Neural Model for Biological and Chemical Transformation in Sewage Treatment Plant

Sewage treatment plant, neural networks
estimation, process control

Krzysztof SZABAT, Marcin KAMIŃSKI*

NEURAL MODEL FOR BIOLOGICAL AND CHEMICAL TRANSFORMATION IN SEWAGE TREATMENT PLANT

In the work issues related to modeling of the biological-chemical wastewater treatment plant are presented. A commonly-used control strategy including model predictive control of the wastewater treatment plant are presented. Then a model of biological and chemical transformations AMS1 is described and its limitations are pointed out. Then the feed-forward neural networks are introduced. The performance of the obtaining models are shown and discussed. The designed models can be used in the model predictive control structure (MPC) of the savage treatment plant and ensure the optimal control of the electrical blowers.

1. INTRODUCTION

Control processes in sewage treatment plants is a rapidly growing field of knowledge. This follows on the one hand the universality of using urban wastewater treatment and industrial processes, striving for continuous reduction of maintenance costs of wastewater treatment as well as a complicated and non-linear model of the transformation of compounds in the treatment process. Optimal control strategy should be the one hand, easy to install on the other lead to a reduction of operating costs and prevent exceeding the limits on the outflow of sewage. In the literature there are a number of control structures for control processes used in the sewage treatment plant [1]-[4].

The simplest control strategies based on maintaining constant air flow to the zones of nitrification in the biological reactor. The advantage of this approach is its simplicity. It does not require any installed sensors in the bioreactor. Disadvantages of this strategy stem from the constant flow of air from the electrical blowers into the bioreactor.Under conditions of low water flow the bioreactor is over-aerated. Even the slow increase in the flow of pollution can be brought to exceed the limits in the outflow. Despite these drawbacks, this strategy is still in use [1].

More advanced control strategy commonly used for small wastewater treatment plants in Poland consists in controlling the dissolved oxygen level in bioreactor. It requires the installation of oxygen sensors in the selected parts of the bioreactor. Usually the PI (in the case of the variable speed drive) or hysteresis (in the case of the direct switched motors without converter) controllers are applied. Depending on the value of the control signal the number of the electrical blowers can be set on or off. This control strategy do not take into consideration the input of the bioreactor which is it drawback. The large values of the inflow can lead the dissolved oxygen level below the required value. This may result in exceeding the limit values at the effluent [1]-[2].

Another strategy, also based on controlling the level of oxygen, introduced into the system additional information about the state of system input. Usually this is done by adding to the classic structure the feed-forward controller. Different functions of the activation of additional controller may be used, such as linear, fuzzy, etc. The use of the additional information from the system input results in increase the level of oxygen in bioreactor which prepare a system to adopt more west water. The disadvantage of this strategy is the need to applied the additional sensors at the input of the system. This strategy can be especially effectively applied in the waste-water treatment plant in the case of the variable-speed electrical blowers driven by the power converter [1]-[2].

One of the most modern methods of control used at wastewater treatment plants is MPC [4]-[5]. It ensures the optimal control of wastewater treatment process. This allows reductions in operating costs for wastewater treatment and avoid the exceed the limits in the outflow. Its drawback is a very complicated algorithm which require the information of all state variables and disturbances of the process. These are usually estimated by the Kalman filter [5]. Based on the current state of the system a control algorithm (using a model biological transformations of chemical state of the system) calculate within a specified time horizon the state of the plant. The system behavior is calculated for various control options and then the optimal strategy is selected (which guarantee the minimal value of the cost function). This requires the implementation of many simulations in a finite period of time. It is therefore necessary to install a unit for high-power computing. One way to simplify the calculation is to replace the complex mathematical model of the plant by a simpler and faster neural model. Also this strategy needs modern electrical drives to be installed in the sewage treatment plant.

In the literature many applications of the neural networks to control and diagnosis of biological-chemical processes in wastewater treatment plants can be found [2], [4], [6], [7]. Neural networks have been used to predict a variable inflow of sewage treatment, model the entire system, estimation of individual parameters or in the control system application. In the few works issues related to the replacement of the analytical model by neural network are raised.

The main objective of the work is to present the neural model transformations of biological chemistry in sewage. After a short introduction the analytical model for biological-chemical transformation ASM1 is introduced [2]-[4]. Successively the structure of the feed-forward neural networks are described. Then the training and the validation data are discussed. Next the obtained results are presented and described. The scope of future research is clarified. The obtained model can be used in order to provide the optimal control signal for the electrical machines installed in the waste-water treatment plant using MPC strategy.

2. ASM1 MODEL

ASM1 model describes the transformation of organic compounds and nitrogen in sewage treatment plant. Its original form was proposed in 1987 in [8]. It consisted of eight equations that describe the kinematics of change by manipulating the 13 state variables. ASM1 model was based on mass balance equations and stoichiometric relationships of kinematics. Currently used form consists of ten equations which describing the transformation of the fourteen variables.

ASM1 model operates on the following state variables:

-SSeasily degradable organic compounds considered as dissolved; simple organic compounds, which are the source of energy and raw materials for the heterotrophic growth of microorganisms.

-SIdissolved organic compounds biologically nondegradable treated as dissolved organic compounds do not take part in biological processes, their do not change their composition or character.

-SNH – ammonium nitrogen, expressed as the sum of ammonia (NH3) and ammonium (NH4+).

-SNOnitrate nitrogen, expressed as an aggregate concentration of nitrates and nitrites in the model because it does not take into account the fraction of nitrite.

-SNDDissolved organic nitrogen, nitrogen being the combinations of readily biodegradable organic compounds

-SOdissolved oxygen .

-SALKalkalinity .

-XSslowly biodegradable organic compounds, organic compounds of large size; it is assumed that they are suspended, although some may be present in dissolved form.

-XIorganic compounds in a suspension of biologically non degradable, suspensions and colloids of organic compounds that are resistant to biodegradation biodiversity, they do not change their composition or character.

-XBH– heterotrophic bacteria, microorganisms, which in carry out the biodegradation in aerobic and anaerobic zones, as well as the hydrolysis and ammonification of XS

-XBAautotrophic bacteria, microorganisms that carry out the process of nitrification - derive energy from oxidation of ammonia; this fraction express at the same time the microorganisms which oxidizing of nitrite and ammonia

-XPproducts of biomass death, organic compounds in the suspension resulting from the withering away of biomass, resistant to biodegradation.

-XNDorganic nitrogen in the suspension, Organic nitrogen which is connected with the fraction XS. Together with XS hydrolyzes to dissolved organic nitrogen (SND)

-XMIN– mineral slurry, a suspension, which is not included in the COD and do not undergo any treatment.

The expression (1) describe the change in a particular state vector.

(1)

where: j- the stoichiometric coefficient [2]-[4], j- the kinematics equation.

The following kinematics equations describe the transformation in the ASM model:

(2)

(3)

(4)

(5)

(6)

(7)

(8)

(9)

(10)

(11)

The equation (2) determines the rate of increase oxygen assimilation of heterotrophy with the transformation of the ammonium as a source of nitrogen. This process is limited by the availability of organic compounds resulting of maintaining of the system a large number of microorganisms (directly proportional to the age of the sediment). Aerobic growth of heterotrophy (3) with the use of nitrates as a source of nitrogen is an alternative to the previous process. Linking between those processes occurs in the absence of ammonium nitrogen in the zones by a factor, which seeks to value of one at high concentrations of ammonium nitrogen. Limiting process is associated with the availability of nitrate. Anoxic increase of heterotrophy (4) with the assimilation of ammonium nitrogenoccurs when oxygen concentration is close or equal to zero. This process compared to oxygen (2) is slower which is expressed in the equation by a constant .Anoxic increase of the heterotrophy (5), using nitrate as nitrogen source take place in a anaerobic conditions or at a minimum concentration of oxygen. The required condition is the availability of nitrate in the absence of ammonium nitrogen. Oxygen autotroph growth is described by the equation (6). It is limited by the concentration of ammonia nitrogen, alkalinity and dissolved oxygen. In the process of dying heterotrophy (7), it is assumed that a part of the biomass is transformed to the fraction - XP. The rest of it is convert to the slowly degradable compoundsXS, which contain organic nitrogen in suspension– XND. The equation (7) express the same process as in (1). The same fractions with the same stoichiometric ratios are formed; the speed of the process is proportional to the concentration of microorganisms. Ammonification of the dissolved organic nitrogen (9) is carried out by the heterotrophy with the speed proportional to the concentrations of organic nitrogen and microorganisms. It results in the increase in alkalinity. The rate of hydrolysis of organic compounds (10) is express by (11). The hydrolysis of organic nitrogen in suspension is described by (12). Tables of the stoichiometric coefficient are provided in [2]-[4].

3. THE NEURAL MODEL OF THE ASM1

4.1feedforward NEURAL-NETWORKS

A lot of different types of the neural networks can be found in the scientific papers [9]-[10]. One of the most commonly-used is called a multi-layer perceptron. Networks of this type are the construction of feed-forward type. It means that in this neural network there is one direction of flow of data between layers (Fig. 1). The structure of the network includes interconnected neurons arranged in layers (input and the starting-in, and hidden layers - with no direct connection to external signals). There are no connections between neurons of the same layer. Layers are arranged in the pattern of neural networks to each other in series, while the neurons in the layers are arranged in parallel.

Fig. 2. The structure of the feed-forward neural networks

A single neuron carries out the operation of aggregation of input signals multiplied by the corresponding weighting factors w. This value is determined in the learning process. The result of this argument is the activation function. In the hidden layer of the neural networks the bipolar sigmoidal activation function (hyperbolic tangent) is used. In the output layer a linear functions are applied. The output signal of the individual neuron is expressed by the following equation:

(12)

(13)

where: f – hypothetical activation function, wjk– weight coefficients, xk– input signal, - activation function correction coefficient, u – the argument of the activation function, wj0– the value of the bias,

In the case of the classical MLP network is often necessary to processing of input data for linear scaling. The goal of such action is to adjust the input to the compartments in which there is significant variation in activation function. This allows to increase the efficiency of the obtained results. Effect of scaling the input vector to the accuracy of the neural network in the wastewater treatment plant modelling has been presented in the next part of the work.

4.2SIMULATION RESULTS

The research with the for feed-forward neural networks with one hidden layer are performed. As the activation function the hyperbolic tangent is adopted. The number of neurons in the input and output layers depends on the dimensions of the learning data. Input vector includes: parameters of flow, flow volume and the parameters of the zone from the previous measurement (total 29). Also the dissolved oxygen level has been assumed as know due to the fact that this variable is accessible in almost every sewage treatment plant and can be considered us a control variable (in MPC algorithm). The intensity of the flow is divided by the volume of the zone (195 m3). Different data has been used for the training and testing procedures. The data obtained from the analytical model ASM1 has been down-sampling (every 100ts sample is used). Neural networks used in the proposed models were trained using one of the methods of gradient - Levenberg-Marquardt algorithm. Number of hidden neurons and the number of iterations of the algorithm has been selected experimentally.

Firstly, the single neural network used to estimated all parameters of the ASM model has been tested. The input data have been previously scaled in order to minimise the relative difference between the variables of the system. The selected network structure is (29-5-14). Number of the training period is 30.

a) / b)
c) / d)
e) / f)
g) / h)
i) / j)
k) / l)
i) / j)

Fig. 3. The real and the estimated transients of the state variables and its estimation errors for the case of single NN used to estimated all variables for the testing data

The transients of the real and estimated variables and the estimated errors is presented in Fig. 2. The application of the single NN to estimation all state variables of the system is possible. As can be seen from the transients presented in Fig. 3. the neural estimator works in a stable way. However, due to the completely different behaviour of the selected state variables in some estimates the steady stay level of the error is relatively high (especially in SS, SNH, SND, SO, SALK and XND). In order to minimise it, the following modification has been proposed. The state variable has been divided into two groups. The dynamic behaviour of particular variables has been used as the criterion. The first group includes the following states: SS, SI, SNH, SNO, SND, SO, SALK, XS and XI. The second group consists of the following states: XBH, XBA, XP, XND and XMIN.The structure of the first an the second networks have the following structure {29-12-9} and {29-4-5} respectively.

The transients of the state variables of the system as well as the estimation error in the case of the application of the two separate estimators are presented in Fig. 4 (for the testing data).

The used of the two parallel NN for the estimation of the state variables of the ASM1system increase the accuracy of the neural models significantly. The level of the errors of almost all variables decreased. For example for the variable SS the error has been reduced from about 0.8 to almost 0 value.

a) / b)
c) / d)
e) / f)
g) / h)
i) / j)
k) / l)
i) / j)

Fig. 4. The real and the estimated transients of the state variables and its estimation errorsfor the case of two parallel NN used to estimated separate variables for the testing data

4. CONCLUSIONS

The characteristic property of the ASM1 model is its large computational effort. It requires a relatively small step calculation to ensure its correct operation. An additional problem is its high dependence on the hard-identifiable parameters. These constraints of the ASM1 model lead designers to search for better models of biological and chemical transformations. One of them are the neural models. The neural networks have the ability of generalization which allows the optimal modeling of processes in terms of parametric uncertainty. They can also work with the increased computational step, which accelerates the execution of the simulation. Reduce the amount of calculation is very important in the MPC control strategy. It requires the implementation of a series of calculations for various values of the control signals in a relatively small unit of time. Reduce the complexity of the model enables the use of cheaper computing unit which facilitates the implementation of the real object.

The neural models presented in the paper can be used in the MPC control structure. This control strategy allows the optimal way to control electrical blowers. This enables reductions of the electrical energy consumed the sewage treatment plant and the greater security of the crossing limits in the effluent. The future works will be concern of the application of the designed neural networks in the MPC strategy. Also the application of the different types of the neural networks (e.g. recurrent, neuro-fuzzy) to model of the ASM1 will be investigated.

REFERENCES

[1]STARE A., VRECKO D., HVALA N., STRMCNIK S., Comparison of control strategies for nitrogen removal in an activated sludge process in terms of operating costs: simulation study, Water research, vol. 41, pp 2004-2014, 2007.