Biochemical oxygen demand online prediction method for sewage treatment of thermal power plant

Abstract

The present application relates to a kind of biochemical oxygen demand online prediction method of thermal power plant sewage treatment, comprising: determining the main parameter of influencing biochemical oxygen demand as auxiliary variable;Dimension reduction processing is carried out;Using the auxiliary variable after dimension reduction as input variable, using improved case reasoning method to carry out the case matching of biochemical oxygen demand soft measurement model, using nearest neighbor search strategy to carry out the case adjustment of biochemical oxygen demand soft measurement model, establish biochemical oxygen demand soft measurement model;Collecting thermal power plant sewage treatment data input biochemical oxygen demand soft measurement model and training learning;The data to be predicted of thermal power plant sewage treatment is input into trained biochemical oxygen demand soft measurement model, and the prediction value of biochemical oxygen demand is output.The present application uses RBF neural network to carry out dimension reduction processing to auxiliary variable, reduce the time and space complexity of system calculation;Effectively improve the prediction accuracy of biochemical oxygen demand, with good real-time online prediction effect.

Description

TECHNICAL FIELD

[0001] The present application relates to the field of power plant wastewater treatment technology, and particularly relates to a biochemical oxygen demand online prediction method for power plant wastewater treatment. BACKGROUND

[0002] Power plants are one of the main sources of industrial wastewater, so strengthening power plant wastewater treatment and recycling is of great significance for protecting the water environment and realizing sustainable development.

[0003] The activated sludge method is widely used in the chemical treatment process of power plant wastewater, and its essence is to remove organic matter in wastewater through a series of operations such as oxidation and decomposition. The current quality indicators of power plant wastewater treatment effluent include BOD (biochemical oxygen demand), T-N (total nitrogen content), COD (chemical oxygen demand), and T-P (total phosphorus content) and other parameters. These parameters are mainly determined by chemical analysis sensors or manual sampling detection, but such chemical analysis sensors have defects such as high cost, short service life, small range, and poor stability, and manual sampling detection is still the main method in the field. In addition, since the BOD test result cycle is relatively long, it greatly lags behind the wastewater treatment process, and when it is found that the effluent quality indicator BOD is unqualified, a large amount of unqualified wastewater has been discharged into the environment, so timely online detection of the BOD parameter is the key to improving the quality of the wastewater treatment effluent.

[0004] At present, neural networks and case-based reasoning methods are usually used for online detection of the BOD parameter of wastewater treatment. For single neural network prediction of BOD, a larger network structure is required, which leads to a decrease in network learning rate and causes a delay in the output of the BOD online detection result. The conventional case-based reasoning method based on the nearest neighbor strategy is only suitable for solving simple problems where the weights of attributes are easy to determine, but for complex problems such as wastewater treatment quality indicator soft measurement, it is difficult to reasonably determine the weights of the attributes due to the strong coupling and nonlinear relationship between the attributes, which leads to a decrease in the prediction accuracy of the conventional case-based reasoning method for BOD. SUMMARY

[0005] To solve the technical problems of high cost, low precision, and poor real-time performance of the existing biochemical oxygen demand detection of power plant wastewater treatment, the present application aims to provide a biochemical oxygen demand online prediction method for power plant wastewater treatment with low detection cost, high precision, and real-time detection.

[0006] To achieve the above-mentioned purpose, the present application adopts the following technical solution: a biochemical oxygen demand online prediction method for power plant wastewater treatment, which comprises the following sequential steps:

[0007] (1) Determine the main parameters affecting biochemical oxygen demand during the biochemical treatment of wastewater in thermal power plants, and select these main parameters as auxiliary variables related to biochemical oxygen demand;

[0008] (2) Use RBF neural network to reduce the dimensionality of the n inlet pump motor currents in the auxiliary variables, thereby reducing the number of auxiliary variables, while keeping the other auxiliary variables unchanged;

[0009] (3) The auxiliary variables after dimensionality reduction are used as input variables of the biochemical oxygen demand soft measurement model. An improved case reasoning method is used, namely, BP neural network is used for case matching of the biochemical oxygen demand soft measurement model. The nearest neighbor retrieval strategy is used to adjust the cases of the biochemical oxygen demand soft measurement model and establish the biochemical oxygen demand soft measurement model.

[0010] (4) Initialize the parameters of the biochemical oxygen demand soft measurement model, and collect wastewater treatment data from thermal power plants and input them into the biochemical oxygen demand soft measurement model for training and learning;

[0011] (5) Input the wastewater treatment data of the thermal power plant to be predicted into the trained biochemical oxygen demand soft measurement model and output the predicted value of biochemical oxygen demand.

[0012] In step (1), the main parameters include: MLSS activated sludge concentration, temperature, pH value, influent COD chemical oxygen demand, DO dissolved oxygen, current of n influent pump motors, NH4+-N nitrogen source, and HRT hydraulic retention time.

[0013] Step (2) specifically includes the following steps:

[0014] (2a) The input variables of the RBF neural network are selected as the currents of n water pump motors, and the output variable is the total water intake. ;

[0015] (2b) Taking the Gaussian function as the hidden layer node function of the RBF neural network, we obtain the RBF neural network's... Output of each hidden layer node:

[0016] (1)

[0017] in: For the first The output of each hidden layer node For the RBF neural network The center vector of each node For the input vector of the RBF neural network, The first of the basis width vectors in the RBF neural network One value, , This represents the number of hidden layer nodes;

[0018] Therefore, the total water inflow prediction model obtained by the RBF neural network is as follows:

[0019] (2)

[0020] in: These are the connection weights from the hidden layer to the output layer of the RBF neural network. This represents the number of nodes in the output layer.

[0021] (2c) Define the objective function:

[0022] (3)

[0023] Using the actual total water intake volume of the inlet pump Total water inflow predicted by RBF neural network Difference between By continuously adjusting the connection weights of the RBF neural network, the output of the RBF neural network gradually approximates the actual total water intake of the object. ;

[0024] After dimensionality reduction, the auxiliary variables obtained are MLSS activated sludge concentration, temperature, pH value, influent COD, DO, and total influent volume. NH4+-N nitrogen source and HRT (hydraulic retention time).

[0025] Step (3) specifically includes the following steps:

[0026] (3a) The case of the soft measurement model for biochemical oxygen demand (BOD) is represented by three parts: operating condition variables, solution time, and the solution to the problem. The operating condition variables are auxiliary variables related to the effluent BOD, represented as X = x1, x2, ... x8. The solution to the problem is the predicted value of BOD, represented as... ;

[0027] (3b) Case matching: A BP neural network is used to train the input sample (i.e., the working condition variable sample) and the output sample (i.e., the solution to the problem). After convergence, the connection weight data is saved for use in the case matching process. The similarity is compared with the samples in the case library to improve the efficiency and accuracy of case matching.

[0028] (3c) Adjust the case study:

[0029] Suppose adjustments are needed The cases are respectively , … Then the estimated value of the current operating condition variable information for:

[0030] (4)

[0031] wherein: is the nearest neighbor search strategy based on Euclidean distance, representing the similarity between the current working condition variable and the case in the case base, and respectively represent the working condition variables of the current working condition variable and the th case in the case base, represent the weighting coefficients of the working condition description characteristics, the adjusted case outputs the result, and saves the solution, working condition variable and solving time of the problem of the case this time;

[0032] (3d) The adjusted case is stored in the case base in time. If the case base has the same case as the target case, the target case will be discarded directly, otherwise, the adjusted case is stored in the case base.

[0033] The step (3b) specifically comprises the following steps:

[0034] (3b1) initialize the BP neural network parameters;

[0035] (3b2) provide the first group of input and output (X k , Y k ) to the BP neural network;

[0036] (3b3) determine the input and output of the hidden layer and the output layer respectively, as shown in formula (5) to formula (8):

[0037] (5)

[0038] (6)

[0039] (7)

[0040] (8)

[0041] wherein: , is the connection weight from the input layer to the hidden layer, is the connection weight from the hidden layer to the output layer, is the input of the neural network, , are respectively the threshold values of the hidden layer and the output layer, i=1,2,...,p1, j=1,2,...,p2; is the input of the hidden layer, is the output of the hidden layer, is the input of the output layer, For the output layer output, p1 is the input layer node number, and p2 is the hidden layer node number;

[0042] (3b4) Construct a random motion mechanism for the connection weight parameter value, formula (9) is the connection weight correction error of the input layer, formula (10) is the connection weight correction error of the output layer, adjust the connection weight between the hidden layer and the output layer, first generate a random number between 0 and 1 , For the nonlinear self-feedback introduces probability, when , adjust according to formula (12), when , adjust according to formula (11); and adjust the output layer threshold value according to formula (13), similarly, the connection weight between the input layer and the hidden layer and the threshold value of the hidden layer are adjusted according to the above method, as shown in formula (14) to formula (16);

[0043] d k =(Y k -c)c(1-c)(9)

[0044] e j =d k v j b j (1-b j ) (10)

[0045] (11)

[0046] (12)

[0047] (13)

[0048] (14)

[0049] (15)

[0050] (16)

[0051] In the formula, and are learning rates, is the number of learning times; d k is the connection weight correction error of the input layer, and e j is the connection weight correction error of the output layer.

[0052] As can be seen from the above technical solution, the beneficial effects of the present invention are as follows: First, since there are many auxiliary variables affecting the accuracy of the biochemical oxygen demand (BOD) soft measurement model, the present invention uses an RBF neural network to reduce the dimensionality of the auxiliary variables, thereby reducing the time and space complexity of the system calculation; Second, since there are strong couplings and complex nonlinear relationships among the input variables, the BP neural network is combined with the nearest neighbor retrieval strategy for case matching and case adjustment, improving the efficiency and accuracy of case matching, and using this improved case reasoning method to establish a BOD soft measurement model for BOD prediction; Third, the present invention effectively improves the prediction accuracy of BOD and has a good real-time online prediction effect. Attached Figure Description

[0053] Figure 1 This is a flowchart of the method of the present invention;

[0054] Figure 2 Example 1 shows the biochemical oxygen demand (BOD) curve of wastewater treatment in a thermal power plant predicted using the RBF neural network soft measurement model.

[0055] Figure 3 Example 1 shows the biochemical oxygen demand (BOD) curve of wastewater treatment in a thermal power plant, predicted using a conventional case-based soft measurement model.

[0056] Figure 4 This is a graph showing the biochemical oxygen demand (BOD) curve of wastewater treatment from a thermal power plant, predicted using a soft measurement model for BOD in Example 1. Detailed Implementation

[0057] like Figure 1 As shown, an online prediction method for biochemical oxygen demand (BOD) in wastewater treatment from thermal power plants is provided. This method includes the following sequential steps:

[0058] (1) Determine the main parameters affecting biochemical oxygen demand during the biochemical treatment of wastewater in thermal power plants, and select these main parameters as auxiliary variables related to biochemical oxygen demand;

[0059] (2) Use RBF neural network to reduce the dimensionality of the n inlet pump motor currents in the auxiliary variables, thereby reducing the number of auxiliary variables, while keeping the other auxiliary variables unchanged;

[0060] (3) The auxiliary variables after dimensionality reduction are used as input variables of the biochemical oxygen demand soft measurement model. An improved case reasoning method is used, namely, BP neural network is used for case matching of the biochemical oxygen demand soft measurement model. The nearest neighbor retrieval strategy is used to adjust the cases of the biochemical oxygen demand soft measurement model and establish the biochemical oxygen demand soft measurement model.

[0061] (4) Initialize the parameters of the biochemical oxygen demand soft measurement model, and collect wastewater treatment data from thermal power plants and input them into the biochemical oxygen demand soft measurement model for training and learning;

[0062] (5) Input the wastewater treatment data of the thermal power plant to be predicted into the trained biochemical oxygen demand soft measurement model and output the predicted value of biochemical oxygen demand.

[0063] In step (1), the main parameters include: MLSS activated sludge concentration, temperature, pH value, influent COD chemical oxygen demand, DO dissolved oxygen, current of n influent pump motors, NH4+-N nitrogen source, and HRT hydraulic retention time.

[0064] Step (2) specifically includes the following steps:

[0065] (2a) The input variables of the RBF neural network are selected as the currents of n water pump motors, and the output variable is the total water intake. ;

[0066] (2b) Taking the Gaussian function as the hidden layer node function of the RBF neural network, we obtain the RBF neural network's... Output of each hidden layer node:

[0067] (1)

[0068] in: For the first The output of each hidden layer node For the RBF neural network The center vector of each node For the input vector of the RBF neural network, The first of the basis width vectors in the RBF neural network One value, , This represents the number of hidden layer nodes;

[0069] Therefore, the total water inflow prediction model obtained by the RBF neural network is as follows:

[0070] (2)

[0071] in: These are the connection weights from the hidden layer to the output layer of the RBF neural network. This represents the number of nodes in the output layer.

[0072] (2c) Define the objective function:

[0073] (3)

[0074] Using the actual total water intake volume of the inlet pump The difference between the total influent water quantity and the total influent water quantity predicted by the RBF neural network The difference between the total influent water quantity and the total influent water quantity predicted by the RBF neural network is continuously corrected The connection weight value of the RBF neural network is continuously corrected, so that the output of the RBF neural network gradually approaches the actual total influent water quantity of the object ;

[0075] After dimension reduction, the obtained auxiliary variables are MLSS active sludge concentration, temperature, pH value, influent COD chemical oxygen demand, DO dissolved oxygen, total influent water quantity, NH4+-N nitrogen source substance and HRT hydraulic retention time .

[0076] The step (3) specifically comprises the following steps:

[0077] (3a) The case of the biochemical oxygen demand soft measurement model adopts three parts of working condition variables, solution time and problem solution, wherein the working condition variables are auxiliary variables related to the effluent biochemical oxygen demand, and are represented as X=x1,x2,...x8, and the problem solution is the predicted value of the biochemical oxygen demand, and is represented as .

[0078] (3b) Case matching is performed: a BP neural network is adopted, and through training of input samples, i.e., working condition variable samples, and output samples, i.e., problem solutions, connection weight data is saved after convergence for calling in the case matching link, and similarity comparison is performed with samples in the case library, so as to achieve the purpose of improving case matching efficiency and accuracy;

[0079] (3c) Case adjustment is performed:

[0080] The case retrieval process is to find the most similar case from the case library, and since it is difficult to find a past case that completely matches the current problem, the past case needs to be properly adjusted. The adjustment process of the case mainly involves content selection, solution conversion and solution evaluation. The adjusted case can be put into practice if it passes the quality judgment of the case evaluation, and can also be stored in the case library after examination by the case learning technology. In order to eliminate the inconsistency of the case time and prevent the time invalidation of the case, the old case needs to be deleted.

[0081] Suppose that the cases to be adjusted are , , , , ,

[0082] , , ,

[0083] , The nearest neighbor search strategy based on Euclidean distance represents the similarity between the current working condition variable and the case in the case base, and respectively represent the working condition variables of the current working condition variable and the first case in the case base, represent the weighting coefficients of the working condition description characteristics, the adjusted case outputs the result and saves the solution, working condition variable and solving time of the problem of the current case;

[0084] (3d) The adjusted case is stored in the case base in time. If the case base has the same case as the target case, the target case will be discarded directly, otherwise, the adjusted case is stored in the case base. The increase of the number of case bases can improve the accuracy of the system, but will slow down the operation of the system, so the case base should be maintained in time. Usually, the case maintenance adopts appropriate increasing and decreasing of the number of case samples, pruning of interference case samples, adjusting the structure and index device of the case base, etc. to ensure that the cases in the case base have high typicality, timeliness, consistency and non-redundancy.

[0085] The step (3b) specifically comprises the following steps:

[0086] (3b1) initializing the BP neural network parameters; the connection weights and thresholds between layers of the BP neural network are random values between [-1, +1], and are ensured to be uniformly distributed to avoid saturation of the BP neural network at large weights;

[0087] (3b2) providing the first group of input and output (X k , Y k ) to the BP neural network;

[0088] (3b3) determining the input and output of the hidden layer and the output layer respectively, as shown in formula (5) to formula (8):

[0089] (5)

[0090] (6)

[0091] (7)

[0092] (8)

[0093] Wherein: , is the connection weight from the input layer to the hidden layer, is the connection weight from the hidden layer to the output layer, is the input of the neural network, , are the threshold values of the hidden layer and the output layer, respectively, i = 1, 2,..., pi, j = 1, 2,..., p2; is the input of the hidden layer, is the output of the hidden layer, is the input of the output layer, is the output of the output layer, pi is the number of nodes of the input layer, and p2 is the number of nodes of the hidden layer;

[0094] (3b4) Construct a random motion mechanism for the connection weight parameter value, formula (9) is the correction error of the connection weight of the input layer, formula (10) is the correction error of the connection weight of the output layer, and the connection weight between the hidden layer and the output layer is adjusted. First, a random number between 0 and 1 is generated , is the nonlinear self-feedback introduction probability, when , it is adjusted according to formula (12), when , it is adjusted according to formula (11); and the threshold value of the output layer is adjusted according to formula (13). Similarly, the connection weights between the input layer and the hidden layer and the threshold value of the hidden layer are adjusted according to the above method, as shown in formula (14) to formula (16);

[0095] d k =(Y k -c)c(1-c)(9)

[0096] e j =d k v j b j (1-b j ) (10)

[0097] (11)

[0098] (12)

[0099] (13)

[0100] (14)

[0101] (15)

[0102] (16)

[0103] In the formula, and are learning rates, is the number of learning times; d k is the correction error of the connection weight of the input layer, and ej The connection weight correction error of the output layer.

[0104] Embodiment One

[0105] The biochemical oxygen demand parameter of the effluent in a sewage treatment site of a certain 660 MW thermal power plant was detected, and the changes of the water quality parameters at various positions in the sewage treatment process, such as the horizontal flow oil separation tank, the vertical flow sedimentation tank, the dosing tank, the filter tank, and the desalination tank, were recorded and analyzed. The auxiliary variables related to the biochemical oxygen demand of the effluent were preliminarily determined as: MLSS (activated sludge concentration), temperature, pH value, influent COD (chemical oxygen demand), DO (dissolved oxygen), 8 influent pump motor currents, NH4+-N (nitrogen source substance), and HRT (hydraulic retention time). 1000 groups of auxiliary variable data after dissimilation and normalization were collected as sample data.

[0106] The total influent volume was predicted by the RBF neural network through the 8 influent pump motor currents To reduce the number of auxiliary variables, the node numbers of the input layer, the hidden layer, and the output layer of the neural network were taken as 8, 4, and 1 respectively, the hidden layer node function was taken as the Gaussian basis function, the input variable was the 8 influent pump motor currents, and the output variable was the total influent volume The first 800 groups of sample data were used to train the total influent volume prediction model, and the last 200 groups of data were used to verify the total influent volume prediction model.

[0107] MLSS, temperature, pH value, influent COD, DO, Q, NH4+-N, and HRT were used as the auxiliary variables of the biochemical oxygen demand soft measurement model, the first 800 groups of sample data were used to train the biochemical oxygen demand soft measurement model, and the last 200 groups of data were used to verify the biochemical oxygen demand soft measurement model.

[0108] The case of the biochemical oxygen demand soft measurement model was represented by three parts: the working condition variable, the solving time, and the solution of the problem. The working condition variable was the 8 auxiliary variables related to the biochemical oxygen demand of the effluent, represented as The solution of the problem was the predicted value of the biochemical oxygen demand, represented as The storage structure of the case representation is shown in Table 1:

[0109] Table 1 Storage structure of case representation

[0110]

[0111] Figures 2 to 4 The curve graphs of the biochemical oxygen demand of the sewage treatment in the thermal power plant predicted by the RBF neural network soft measurement model, the conventional case-based reasoning soft measurement model, and the biochemical oxygen demand soft measurement model of this embodiment one, respectively. In order to evaluate the models, the following three performance index functions were used to evaluate the prediction results. Figure 3, 4 Case-based Reasoning (CBR) is case-based reasoning.

[0112] (1) Root mean-square error (MSE)

[0113] (17)

[0114] (2) mean-absolute relatively error (EAE)

[0115] (18)

[0116] (3) maximize-absolute relatively error (EME)

[0117] (19)

[0118] Wherein: is the number of test samples, is the actual value of data, is the model prediction value. The prediction performance of the three models is shown in Table 2:

[0119] Table 2 Performance evaluation index table

[0120]

[0121] As shown in Table 2, compared with the RBF neural network and the conventional case-based reasoning soft measurement prediction model, the MSE value of the biochemical oxygen demand soft measurement model is reduced by 0.083 and 0.017 respectively, the EAE value is reduced by 0.749% and 0.222% respectively, and the EME value is reduced by 0.033 and 0.018 respectively, and the prediction accuracy of the biochemical oxygen demand soft measurement model is significantly improved. Therefore, the biochemical oxygen demand soft measurement model can predict the biochemical oxygen demand of the effluent of the sewage treatment of the thermal power plant in real time and accurately.

[0122] In summary, the RBF neural network is adopted to reduce the dimension of auxiliary variables, so that the time and space complexity of system calculation is reduced; since there is strong coupling and complex nonlinear relationship between input variables, the BP neural network is combined with the nearest neighbor search strategy and applied to case matching and case adjustment, so that the case matching efficiency and accuracy are improved, and the improved case-based reasoning method is adopted to establish the biochemical oxygen demand soft measurement model to predict the biochemical oxygen demand; the prediction accuracy of the biochemical oxygen demand is effectively improved, and the real-time online prediction effect is good.

Claims

1. A method for on-line prediction of biochemical oxygen demand in sewage treatment of a thermal power plant, characterized in that: The method comprises the following steps in sequence: (1) determining main parameters affecting biochemical oxygen demand in the sewage biochemical treatment process of the thermal power plant, and selecting the main parameters as auxiliary variables related to the biochemical oxygen demand; (2) reducing the dimension of n water inlet pump motor currents in the auxiliary variables by using the RBF neural network, reducing the number of auxiliary variables, and keeping the rest of the auxiliary variables unchanged; (3) using the auxiliary variables after dimension reduction as input variables of the biochemical oxygen demand soft measurement model, using the improved case-based reasoning method, that is, using the BP neural network to perform case matching of the biochemical oxygen demand soft measurement model, using the nearest neighbor search strategy to perform case adjustment of the biochemical oxygen demand soft measurement model, and establishing the biochemical oxygen demand soft measurement model; (4) initializing parameters of the biochemical oxygen demand soft measurement model, collecting sewage treatment data of the thermal power plant, and inputting the sewage treatment data into the biochemical oxygen demand soft measurement model for training and learning; (5) inputting sewage treatment data of the thermal power plant to be predicted into the trained biochemical oxygen demand soft measurement model, and outputting a predicted value of the biochemical oxygen demand; The step (3) specifically comprises the following steps: (3a) The case of biochemical oxygen demand soft-sensing model is expressed by three parts of working condition variables, solving time and solution of the problem, wherein the working condition variables are auxiliary variables related to effluent biochemical oxygen demand, expressed as X=x1, x2,... x8, and the solution of the problem is the predicted value of biochemical oxygen demand, expressed as ; (3b) performing case matching: using the BP neural network, training the input sample, that is, the working condition variable sample, and the output sample, that is, the solution of the problem, until convergence is achieved, saving the connection weight data for calling in the case matching link, comparing with the samples in the case library, and achieving the purpose of improving the case matching efficiency and accuracy; (3c) Case adjustment: Let the cases to be adjusted be Case 1, Case 2, …, Case n, respectively. , , The estimated value of the current operating variable information is ​ （4） wherein: is a nearest neighbor search strategy based on Euclidean distance, representing the similarity between the current working condition variable and the case in the case base, and respectively represent the working condition variables of the current working condition variable and the th case in the case base, represents the weighting coefficient of the working condition description feature, the case is adjusted, the result is output, and the solution, working condition variable and solving time of the problem of this case are saved. (3d) storing the adjusted case in the case library in time, if the case library has the same case as the target case, the target case will be discarded directly, otherwise, the adjusted case is stored in the case library; The step (3b) specifically comprises the following steps: (3b1) initializing the BP neural network parameters; (3b2) providing the first set of input and output (X k , Y k ) to the BP neural network; (3b3) determining the input and output of the hidden layer and the output layer respectively, as shown in formulas (5) to (8): （5） （6） （7） （8） wherein: , are connection weights from the input layer to the hidden layer, are connection weights from the hidden layer to the output layer, is an input of the neural network, , are thresholds of the hidden layer and the output layer, respectively, i = 1, 2,..., pi, j = 1, 2,..., p2; is a hidden layer input, is a hidden layer output, is an output layer input, is an output layer output, pi is the number of input layer nodes, and p2 is the number of hidden layer nodes. (3b4) Construct a random motion mechanism for the connection weight parameters. Equation (9) is the connection weight correction error of the input layer, and Equation (10) is the connection weight correction error of the output layer. Adjust the connection weights between the hidden layer and the output layer. First, generate a random number between 0 and 1. , for The nonlinear self-feedback introduces probability, when Adjust according to formula (12) when Adjust the threshold according to formula (11); and adjust the output layer threshold according to formula (13). Similarly, adjust the connection weights between the input layer and the hidden layer. and the threshold of the hidden layer Adjustments are also made in accordance with the above methods, as shown in equations (14) to (16); d k = (Y k -c)c(1-c) (9) e j =d k v j b j (1-b j ) (10) （11） （12） （13） （14） （15） （16） wherein and is the learning rate, is the number of learning times; d k is the connection weight correction error of the input layer, e j is the connection weight correction error of the output layer.

2. The method of claim 1, wherein the method is characterized by: In step (1), the main parameters include: MLSS activated sludge concentration, temperature, pH value, influent COD chemical oxygen demand, DO dissolved oxygen, n water inlet pump motor currents, NH4+-N nitrogen source material, and HRT hydraulic retention time.

3. The method of claim 1, wherein the method is characterized by: The step (2) specifically comprises the following steps: (2a) The input variables of the RBF neural network are selected as n water inlet pump motor currents, and the output variable is the total water inlet amount ; (2b) Take the Gaussian function as the RBF neural network hidden layer node function, get the output of the first hidden layer node of the RBF neural network: ​ （1） wherein: is the output of the th hidden layer node, is the center vector of the th node of the RBF neural network, is the input vector of the RBF neural network, is the th value in the basis width vector of the RBF neural network, , is the number of hidden layer nodes; Further, the total water inflow prediction model of the RBF neural network is: （2） wherein: are RBF neural network hidden layer to output layer connection weights, is the number of output layer nodes; (2c) defining a target function: （3） Using the actual total water intake volume of the inlet pump Total water inflow predicted by RBF neural network Difference between By continuously adjusting the connection weights of the RBF neural network, the output of the RBF neural network gradually approximates the actual total water intake of the object. ; After dimension reduction, the auxiliary variables are MLSS activated sludge concentration, temperature, pH value, influent COD chemical oxygen demand, DO dissolved oxygen, total influent volume, NH4+-N nitrogen source, and HRT hydraulic retention time. ​

Images