Method for predicting concentration of sulfuric acid discharged from electronic grade sulfuric acid absorption tower

Abstract

The invention relates to the technical field of electronic-grade sulfuric acid production, and particularly discloses a method for predicting the concentration of sulfuric acid discharged from an electronic-grade sulfuric acid absorption tower, which comprises the following steps of: collecting data, performing grouping preprocessing, performing nondimensionalization on the data by using a normalization method, and improving the stability of a model during operation; constructing a regression model based on a feed-forward neural network, and performing hyper-parameter optimization and training on the hidden layer number and the neuron number of the model by taking the feature data as input and the concentration value as output, so that prediction model parameters are continuously learned and updated; and finally obtaining a concentration prediction model of the electronic-grade sulfuric acid out of the absorption tower. When a follow-up device operates, a model prediction concentration value and an actual measurement value can be synchronously compared, so that the operation state of the absorption tower is predicted in advance, the production process is optimized, and a certain guiding effect on the current production process is achieved.

Description

Technical Field

[0001] The invention belongs to the technical field of electronic-grade sulfuric acid production, and in particular relates to a method for predicting sulfuric acid concentration at an outlet of an electronic-grade sulfuric acid absorption tower. Background Art

[0002] In the process of producing electronic-grade sulfuric acid by hydration, the absorption tower is the key place and equipment for the reaction and generation of electronic-grade sulfuric acid. The efficiency and performance of the absorption tower are directly related to the quality of the product and the stability of the production process, and the outlet concentration is the key indicator to measure the performance of the absorption tower. The outlet concentration refers to the concentration of electronic-grade sulfuric acid leaving the absorption tower. Accurately controlling the outlet concentration of electronic-grade sulfuric acid is of great significance for ensuring product quality, improving product yield and reducing waste emissions. The concentration of electronic-grade sulfuric acid in the outlet of the absorption tower is an important control parameter in production. The concentration reflects the amount of water replenishment and circulation in the production process of electronic-grade sulfuric acid, and also reflects the temperature in the tower. This is crucial to controlling the operation of the absorption tower within a safe range and the quality of the product.

[0003] However, in actual industrial production, due to technical or economic reasons, such as equipment cost, installation difficulty, maintenance cost, etc., the out-tower concentration is often not directly measured online by traditional sensors. This leads to a lack of real-time and accurate feedback information in the production process, which affects the optimization of the production process and the stability of product quality. Some existing online concentration meters also have large errors in actual measurements compared with offline manual detection. Therefore, the development of a method that can predict the concentration of electronic-grade sulfuric acid out of the absorption tower in real time has important industrial application value. Summary of the invention

[0004] The invention provides a method for predicting sulfuric acid concentration at an outlet of an electronic grade sulfuric acid absorption tower, which can predict the operating state of the absorption tower in advance and provide a basis for the stable operation of a production process and the control of product quality.

[0005] The technical solution of the present invention is to provide a method for predicting sulfuric acid concentration at the outlet of an electronic grade sulfuric acid absorption tower, comprising the following steps:

[0006] S1. Collect historical operation data of the absorption tower, clarify input features (or input variables) and output labels (or output variables), the input features are the operating conditions and raw material properties that affect the sulfuric acid concentration out of the tower, and the output label is the sulfuric acid concentration out of the tower;

[0007] S2, distinguish the data obtained in S1 according to the working conditions, retain the data under normal full-load operating conditions, and normalize and process the abnormal data for this part of the data, and divide the obtained data into a training set, a validation set, and a test set;

[0008] S3. Construct a prediction model based on a feedforward neural network. The model includes an input layer, a hidden layer, and an output layer. Then, use the training set to train the optimization model, use the validation set to validate the optimization model, and finally use the test set to test the results of the optimization model to obtain the predicted value of the model. The predicted value of the model is denormalized to obtain the tower sulfuric acid concentration.

[0009] Furthermore, the input variables in S1 include, but are not limited to: temperature, pressure, and flow parameters of gaseous sulfur trioxide, temperature and flow parameters of ultrapure water, and temperature and flow parameters of electronic grade sulfuric acid circulating back to the absorption tower.

[0010] Furthermore, the operating conditions in S2 at least include parking conditions, starting conditions, full-load operating conditions, high-load operating conditions and low-load operating conditions.

[0011] Furthermore, in S2, during normalization, the original data is converted into a data set with a mean of 0 and a standard deviation of 1, and the transformation formula is:

[0012]

[0013] Where μ is the mean of the original data and σ is the standard deviation of the original data.

[0014] Furthermore, in S2, when processing abnormal data, identification is first performed using the 3σ principle, and any data points outside the range of [μ-3σ, μ+3σ] are marked as outliers and removed.

[0015] Furthermore, when dividing the data in S2, a random seed is set to randomly divide the data, where the ratio of the training set, validation set, and test set is 7:2:1.

[0016] Furthermore, when constructing a feedforward neural network model in S3, the input layer contains a certain number of neurons, which is equal to the number of input variables. The neurons receive the feature data of the training set and pass it to the hidden layer;

[0017] The hidden layer performs a linear transformation on the data transmitted by the input layer, where each hidden layer is a linear transformation, which is defined by a weight matrix and a bias vector. For the neuron i in the lth layer, its output is expressed as:

[0018]

[0019] In the formula, is the weight connecting neuron j in layer l-1 and neuron i in layer l, is the output of layer l-1, is the bias of the lth layer;

[0020] After the hidden layer is linearly transformed, an activation function is used to introduce nonlinearity so that the neural network can learn and simulate complex functions. The activation function is one of the Sigmoid function, ReLU function or Tanh function. The specific activation method is:

[0021]

[0022] ReLU: f(z) = max(0,z);

[0023]

[0024] After the data is transformed by linear transformation and activation function in the hidden layer, it continues to propagate forward from the input layer to the output layer. The output of each layer is used as the input of the next layer. For the neuron i in the lth layer, its output is:

[0025]

[0026] After the forward propagation is completed, reverse gradient propagation is performed to calculate the network parameter gradient (weight and bias). Specifically, the chain rule is used to start from the output layer and calculate the gradient layer by layer. For the weight of the lth layer and bias Its gradient can be expressed as:

[0027]

[0028] E is a loss function used to measure the difference between the predicted value and the actual value of the model; the loss function is the MSE function; is the error of the lth layer, calculated as:

[0029]

[0030] in, is the linear combination output of the lth layer, is the derivative of the activation function;

[0031] After the reverse gradient propagation, the network weights and biases are updated, and the update method is expressed as:

[0032]

[0033] The output layer produces the final output result. The neurons in this layer are 1, and the output result is the predicted value after the model learns based on the input features.

[0034] Furthermore, the number of hidden layers of the neural network ranges from [1, 4], and the number of neurons ranges from [8, 256].

[0035] Furthermore, the denormalization process is the opposite of the normalization process, that is, the reverse operation of the normalization process, and the transformation formula is:

[0036] X=σ·X NORM +μ;

[0037] The output value after the inverse normalization operation is the final predicted concentration of electronic-grade sulfuric acid out of the absorption tower by the model.

[0038] The invention also relates to the application of the prediction method in the production of electronic grade sulfuric acid.

[0039] The present invention has the following beneficial effects:

[0040] The method for predicting sulfuric acid concentration out of an electronic-grade sulfuric acid absorption tower provided by the present invention can collect the temperature, pressure, flow rate of each stream of materials entering the absorption tower and the concentration history data of materials out of the absorption tower through a DCS system; clarify input features and output labels; group the data and filter out the data when running at full load; use a normalization method to make the data dimensionless and improve the stability of the model when running; construct a regression model based on a feedforward neural network, take feature data as input and label data (concentration value) as output, optimize and train the number of hidden layers and neurons of the model, so that the parameters of the prediction model are continuously learned and updated; and finally obtain a concentration prediction model of electronic-grade sulfuric acid out of the absorption tower.

[0041] By predicting the sulfuric acid concentration at the outlet of the absorption tower, the model-predicted concentration value and the actual measured value can be compared synchronously during the operation of subsequent devices, so as to predict the operating status of the absorption tower in advance and optimize the production process, provide a basis for the stable operation of the production process and the control of product quality, improve the efficiency and performance of the absorption tower, reduce production costs, and play a certain guiding role in the production process. BRIEF DESCRIPTION OF THE DRAWINGS

[0042] Figure 1 The figure is a flow chart of the prediction method of the present invention. DETAILED DESCRIPTION

[0043] The present invention is described in detail below in conjunction with specific embodiments. This embodiment is implemented based on the technical solution of the present invention, and provides a detailed implementation method and a specific operation process, but the protection scope of the present invention is not limited to the following embodiments.

[0044] The method for predicting sulfuric acid concentration at the outlet of an electronic-grade sulfuric acid absorption tower provided by the present invention comprises the following steps:

[0045] 1) Collect a large amount of historical production data, which should cover the production process of the DCS (distributed control system) absorption tower. These data usually include but are not limited to: the operating parameters of the absorption tower (such as temperature, pressure, flow, etc.), the composition and properties of the raw materials, and the concentration of the tower product.

[0046] Specify parameter indicators as input variables (features) and as output variables (labels). Input variables are operating conditions and raw material properties that affect the out-of-tower concentration, such as the concentration, flow rate, temperature of sulfuric acid entering the tower, and the use of absorbents. The output variable is the out-of-tower concentration of electronic-grade sulfuric acid to be predicted; in a specific embodiment, the collected parameters are seven parameters: temperature, pressure, and flow rate parameters of gaseous sulfur trioxide, temperature and flow rate parameters of ultrapure water, and temperature and flow rate parameters of electronic-grade sulfuric acid circulating back to the absorption tower as input variables or input features. The out-of-tower concentration parameter of the electronic-grade sulfuric acid absorption tower is used as an output variable or label.

[0047] Data preprocessing. The data recorded in the DSC system includes the operating data of the absorption tower under all working conditions, involving various working conditions: parking conditions, starting conditions, full-load operating conditions, high-load operating conditions, low-load operating conditions, etc. All data need to be distinguished, and the data under normal full-load operation is retained, and other data are discarded. The original data of normal full-load operation is normalized and abnormal data is processed.

[0048] Specifically, the normalization method for the full-load operation group data is the Z-score normalization method, which converts the original data into a data set with a mean of 0 and a standard deviation of 1. The transformation formula is:

[0049]

[0050] Where μ is the mean of the original data and σ is the standard deviation of the original data.

[0051] The abnormal data identification method is the 3σ principle. Any data point beyond the range of [μ-3σ,μ+3σ] can be regarded as a potential outlier. Each data point in the data set is checked. If the value of the data point is less than μ-3σ or greater than μ+3σ, it is marked as an outlier and removed.

[0052] The last retained data is divided into training set, validation set and test set; in the specific case, the ratio of data division into training set, validation set and test set is 7:2:1. When dividing the data, a random seed is set to make the data randomly divided.

[0053] 3) First build a feedforward neural network model; then build an optimization model, use the training set data to optimize the model, and use the validation set data to verify the model; then use the test set data to test the constructed model; finally, the test data is normalized to the real prediction data.

[0054] The constructed prediction model is a feed-forward neural network, which includes an input layer, a hidden layer and an output layer.

[0055] The input layer contains a certain number of neurons, in some specific cases, the number is equal to the number of features, which is 7. The neurons receive the training set feature data and pass it to the hidden layer.

[0056] The hidden layer performs a linear transformation on the data transmitted by the input layer. Each hidden layer can be regarded as a linear transformation, which is defined by a weight matrix and a bias vector. For the neuron i in the lth layer, its output can be expressed as:

[0057]

[0058] in, is the weight connecting neuron j in layer l-1 and neuron i in layer l, is the output of layer l-1, is the bias of the lth layer.

[0059] After the hidden layer is linearly transformed, an activation function is used to introduce nonlinearity so that the neural network can learn and simulate complex functions. The activation function is one of the Sigmoid function, the ReLU function, and the Tanh function.

[0060] The specific activation method is:

[0061]

[0062] ReLU: f(z) = max(0,z)

[0063]

[0064] After the data is transformed by linear transformation and activation function in the hidden layer, it continues to propagate forward from the input layer to the output layer, and the output of each layer serves as the input of the next layer. For neuron i in layer l, its output is:

[0065]

[0066] After the forward propagation, the reverse gradient propagation is performed to calculate the network parameter gradient (weight and bias), which uses the chain rule to start from the output layer and calculate the gradient layer by layer. and bias Its gradient can be expressed as:

[0067]

[0068] E is a loss function used to measure the difference between the predicted value and the actual value of the model. The loss function is the MSE function. is the error of the lth layer, calculated as:

[0069]

[0070] in, is the linear combination output of the lth layer, is the derivative of the activation function.

[0071] After back propagation, the network weights and biases are updated, and the update method can be expressed as:

[0072]

[0073] The output layer produces the final output result. The neurons in this layer are 1, and the output result is the predicted value after the model learns based on the input features.

[0074] The constructed neural network model is optimized to obtain the optimal neural network model. After the test set data is input into the optimal neural network model, the predicted value of the model is obtained;

[0075] The model prediction value needs to be denormalized to return to its original value. The denormalization method is the inverse operation of the Z-score normalization method:

[0076] X=σ·X NORM +μ

[0077] The output value after the inverse normalization operation is the final predicted concentration of electronic-grade sulfuric acid out of the absorption tower by the model.

[0078] For example, some of the predicted values ​​and true values ​​after prediction using the above method are shown in Table 1 below.

[0079] Table 1. Some prediction values ​​and errors of the model

[0080]

[0081] The above embodiments describe the preferred implementation of the present invention, but the present invention is not limited thereto. Within the technical concept of the present invention, the technical solution of the present invention can be subjected to a variety of simple modifications, including the combination of various technical features in any other manner, and these simple modifications and combinations should also be regarded as the contents disclosed by the present invention and belong to the protection scope of the present invention. Therefore, the protection scope of the patent of the present invention shall be subject to the attached claims.

Claims

1. A method for predicting sulfuric acid concentration at the outlet of an electronic grade sulfuric acid absorption tower, characterized in that: The following steps are involved: S1. Collect historical operation data of the absorption tower, and clarify input features and output labels. The input features are the operating conditions and raw material properties that affect the sulfuric acid concentration out of the tower, and the output label is the sulfuric acid concentration out of the tower; S2, distinguish the data obtained in S1 according to the working conditions, retain the data under normal full-load operating conditions, and normalize and process the abnormal data for this part of the data, and divide the obtained data into a training set, a validation set, and a test set; S3. Construct a prediction model based on a feedforward neural network. The model includes an input layer, a hidden layer, and an output layer. Then, the model is trained using a training set, and then the model is verified using a validation set. Finally, the model result is tested using a test set to obtain the predicted value of the model. The predicted value of the model is denormalized to obtain the tower sulfuric acid concentration.

2. The method according to claim 1, characterized in that The input variables in S1 include, but are not limited to: temperature, pressure, and flow parameters of gaseous sulfur trioxide, temperature and flow parameters of ultrapure water, and temperature and flow parameters of electronic-grade sulfuric acid circulating back to the absorption tower.

3. The method according to claim 1, characterized in that: The data in S2 belongs to different operating conditions, including but not limited to parking conditions, starting conditions, full-load operating conditions, high-load operating conditions, and low-load operating conditions.

4. The method according to claim 1, characterized in that: In S2, during normalization, the original data is converted into a data set with a mean of 0 and a standard deviation of 1. The transformation formula is: Where μ is the mean of the original data and σ is the standard deviation of the original data.

5. The method according to claim 1, characterized in that: In S2, when processing abnormal data, identification is performed first. The identification method is the 3σ principle. Any data point outside the range of [μ-3σ, μ+3σ] is marked as an outlier and eliminated.

6. The method according to claim 1, characterized in that: When dividing the data in S2, a random seed is set to randomly divide the data, where the ratio of the training set, validation set, and test set is 7:2:

1.

7. The method according to any one of claims 1 to 6, characterized in that: When constructing a feedforward neural network model in S3, the input layer contains a certain number of neurons, which is equal to the number of input variables. The neurons receive the feature data of the training set and pass it to the hidden layer; The hidden layer performs a linear transformation on the data transmitted by the input layer, where each hidden layer acts as a linear transformation. The linear transformation is defined by the weight matrix and the bias vector. For the neuron i in the lth layer, its output is expressed as: In the formula, is the weight connecting neuron j in layer l-1 and neuron i in layer l, is the output of layer l-1, is the bias of the lth layer; After the hidden layer is linearly transformed, an activation function is used to introduce nonlinearity so that the neural network can learn and simulate complex functions. The activation function is one of the Sigmoid function, ReLU function or Tanh function. The specific activation method is: Sigmoid: ReLU: f(z) = max(0,z); Fishy: In the formula, e is a natural constant, and the max function is a maximum value function; After the data is transformed by linear transformation and activation function in the hidden layer, it continues to propagate forward from the input layer to the output layer. The output of each layer is used as the input of the next layer. For the neuron i in the lth layer, its output is: After the forward propagation is completed, reverse gradient propagation is performed to calculate the network parameter gradient (weight and bias). Specifically, the chain rule is used to start from the output layer and calculate the gradient layer by layer. For the weight of the lth layer and bias Its gradient can be expressed as: E is a loss function used to measure the difference between the predicted value and the actual value of the model; the loss function is the MSE function; is the error of the lth layer, calculated as: in, is the linear combination output of the lth layer, is the derivative of the activation function; After the reverse gradient propagation, the network weights and biases are updated, and the update method is expressed as: The output layer produces the final output result. The neurons in this layer are 1, and the output result is the predicted value after the model learns based on the input features.

8. The method according to claim 7, characterized in that: The number of hidden layers in a neural network ranges from [1, 4], and the number of neurons ranges from [8, 256].

9. The method according to claim 1, characterized in that: The denormalization operation is opposite to the normalization operation, that is, the reverse operation of the normalization operation. The transformation formula is: X=σ·X NORM +m; Where μ is the mean of the original data, σ is the standard deviation of the original data; The output value after the inverse normalization operation is the final predicted concentration of electronic-grade sulfuric acid out of the absorption tower by the model.

10. Use of the method according to any one of claims 1 to 9 in the production of electronic grade sulfuric acid.

Images