An online compensation method for missing data in urban sewage treatment process based on autoencoder neural network

Abstract

An online compensation method for missing data in the urban sewage treatment process based on an autoencoder neural network realizes real-time compensation for random missing data in the urban sewage treatment process. In view of the nonlinear and high-noise characteristics of urban sewage treatment process data, which makes it difficult to establish an accurate mathematical model, this compensation method uses an autoencoder neural network to automatically extract the nonlinear relationship between variables, judges the missing situation of real-time samples according to the event trigger mechanism, and uses a gradient descent algorithm to adaptively adjust the model parameters online, thereby reducing the impact of high noise on the compensation results and improving the accuracy of real-time compensation for missing data in the urban sewage treatment process.

Description

Technical Field

[0001] The present invention is based on the analysis of the operation characteristics of the urban sewage treatment process, and establishes an online compensation model for missing data in the urban sewage treatment process based on an autoencoder neural network. In the actual process, the online compensation model can compensate for the random missing data in the urban sewage treatment process data, improve the credibility of the urban sewage treatment process data, and belongs to the field of sewage treatment. Background Art

[0002] In recent years, with the large-scale installation of sensors and distributed systems in urban sewage treatment plants, urban sewage treatment process data is collected, transmitted and saved in real time. Real-time data is an important basis for realizing the operation monitoring, optimization control and fault prediction of urban sewage treatment process, and is the foundation for building sewage treatment informatization. However, due to sensor failure, transmission network interruption or human factors, the loss of real-time data has become an inevitable problem, which has brought harm to the smooth and efficient operation of urban sewage treatment systems. Therefore, real-time compensation of missing data has important theoretical significance and application value to ensure the normal and stable operation of sewage treatment plants.

[0003] Since the urban sewage treatment process is a biochemical process involving many complex treatment processes, the working environment is complex and the interference is large, which makes it difficult to build an accurate mathematical model to accurately compensate for missing data, seriously reducing the reliability of data and bringing difficulties to the analysis, processing and application of urban sewage treatment process data. How to compensate for missing data in real time has become a challenge faced by the application of urban sewage treatment process data. At present, sewage treatment plants mainly compensate through statistical methods, but do not consider the changes in data distribution and the relationship between variables, resulting in unsatisfactory data compensation effect. Therefore, artificial neural networks have become an effective technology for online compensation. Artificial neural networks have the ability to extract nonlinear relationships between variables without the need to build accurate mathematical models. Therefore, the present invention has important practical significance in improving the effect of online missing data compensation.

[0004] The present invention designs an online compensation method for missing data in the urban sewage treatment process based on an autoencoder neural network. The method extracts input features, establishes an online compensation model based on the autoencoder neural network, uses an event trigger mechanism to judge the missing situation of input samples, and uses a gradient descent algorithm to adjust the model parameters online, thereby achieving accurate compensation for missing data in the urban sewage treatment process, improving data quality, and meeting the actual needs of sewage treatment plants. Summary of the invention

[0005] The present invention obtains an online compensation method for missing data in the urban sewage treatment process based on an autoencoder neural network. The method first establishes a compensation model based on the autoencoder neural network to extract the correlation between variables in the urban sewage treatment process data; secondly, the missing situation of real-time samples is identified according to the event trigger mechanism; finally, the gradient descent algorithm is used to adjust the model parameters online to improve the online compensation accuracy;

[0006] The present invention adopts the following technical solutions and implementation steps:

[0007] An online compensation method for missing data in a municipal sewage treatment process based on an autoencoder neural network is characterized in that a compensation model based on an autoencoder neural network is designed, missing samples are identified according to an event trigger mechanism, and compensation model parameters are updated using a gradient descent algorithm to achieve online compensation for missing data in a municipal sewage treatment process, and the method comprises the following steps:

[0008] (1) Collecting operation data

[0009] Taking the urban sewage treatment process as the research object, the operation data are collected in real time through the instruments installed on site, including inlet flow, inlet chemical oxygen demand, inlet total phosphorus, effluent total nitrogen, effluent total phosphorus, inlet ammonia nitrogen, inlet pH, aeration tank temperature and return sludge volume; suppose the collected operation data set D has a total of Z samples, and Z is an integer between [1500,10000];

[0010] (2) Establishing a compensation model

[0011] The compensation model is based on an autoencoder neural network and consists of three layers: input layer, hidden layer and output layer. Its structure is a 9-5-9 connection mode, that is, there are 9 neurons in the input layer, 5 neurons in the hidden layer and 9 neurons in the output layer. The input variables of the compensation model are x1(t), x2(t), x3(t), x4(t), x5(t), x6(t), x7(t), x8(t), x9(t), t = 1, 2, ..., Z, x1(t) is the water flow at time t, in liters / cubic meters, x2(t) is the water inlet concentration at time t, and x1(t) is the water inlet concentration at time t. x3(t) is the influent total phosphorus concentration at time t, in mg / L, x4(t) is the effluent total nitrogen concentration at time t, in mg / L, x5(t) is the effluent total phosphorus concentration at time t, in mg / L, x6(t) is the influent ammonia nitrogen concentration at time t, in mg / L, x7(t) is the influent pH at time t, x8(t) is the aeration tank temperature at time t, in degrees Celsius, x9(t) is the biological oxygen demand at time t, in mg / L, and the layers in the compensation model based on the autoencoder neural network are expressed as follows:

[0012] Input layer: It consists of 9 neurons. The input variables are x1(t), x2(t), x3(t), x4(t), x5(t), x6(t), x7(t), x8(t), x9(t). The output of each neuron is:

[0013] u p (t) = x p (t) (1)

[0014] Among them, u p (t) is the output of the pth neuron at time t, p = 1, 2, ..., 9;

[0015] Hidden layer: consists of 5 neurons, and the output of each neuron is:

[0016]

[0017] Among them, h j (t) is the output of the jth hidden layer neuron at time t, W epj (t) is the connection weight of the jth hidden layer neuron of the pth variable at time t, b ej (t) is the offset of the jth hidden layer neuron at time t, j = 1,…,5;

[0018] Output layer: It consists of 9 neurons, and the output of each neuron is:

[0019]

[0020] in, is the output of the pth neuron at time t, W djp (t) is the connection weight of the j-th input and p-th output layer neuron at time t, b dp (t) is the offset of the p-th output layer neuron at time t;

[0021] (3) Update compensation model parameters

[0022] ① Assume that the initial number of iterations s = 1, the maximum number of iterations S = 100, the starting sample of the running data set D is the first input sample, that is, t = 1, and the connection weights of the hidden layer neurons at the first iteration at the initial moment are Randomly select a value in the interval [-1,1], the hidden layer offset of the first iteration at the initial moment Randomly select values ​​in the interval [-1,1], the output layer connection weights of the first iteration at the initial moment Randomly select a value in the interval [-1,1], the output layer offset of the first iteration at the initial moment Randomly select a value in the interval [-1,1];

[0023] ②Identify missing samples

[0024] The input sample of the running data set D at time t is x(t)=[x1(t),x2(t),…,x p (t),…,x9(t)] T , T is the transpose calculation, the missing mask vector of the input sample x(t) is:

[0025] m(t)=[m1(t),m2(t),...,m p (t),...,m9(t)] (4)

[0026] Among them, m1(t) is the missing mask of the influent flow rate at time t, m2(t) is the missing mask of the influent chemical oxygen demand at time t, m3(t) is the missing mask of the influent total phosphorus concentration at time t, m4(t) is the missing mask of the effluent total nitrogen concentration at time t, m5(t) is the missing mask of the effluent total phosphorus concentration at time t, m6(t) is the missing mask of the influent ammonia nitrogen at time t, m7(t) is the missing mask of the influent pH at time t, m8(t) is the missing mask of the aeration temperature at time t, m9(t) is the missing mask of the return sludge amount at time t, m p (t) is the missing mask of the pth variable at time t; if x p (t) has actual value, m p (t)=0, if x p (t) No actual value, m p (t) = 1;

[0027] ③If Save the input sample x(t) to the sample set D′, and when time t increases by 1, return to step ②, where ∑ is the sum calculation; if Then execute step ④;

[0028] ④ Extract the training set X = [x(tN),…,x(t-1)] at time t from the sample set D′, is rounded up; the training set X is sequentially substituted into the compensation model, and the i-th training sample is According to formulas (1)-(3), calculate the i-th training sample x i The output of (t) is

[0029] ⑤The parameter update formula of the compensation model at time t is:

[0030]

[0031] in, is the connection weight of the jth hidden layer neuron at the sth iteration at time t, is the offset of the hidden layer neurons at the s-th iteration at time t, is the connection weight of the p-th output layer neuron at the s-th iteration at time t, is the offset of the output layer neurons at the s-th iteration at time t, α is the learning rate and its value range is (0, 0.1), and E(t) is the error between the output of the compensation model and the input:

[0032]

[0033] where, is the u-th parameter of the parameter vector θ s (t) at the s-th iteration at time t, is the u-th parameter of the parameter vector θ S (t - 1) at the S-th iteration at time t - 1, Ω u (t) is the importance of the u-th parameter at time t and is randomly taken in [0.1, 0.5], u = 1, …, 20×5, and λ is the regularization parameter and is randomly taken in (0, 1);

[0034] ⑥ If the number of iterations s < S, increase the number of iterations s by 1, then return to step ⑤; if the number of iterations s ≥ S, then execute step ⑦;

[0035] ⑦ Using the trained compensation model, take the input sample x(t) as the input of the compensation model, and calculate the output according to formulas (1)-(3) Using the output compensate the input sample x(t):

[0036]

[0037] where, is the final compensation result of the input sample x(t), is the multiplication of the corresponding elements in the two vectors, I is the unit vector, and save the compensation result to the sample set D′;

[0038] ⑧ Increase the time t by 1, if t < Z, then return to step ②; if t ≥ Z, then stop the compensation;

[0039] (4) Output the compensation result

[0040] Output the sample set D′, which is the operation data sample set of the urban sewage treatment process after online compensation.

[0041] The creativity of the present invention is mainly reflected in:

[0042] (1) Aiming at the real-time missing data problem in the urban sewage treatment process, the present invention proposes an online missing data compensation method, which identifies the missing situation of real-time samples based on an event trigger mechanism and improves the missing data compensation efficiency;

[0043] (2) In order to take into account the nonlinear and high-noise characteristics of urban sewage treatment process data, which are difficult to represent with an accurate data model, the present invention adopts an autoencoder neural network to extract the nonlinear relationship between variables in an unsupervised manner based on the measured data of actual urban sewage treatment plants. At the same time, a perceptual synaptic weighted regularization term is designed to achieve a balance between new and old data, and has good adaptability to high-noise data. BRIEF DESCRIPTION OF THE DRAWINGS

[0044] Figure 1 This is a diagram showing the effect of the missing sample compensation of ammonia nitrogen in the water inlet chamber of the present invention;

[0045] Figure 2 It is the filling error diagram of ammonia nitrogen in the water inlet chamber of the present invention;

[0046] Figure 3 This is a diagram showing the effect of missing sample compensation for the pH value of the water inlet chamber of the present invention;

[0047] Figure 4 It is the filling error diagram of the pH value of the effluent of the present invention; DETAILED DESCRIPTION

[0048] The experimental data comes from the water quality analysis table of a sewage treatment plant in 2022; the actual test data of inlet flow, inlet chemical oxygen demand, inlet total phosphorus, effluent total nitrogen, effluent total phosphorus, inlet ammonia nitrogen, inlet pH, aeration tank temperature and return sludge volume are taken as 2000 groups of experimental sample data, of which the first 600 groups are complete samples and the remaining 1400 groups are missing samples.

[0049] The present invention adopts the following technical solutions and implementation steps:

[0050] An online compensation method for missing data in a municipal sewage treatment process based on an autoencoder neural network is characterized in that a compensation model based on an autoencoder neural network is designed, missing samples are identified according to an event trigger mechanism, and compensation model parameters are updated using a gradient descent algorithm to achieve online compensation for missing data in a municipal sewage treatment process, and the method comprises the following steps:

[0051] (1) Collecting operation data

[0052] Taking the urban sewage treatment process as the research object, the operation data are collected in real time through the instruments installed on site, including inlet flow, inlet chemical oxygen demand, inlet total phosphorus, effluent total nitrogen, effluent total phosphorus, inlet ammonia nitrogen, inlet pH, aeration tank temperature and return sludge volume; suppose the collected operation data set D has a total of Z samples, Z = 2000;

[0053] (2) Establishing a compensation model

[0054] The compensation model is based on an autoencoder neural network and consists of three layers: input layer, hidden layer and output layer. Its structure is a 9-5-9 connection mode, that is, there are 9 neurons in the input layer, 5 neurons in the hidden layer and 9 neurons in the output layer. The input variables of the compensation model are x1(t), x2(t), x3(t), x4(t), x5(t), x6(t), x7(t), x8(t), x9(t), t = 1, 2, ..., Z, x1(t) is the water flow at time t, in liters / cubic meters, x2(t) is the water inlet concentration at time t, and x1(t) is the water inlet concentration at time t. x3(t) is the influent total phosphorus concentration at time t, in mg / L, x4(t) is the effluent total nitrogen concentration at time t, in mg / L, x5(t) is the effluent total phosphorus concentration at time t, in mg / L, x6(t) is the influent ammonia nitrogen concentration at time t, in mg / L, x7(t) is the influent pH at time t, x8(t) is the aeration tank temperature at time t, in degrees Celsius, x9(t) is the biological oxygen demand at time t, in mg / L, and the layers in the compensation model based on the autoencoder neural network are expressed as follows:

[0055] Input layer: It consists of 9 neurons. The input variables are x1(t), x2(t), x3(t), x4(t), x5(t), x6(t), x7(t), x8(t), x9(t). The output of each neuron is:

[0056] u p (t) = x p (t) (11)

[0057] Among them, u p (t) is the output of the pth neuron at time t, p = 1, 2, ..., 9;

[0058] Hidden layer: consists of 5 neurons, and the output of each neuron is:

[0059]

[0060] Among them, h j (t) is the output of the jth hidden layer neuron at time t, W epj (t) is the connection weight of the jth hidden layer neuron of the pth variable at time t, b ej(t) is the offset of the jth hidden layer neuron at time t, j = 1,…,5;

[0061] Output layer: It consists of 9 neurons, and the output of each neuron is:

[0062]

[0063] in, is the output of the pth neuron at time t, W djp (t) is the connection weight of the j-th input and p-th output layer neuron at time t, b dp (t) is the offset of the p-th output layer neuron at time t;

[0064] (3) Update compensation model parameters

[0065] ① Assume that the initial number of iterations s = 1, the maximum number of iterations S = 100, the starting sample of the running data set D is the first input sample, that is, t = 1, and the connection weights of the hidden layer neurons at the first iteration at the initial moment are Randomly select a value in the interval [-1,1], the hidden layer offset of the first iteration at the initial moment Randomly select values ​​in the interval [-1,1], the output layer connection weights of the first iteration at the initial moment Randomly select a value in the interval [-1,1], the output layer offset of the first iteration at the initial moment Randomly select a value in the interval [-1,1];

[0066] ②Identify missing samples

[0067] The input sample of the running data set D at time t is x(t)=[x1(t),x2(t),…,x p (t),…,x9(t)] T , T is the transpose calculation, the missing mask vector of the input sample x(t) is:

[0068] m(t)=[m1(t),m2(t),...,m p (t),...,m9(t)] (14)

[0069] Among them, m1(t) is the missing mask of the influent flow rate at time t, m2(t) is the missing mask of the influent chemical oxygen demand at time t, m3(t) is the missing mask of the influent total phosphorus concentration at time t, m4(t) is the missing mask of the effluent total nitrogen concentration at time t, m5(t) is the missing mask of the effluent total phosphorus concentration at time t, m6(t) is the missing mask of the influent ammonia nitrogen at time t, m7(t) is the missing mask of the influent pH at time t, m8(t) is the missing mask of the aeration temperature at time t, m9(t) is the missing mask of the return sludge amount at time t, m p(t) is the missing mask of the pth variable at time t; if x p (t) has actual value, m p (t)=0, if x p (t) No actual value, m p (t) = 1;

[0070] ③If Save the input sample x(t) to the sample set D′, and when time t increases by 1, return to step ②.

[0071] ∑ is the summation calculation; if Then execute step ④;

[0072] ④ Extract the training set X = [x(tN),…,x(t-1)] at time t from the sample set D′, is rounded up; the training set X is sequentially substituted into the compensation model, and the i-th training sample is According to formulas (11)-(13), calculate the i-th training sample x i The output of (t) is

[0073] ⑤The parameter update formula of the compensation model at time t is:

[0074]

[0075]

[0076] in, is the connection weight of the jth hidden layer neuron at the sth iteration at time t, is the offset of the hidden layer neuron at the sth iteration at time t, is the connection weight of the pth output layer neuron at the sth iteration at time t, is the offset of the output layer neuron at the sth iteration at time t, α is the learning rate and its value range is (0,0.1), and E(t) is the error between the compensation model output and input:

[0077]

[0078] in, is the sth iteration parameter vector θ at time t s The uth parameter of (t), is the Sth iteration parameter vector θ at time t-1 S The uth parameter of (t-1), Ω u (t) is the importance degree of the u-th parameter at time t and randomly takes values in [0.1, 0.5], where u = 1, …, 20×5, and λ is the regularization parameter and randomly takes values in (0, 1);

[0079] ⑥ If the number of iterations s < S, the number of iterations s is incremented by 1, and then return to step ⑤; if the number of iterations s ≥ S, then execute step ⑦;

[0080] ⑦ Using the trained compensation model, take the input sample x(t) as the input of the compensation model, and calculate the output according to formulas (11)-(13) Using the output Compensate the input sample x(t):

[0081]

[0082] where, is the final compensation result of the input sample x(t), is the multiplication of the corresponding elements in the two vectors, I is the unit vector, and the compensation result is saved to the sample set D′;

[0083] ⑧ Increment time t by 1. If t < Z, then return to step ②; if t ≥ Z, then stop the compensation;

[0084] (4) Output the compensation result

[0085] Output the sample set D′, which is the operation data sample set of the urban sewage treatment process after online compensation.

Claims

1. An online compensation method for missing data in urban sewage treatment process based on autoencoder neural network, characterized in that: Including the following steps: (1) Collect operation data Taking the urban sewage treatment process as the research object, the operation data is collected in real time through the instruments installed on site, including influent flow rate, influent chemical oxygen demand, influent total phosphorus, effluent total nitrogen, effluent total phosphorus, influent ammonia nitrogen, influent pH value, aeration tank temperature and return sludge volume; it is assumed that the collected operation data set D has a total of Z samples, and Z takes an integer between [1500, 10000]. (2) Establish a compensation model The compensation model is based on an autoencoder neural network and consists of the following three layers: input layer, hidden layer and output layer; its structure is a 9-5-9 connection method, that is, there are 9 neurons in the input layer, 5 neurons in the hidden layer, and 9 neurons in the output layer; the input variables of the compensation model are x1(t), x2(t), x3(t), x4(t), x5(t), x6(t), x7(t), x8(t), x9(t), t = 1, 2,..., Z, x1(t) is the influent flow rate at time t, with the unit of liters per cubic meter, x2(t) is the influent chemical oxygen demand at time t, with the unit of milligrams per liter, x3(t) is the influent total phosphorus concentration at time t, with the unit of milligrams per liter, x4(t) is the effluent total nitrogen concentration at time t, with the unit of milligrams per liter, x5(t) is the effluent total phosphorus concentration at time t, with the unit of milligrams per liter, x6(t) is the influent ammonia nitrogen concentration at time t, with the unit of milligrams per liter, x7(t) is the influent pH value at time t, x8(t) is the aeration tank temperature at time t, with the unit of degrees Celsius, x9(t) is the biological oxygen demand at time t, with the unit of milligrams per liter. The representations of each layer in the compensation model based on the autoencoder neural network are as follows: Input layer: Consisting of 9 neurons, the input variables are x1(t), x2(t), x3(t), x4(t), x5(t), x6(t), x7(t), x8(t), x9(t), and the output of each neuron is: u p (t)=x p (t) (1) Among them, u p (t) is the output of the pth neuron at time t, p = 1, 2, ..., 9; Hidden layer: Consisting of 5 neurons, the output of each neuron is: Among them, h j (t) is the output of the jth hidden layer neuron at time t, W epj (t) is the connection weight of the jth hidden layer neuron of the pth variable at time t, b ej (t) is the offset of the jth hidden layer neuron at time t, j = 1,…,5; Output layer: Consisting of 9 neurons, the output of each neuron is: in, is the output of the pth neuron at time t, W djp (t) is the connection weight of the j-th input and p-th output layer neuron at time t, b dp (t) is the offset of the p-th output layer neuron at time t; (3) Update the parameters of the compensation model ① Assume that the initial number of iterations s = 1, the maximum number of iterations S = 100, the starting sample of the running data set D is the first input sample, that is, t = 1, and the connection weights of the hidden layer neurons of the first iteration at the initial moment are Randomly select a value in the interval [-1,1], the hidden layer offset b of the first iteration at the initial moment 1ep (1) Randomly select values ​​in the interval [-1, 1], and the output layer connection weights of the first iteration at the initial moment Randomly select a value in the interval [-1,1], the output layer offset b of the first iteration at the initial moment 1dp (1) Randomly select a value in the interval [-1,1]; ② Identify missing samples The input sample of the running data set D at time t is x(t)=[x1(t),x2(t),…,x p (t),…,x9(t)] T , T is the transpose calculation, the missing mask vector of the input sample x(t) is: m(t)=[m1(t),m2(t),...,m p (t),...,m9(t)] (4) Among them, m1(t) is the missing mask of the influent flow rate at time t, m2(t) is the missing mask of the influent chemical oxygen demand at time t, m3(t) is the missing mask of the influent total phosphorus concentration at time t, m4(t) is the missing mask of the effluent total nitrogen concentration at time t, m5(t) is the missing mask of the effluent total phosphorus concentration at time t, m6(t) is the missing mask of the influent ammonia nitrogen at time t, m7(t) is the missing mask of the influent pH at time t, m8(t) is the missing mask of the aeration temperature at time t, m9(t) is the missing mask of the return sludge amount at time t, m p (t) is the missing mask of the pth variable at time t; if x p (t) has actual value, m p (t)=0, if x p (t) No actual value, m p (t) = 1; ③If Save the input sample x(t) to the sample set D′, and when time t increases by 1, return to step ②, where ∑ is the sum calculation; if Then execute step ④; ④ Extract the training set X = [x(tN),…,x(t-1)] at time t from the sample set D′, is rounded up; the training set X is sequentially substituted into the compensation model, and the i-th training sample is i=1,…,N, according to formulas (1)-(3), calculate the i-th training sample x i The output of (t) is ⑤ The parameter update formula of the compensation model at time t is: in, is the connection weight of the jth hidden layer neuron at the sth iteration at time t, is the offset of the hidden layer neuron at the sth iteration at time t, is the connection weight of the pth output layer neuron at the sth iteration at time t, is the offset of the output layer neuron at the sth iteration at time t, α is the learning rate and its value range is (0,0.1), and E(t) is the error between the compensation model output and input: Among them, θs u (t) is the sth iteration parameter vector θ at time t s The uth parameter of (t), is the Sth iteration parameter vector θ at time t-1 S The uth parameter of (t-1), Ω u (t) is the importance of the u-th parameter at time t and takes a random value in [0.1, 0.5], u = 1, ..., 20 × 5, λ is the regularization parameter and takes a random value in (0, 1); ⑥ If the iteration number s < S, the iteration number s is increased by 1, and then return to step ⑤; if the iteration number s ≥ S, then execute step ⑦; ⑦ Using the trained compensation model, take the input sample x(t) as the input of the compensation model and calculate the output according to formulas (1)-(3) Utilize output Compensate the input sample x(t): in, is the final compensation result of the input sample x(t), is the multiplication of the corresponding position elements in the two vectors, I is the unit vector, and the compensation result is Save to sample set D′; ⑧ The time t is increased by 1. If t < Z, then return to step ②; if t ≥ Z, then stop compensation; (4) Output the compensation result Output the sample set D′, which is the operation data sample set of the urban sewage treatment process after online compensation.

Images