, H., Soeberg, H., & K ummel, M. (1995). An analysis of nitrogen removal and control strategies in an alternating activated sludge process. Water Research, 29(2), 535544.

Potrebbero piacerti anche