Water Environment Quality Analysis Based on Information Diffusion Theory and Fuzzy Neural Network

Reservoirs play a key role in many infrastructure functions for people like flood control, irrigation, and water supply. In this work, we focused on the water quality evaluation model for Shimen Reservoir. Based on the monthly changes of factors such as pH, nitrate, ammonia nitrogen (NH3-N) and total nitrogen (TN) in 2013 and 2014, the information diffusion theory and fuzzy neural network technology were utilized to evaluate the water quality comprehensively. The probability distribution of these four factors in the reservoir was analysed and the water quality of the reservoir evaluated. The results show its reliability and these two methods can provide a basis for water quality control of Shimen Reservoir. Furthermore, the methods can be universally applied to the analysis and research of water quality in other regions..


INTRODUCTION
The Shimen Reservoir is a large-scale water conservancy project with flood control, irrigation, and urban water supply as its main function, taking into account the functions of power generation and fish farming. Therefore, it is important to ensure the basic stability of the water quality of the reservoir. At present, there are many methods for water quality evaluation, and more methods such as single factor evaluation method, comprehensive evaluation method and Nemerow index method are used. The single-factor index method is simple to calculate, but it can only reflect the change of a single factor, so many scholars combine single-factor evaluation with other methods to get better results. Song (2018) used a single factor evaluation method combined with an improved grey correlation to evaluate water quality. Wu (2019) evaluates the groundwater environment in Pinghai Bay, Putian City based on a single factor and fuzzy mathematics comprehensive evaluation method.
Information diffusion theory can dig out more information when the amount of data is insufficient and the samples are insufficient. It can be well combined with artificial neural network models and applied to many fields. Currently, mainly involved in meteorological disasters, signal processing, geological disasters, surveying and mapping, etc. Liu et al. (2019) applied entropy information diffusion theory to the risk assessment of agricultural drought and flood disasters in the middle and lower reaches of the Yangtze River. Zhong et al. (2019) evaluation of flash flood risk was based on information diffusion method. Wang et al. (2016) used information diffusion theory to study flood and drought disaster risk characteristics in southern China. Lu et al. (2014) applied information diffusion technology in the probability analysis of grassland biological disaster risk. Besides, it is involved in some aspects such as thunderstorms, pests, typhoons and crop yields.
Regarding water environment issues, some scholars have applied the information diffusion theory. Li (2007) used the information diffusion technology to study the river health risk estimation model under the condition of incomplete information. The inspection of reservoir water quality is very complicated. Therefore, grasping the changes in the content of each element in the reservoir can better analyse the water quality and maintain the normal use of the reservoir. This paper analyses the risk probability of each element's pollution index through a combination of single factor index and information diffusion theory. It is of great significance to prevent the pollution index from exceeding the standard and the water quality to be stable.
The T-S fuzzy neural network model is an organic combination of fuzzy logic and neural networks. It inherits the advantages of both fuzzy logic and neural networks. It can represent highly nonlinear complex systems with …(3 Where, PHsd is the lower limit of the evaluation standard; PHsu is the upper limit of the evaluation standard; and e measured value of PH. Table 1：Standard value of basic items for environmental quality standard for surface water (unit in mg/L).

Classification
Class Ⅰ Class Ⅱ Class Ⅲ Class Ⅳ Class Ⅴ standard value items pH (dimensionless) 6～9 NH 3 -N 0.15 0.5 1.0 1.5 2.0 TN (N for lakes and reservoirs) 0.2 0.5 1.0 1.5 2.0 According to the standard limit value of the supplementary project for centralized drinking water and surface water the standard value of nitrate (in N) is 10 mg/L.

Information Diffusion Theory
Information diffusion is a kind of fuzzy mathematical processing method for set-valued samples. In order to mak the lack of information, it is considered to preferentially use the fuzzy information of samples, so as to set-valued The information diffusion method can turn a sample with observations into a fuzzy set, that is, a single-valued sam a set-valued sample. The most commonly used model is the normal diffusion model. Suppose that the set of actual observation samples is X, the sample series is X = {x 1 , x 2 , ⋯ , x n }, and the discourse …(2) in southern China. Lu et al. (2014) applied information diffusion technology in the probability analysis of logical disaster risk. Besides, it is involved in some aspects such as thunderstorms, pests, typhoons and cro Regarding water environment issues, some scholars have applied the information diffusion theory. Li (2 information diffusion technology to study the river health risk estimation model under the condition of inc mation. The inspection of reservoir water quality is very complicated. Therefore, grasping the changes in each element in the reservoir can better analyse the water quality and maintain the normal use of the reserv analyses the risk probability of each element's pollution index through a combination of single factor index a diffusion theory. It is of great significance to prevent the pollution index from exceeding the standard and th to be stable. The T-S fuzzy neural network model is an organic combination of fuzzy logic and neural networks. It i vantages of both fuzzy logic and neural networks. It can represent highly nonlinear complex systems wi rules, which is very suitable for water quality evaluation. Mo et al. (2017) evaluated the water quality of the Qinzhou based on the T-S fuzzy network model. Zhang et al. (2018) combined fuzzy neural network with L predict comprehensive water quality. Zhao (2018) uses neural networks to study early warning of aquacul portation environment. This paper uses a fuzzy neural network to evaluate the water quality of the Shimen

Single Factor Index
The single factor evaluation method is to determine the category of comprehensive water quality of the wa category of the single index with the worst water quality. This method is simple to calculate and can dire pollution of a single factor. With reference to the "Environmental Quality Standard for Surface Water" (G Category V water standard (2002), the formula is: Where, C i is the measured value of type i pollutant, and S i is the evaluation standard of type i pollutant. When P i ≤ 1 , it means that the water body is not polluted; when P i > 1 , it means that the water body is po The standard index for PH value is: Where, PHsd is the lower limit of the evaluation standard; PHsu is the upper limit of the evaluation standard e measured value of PH. According to the standard limit value of the supplementary project for centralized drinking water and surface the standard value of nitrate (in N) is 10 mg/L.

Information Diffusion Theory
Information diffusion is a kind of fuzzy mathematical processing method for set-valued samples. In order the lack of information, it is considered to preferentially use the fuzzy information of samples, so as to set-v The information diffusion method can turn a sample with observations into a fuzzy set, that is, a single-valu Where, PH sd is the lower limit of the evaluation standard; PH su is the upper limit of the evaluation standard; and PH j is the measured value of PH.
According to the standard limit value of the supplementary project for centralized drinking water and surface water sources, the standard value of nitrate (in N) is 10 mg/L.

Information Diffusion Theory
Information diffusion is a kind of fuzzy mathematical processing method for set-valued samples. In order to make up for the lack of information, it is considered to preferentially use the fuzzy information of samples, so as to set-valued samples.
The information diffusion method can turn a sample with observations into a fuzzy set, that is, a single-valued sample into a set-valued sample. The most commonly used model is the normal diffusion model.
Suppose that the set of actual observation samples is X, the sample series is X={x 1 ,x 2 ,...,x n }, and the discourse domain is U={u 1 ,u 2 ,...,u m }. A single-valued observation sample can carry the information spread to all points in U. pH (dimensionless) 6～ NH 3 -N 0.15 0.5 1.0 TN (N for lakes and reservoirs) 0.2 0.5 1.0 According to the standard limit value of the supplementary the standard value of nitrate (in N) is 10 mg/L.

Information Diffusion Theory
Information diffusion is a kind of fuzzy mathematical pro the lack of information, it is considered to preferentially u The information diffusion method can turn a sample with a set-valued sample. The most commonly used model is t Suppose that the set of actual observation samples is X, th is U = {u 1 , u 2 , ⋯ , u m }. A single-valued observation samp Where, h is the information diffusion coefficient, which is determined according to the maximum value b, the minimum Where, h is the information diffusion coefficient, which is determined according to the maximum valu value a, and the number of samples n in the sample set. ℎ = Where, C i is the normal diffusion information sum of the observation sample x i . The membership function of the corresponding fuzzy subset is: Assume that after performing the above processing on all the samples, the number of samples with an u j is inferred to be q(u j ). Let Q is the sum of the number of sample points at each u j point.
Where p(u j ) is the frequency value at which the sample point falls at uj, which can be used as an estimate Its transcendence probability is: Where, P(u ≥ u j ) is the probability of exceeding u j .

T-S Fuzzy Neural Network
The T-S fuzzy system can be defined by the "if-then" rule form. In the case of the rule R i , the fuzzy reaso Where A j i is the fuzzy set of the fuzzy system; p j i is the fuzzy system parameter; yi is the output obtaine fuzzy rules, the input part (the if part) is fuzzy, and the output part (the then part) is determined. The fuzz is a linear combination of the inputs. T-S fuzzy system is a kind of fuzzy system with strong adaptive ability. The model can not only update also continuously modify the membership function of the fuzzy subset. It consists of an antecedent ne ware network.

Antecedent network
The first layer is the input layer. Assume that the input quantity x = [x 1 , x 2 , ⋯ x k ] T , and the number of is k. The second layer is the fuzzification layer. The membership degree of each input variable x j is calculate fuzzy rules.
Where, c j i and b j i are the center and width of the membership function respectively; k is the input parame of fuzzy subsets. The third layer is the fuzzy rule layer. The fuzzy operator is used as the multiplication operator to perf tions for each membership degree.
(1 The fourth layer is used to normalize the applicability of each rule. The calculation expression is: Where, h is the information diffusion coefficient, which value a, and the number of samples n in the sample set. ℎ Assume that after performing the above processing on al u j is inferred to be q(u j ). Let Q is the sum of the number of sample points at each u j po Where p(u j ) is the frequency value at which the sample p Its transcendence probability is: Where, P(u ≥ u j ) is the probability of exceeding u j .

T-S Fuzzy Neural Network
The T-S fuzzy system can be defined by the "if-then" rule Where A j i is the fuzzy set of the fuzzy system; p j i is the fu fuzzy rules, the input part (the if part) is fuzzy, and the out is a linear combination of the inputs. T-S fuzzy system is a kind of fuzzy system with strong ad Where, C i is the normal diffusion information sum of the observation sample x i .
The membership function of the corresponding fuzzy subset is: Where, h is the information diffusion coefficient, which is Assume that after performing the above processing on all t u j is inferred to be q(u j ). Let Q is the sum of the number of sample points at each u j poin Where p(u j ) is the frequency value at which the sample poi Its transcendence probability is: Where, P(u ≥ u j ) is the probability of exceeding u j .

T-S Fuzzy Neural Network
The T-S fuzzy system can be defined by the "if-then" rule fo Where A j i is the fuzzy set of the fuzzy system; p j i is the fuzz fuzzy rules, the input part (the if part) is fuzzy, and the outpu …(7) Assume that after performing the above processing on all the samples, the number of samples with an observed value of u j is inferred to be q(u j ).

Let
Where, h is the information diffusion coefficient, w value a, and the number of samples n in the sample Assume that after performing the above processing u j is inferred to be q(u j ). Let Where p(u j ) is the frequency value at which the sam Its transcendence probability is: Where, P(u ≥ u j ) is the probability of exceeding u

T-S Fuzzy Neural Network
The T-S fuzzy system can be defined by the "if-then Where A i is the fuzzy set of the fuzzy system; p i is

…(8)
Where, h is the information diffusion coefficient, Assume that after performing the above processin u j is inferred to be q(u j ). Let Q is the sum of the number of sample points at ea Where p(u j ) is the frequency value at which the s Its transcendence probability is: Where, P(u ≥ u j ) is the probability of exceeding

T-S Fuzzy Neural Network
The T-S fuzzy system can be defined by the "if-the Q is the sum of the number of sample points at each u j point. Table 1：Standard value of basic items for environmental quality standard for surface water (unit in mg/L).

Classification standard valueitems
Class Ⅰ Class Ⅱ Class Ⅲ Class Q is the sum of the number of sample points at each u j point.
Where p(u j ) is the frequency value at which the sample point falls at u j , which can be used as an estimate of the pro Its transcendence probability is: is the probability of exceeding u j .

T-S Fuzzy Neural Network
The T-S fuzzy system can be defined by the "if-then" rule form. In the case of the rule R i , the fuzzy reasoning is as Where A j i is the fuzzy set of the fuzzy system; p j i is the fuzzy system parameter; yi is the output obtained accordi fuzzy rules, the input part (the if part) is fuzzy, and the output part (the then part) is determined. The fuzzy inferen is a linear combination of the inputs. T-S fuzzy system is a kind of fuzzy system with strong adaptive ability. The model can not only update automati also continuously modify the membership function of the fuzzy subset. It consists of an antecedent network and ware network.

Antecedent network
The first layer is the input layer. Assume that the input quantity x = [x 1 , x 2 , ⋯ x k ] T , and the number of nodes in is k. The second layer is the fuzzification layer. The membership degree of each input variable x j is calculated accordi fuzzy rules.
Where, c j i and b j i are the center and width of the membership function respectively; k is the input parameter; n is th of fuzzy subsets. The third layer is the fuzzy rule layer. The fuzzy operator is used as the multiplication operator to perform fuzzy tions for each membership degree.
The fourth layer is used to normalize the applicability of each rule. The calculation expression is: Where p(u j ) is the frequency value at which the sample point falls at u j , which can be used as an estimate of the probability. Its transcendence probability is: Q is the sum of the number of sample points at each u j point.
Where p(u j ) is the frequency value at which the sample point falls at u j , which can be used as an estimate of the proba Its transcendence probability is: Where, P(u ≥ u j ) is the probability of exceeding u j .

T-S Fuzzy Neural Network
The T-S fuzzy system can be defined by the "if-then" rule form. In the case of the rule R i , the fuzzy reasoning is as fo Where A j i is the fuzzy set of the fuzzy system; p j i is the fuzzy system parameter; yi is the output obtained according fuzzy rules, the input part (the if part) is fuzzy, and the output part (the then part) is determined. The fuzzy inference o is a linear combination of the inputs. T-S fuzzy system is a kind of fuzzy system with strong adaptive ability. The model can not only update automatical also continuously modify the membership function of the fuzzy subset. It consists of an antecedent network and an ware network.

Antecedent network
The first layer is the input layer. Assume that the input quantity x = [x 1 , x 2 , ⋯ x k ] T , and the number of nodes in this is k. The second layer is the fuzzification layer. The membership degree of each input variable x j is calculated according fuzzy rules.
Where, c j i and b j i are the center and width of the membership function respectively; k is the input parameter; n is the nu of fuzzy subsets. The third layer is the fuzzy rule layer. The fuzzy operator is used as the multiplication operator to perform fuzzy ca tions for each membership degree.
The fourth layer is used to normalize the applicability of each rule. The calculation expression is: Where, P(u ≥ u j ) is the probability of exceeding .

T-S Fuzzy Neural Network
The T-S fuzzy system can be defined by the "if-then" rule form. In the case of the rule R i , the fuzzy reasoning is as follows: Assume that after performing the above processing on all the samples, the number of samples with an observed value of u j is inferred to be q(u j ). Let Q is the sum of the number of sample points at each u j point.
Where p(u j ) is the frequency value at which the sample point falls at u j , which can be used as an estimate of the probability. Its transcendence probability is: Where, P(u ≥ u j ) is the probability of exceeding u j .

T-S Fuzzy Neural Network
The T-S fuzzy system can be defined by the "if-then" rule form. In the case of the rule R i , the fuzzy reasoning is as follows: Where A j i is the fuzzy set of the fuzzy system; p j i is the fuzzy system parameter; yi is the output obtained according to the fuzzy rules, the input part (the if part) is fuzzy, and the output part (the then part) is determined. The fuzzy inference output is a linear combination of the inputs. T-S fuzzy system is a kind of fuzzy system with strong adaptive ability. The model can not only update automatically but also continuously modify the membership function of the fuzzy subset. It consists of an antecedent network and an after ware network.

Antecedent network
The first layer is the input layer. Assume that the input quantity x = [x 1 , x 2 , ⋯ x k ] T , and the number of nodes in this layer is k. The second layer is the fuzzification layer. The membership degree of each input variable x j is calculated according to the fuzzy rules.
Where, c j i and b j i are the center and width of the membership function respectively; k is the input parameter; n is the number of fuzzy subsets. The third layer is the fuzzy rule layer. The fuzzy operator is used as the multiplication operator to perform fuzzy calculations for each membership degree.
The fourth layer is used to normalize the applicability of each rule. The calculation expression is: Where Assume that after performing the above processing on all the samples, the number of samples with an observed value of u j is inferred to be q(u j ). Let Q is the sum of the number of sample points at each u j point.
Where p(u j ) is the frequency value at which the sample point falls at u j , which can be used as an estimate of the probability. Its transcendence probability is: Where, P(u ≥ u j ) is the probability of exceeding u j .

T-S Fuzzy Neural Network
The T-S fuzzy system can be defined by the "if-then" rule form. In the case of the rule R i , the fuzzy reasoning is as follows: Where A j i is the fuzzy set of the fuzzy system; p j i is the fuzzy system parameter; yi is the output obtained according to the fuzzy rules, the input part (the if part) is fuzzy, and the output part (the then part) is determined. The fuzzy inference output is a linear combination of the inputs. T-S fuzzy system is a kind of fuzzy system with strong adaptive ability. The model can not only update automatically but also continuously modify the membership function of the fuzzy subset. It consists of an antecedent network and an after ware network.

Antecedent network
The first layer is the input layer. Assume that the input quantity x = [x 1 , x 2 , ⋯ x k ] T , and the number of nodes in this layer is k. The second layer is the fuzzification layer. The membership degree of each input variable x j is calculated according to the fuzzy rules.
Where, c j i and b j i are the center and width of the membership function respectively; k is the input parameter; n is the number of fuzzy subsets. The third layer is the fuzzy rule layer. The fuzzy operator is used as the multiplication operator to perform fuzzy calculations for each membership degree.
The fourth layer is used to normalize the applicability of each rule. The calculation expression is: is the fuzzy set of the fuzzy system; Assume that after performing the above processing on all the samples, the number of samples with an observed value of u j is inferred to be q(u j ). Let Q is the sum of the number of sample points at each u j point.
Where p(u j ) is the frequency value at which the sample point falls at u j , which can be used as an estimate of the probability. Its transcendence probability is: Where, P(u ≥ u j ) is the probability of exceeding u j .

T-S Fuzzy Neural Network
The T-S fuzzy system can be defined by the "if-then" rule form. In the case of the rule R i , the fuzzy reasoning is as follows: Where A j i is the fuzzy set of the fuzzy system; p j i is the fuzzy system parameter; yi is the output obtained according to the fuzzy rules, the input part (the if part) is fuzzy, and the output part (the then part) is determined. The fuzzy inference output is a linear combination of the inputs. T-S fuzzy system is a kind of fuzzy system with strong adaptive ability. The model can not only update automatically but also continuously modify the membership function of the fuzzy subset. It consists of an antecedent network and an after ware network.

Antecedent network
The first layer is the input layer. Assume that the input quantity x = [x 1 , x 2 , ⋯ x k ] T , and the number of nodes in this layer is k. The second layer is the fuzzification layer. The membership degree of each input variable x j is calculated according to the fuzzy rules.

 
Where, c j i and b j i are the center and width of the membership function respectively; k is the input parameter; n is the number of fuzzy subsets. The third layer is the fuzzy rule layer. The fuzzy operator is used as the multiplication operator to perform fuzzy calculations for each membership degree.
The fourth layer is used to normalize the applicability of each rule. The calculation expression is: is the fuzzy system parameter; y i is the output obtained according to the fuzzy rules, the input part (the if part) is fuzzy, and the output part (the then part) is determined. The fuzzy inference output is a linear combination of the inputs.
T-S fuzzy system is a kind of fuzzy system with strong adaptive ability. The model can not only update automatically but also continuously modify the membership function of the fuzzy subset. It consists of an antecedent network and an after ware network.

Antecedent network
The first layer is the input layer. Assume that the input quantity x = [x 1 , x 2 , ... x k ] T , and the number of nodes in this layer is k.
The second layer is the fuzzification layer. The membership degree of each input variable x j is calculated according to the fuzzy rules. Where, C i is the normal diffusion information sum of the observation sample x i . The membership function of the corresponding fuzzy subset is: Assume that after performing the above processing on all the samples, the number of samples with an observed valu u j is inferred to be q(u j ). Let Q is the sum of the number of sample points at each u j point.
Where p(u j ) is the frequency value at which the sample point falls at u j , which can be used as an estimate of the probab Its transcendence probability is: Where, P(u ≥ u j ) is the probability of exceeding u j .

T-S Fuzzy Neural Network
The T-S fuzzy system can be defined by the "if-then" rule form. In the case of the rule R i , the fuzzy reasoning is as follo Where A j i is the fuzzy set of the fuzzy system; p j i is the fuzzy system parameter; yi is the output obtained according to fuzzy rules, the input part (the if part) is fuzzy, and the output part (the then part) is determined. The fuzzy inference ou is a linear combination of the inputs. T-S fuzzy system is a kind of fuzzy system with strong adaptive ability. The model can not only update automatically also continuously modify the membership function of the fuzzy subset. It consists of an antecedent network and an a ware network.

Antecedent network
The first layer is the input layer. Assume that the input quantity x = [x 1 , x 2 , ⋯ x k ] T , and the number of nodes in this l is k. The second layer is the fuzzification layer. The membership degree of each input variable x j is calculated according to fuzzy rules.

 
Where, c j i and b j i are the center and width of the membership function respectively; k is the input parameter; n is the num of fuzzy subsets. The third layer is the fuzzy rule layer. The fuzzy operator is used as the multiplication operator to perform fuzzy calc tions for each membership degree.
The fourth layer is used to normalize the applicability of each rule. The calculation expression is: Where, C i is the normal diffusion information sum of the observation sample x i . The membership function of the corresponding fuzzy subset is: Assume that after performing the above processing on all the samples, the number of samples with an observed value of u j is inferred to be q(u j ). Let Q is the sum of the number of sample points at each u j point.
Where p(u j ) is the frequency value at which the sample point falls at u j , which can be used as an estimate of the probability. Its transcendence probability is: Where, P(u ≥ u j ) is the probability of exceeding u j .

T-S Fuzzy Neural Network
The T-S fuzzy system can be defined by the "if-then" rule form. In the case of the rule R i , the fuzzy reasoning is as follows: Where A j i is the fuzzy set of the fuzzy system; p j i is the fuzzy system parameter; yi is the output obtained according to the fuzzy rules, the input part (the if part) is fuzzy, and the output part (the then part) is determined. The fuzzy inference output is a linear combination of the inputs. T-S fuzzy system is a kind of fuzzy system with strong adaptive ability. The model can not only update automatically but also continuously modify the membership function of the fuzzy subset. It consists of an antecedent network and an after ware network.

Antecedent network
The first layer is the input layer. Assume that the input quantity x = [x 1 , x 2 , ⋯ x k ] T , and the number of nodes in this layer is k. The second layer is the fuzzification layer. The membership degree of each input variable x j is calculated according to the fuzzy rules.

 
Where, c j i and b j i are the center and width of the membership function respectively; k is the input parameter; n is the number of fuzzy subsets. The third layer is the fuzzy rule layer. The fuzzy operator is used as the multiplication operator to perform fuzzy calculations for each membership degree.
The fourth layer is used to normalize the applicability of each rule. The calculation expression is: are the center and width of the membership function respectively; k is the input parameter; n is the number of fuzzy subsets.
The third layer is the fuzzy rule layer. The fuzzy operator is used as the multiplication operator to perform fuzzy calculations for each membership degree.
Where, h is the information diffusion coefficient, which is determined according to the maximum value b, the minimum value a, and the number of samples n in the sample set. ℎ = Assume that after performing the above processing on all the samples, the number of samples with an observed value of u j is inferred to be q(u j ). Let Q is the sum of the number of sample points at each u j point.
Where p(u j ) is the frequency value at which the sample point falls at uj, which can be used as an estimate of the probability. Its transcendence probability is: Where, P(u ≥ u j ) is the probability of exceeding u j .

T-S Fuzzy Neural Network
The T-S fuzzy system can be defined by the "if-then" rule form. In the case of the rule R i , the fuzzy reasoning is as follows: Where A j i is the fuzzy set of the fuzzy system; p j i is the fuzzy system parameter; yi is the output obtained according to the fuzzy rules, the input part (the if part) is fuzzy, and the output part (the then part) is determined. The fuzzy inference output is a linear combination of the inputs. T-S fuzzy system is a kind of fuzzy system with strong adaptive ability. The model can not only update automatically but also continuously modify the membership function of the fuzzy subset. It consists of an antecedent network and an after ware network.

Antecedent network
The first layer is the input layer. Assume that the input quantity x = [x 1 , x 2 , ⋯ x k ] T , and the number of nodes in this layer is k. The second layer is the fuzzification layer. The membership degree of each input variable x j is calculated according to the fuzzy rules.

 
Where, c j i and b j i are the center and width of the membership function respectively; k is the input parameter; n is the number of fuzzy subsets. The third layer is the fuzzy rule layer. The fuzzy operator is used as the multiplication operator to perform fuzzy calculations for each membership degree.
The fourth layer is used to normalize the applicability of each rule. The calculation expression is: The fourth layer is used to normalize the applicability of each rule. The calculation expression is:

After ware network
The first layer is the input layer, which is used to The role of the second layer is to calculate each The third layer calculates the output value of t calculations.

Algorithm learning
The model's error analysis, coefficient correction Error calculation: Where y d is the expected output of the network; output and the actual output.
Coefficient correction: and ω i is the membership product of input param Parameter correction: Where, c j i and b j i are the centre and width of the Empirical formula for the number of nodes in th = √ + + Where, M is the number of nodes in the hidden l in the output layer

Survey of Research Area
Shimen Reservoir is located at 122°45′00″ east lo project with flood control, irrigation, and urban

After ware network
The first layer is the input layer, which is used to provide the constant term of the fuzzy rule follower.
The role of the second layer is to calculate each postscript of the rule.

After ware network
The first layer is the input layer, which is used to provi The role of the second layer is to calculate each postscr The third layer calculates the output value of the fuz calculations.

Algorithm learning
The model's error analysis, coefficient correction, and p Error calculation: Where y d is the expected output of the network; y c is t output and the actual output.
Coefficient correction: Where, p j i is the neural network coefficient, α is the lea and ω i is the membership product of input parameters.
Parameter correction: Where, c j i and b j i are the centre and width of the membe Empirical formula for the number of nodes in the hidde = √ + + Where, M is the number of nodes in the hidden layer; I in the output layer The third layer calculates the output value of the fuzzy model based on the fuzzy calculation results to achieve clear calculations.

After ware network
The first layer is the input layer, which is used to provide t The role of the second layer is to calculate each postscript The third layer calculates the output value of the fuzzy calculations.

Algorithm learning
The model's error analysis, coefficient correction, and para Error calculation: Where y d is the expected output of the network; y c is the a output and the actual output.
Coefficient correction: Where, p j i is the neural network coefficient, α is the learni and ω i is the membership product of input parameters.
Parameter correction: Where, c j i and b j i are the centre and width of the membersh Empirical formula for the number of nodes in the hidden l = √ + + Where, M is the number of nodes in the hidden layer; I is t in the output layer

Algorithm learning
The model's error analysis, coefficient correction, and parameter correction methods are as follows: Error calculation:

After ware network
The first layer is the input layer, which is used to p The role of the second layer is to calculate each po The third layer calculates the output value of th calculations.

Algorithm learning
The model's error analysis, coefficient correction, Error calculation: Where y d is the expected output of the network; y output and the actual output.
Coefficient correction: and ω i is the membership product of input parame Parameter correction: Where, c j i and b j i are the centre and width of the m Empirical formula for the number of nodes in the = √ + + Where, M is the number of nodes in the hidden lay in the output layer Table 2: T-S fuzzy neural network construction pa Where y d is the expected output of the network; y c is the actual output of the network; e is the error between the expected output and the actual output.

After ware network
The first layer is the input layer, which is used to provid The role of the second layer is to calculate each postscr The third layer calculates the output value of the fuzz calculations.

Algorithm learning
The model's error analysis, coefficient correction, and p Error calculation: Where y d is the expected output of the network; y c is th output and the actual output.
Coefficient correction: Where, p j i is the neural network coefficient, α is the lea and ω i is the membership product of input parameters.
Parameter correction: Where, c j i and b j i are the centre and width of the membe Empirical formula for the number of nodes in the hidde = √ + + Where, M is the number of nodes in the hidden layer; I i in the output layer Table 2: T-S fuzzy neural network construction paramet

After ware network
The first layer is the input layer, which is used to provide The role of the second layer is to calculate each postscrip The third layer calculates the output value of the fuzzy calculations.

Algorithm learning
The model's error analysis, coefficient correction, and pa Error calculation: Where y d is the expected output of the network; y c is the output and the actual output.
Coefficient correction: Where, p j i is the neural network coefficient, α is the learn and ω i is the membership product of input parameters.
Parameter correction: Where, c j i and b j i are the centre and width of the members Empirical formula for the number of nodes in the hidden = √ + + Where, M is the number of nodes in the hidden layer; I is in the output layer

After ware network
The first layer is the input layer, which is used to provide the con The role of the second layer is to calculate each postscript of the The third layer calculates the output value of the fuzzy model calculations.

Algorithm learning
The model's error analysis, coefficient correction, and parameter Error calculation: Where y d is the expected output of the network; y c is the actual o output and the actual output.
Coefficient correction: Where, p j i is the neural network coefficient, α is the learning effi and ω i is the membership product of input parameters.
Parameter correction: Where, c j i and b j i are the centre and width of the membership fun Empirical formula for the number of nodes in the hidden layer: = √ + + Where, M is the number of nodes in the hidden layer; I is the num in the output layer is the neural network coefficient, x j is the learning efficiency of the network, x j is the network input parameter, and w i is the membership product of input parameters.

After ware network
The first layer is the input layer, which is used to prov The role of the second layer is to calculate each postsc The third layer calculates the output value of the fuz calculations.

Algorithm learning
The model's error analysis, coefficient correction, and Error calculation: Where y d is the expected output of the network; y c is output and the actual output.
Coefficient correction: Where, p j i is the neural network coefficient, α is the le and ω i is the membership product of input parameters Parameter correction: Where, c j i and b j i are the centre and width of the memb Empirical formula for the number of nodes in the hidd = √ + + Where, M is the number of nodes in the hidden layer; I in the output layer Table 2: T-S fuzzy neural network construction param

After ware network
The first layer is the input layer, which is used to provid The role of the second layer is to calculate each postscr The third layer calculates the output value of the fuzz calculations.

Algorithm learning
The model's error analysis, coefficient correction, and p Error calculation: Where y d is the expected output of the network; y c is th output and the actual output.
Coefficient correction: Where, p j i is the neural network coefficient, α is the lea and ω i is the membership product of input parameters.
Parameter correction: Where, c j i and b j i are the centre and width of the membe Empirical formula for the number of nodes in the hidde = √ + + Where, M is the number of nodes in the hidden layer; I i in the output layer

After ware network
The first layer is the input layer, which is used to provide the con The role of the second layer is to calculate each postscript of the The third layer calculates the output value of the fuzzy model calculations.

Algorithm learning
The model's error analysis, coefficient correction, and parameter Error calculation: Where y d is the expected output of the network; y c is the actual output and the actual output.
Coefficient correction: Where, p j i is the neural network coefficient, α is the learning effi and ω i is the membership product of input parameters.
Parameter correction: Where, c j i and b j i are the centre and width of the membership fun Empirical formula for the number of nodes in the hidden layer: = √ + + Where, M is the number of nodes in the hidden layer; I is the num in the output layer Empirical formula for the number of nodes in the hidden layer:

After ware network
The first layer is the input layer, which is used to pro The role of the second layer is to calculate each posts The third layer calculates the output value of the fu calculations.

Algorithm learning
The model's error analysis, coefficient correction, and Error calculation: Where y d is the expected output of the network; y c is output and the actual output.
Coefficient correction: Where, p j i is the neural network coefficient, α is the l and ω i is the membership product of input parameter Parameter correction: Where, c j i and b j i are the centre and width of the mem Empirical formula for the number of nodes in the hid = √ + + Where, M is the number of nodes in the hidden layer; in the output layer Where, M is the number of nodes in the hidden layer; I is the number of nodes in the input layer; O is the number of nodes in the output layer

Survey of Research Area
Shimen Reservoir is located at 122°45'00"″ east longitude and 40°22'00"″north latitude. It is a large-scale water conservancy project with flood control, irrigation, and urban water supply as its main function while taking into account power generation and fish farming functions. The maximum dam height is 47 meters, the dam is 350 meters long, the annual runoff is 117 million cubic meters, and the total storage capacity is 102.2 million cubic meters. The reservoir provides an average of 25 million cubic meters of industrial and urban domestic water and 36 million cubic meters of agricultural water to Yingkou and Gaizhou. Reservoir operation has played a huge benefit and harm removal benefits

RESULTS AND ANALYSIS
In this paper, the monthly pH, nitrate, ammonia nitrogen, and total nitrogen contents of Shimen Reservoir in 2013 and 2014 were selected to ensure the stability of the water environment of the reservoir. The feasibility of the information diffusion theory and T-S neural network in the analysis of reservoir water quality was verified. The main research contents are as follows: (1) After determining the single factor index of each element, use information diffusion technology to process each single factor index to analyse the occurrence probability of pH, nitrate, ammonia nitrogen and total nitrogen in Shimen Reservoir.
(2) The T-S fuzzy neural network was used to evaluate the water quality of the Shimen Reservoir in 2013 and 2014.

Water Quality Analysis of Shimen Reservoir Based on Information Diffusion Theory and Single Factor Index
The single factor index corresponding to the four factors was calculated from Equations 1-3. According to the information diffusion theory, each single factor index is used as the information diffusion sample for diffusion. X 1 , X 2 , X 3 and X 4 are information diffusion samples of pH, nitrate, ammonia nitrogen, and total nitrogen, respectively. It can be seen from the Figs. 1-4 that the probability of the corresponding single factor index of four elements in Shimen Reservoir appears. The maximum probability of pH is SPH j = 0.1, which is 32.93%. Nitrate has the highest probability at 46.13%; NH 3 -N has the highest probability at 18.99% and TN has the highest probability at 23.78%.
It can be seen from Fig. 1 and Table 3 that the transcendental probability curve of pH is on the side less than 1. Therefore, the pH value of the Shimen Reservoir is relatively stable.
It can be known from Table 5 and Fig. 2 that the single factor exponential distribution of nitrate is distributed on the side of Pi≤1.  That is, the nitrate content of the water body meets the standard.
From Table 6 and Fig. 3, when P i ″ ≤ 0.075, the content of NH 3 -N conforms to type I water, and the probability of occurrence is P ≥ 76.17%; when 0.075 < P i ≤ 0.25, the NH 3 -N content accords with the water of class II, and the probability of occurrence at this time is 76.17%″ ≤ P″ ≤ 0.09%. That is, most of the ammonia nitrogen content in the water body meets the standards of Class I and II water.
It is clear from Table 7 and Fig. 3 that the single factor index of total nitrogen in the water body is greater than 1. Therefore, the biggest cause of water pollution may be excessive nitrogen content.

Water Quality Evaluation of Shimen Reservoir Based on T-S Fuzzy Neural Network
The basic steps and evaluation results of Shimen reservoir water quality evaluation based on T-S fuzzy neural network are as follows: (1) Selected network structure: The construction of the fuzzy neural network determines the number of input and output points of the fuzzy neural network according to the dimensions of the training samples. Selected four indicators of pH, nitrate, NH 3 -N and TN, so the number of input nodes is 4. Water quality levels I-V are represented by numbers 1-5, and the number of output nodes is 1. It can be known from Table 2 that the number of nodes in the hidden layer is 10, so a 4-10-1 network structure is formed.
(2) Generate training samples: In this paper, 400 sets of training samples are generated by interpolating water quality index standard data "Environmental Quality Standard for Surface Water" (GB3838-2002) with the evenly spaced distribution. Obtain water quality level indicators based on network predictions. When the predicted value is less than 1.5, the water quality level is Class I; when the predicted value is 1.5 to 2.5, the water quality level is Class II; when the predicted value is 2.5 to 3.5, the water quality level is Class III; when the predicted value is 3.5 to 4.5 When the predicted value is greater than 4.5, the water quality level is Category V.
(3) Network training and testing: The training sample is used to train the fuzzy neural network 100 times. The training result is shown in Fig. 5. Fifty sets of data were drawn from a random sample to verify the accuracy of the model. The verification results are shown in Fig. 6 It can be seen from Fig. 6 that the error between the actual water quality level and the model output water quality level is small, and it can play an accurate prediction role.
(4) Water quality evaluation: Select the content of pH, nitrate, NH 3 -N and TN from month to month from 2013 to 2014. The trained T-S fuzzy neural network was used to evaluate the Shimen reservoir. The evaluation results are shown in Fig. 7. It can be seen from the figure that the water quality level of Shimen Reservoir is basically maintained at Class III and IV, and the water environment is relatively stable.

CONCLUSIONS
(1) 24 sample data are selected in this paper. The nitrate, NH 3 -N, TN and pH in Shimen Reservoir were evaluated based on single factor index and information diffusion theory. The results show that the pH, nitrate and ammonia nitrogen in Shimen Reservoir are stable and meet       It can be seen from the Figs. 1-4 that the probability of the corresponding single factor in Reservoir appears. The maximum probability of pH is SPH j = 0.1, which is 32.93%; nitra 46.13% when P i = 0.3; NH 3 -N has the highest probability at 18.99% when P i = 0.075; T P i = 1.9, 23.78%. It can be seen from Fig. 1 and Table 3 that the transcendental probability curve of pH is on  It can be seen from the Figs. 1-4 that the probability of the corresponding single factor in Reservoir appears. The maximum probability of pH is SPH j = 0.1, which is 32.93%; nitr 46.13% when P i = 0.3; NH 3 -N has the highest probability at 18.99% when P i = 0.075; T P = 1.9, 23.78%.   (4) Water quality evaluation: Select the content of pH, nitrate, NH 3 -N and TN from mo 2014. The trained T-S fuzzy neural network was used to evaluate the Shimen reserv are shown in Fig. 7. It can be seen from the figure that the water quality level of Shi maintained at Class III and IV, and the water environment is relatively stable. Fig. 7: Evaluation of water quality in Shimen Reservoir by the fuzzy neural network.

CONCLUSIONS
(1) 24 sample data are selected in this paper. The nitrate, NH 3 -N, TN and pH in Shimen Res on single factor index and information diffusion theory. The results show that the pH, nitra Shimen Reservoir are stable and meet national standards, but the total nitrogen contents are standard.
(2) Analysis of water quality of Shimen Reservoir from 2013 to 2014 based on T-S fuzzy neur   (4) Water quality evaluation: Select the content of pH, nitrate, NH 3 -N and TN from mont 2014. The trained T-S fuzzy neural network was used to evaluate the Shimen reservoi are shown in Fig. 7. It can be seen from the figure that the water quality level of Shim maintained at Class III and IV, and the water environment is relatively stable. Fig. 7: Evaluation of water quality in Shimen Reservoir by the fuzzy neural network.

CONCLUSIONS
(1) 24 sample data are selected in this paper. The nitrate, NH 3 -N, TN and pH in Shimen Rese on single factor index and information diffusion theory. The results show that the pH, nitrate Shimen Reservoir are stable and meet national standards, but the total nitrogen contents are s standard.  (4) Water quality evaluation: Select the content of pH, nitrate, NH 3 -N and TN from month 2014. The trained T-S fuzzy neural network was used to evaluate the Shimen reservoir. are shown in Fig. 7. It can be seen from the figure that the water quality level of Shime maintained at Class III and IV, and the water environment is relatively stable.

CONCLUSIONS
(1) 24 sample data are selected in this paper. The nitrate, NH 3 -N, TN and pH in Shimen Reserv on single factor index and information diffusion theory. The results show that the pH, nitrate Shimen Reservoir are stable and meet national standards, but the total nitrogen contents are sig standard.
(2) Analysis of water quality of Shimen Reservoir from 2013 to 2014 based on T-S fuzzy neural that the comprehensive water quality of Shimen Reservoir is good and stable.
(3) It is feasible to use information diffusion technology combined with single factor index to e analysis results are clear and have guiding significance for the water quality control of Shimen R national standards, but the total nitrogen contents are significantly higher than the standard.
(2) Analysis of water quality of Shimen Reservoir from 2013 to 2014 based on T-S fuzzy neural network. The results show that the comprehensive water quality of Shimen Reservoir is good and stable.
(3) It is feasible to use information diffusion technology combined with single factor index to evaluate water quality. The analysis results are clear and have guiding significance for the water quality control of Shimen Reservoir. T-S fuzzy neural network fuzzy takes single factor prediction as input and comprehensive evaluation