A Method for Effluent Total Nitrogen-based on a Recurrent Self-organizing RBF Neural Network

Abstract

In this present disclosure, a computing implemented method is designed for predicting the effluent total nitrogen concentration (TN) in an urban wastewater treatment process (WWTP). The technology of this present disclosure is part of advanced manufacturing technology and belongs to both the field of control engineer and environment engineer. To improve the predicting efficiency, a recurrent self-organizing RBF neural network (RSORBFNN) can adjust the structure and parameters simultaneously. This RSORBFNN is developed to implement this method, and then the proposed RSORBFNN-based method can predict the effluent TN with acceptable accuracy. Moreover, online information of effluent TN may be predicted by this computing implemented method to enhance the quality monitoring level to alleviate the current situation of wastewater and to strengthen the management of WWTP.

Description

CROSS REFERENCE TO RELATED PATENT APPLICATIONS

This application claims priority to Chinese Patent Application No. 201610606146.X, filed on Jul. 28, 2016, entitled “a method for effluent total nitrogen based on a recurrent self-organizing RBF neural network,” which is hereby incorporated by reference in its entirety.

Technical Field

In this present disclosure, a computing implemented method is designed for predicting the effluent total nitrogen TN concentration (TN) in the urban wastewater treatment process (WWTP) by a recurrent self-organizing RBF neural network (RSORBFNN). To improve the measurement efficiency, the RSORBFNN can adjust the structure and parameters concurrently: a growing and pruning algorithm is proposed to design the structure, and an adaptive second-order algorithm is utilized to train the parameters. The technology of this present disclosure is part of advanced manufacturing technology and belongs to both the field of control engineer and environment engineer.

Background

The urban WWTP not only guarantees the reliability and stability of the wastewater treatment system but also meets the water quality national discharge standard. However, the influence factors are various for effluent TN of wastewater treatment process and the relationship between different influencing factors are complex. Therefore, it is hard to make real-time detecting for effluent TN, which seriously affected the stable operation of the urban WWTP. The computing implemented a method for effluent TN, based on RSORBFNN, is helpful to improve the efficiency, strengthen delicacy management and ensure water quality effluent standards of urban WWTP. It has better economic benefit as well as significant environmental and social benefits. Thus, the research achievements have wide application prospect in this present disclosure.

The control target of urban WWTP is to make the water quality meet the national discharge standards, mainly related to the parameters of effluent TN, chemical oxygen demand (COD), effluent suspended solids (SS), ammonia nitrogen (NH4-N), biochemical oxygen demand (BOD) and effluent total phosphorus (TP). Effluent TN refers to the sum of all the nitrogen pollution of the water after dealing with the sewage treatment plant process facilities, mainly for the ammonia nitrogen, nitrate nitrogen, inorganic nitrogen, protein, amino acid and organic amine organic nitrogen combined. According to statistics, nitrogen fixation rate of about 150 million tons per year in nature and chemical nitrogen fertilizer production rate of about 5000˜6000 tons a year. If nature denitrification reaction failed to complete the nitrogen cycle, too much nitrogen compounds and the ammonia nitrogen nutrient caused a significant number of algae in the water, the plants breeding, appearance of eutrophication status. To curb the trend of worsening of water environment, many sewage treatment facilities have spent a large sum of money to build and put into operation in the country, the cities, and towns. The general method for determination is the alkaline potassium persulfate UV spectrophotometry and molecular absorption spectrometry. However, the determination of total nitrogen TN is often offline and can't realize the effluent TN real-time measurement, which led directly to the sewage treatment process is hard to achieve closed loop control. Moreover, it is a big challenge for detection due to a significant amount of pollutants in wastewater and different content. Developing new hardware measuring instrument, although directly solving various wastewater treatment process variables and the detection problem of water quality parameters, due to the very complex organic matter in sewage, research and development of the new sensor will be a significant cost and a time-consuming project. Hence, the new method presented to solve the problem of the real-time measurement of the process parameters of WWTP has become an important topic to research in the field of wastewater control engineering and has important practical significance.

To obtain more reliable information on effluent TN in urban WWTP, we have investigated a computing implemented method based on the RSORBFNN. The neural network uses competitiveness of the hidden neuron to determine whether to add or delete the hidden neurons and to use an adaptive second order algorithm to ensure the accuracy of RSORBFNN. The objective of this present disclosure is to develop a computing implemented method for estimating the effluent TN online and with high precision.

SUMMARY

A computing implemented method is designed for the effluent TN based on an RSORBFNN in this present disclosure. For this computing implemented method, the inputs are those variables that are easy to measure and the outputs are estimates of the effluent TN. By constructing the RSORBFNN, it realizes the mapping between auxiliary variables and effluent TN. Also, the method can obtain a real-time measurement of effluent TN, solve the problems of long measurement cycle for effluent TN.

A computing implemented method for the effluent TN based on an RSORBFNN, its characteristic and steps include the following steps:

(1) Determine the input and output variables of effluent TN:

For sewage treatment process of activated sludge system, the variables of sewage treatment process are analyzed and select the input variables of effluent TN soft-computing model: ammonia nitrogen —NH₄—N, nitrate nitrogen —NO₃—N, effluent suspended solids—SS, biochemical oxygen demand—BOD, total phosphorus—TP, The output value of soft-computing model is detected effluent TN.

(2) Initialize RSORBFNN

The initial structure of RSORBFNN consists of three layers: an input layer, hidden layer, and an output layer. There are 5 neurons in the input layer, J neurons in the hidden layer and 1 neuron in the output layer; J>2 is a positive integer. Connection weights between input layer and hidden layer are assigned 1, the feedback weights between hidden layer and output layer randomly assign values, the assignment internal is [1, 1]; the number of the training sample is P, and the input vector of RSORBFNN is x(t)=[x₁(t), x₂(t), x₃(t), x₄(t), x₅(t)] at time t; y(t) is the output of RSORBFNN, and y_d(t) is the real value of effluent TN at time t, respectively; The output of RSORBFNN can be described:

$\begin{array}{cc}y\left(t\right)=\sum _{j=1}^Jw_j^2\left(t\right){\theta }_j\left(t\right),& \left(1\right)\end{array}$

wherein w2 j(t) is the output weight between the jth hidden neuron and the output neuron, w²(t)=[w2 1(t), w2 2(t), . . . , w2 J(t)]^T is the output weight vector between hidden neurons and output neuron, j=1, 2, . . . , J, J is the number of hidden neurons, and θ_j(t) is the output value of the jth hidden neuron which is usually defined by a normalized Gaussian function:

$\begin{array}{cc}{\theta }_j\left(t\right)=e^{-{h_j\left(t\right)-e_j\left(t\right)}^2/2{\sigma }_j^2\left(t\right)},& \left(2\right)\end{array}$

wherein ||h_j-c_j|| represents the Euclidean distance between h_j and c_j, c_j(t)=[c_1j(t), c_2j(t), . . . , c_5j(t)]^T and σ_j represent the center vector and radius of the jth hidden neuron, respectively; c_ij(t) is ith element of jth hidden neuron, and h_j is the input vector of jth hidden neuron

h _j(t)=[h_j1(t), h_j2(t), . . . , h_j1(t)],   (3)

h _ij(t)=u_j(t),   (4)

wherein w1j(t) is the feedback weight connecting the jth hidden neuron with the output neuron, w¹(t)=[w1 1(t), w1 2(t), . . . , w1 J(t)]^T is the feedback weight vector connecting the jth hidden neuron with the output neuron and y(t-1) is the output value of RSORBFNN at time t-1.

The training error function of RSORBFNN is defined

$\begin{array}{cc}E\left(t\right)=\frac{1}{P}\sum _{t=1}^p{\left(Y_d\left(t\right)-y\left(t\right)\right)}^2,& \left(5\right)\end{array}$

wherein P is the number of the training samples.

(3) Train RSORBFNN

1) Given RSORBFNN, the initial number of hidden layer neurons is J; J>2 is a positive integer. The input of RSORBFNN is x(1), x(2), . . . , x(t), . . . , x(P), the desired output is y_d(1), y_d(2), . . . , y_d(t), . . . , y_d(P); the desired error value is set to E_d, E_d∈(0, 0.01), the initial center is c_j(1)∈(−2, 2), the initial width value σ_j(1)∈(0, 1), the initial feedback weight is w1 j(1)∈(0, 1), and the initial weight is w2 j(1)∈(0, 1), j=1, 2, . . . , J;

2) Set the learning step s=1;

3) t=s, calculate the output y(t) of RSORBFNN, update the weight, width, and center of RSORBFNN using the rule:

Θ(t+1)=Θ(t)+(Ψ(t)+η(t)×I)⁻¹×Ω(t),   (6)

where Θ(t)=[w¹(t), w²(t), C(t), σ(t)] is the variable vector at time t, Ψ(t) is quasi

Hessian matrix at time t, I is the identity matrix, η(t) is the adaptive learning rate defined as:

$\begin{array}{cc}\eta \left(t\right)=\mu \left(t\right)\eta \left(t-1\right),& \left(7\right)\\ \mu \left(t\right)=\frac{\left({\beta }^{\max }\left(t\right)+\eta \left(t-1\right)\right)/\left(1+{\beta }^{\max }\left(t-1\right)\right)-{\beta }^{\min }\left(t-1\right)}{\eta \left(t-1\right)},& \left(8\right)\end{array}$

wherein μ(t) is the adapting factor at time t, and the initial value of μ(t) is μ(1)=1, β^max(t) and β^min(t) are the maximum and minimum eigenvalues of Ψ(t), respectively; 0<β^min(t)<β^max(t), 0<η(t)<1 and η(1)=1. Θ(t) contains four kinds of variables: the feedback connection weight vector w¹(t) at time t, the connection weight vector w²(t) at time t, the centre matrix C(t)=[c₁(t), c₂(t), . . . , c_j(t)]^T and width vector σ(t)=[σ₁(t), σ₂(t), . . . , σ_j(t)]^T at time t.

Θ(1)=[w¹(1), w²(1), C(1), σ(1)],   (9)

the quasi Hessian matrix Ψ(t) and the gradient vector Ω(t) are accumulated as the sum of related submatrices and vectors:

Ψ(t)=j^T(t)j(t),   (10)

Ω(t)=j^Te(t),   (11)

e(t)=y_d(t)−y(t),   (12)

e(t) is the approximating error at time t, y_d(t) is the desired output and y(t) is the network output at time t, and the Jacobian-vector j(t) is calculated as:

$\begin{array}{cc}j\left(t\right)=\left[\frac{\partial e\left(t\right)}{\partial w^1\left(t\right)},\frac{\partial e\left(t\right)}{\partial w^2\left(t\right)},\frac{\partial e\left(t\right)}{\partial C\left(t\right)},\frac{\partial e\left(t\right)}{\partial \sigma \left(t\right)}\right],& \left(13\right)\end{array}$

4) t>3, calculate competitiveness of the jth hidden neuron:

wherein cp_j(t) is the competitiveness of the jth hidden neuron, ρ denotes the correlation coefficient between the hidden layer output and network output, ρ∈(0, 1), f_j(t) is the active state of the jth hidden neuron, σ_j(t) is the width of the jth hidden neuron; the active state f_j(t) is defined as

f_j(t)=χ^−|_j(t)−c_j(t)|,   (15)

wherein χ∈(1,2), and f(t)=[f₁(t), f₂(t), . . . , f_j(t)], the correlation coefficient ρ_j(t) at time t is calculated as

$\begin{array}{cc}{\rho }_j\left(t\right)=\frac{\sum _{k=0}^3\left[A_j\left(t-k\right)-\stackrel{\_}{A}\left(t\right)\right]\left[B\left(t-k\right)-\stackrel{\_}{B}\left(t\right)\right]}{\sqrt{\sum _{k=0}^3{\left[A_j\left(t-k\right)-\stackrel{\_}{A}\left(t\right)\right]}^2\sum _{k=0}^3{\left[B\left(t-k\right)-\stackrel{\_}{B}\left(t\right)\right]}^2}},& \left(16\right)\end{array}$

wherein the correlation coefficient of hidden neurons A_j(t)=w2 j(t)θ_j (t), the correlation coefficient of output layer B(t)=y(t), Ā(t) is the average value of correlation coefficient of hidden neurons at time t, ^B(t) is the average value of correlation coefficient of output layer at time t;

5) Adjust the structure of RSORBFNN:

If the competitiveness of the jth hidden neuron and training error at time t and t+τ satisfy

$\begin{array}{cc}E\left(t\right)-E\left(t+\tau \right)\le \varepsilon ,& \left(17\right)\\ j=\arg \underset{1\le j\le J}{\max }\left({\mathrm{cp}}_j\left(t\right)\right),\mathrm{where}\arg \underset{1\le j\le J}{\max }\left({\mathrm{cp}}_j\left(t\right)\right)& \left(18\right)\end{array}$

denotes the value of j when cp_j(t) owns the maximum value. E(t) and E(t+τ) are the training errors at times t and t+τ, respectively, τ is a time interval, τ>2, and ε is the preset threshold, ε∈(0,0.01). Add one hidden neuron, and the number of hidden neurons is M₁=J+1. Otherwise, the structure of RSORBFNN will be not adjusted, M₁=J.

When the competitiveness of the jth hidden neuron satisfies

cp _j(t)<ξ,   (19)

wherein ξ is the preset pruning threshold, ξ∈(0, E_d), E_d is the preset error, E_d∈(0,0.01]. The jth hidden neuron will be pruned, the number of hidden neurons will be updated M₂=M₁−1. Otherwise, the structure of RSORBFNN will be not adjusted, M₂=M₁.

6) Increase 1 learning step for s, if s<P, go to step 3); if s=N, proceed to step 7).

7) Per Eq. (5), calculate the performance of RSORBFNN. If E(t)≧E_d, proceed to step 3); if E(t)<E_d, stop the training process.

(4) Effluent TN concentration prediction;

The testing samples are used as the input of RSORBFNN, the output of RSORBFNN is the soft-computing values of effluent TN.

The Novelties of this Present Disclosure Contain:

(1) To detect the effluent TN online and with acceptable accuracy, a computing implemented method is developed in this present disclosure. The results demonstrate that the effluent TN trends in WWTP can be predicted with acceptable accuracy using the NH₄—N, NO₃—N, effluent SS, BOD, TP as input variables. This computing implemented method can predict the effluent TN with acceptable accuracy and solve the problem that the effluent TN ‘s hard to be measured online.

(2) Since wastewater treatment process has the features of a complicated mechanism, and many influential factors, it was difficult to build a precise mathematical model to predict the effluent TN. Hence, the computing implemented method is based on the RSORBFNN in this present disclosure, which is proposed to predict it. The advantages of the proposed RSORBFNN are that it can simplify and accelerate the structure optimization process of the recurrent neural network, and can predict the effluent TN accurately. Moreover, the predicting performance shows that the RSORBFNN-based computing implemented method can adapt well to environment change. Therefore, this computing implemented method performs well in the whole operating space.

Attention: this present disclosure utilizes five input variables in this RSORBFNN method to predict the effluent TN. In fact, it is in the scope of this present disclosure that any of the variables: the NH₄—N, NO₃—N, effluent SS, BOD, TP are used to predict the effluent TN. Moreover, this RSORBFNN method is also able to predict the others variables in urban WWTP.

BRIEF DESCRIPTION OF THE DRAWINGS

FIG. 1 shows the structure of computing implemented method based on the RSORBFNN in this present disclosure.

FIG. 2 shows the training result of the computing implemented method.

FIG. 3 shows the training error of the computing implemented method.

FIG. 4 shows the predicting result of the computing implemented method.

FIG. 5 shows the predicting error of the computing implemented method.

DETAILED DESCRIPTION

Various A computing implemented method is developed to predict the effluent TN based on an RSORBFNN in this present disclosure. For this computing implemented method, the inputs are those variables that are easy to measure and the outputs are estimates of the effluent TN. In general, the procedure of computing implemented method consists of three parts: data acquisition, data pre-processing and model design. For this present disclosure, an experimental hardware is set up as shown in FIG. 1. The historical process data are routinely acquired and stored in the data acquisition system. The input-output water quality data can be easily retrieved, measured during the year 2011. The variables whose data are easy to measure by the instruments consist of NH₄—N, NO₃—N, effluent SS, BOD, TP and effluent TN were used as experimental samples. After deleting abnormal data, 100 groups were obtained and normalized, 60 groups were used as training data, whilst the remaining 40 were used as testing data.

This present disclosure adopts the following technical scheme and implementation steps:

A computing implemented method for the effluent TN based on an RSORBFNN, its characteristic and steps include the following steps:

(1) Determine the input and output variables of effluent TN:

For sewage treatment process of activated sludge system, the variables of sewage treatment process are analyzed and select the input variables of effluent TN soft-computing model: ammonia nitrogen—NH₄—N, nitrate nitrogen—NO₃—N, effluent suspended solids—SS, biochemical oxygen demand—BOD, total phosphorus—TP, The output value of soft-computing model is detected effluent TN.

(2) Initialize RSORBFNN

The initial structure of RSORBFNN consists of three layers: input layer, hidden layer, and output layer. There are 5 neurons in the input layer, J neurons in the hidden layer and 1 neuron in the output layer; J>2 is a positive integer. Connection weights between input layer and hidden layer are assigned 1, the feedback weights between hidden layer and output layer randomly assign values, the assignment internal is [1, 1]; the number of the training sample is P, and the input vector of RSORBFNN is x(t)=[x₁(t), x₂(t), x₃(t), x₄(t), x₅(t)] at time t; y(t) is the output of RSORBFNN, and y_d(t) is the real value of effluent TN at time t, respectively; The output of RSORBFNN can be described:

$\begin{array}{cc}y\left(t\right)=\sum _{j=1}^Jw_j^2\left(t\right){\theta }_j\left(t\right),& \left(20\right)\end{array}$

wherein w2 j(t) is the output weight between the jth hidden neuron and the output neuron, w²(t)=[w2 1(t), w2 2(t), . . . , w2 J(t)]^T is the output weight vector between hidden neurons and output neuron, j=1, 2, . . . , J, J is the number of hidden neurons, and θ_j(t) is the output value of the jth hidden neuron which is usually defined by a normalized Gaussian function:

$\begin{array}{cc}{\theta }_j\left(t\right)=e^{-{h_j\left(t\right)-c_j\left(t\right)}^2/2{\sigma }_j^2\left(t\right)},& \left(21\right)\end{array}$

wherein ||h_j-c_j|| represents the Euclidean distance between h_j and c_j, c_j(t)=[c_1j(t), c_2j(t), . . . , c_5j(t)]^T and σ_j represent the center vector and radius of the jth hidden neuron, respectively; c_ij(t) is ith element of jth hidden neuron, and h_j is the input vector of jth hidden neuron

h _j(t)=[h_j1(t), h_j2(t), . . . , h_j1(t)],   (22)

h _ij(t)=u_i(t),   (23)

wherein w1 j(t) is the feedback weight connecting the jth hidden neuron with the output neuron, w¹(t)=[w1 1(t), w1 2(t), . . . , w1 J(t)]^T is the feedback weight vector connecting the jth hidden neuron with the output neuron and y(t-1) is the output value of RSORBFNN at time t-1.

The training error function of RSORBFNN is defined

$\begin{array}{cc}E\left(t\right)=\frac{1}{P}\sum _{t=1}^P{\left(y_d\left(t\right)-y\left(t\right)\right)}^2,& \left(24\right)\end{array}$

wherein P is the number of the training samples.

(3) Train RSORBFNN

1) Given RSORBFNN, the initial number of hidden layer neurons isJ;J>2 is a positive integer. The input of RSORBFNN is x(1), x(2), . . . , x(t), . . . , x(P), the desired output is y_d(1), y_d(2), . . . , y_d(t), . . . , y_d(P); the desired error value is set to E_d, E_d∈(0, 0.01), the initial center is CJ(1)∈(−2, 2), the initial width value σ_j(1) ∈(0, 1), the initial feedback weight is w1 j(1)∈(0, 1), and the initial weight is w2 j(1)∈(0, 1), j=1, 2, . . . , J;

2) Set the learning step s=1;

3) t=s, calculate the output y(t) of RSORBFNN, update the weight, width, and center of RSORBFNN using the rule:

Θ(t+1)=Θ(t)+(Ψ(t)+η(t)×I)⁻¹×Ω(t),   (25)

where Θ(t+1)=[w¹(t), w²(t), C(t), σ(t)[ is the variable vector at time t, Ψ(t) is quasi

Hessian matrix at time t, I is the identity matrix, η(t) is the adaptive learning rate defined as:

$\begin{array}{cc}\eta \left(t\right)=\mu \left(t\right)\eta \left(t-1\right),& \left(26\right)\\ \mu \left(t\right)=\frac{\left({\beta }^{\max }\left(t\right)+\eta \left(t-1\right)\right)/\left(1+{\beta }^{\max }\left(t-1\right)\right)-{\beta }^{\min }\left(t-1\right)}{\eta \left(t-1\right)},& \left(27\right)\end{array}$

wherein μ(t) is the adapting factor at time t, and the initial value of μ(t) is μ(1)=1, β^max(t) and β^min(t) are the maximum and minimum eigenvalues of Ψ(t), respectively; 0<β^min(t)<β^max(t), 0<η(t)<1 and η(1)=1. Θ(t) contains four kinds of variables: the feedback connection weight vector w¹(t) at time t, the connection weight vector w²(t) at time t, the centre matrix C(t)=[c₁(t), c₂(t), . . . , c_j(t)]^T and width vector σ(t)=[σ₁(t), σ₂(t), . . . , σ_j(t)]^T at time t.

Θ(1)=[w¹(1), w²(1), C(1), σ(1)],   (28)

the quasi Hessian matrix Ψ(t) and the gradient vector Ω(t) are accumulated as the sum of related submatrices and vectors:

Ψ(t)=j^T(t)j(t),   (29)

Ω(t)=j^Te(t),   (30)

e(t)=y_d(t)−y(t),   (31)

e(t) is the approximating error at time t, y_d(t) is the desired output and y(t) is the network output at time t, and the Jacobian-vector j(t) is calculated as:

$\begin{array}{cc}j\left(t\right)=\left[\frac{\partial e\left(t\right)}{\partial w^1\left(t\right)},\frac{\partial e\left(t\right)}{\partial w^2\left(t\right)},\frac{\partial e\left(t\right)}{\partial C\left(t\right)},\frac{\partial e\left(t\right)}{\partial \sigma \left(t\right)}\right],& \left(32\right)\end{array}$

4) t>3, calculate competitiveness of the jth hidden neuron:

cp _j(t)=ρf_j(t)σ_j(t), j=1, 2, . . . ,J,   (33)

wherein cp_j(t) is the competitiveness of the jth hidden neuron, ρ denotes the correlation coefficient between the hidden layer output and network output, ρ∈(0, 1),f_i(t) is the active state of the jth hidden neuron, σ_j(t) is the width of the jth hidden neuron; the active state f_j(t) is defined aswherein χ∈(1,2), and f(t)=[f₁(t), f₂(t), . . . , f_j(t)], the correlation coefficient _AM at time t is calculated as

$\begin{array}{cc}{\rho }_j\left(t\right)=\frac{\sum _{k=0}^3\left[A_j\left(t-k\right)-\stackrel{\_}{A}\left(t\right)\right]\left[B\left(t-k\right)-\stackrel{\_}{B}\left(t\right)\right]}{\sqrt{\sum _{k=0}^3{\left[A_j\left(t-k\right)-\stackrel{\_}{A}\left(t\right)\right]}^2\sum _{k=0}^3{\left[B\left(t-k\right)-\stackrel{\_}{B}\left(t\right)\right]}^2}},& \left(35\right)\end{array}$

wherein the correlation coefficient of hidden neurons A_j(t)=w2 j(t) θ_j(t), the correlation coefficient of output layer B(t)=y(t), Ā(t) is the average value of correlation coefficient of hidden neurons at time t, ^B(t) is the average value of correlation coefficient of output layer at time t;

5) Adjust the structure of RSORBFNN:

If the competitiveness of the jth hidden neuron and training error at time t and t+τ satisfy

$\begin{array}{cc}E\left(t\right)-E\left(t+\tau \right)\le \varepsilon ,& \left(36\right)\\ j=\arg \underset{1\le j\le J}{\max }\left({\mathrm{cp}}_j\left(t\right)\right),\mathrm{where}\arg \underset{1\le j\le J}{\max }\left({\mathrm{cp}}_j\left(t\right)\right)& \left(37\right)\end{array}$

denotes the value of jwhen cp_j(t) obtain the maximum value. E(t) and E(t+τ) are the training errors at times t and t+τ, respectively, τ is a time interval, τ=5, and ε is the preset threshold, ε=0.001. Add one hidden neuron, and the number of hidden neurons is M₁=J+1. Otherwise, the structure of RSORBFNN will be not adjusted, M₁=J.

When the competitiveness of the jth hidden neuron satisfies

cp _j(t)<ξ,   (38)

wherein ξ is the preset pruning threshold, ξ∈(0, E_d), E_d is the preset error, E_d=0.002. The jth hidden neuron will be pruned, the number of hidden neurons will be updated M₂=M₁−1. Otherwise, the structure of RSORBFNN will be not adjusted, M₂=M₁.

6) Increase 1 learning step for s, if s<P, go to step 3); if s=N, proceed to step 7).

7) According to Eq. (24), calculate the performance of RSORBFNN. If E(t)≧E_d, proceed to step 3); if E(t)<E_d, stop the training process.

The training result of the computing implemented method for effluent TN is shown in FIG. 2. X-axis indicates the number of samples. Y axis shows the effluent TN. The unit of Y axis is mg/L. The solid line presents the real values of effluent TN. The dotted line shows the outputs of computing implemented method in the training process. The errors between the true values and the outputs of intelligent detecting method in the training process are shown in FIG. 3. X-axis indicates the number of samples. Y axis shows the training error. The unit of Y axis is mg/L.

(4) Effluent TN concentration prediction;

The testing samples are used as the input of RSORBFNN, and the output of RSORBFNN is the soft-computing values of effluent TN. The predicting result is shown in FIG. 4. X-axis indicates the number of testing samples. Y axis shows the effluent TN. The unit of Y axis is mg/L. The solid line presents the real values of effluent TN. The dotted line shows the outputs of intelligent detecting method in the testing process. The errors between the true values and the outputs of intelligent detecting method in the testing process are shown in FIG. 5. X-axis shows the number of samples. Y axis shows the testing error. The unit of Y axis is mg/L.

Tables 1-14 show the experimental data in this present disclosure. Tables 1-6 show the training samples of biochemical oxygen demand - BOD, ammonia nitrogen—NH₄—N, nitrate nitrogen—NO₃—N, effluent suspended solids —SS, total phosphorus—TP real effluent TN. Table 7 shows the outputs of the RSORBFNN in the training process. Tables 8-14 show the testing samples of biochemical oxygen demand—BOD, ammonia nitrogen—NH₄—N, nitrate nitrogen —NO₃—N, effluent suspended solids—SS, total phosphorus—TP and real effluent TN. Table 14 shows the outputs of the RSORBFNN in the predicting process

Training samples are provided as follow.

TABLE 1
The training samples of biochemical oxygen demand-BOD (mg/L)
192	222	201	264	195	209	260	197	206	289
188	350	210	204	200	180	230	338	200	330
320	232	260	240	218	316	310	172	210	316
310	244	248	168	204	145	170	142	190	260
200	240	280	174	250	136	222	204	239	242
310	232	290	210	144	214	251	158	262	290

TABLE 2
The training samples of ammonia nitrogen-NH4—N (mg/L)
64.3	69.4	72.6	71.7	71.5	63.5	70.7	68.4	64.3	68.3
71.9	64.3	63.8	56.9	44.6	64.9	68.9	76.9	63.5	70
60.3	60	72.1	69.7	70.5	66.1	62.2	58.8	60.5	63.5
65.7	59.4	54.8	60	59.1	63.7	64.5	58.1	61.9	66.7
57.6	70.7	61.3	57.8	55.3	65.8	65.1	61.3	72	62.8
63.4	61.4	71.3	61.2	58.7	55.7	67.7	58.5	61.5	73.2

TABLE 3
The training samples of nitrate nitrogen-NO3—N (mg/L)
13.8325	13.7215	13.6408	13.6666	13.7288	13.8617	13.8873	13.9157	13.9758	14.1119
14.4164	14.4829	15.2031	15.2791	15.6909	16.1498	16.6379	16.9443	16.8975	16.8101
16.5498	16.2205	15.7517	15.3732	14.5885	13.9968	13.5851	12.9808	12.6256	12.2428
11.9133	11.6286	11.4642	10.7946	10.3934	10.4852	10.9491	11.5281	12.2201	12.8419
13.3324	13.0934	12.8794	12.9103	12.5906	12.3108	12.0798	11.9742	11.8102	11.6730
11.6093	11.4942	11.4940	11.5036	11.4617	11.4878	11.3927	11.3851	11.4866	11.7895

TABLE 4
The training samples of effluent suspended solids-SS (mg/L)
146	192	226	208	154	264	276	208	178	250
204	288	210	172	200	170	214	324	186	422
168	238	232	260	184	330	312	230	162	300
268	231	270	132	252	204	148	116	182	292
210	210	350	214	212	170	262	178	228	164
296	308	240	170	140	178	196	312	164	320

TABLE 5
The training samples of total phosphorus-TP (mg/L)
6.38	6.71	7.15	7.29	6.31	7.03	7.35	7.05	6.66	7.28
7.06	7.73	6.92	6.7	6.91	6.38	7.18	7.81	7.39	8.21
6.56	6.83	6.95	7.41	6.82	9.84	7.91	7.23	6.64	7.3
7.81	7.19	6.63	6	6.65	5.84	5.87	6.15	6.53	7.62
6.9	6.2	8.08	6.47	7.2	5.86	7.69	6.55	6.94	7.01
7.78	6.98	7.55	6.56	5.92	6.17	7.05	6.73	7.65	8.09

TABLE 6
The training samples of real effluent TN (mg/L)
75.3	86	91.3	91.8	88.5	83.9	84.8	82.1	80	84.4
80	89.6	79.9	82.2	77.6	55.5	85.1	85.4	90.4	84.2
80.9	76.1	73.7	86.6	83.1	85.9	81.7	79.6	72	78
79.3	81.77	73.7	62.4	73.2	70.7	72.2	71.1	63	75.3
81.8	72.7	88.9	77.4	74.1	71.2	80.5	76.5	75.8	82.6
80.1	70.3	86.5	71.5	67.9	65.6	68.6	70.9	77.4	87.2

TABLE 7
The effluent TN outputs in the training process (mg/L)
75.09123	85.75465	91.29607	91.6917	88.23302	83.95164	85.46349	82.11712	79.64609	84.5503
79.87456	89.64711	79.92864	81.83561	77.36899	57.73073	84.80773	85.69525	90.44198	82.75301
81.46583	76.12251	73.87198	86.63506	82.91107	85.88516	81.91191	79.37446	72.01563	78.18965
79.34218	81.66961	73.74434	62.82255	73.0666	70.48056	72.29508	71.25872	63.62556	74.98458
81.483	72.48675	88.93721	77.31496	74.22315	70.59969	80.91807	76.37911	75.78082	82.65934
80.05047	71.01168	85.82914	71.58082	67.73245	65.72093	69.74704	69.91498	76.98607	87.36917

Testing Samples:

TABLE 8
The testing samples of biochemical oxygen demand-BOD (mg/L)
217	226	218	390	260	200	248	370	342	347
290	440	289	460	188	318	334	290	341	335
287	346	266	430	294	450	262	372	370	198
347	610	326	283	395	233	331	209	282	174

TABLE 9
The testing samples of ammonia nitrogen-NH4—N (mg/L)
48.6	56.9	64.2	58.9	50.3	61.3	63.7	68.6	54	40.8
53.4	60.2	66.4	60.9	63.4	54.4	40.7	69	63.4	55
66.3	63.2	62.3	52.7	60.5	57	62.1	68.2	64	69
67.2	61.5	66	64.5	62.1	51.4	51	55.5	55.5	58.5

TABLE 10
The testing samples of nitrate nitrogen-NO3—N (mg/L)
12.3085	12.6792	13.0400	13.2389	13.5262	13.4614	13.2849	12.9682	12.7089	12.2269
12.0995	12.1315	12.1361	12.2122	12.2197	12.3499	12.4464	12.4927	12.7326	12.8156
12.9392	13.0438	13.7367	14.1627	14.8751	15.9604	16.7487	17.6572	18.6773	19.1970
19.9069	20.5030	20.9495	21.3475	21.8734	22.4720	22.7922	23.2325	23.4924	23.2459

TABLE 11
The testing samples of effluent suspended solids-SS (mg/L)
154	158	214	204	110	232	226	254	122	538
130	162	142	360	376	231.2	166	118	142	220
266	172	296	235	180	146	206	208	202	146
398	270	328	126	244	218	272	168	262	110

TABLE 12
The testing samples of total phosphorus-TP (mg/L)
5.17	5.39	6.03	5.96	5.24	6.22	5.78	6.17	5.6	5.22
4.75	5.46	6.1	6.48	6.84	5.5	4.06	5.74	5.73	5.8
6.71	5.63	6.18	5.11	5.03	4.6	5.24	5.86	5.62	6.13
7.01	6.11	6.65	5.56	6.52	6.22	6.25	5.2	5.77	6.17

TABLE 13
The testing samples of real effluent TN (mg/L)
62.8	67.4	75.3	70.1	59.4	78.5	75.4	77.3	70.2	54.5
60.7	66.7	74.1	74.9	78.6	66	60.9	65.4	52.3	60.5
72.7	68.2	70	65.1	69.1	61.9	69.3	71.5	70.7	76.7
80.8	73.9	77.3	73.5	76.3	73.4	74.1	64.5	66.6	67.8

TABLE 14
The effluent TN outputs in the testing process (mg/L)
60.43193	67.16412	75.34496	70.96676	63.11076	67.66785	81.3452	79.78831	64.06407	58.64447
63.99991	66.24501	72.44785	72.43734	77.82645	67.75635	57.96904	75.32191	63.95107	56.05289
62.8231	65.67208	71.03243	61.22433	66.2433	65.8583	68.8428	76.71578	67.04345	74.80853
78.61247	75.88474	80.21718	68.98426	77.51966	67.57056	73.42719	71.17669	65.88281	66.41494

Claims

1. A method of detecting the effluent total nitrogen (TN) based on a recurrent self-organizing RBF neural network (RSORBFNN), the method comprising: (1) determining input and output variables of the effluent TN with respect to a sewage treatment process of an activated sludge system by analyzing the variables of the sewage treatment process and selecting the input variables of the effluent TN computing model that include: ammonia nitrogen (NH4—N), nitrate nitrogen (NO3—N), effluent suspended solids (SS), biochemical oxygen demand (BOD), total phosphorus (TP), an output value of the computing model is detected effluent TN;(2) initializing the RSORBFNN of which an initial structure comprises three layers: input layer, hidden layer, and output layer, there are 5 neurons in the input layer, J neurons in the hidden layer, and 1 neuron in the output layer, 1>2 is a positive integer, connection weights between the input layer and hidden layer are assigned 1, feedback weights between hidden layer and output layer are randomly assigned with values, an assignment internal is [1, 1]; the number of the training sample is P, and an input vector of the RSORBFNN is x(t)=[x1(t), x2(t), x3(t), x4(t), x5(t)] at time t; y(t) is an output of the RSORBFNN, and yd(t) is a real value of the effluent TN at time t, respectively; the output of the RSORBFNN is described using the equation (1):