Method for intelligently controlling the amount of ammonia water added to an industrial flue gas absorbent

Abstract

This invention provides a method for intelligently controlling the dosage of ammonia water as an absorbent in industrial flue gas. It includes a real-time data acquisition program, an algorithm calculation program, and a control terminal. The automatic control device rationally determines the ammonia water dosage based on a comparison of previous and current operating conditions. The acquisition program collects desulfurization process data in real time, calculates the appropriate ammonia water dosage in real time using the algorithm program, and feeds back the feedback information to the industrial control computer for real-time control of the ammonia water dosage. Existing automatic adjustment systems for controlling outlet SO2 concentration ignore changes in operating conditions and deviations in system operation, requiring secondary system debugging. Another approach uses deep learning to consider changes in operating conditions and predicts the outlet SO2 concentration more accurately, then adjusts the reducing agent dosage based on the deviation from the set value. However, deep learning itself cannot explain this and is related to the distribution of historical data. This invention combines a mechanistic model and a machine learning model, utilizing the traditional PID control concept and iterating parameters using a deep neural network.

Description

Technical Field

[0001] This invention relates to the field of intelligent environmental protection, specifically to a method for intelligently controlling the dosage of ammonia water as an industrial flue gas absorbent. Background Technology

[0002] Existing automatic regulation systems for controlling SO2 concentration at the outlet fall into two categories: First, those relying on the original PID control system ignore changes in operating conditions and deviations in system operation, requiring secondary system debugging; second, those utilizing deep learning to incorporate changes in operating conditions and predict the outlet SO2 concentration more accurately, then adjusting the amount of reducing agent added based on the deviation from the set value. However, deep learning itself cannot be interpreted and is related to the distribution of historical data. This invention combines mechanistic models and machine learning models, utilizing the traditional PID control concept and iterating parameters using a deep neural network. Summary of the Invention

[0003] The purpose of this invention is to address the shortcomings of existing technologies by proposing a method for intelligently controlling the dosage of ammonia water as an industrial flue gas absorbent.

[0004] To achieve the above objectives, the present invention adopts the following technical solution: a method for intelligently controlling the dosage of ammonia water as an industrial flue gas absorbent, the method comprising the following steps:

[0005] S1: Export the historical operating data of the wet flue gas desulfurization unit as a sample set;

[0006] Historical operating data of wet flue gas desulfurization units were exported from the historical database of the industrial control computer as a sample set.

[0007] S2: Preprocess the historical running data exported in step S1 and populate it with the historical running data;

[0008] Historical operational data is preprocessed using data cleaning methods. For noisy data, unreasonable data is smoothed out using the average sliding method, and historical operational data is filled in using the mean method.

[0009] S3: Perform feature processing on the historical running data after preprocessing in step S2;

[0010] The average values ​​of inlet SO2 concentration, inlet flue gas volume, inlet flue gas humidity, oxygen content, ammonia flow rate, density, pH value, outlet SO2 concentration, outlet flow rate, outlet oxygen content, and humidity in the first 30 minutes are used as the correlation features of ammonia change at that moment. The instantaneous value of outlet SO2 concentration is binned and the data is converted into categorical data.

[0011] S4: The data time periods are divided into 5 categories based on the actual operation of desulfurization;

[0012] The five data time periods are the backflushing stage, the testing equipment calibration stage, the stable operation stage, the operation fluctuation stage, and the low-control state at the outlet.

[0013] S5: Calculation of desulfurization process parameters;

[0014] The inlet SO2 mass during the stable and fluctuating operation phases is compared with the design value of the desulfurization equipment; the liquid-to-gas ratio and the ammonia-to-sulfur molar ratio are calculated as instantaneous operating parameters.

[0015] S6: Construct the network shape and derive the decision formula;

[0016] S7: Make a determination using the determination formula obtained in step S6;

[0017] The determination formula obtained in step S6 is used to determine whether the increase or decrease of ammonia water during the operation fluctuation stage is reasonable; unreasonable changes are eliminated, and the changes of ammonia water under different characteristics are obtained. The distribution of the changes of ammonia water is recorded as a set.

[0018] S8: Using the set in step S7 as the classification result, construct multiple deep network models - convolutional and recurrent neural networks, and train them to select the change in ammonia that minimizes T. After a period of time, the model is retrained to adapt to the new working conditions.

[0019] S9: Determine the final amount of ammonia to be added;

[0020] The average ammonia addition amount U under the stable operating conditions for the first 90 minutes is calculated as the addition amount under stable operating conditions to reduce outlet fluctuations and SO2 concentration fluctuations caused by its own disturbances. The final ammonia addition amount W = U + y is then calculated using the PID adaptive system data.

[0021] Preferably, in step S5, the formula for calculating the inlet SO2 mass during the stable operation phase and the fluctuating operation phase is: M 10 =Q 10 ×C 10 ×10 -6 And the design value of the desulfurization equipment: M 11 =Q 11 ×C 11 ×10 -6 Compare them and set their difference M. t =M 10 - M 11 As a factor affecting its fluctuation, the formula for calculating the liquid-to-gas ratio is: Liquid-to-gas ratio (L / G) = Ammonia water flow rate (m³ / s). 3 / h) / Gas quantity (Nm 3 / h); The formula for calculating the ammonia-sulfur molar ratio is: ammonia-sulfur molar ratio (V0×C2×ρ / 17) / ((Q 10 ×C10 - Q 20 ×C 20 )×10 -6 / 32).

[0022] Preferably, the determination formula in step S6 is T = a1×(C max -C set )-a2×(C min -C set )+a3×C x .

[0023] Preferably, the determination of the change in ammonia in step S8 is as follows:

[0024] S8-1: Construct a simple 3-layer neural network using PyTorch, where the input layer is the previously constructed feature matrix, and the middle hidden layer contains hidden neurons.

[0025] Get the Z score for each category. j ;

[0026] S8-2: Using the softmax function, the change in ammonia concentration is transformed into the probability of a category score. The output value is the one with the highest probability. The loss function is...

[0027] Where N is the number of samples, C is the number of categories, and y ij Let fj(z) represent the probability that the i-th sample belongs to class j.

[0028] S8-3; Using PyTorch, construct partial records and current positively correlated values ​​in a sequence deep network, recording the correlation between preceding and following data, f t =σ(W f *[h t-1 ,x t ] + b f ), D t =tanh(W d *[h t -1,x t ]+bd), C t =C t -1+D t *f t ,

[0029] o t = σ(Wo*[h t -1,x t ] + bo),h t =o t*tanh(C t ), and obtain Ct as the probability of each category;

[0030] S8-4: Calculate the average of the classification probabilities of the two models obtained in steps S8-2 and S8-3 to obtain the final class probability, using the formula: Y=(f j + D t ) / 2, the change in ammonia water y, is the final output value.

[0031] Preferably, in step S9, the average ammonia addition amount U is calculated using an autoregressive model, with the specific formula as follows:

[0032] Where p represents the order of collection, and represents the sequence of independent and identically distributed random variables.

[0033] Compared with the prior art, the beneficial effects of the present invention are: 1. It does not require prediction of SO2 concentration, but directly learns well-functioning strategies and accurately controls the outlet.

[0034] 2. Based on actual working conditions, the data is divided into different working conditions to make the data more interpretable. For situations that the mechanism model cannot explain, such as calibration (when calibrating sensors), the data is useless and cannot be used as samples for training. The control strategies under different working conditions are classified.

[0035] 3. Deep learning networks autonomously learn from historical data to obtain specific data, accurately add raw materials, and the PID control concept is robust, resulting in a more stable model and avoiding system instability. Attached Figure Description

[0036] Figure 1 This is a schematic diagram of the process of the present invention;

[0037] Figure 2 This is a schematic diagram of the soft max model;

[0038] Figure 3 This is a schematic diagram of the Ad Cell model;

[0039] Figure 4 Training loss iteration graph for the Ad Cell model;

[0040] Figure 5 This is a training iteration graph of the Soft max model.

[0041] Figure 6 This is a diagram illustrating the actual effect of an ammonia water regulating valve. Detailed Implementation

[0042] To provide a further understanding of the purpose, structure, features, and functions of the present invention, detailed descriptions are provided below with reference to specific embodiments.

[0043] Please refer to the reference. Figures 1 to 6 This invention provides a method for intelligently controlling the dosage of ammonia water as an industrial flue gas absorbent, characterized by the following steps:

[0044] S1: Export the historical operating data of the wet flue gas desulfurization unit as a sample set;

[0045] Historical operating data of wet flue gas desulfurization units were exported from the historical database of the industrial control computer as a sample set.

[0046] S2: Preprocess the historical running data exported in step S1 and populate it with the historical running data;

[0047] Historical operational data is preprocessed using data cleaning methods. For noisy data, unreasonable data is smoothed out using the average sliding method, and historical operational data is filled in using the mean method.

[0048] S3: Perform feature processing on the historical running data after preprocessing in step S2;

[0049] The average values ​​of inlet SO2 concentration, inlet flue gas volume, inlet flue gas humidity, oxygen content, ammonia flow rate, density, pH value, outlet SO2 concentration, outlet flow rate, outlet oxygen content, and humidity in the first 30 minutes are used as the correlation features of ammonia change at that moment. The instantaneous value of outlet SO2 concentration is binned and the data is converted into categorical data.

[0050] Furthermore, binning, as part of data preprocessing, is also known as discrete binning or data segmentation. Essentially, binning is simply grouping data.

[0051] Binning involves grouping data according to specific rules to discretize the data, enhance data stability, and reduce the risk of overfitting.

[0052] S4: The data time periods are divided into 5 categories based on the actual operation of desulfurization;

[0053] The five data time periods are the backflushing stage, the testing equipment calibration stage, the stable operation stage, the operation fluctuation stage, and the low-control state at the outlet.

[0054] S5: Calculation of desulfurization process parameters;

[0055] The inlet SO2 mass during the stable and fluctuating operation phases is compared with the design value of the desulfurization equipment; the liquid-to-gas ratio and the ammonia-to-sulfur molar ratio are calculated as instantaneous operating parameters.

[0056] S6: Construct the network shape and derive the decision formula;

[0057] Construct a network shape, take the increase or decrease of ammonia water amount y as the target value, and then use the maximum, minimum and variance of the outlet SO2 concentration in the next half hour as the basis for judging whether its increment is reasonable.

[0058] S7: Make a determination using the determination formula obtained in step S6;

[0059] The determination formula obtained in step S6 is used to determine whether the increase or decrease of ammonia water during the operation fluctuation stage is reasonable; unreasonable changes are eliminated, and the changes of ammonia water under different characteristics are obtained. The distribution of the changes of ammonia water is recorded as a set.

[0060] The unreasonable change refers to whether the outlet SO2 concentration is stable; whether the outlet SO2 concentration can be controlled as stably as possible when the inlet SO2 concentration fluctuates, based on the T value, whether it deviates too much;

[0061] S8: Using the set in step S7 as the classification result, construct multiple deep network models - convolutional and recurrent neural networks, and train them to select the change in ammonia that minimizes T. After a period of time, the model is retrained to adapt to the new working conditions.

[0062] S9: Determine the final amount of ammonia to be added;

[0063] The average ammonia addition amount U under the stable operating conditions for the first 90 minutes is calculated as the addition amount under stable operating conditions to reduce outlet fluctuations and SO2 concentration fluctuations caused by its own disturbances. The final ammonia addition amount W = U + y is then calculated using the PID adaptive system data.

[0064] Preferably, in step S2, the data cleaning method includes:

[0065] S2-1: Fill in missing data values: Use regression analysis, Bayesian calculation, or decision tree to infer the most likely value of a specific attribute for the record;

[0066] S2-2: Smoothing Noisy Data: Noise refers to a random error or variation in the measured variable. Data smoothing can be achieved using methods such as Bin, clustering, or regression.

[0067] S2-3: Identify and remove outlier data and resolve data inconsistencies: This can be done manually by leveraging the correlation between the data and external data sources. For example, errors in data entry can generally be corrected by comparing the data with the original document.

[0068] Furthermore, the cleaning methods for noisy data and noiseless data are the same.

[0069] Preferably, in step S5, the formula for calculating the inlet SO2 mass during the stable operation phase and the fluctuating operation phase is: M 10 =Q 10 ×C 10 ×10 -6 And the design value of the desulfurization equipment: M 11 =Q 11 ×C 11 ×10 -6 Compare them and set their difference M. t =M 10 - M 11 As a factor affecting its fluctuation, the formula for calculating the liquid-to-gas ratio is: Liquid-to-gas ratio (L / G) = Ammonia water flow rate (m³ / s). 3 / h) / Gas quantity (Nm 3 / h); The formula for calculating the ammonia-sulfur molar ratio is: ammonia-sulfur molar ratio (V0×C2×ρ / 17) / ((Q 10 ×C 10 - Q 20 ×C 20 )×10 -6 / 32).

[0070] Preferably, the determination formula in step S6 is T = a1×(C max -C set )-a2×(C min -C set )+a3×C x .

[0071] C max The maximum SO2 concentration, C min The minimum SO2 concentration, C set Set the SO2 concentration value, C x Let be the variance of SO2 concentration.

[0072] Preferably, the determination of the change in ammonia in step S8 is as follows:

[0073] S8-1: Construct a simple 3-layer neural network using PyTorch, where the input layer is the previously constructed feature matrix, and the middle hidden layer contains hidden neurons.

[0074] Get the Z score for each category. j ;

[0075] S8-2: Using the softmax function, the change in ammonia concentration is transformed into the probability of a category score. The output value is the one with the highest probability. The loss function is...

[0076] Where N is the number of samples, C is the number of categories, and y ij Let fj(z) represent the probability that the i-th sample belongs to class j.

[0077] S8-3; Using PyTorch, construct partial records and current positively correlated values ​​in a sequence deep network, recording the correlation between preceding and following data, f t =σ(W f *[h t-1 ,x t ] + b f ), D t =tanh(W d *[h t -1,x t ]+bd), C t =C t -1+D t *f t ,

[0078] o t = σ(Wo*[h t -1,x t ] + bo),h t =o t *tanh(C t ), and obtain Ct as the probability of each category.

[0079] Where: f t The parameters are those trained from the operating condition parameters in the previous time period; D t To record the impact of new input feature parameters on the model; C t The calculation uses the multiplication of parameters from the previous operating condition and the new operating condition, resulting in a value of 0 for most states, thus resolving the gradient explosion problem; t To calculate the information parameters that need to be recorded at time t, a partial weight is retained, with values ​​ranging from 0 to 1; h t is the output of the hidden layer at each time t; Wd, Wo, and Wf are the weight matrices between the previous hidden layer and the new input features; σ refers to the sigmoid function, with the formula: F(x)=1 / (1+e^-x) (-x is an exponent), which transforms the value into a value between 0 and 1; bd, bf, and bo are the bias parameters.

[0080] 32S8-4: Calculate the average of the classification probabilities of the two models obtained in steps S8- and S8- to obtain the final class probability, using the formula: Y=(f j + D t ) / 2, the change in ammonia water y, is the final output value.

[0081] Preferably, in step S9, the average ammonia addition amount U is calculated using an autoregressive model, with the specific formula as follows:

[0082] Where p represents the order of collection, and represents the sequence of independent and identically distributed random variables.

[0083] Figure 4 This is a training iteration graph of the AdCell model. Figure 5 This is the training iteration graph of the Softmax model. The deep model is trainable and has a high loss value, which are the evaluation criteria for the model's quality. When training the two models, the loss values ​​of the training set and the test set gradually decrease with each iteration, converge, and the loss values ​​are relatively small, indicating that the model performs well. Based on the trained model, the result of the ammonia water change y is calculated in real time. Through the flow-opening relationship graph of the valve, the opening of the ammonia water regulating valve is directly guided.

[0084] Figure 6 This refers to the actual effect of the ammonia regulating valve. The change in ammonia level, y, is calculated using a model. Based on the calculated value of a stable average ammonia level, the valve opening is guided accordingly. Figure 6 The model output shows that the amount of ammonia added can stably control the SO2 concentration at the outlet within a certain range, with a high average value and no exceedance.

[0085] The present invention has been described in the above-described embodiments; however, these embodiments are merely examples for implementing the present invention. It must be noted that the disclosed embodiments do not limit the scope of the present invention. Conversely, any modifications and refinements made without departing from the spirit and scope of the present invention are within the scope of patent protection of the present invention.

Claims

1. A method for intelligently controlling the dosage of ammonia water as an industrial flue gas absorbent, characterized in that: The method steps are as follows: S1: Export the historical operating data of the wet flue gas desulfurization unit as a sample set; Historical operating data of wet flue gas desulfurization units were exported from the historical database of the industrial control computer as a sample set. S2: Preprocess the historical running data exported in step S1 and populate it with the historical running data; Historical operational data is preprocessed using data cleaning methods. For noisy data, unreasonable data is smoothed out using the average sliding method, and historical operational data is filled in using the mean method. S3: Perform feature processing on the historical running data after preprocessing in step S2; The average values ​​of inlet SO2 concentration, inlet flue gas volume, inlet flue gas humidity, oxygen content, ammonia flow rate, density, pH value, outlet SO2 concentration, outlet flow rate, outlet oxygen content, and humidity in the first 30 minutes are used as the correlation features of ammonia change at that moment. The instantaneous value of outlet SO2 concentration is binned and the data is converted into categorical data. S4: The data time periods are divided into 5 categories based on the actual operation of desulfurization; The five data time periods are the backflushing stage, the testing equipment calibration stage, the stable operation stage, the operation fluctuation stage, and the low-control state at the outlet. S5: Calculation of desulfurization process parameters; The inlet SO2 mass during the stable operation phase and the fluctuating operation phase is calculated and compared with the design value of the desulfurization equipment; The liquid-to-gas ratio and the ammonia-to-sulfur molar ratio are calculated as instantaneous operating parameters. S6: Construct the network shape and derive the decision formula; ； C max The maximum SO2 concentration, C min The minimum SO2 concentration, C set Set the SO2 concentration value, C x The variance of SO2 concentration; S7: Make a determination using the determination formula obtained in step S6; The determination formula obtained in step S6 is used to determine whether the increase or decrease of ammonia water during the operation fluctuation stage is reasonable; unreasonable changes are eliminated, and the changes of ammonia water under different characteristics are obtained. The distribution of the changes of ammonia water is recorded as a set. S8: Using the set in step S7 as the classification result, construct multiple deep network models - convolutional and recurrent neural networks, and train them to select the change in ammonia that minimizes T. After a period of time, the model is retrained to adapt to the new working conditions. S9: Determine the final amount of ammonia to be added; The average ammonia addition amount U under the first 90 minutes of stable operation is calculated as the addition amount under stable operation to reduce outlet fluctuations and SO2 concentration fluctuations caused by its own disturbances. The final ammonia addition amount W = U + y is determined using PID adaptive system data, where y is the change in ammonia amount.

2. The method for intelligently controlling the dosage of ammonia water in industrial flue gas absorbent as described in claim 1, characterized in that: In step S5, the formula for calculating the inlet SO2 mass during the stable operation phase and the fluctuating operation phase is: M 10 =Q 10 ×C 10 ×10 -6 And the design value of the desulfurization equipment: M 11 =Q 11 ×C 11 ×10 -6 Compare them and set their difference M. t = M 10 - M 11 As a factor contributing to its fluctuations; The formula for calculating the liquid-to-gas ratio is: Liquid-to-gas ratio (L / G) = Ammonia water flow rate (m³ / s). 3 / h) / Gas quantity (Nm 3 / h); The formula for calculating the ammonia-sulfur molar ratio is: ammonia-sulfur molar ratio (V0×C2×ρ / 17) / ((Q 10 ×C 10 - Q 20 ×C 20 )×10 -6 / 32).

3. The method for intelligently controlling the dosage of ammonia water in industrial flue gas absorbent as described in claim 1, characterized in that: In step S9, the average ammonia addition amount U is calculated using an autoregressive model, and the specific formula is as follows: ; Where p represents the order of collection, and represents the sequence of independent and identically distributed random variables.

Images