A method and system for controlling the level of a sulfuric acid purification solution based on a dynamic neural network

Abstract

The application discloses a kind of dynamic neural network-based flue gas acidification purification liquid level control method and system, belong to industrial process control and intelligent control technical field.Firstly, power wave scrubber liquid level control process data is collected, the nonlinear liquid level model of power wave scrubber liquid level system is established, and the liquid level estimation value is obtained by online identification of liquid level system;According to the identification error between liquid level estimation value and actual liquid level output, the neural network parameters and weights are updated online;Then the updated identification model is introduced into the indirect adaptive controller, the set liquid level is calculated to calculate the control amount, and the pump and / or valve are driven to act, and the power wave scrubber liquid level is closed-loop regulated.The application can improve the modeling accuracy, control stability and anti-interference ability of power wave scrubber liquid level under complex working conditions.

Description

Technical Field

[0001] This invention relates to the field of industrial process control and intelligent control technology, and in particular to an adaptive control method and system for flue gas acid production and purification liquid level based on parallel combined dynamic neural networks for the liquid level regulation scenario of dynamic wave scrubber in flue gas acid production and purification section. Background Technology

[0002] In the smelting of non-ferrous metals such as copper, lead, and zinc, flue gas containing sulfur dioxide typically requires purification through an acid production process. This process generally includes purification, conversion, dry absorption, and waste acid treatment, with the purification section playing a crucial role in removing dust, droplets, and impurities. The dynamic wave scrubber, as a key piece of equipment in the purification section, directly affects the gas-liquid contact state, impurity removal efficiency, and the stability of subsequent conversion and absorption processes due to its internal purification liquid level.

[0003] When the liquid level in the dynamic wave scrubber is too high, bubble formation and dispersion are hindered, reducing gas-liquid contact efficiency and easily leading to a decrease in purification capacity. When the liquid level is too low, dust-laden flue gas may escape, resulting in insufficient scrubbing and further affecting the efficiency of subsequent catalytic conversion and the quality of sulfuric acid products. Therefore, continuous, stable, and high-precision control of the liquid level in the dynamic wave scrubber is a key issue in the flue gas sulfuric acid purification process.

[0004] In existing technologies, liquid level control in dynamic wave scrubbers mostly employs a combination of mechanistic analysis and PID control. While this approach is effective for linear or weakly nonlinear systems, the liquid level system of a dynamic wave scrubber is affected by factors such as flue gas load fluctuations, liquid inlet changes, backflow disturbances, and time-varying equipment parameters, exhibiting significant nonlinear, time-varying, and strongly coupled characteristics. Traditional models struggle to accurately characterize its dynamic behavior, and PID parameter tuning relies on field experience, making it difficult to maintain long-term stable control under complex operating conditions.

[0005] In recent years, data-driven modeling and advanced control methods have been gradually introduced into the field of industrial process control. Dynamic neural networks have strong nonlinear approximation capabilities and dynamic behavior expression capabilities, while adaptive control can adjust control parameters in real time according to changes in system state. Therefore, the combination of the two is expected to improve the accuracy and robustness of purified liquid level control. However, existing technologies still lack a complete technical solution that can simultaneously achieve online identification of multiple operating conditions, dynamic weight updates, and closed-loop adaptive control of liquid level. Summary of the Invention

[0006] The purpose of this invention is to provide a method and system for adaptive control of flue gas acid production and purification liquid level based on parallel combined dynamic neural network. By preprocessing the liquid level control process data, establishing a nonlinear liquid level model, constructing a parallel combined dynamic neural network identification model, performing online parameter / weight updates, and designing an indirect adaptive control law, stable, rapid, and high-precision control of the liquid level of the dynamic wave scrubber can be achieved.

[0007] To achieve the above objectives, the present invention adopts the following technical solution.

[0008] An adaptive control method for flue gas acid production purification liquid level based on parallel combined dynamic neural network includes the following steps:

[0009] S1, Collect process data of liquid level control of the power wave washer, the process data including at least liquid level sampling value, historical liquid level sequence and control input sequence related to liquid level adjustment;

[0010] S2, perform preprocessing on the collected data, the preprocessing including outlier removal, missing value interpolation and min-max normalization, to obtain normalized time series data;

[0011] S3, Based on the liquid level system of the dynamic wave scrubber, a nonlinear liquid level model is established, and the liquid level state variables and control inputs are mapped into the dynamic equations of the controlled object;

[0012] S4. A parallel combined dynamic neural network is used to identify the nonlinear liquid level model online to obtain the liquid level estimation output.

[0013] S5, based on the identification error between the liquid level estimation output and the actual liquid level output, the neural network parameters and weights are updated online;

[0014] S6 introduces the updated identification model into the indirect adaptive controller, calculates the control quantity based on the set liquid level, and applies it to the pump / valve actuator to adjust the liquid level of the power wave scrubber;

[0015] S7 transmits the latest liquid level feedback results back to the data acquisition and parameter update stage, forming a closed-loop iterative control.

[0016] Furthermore, in step 2, outlier removal is performed based on the three-standard-deviation principle, with the sample mean and standard deviation expressed as follows:

[0017]

[0018] When the observed value satisfies If the observed value is an outlier, it will be removed.

[0019] Furthermore, in step 2, missing values ​​are filled in using interpolation, which can be represented as:

[0020]

[0021] in, and The effective sampling time is adjacent to the missing point.

[0022] Furthermore, in step 2, the minimum-maximum normalization can be expressed as:

[0023]

[0024] By mapping the raw data to This interval improves the convergence speed of network training and enhances the consistency of data under different operating conditions.

[0025] Furthermore, in step 3, the nonlinear dynamic process of the liquid level system of the dynamic wave scrubber can be expressed as:

[0026]

[0027] in, Represents the liquid level state vector. Indicates control input, Represents the nonlinear vector field of the controlled object.

[0028] Furthermore, in step 4, the parallel combined dynamic neural network includes at least two parallel sub-networks, each of which approximates different local dynamic characteristics of the liquid level system. Its overall estimation model can be expressed as:

[0029]

[0030] in, For a stable matrix, , This represents the weight matrix from the input layer to the hidden layer. , This represents the weight matrix from the hidden layer to the output layer. and These represent the activation functions of each parallel subnetwork.

[0031] Furthermore, the activation function can be the Sigmoid function, which has the following form:

[0032]

[0033] in, Indicates the amplitude range parameter. Represents the slope parameter. This represents the vertical translation parameter.

[0034] Furthermore, in step 5, the identification error between the estimated liquid level output and the actual output can be expressed as:

[0035]

[0036] Furthermore, to ensure the stability of parameter updates, the Riccati equation can be introduced:

[0037]

[0038] in, The matrix is ​​a positive definite solution matrix. It is a positive definite matrix. It is a positive definite matrix.

[0039] Furthermore, in step 5, the online weight update law can take the following form:

[0040]

[0041]

[0042]

[0043]

[0044] in, ,and The gain matrix is ​​constructed, and the learning rate of the network can be adjusted by modifying the parameters. To achieve dynamic adjustment, Let P be a positive definite matrix, and let P represent the positive definite solution matrix of the above Riccati equation.

[0045] Furthermore, in step 6, an indirect adaptive control method is used to construct the controller. Let the reference state corresponding to the set liquid level be... Tracking error is defined as:

[0046]

[0047] The corresponding control law can be expressed as:

[0048]

[0049] Where the number of hidden layer nodes l in the network is greater than or equal to the system state dimension n, and the matrix If a row satisfies the full rank condition, then its Moore–Penrose generalized inverse exists.

[0050] Compared with the prior art, the present invention has at least the following beneficial effects:

[0051] 1. By removing outliers, filling in missing values, and normalizing the liquid level control process data, the quality of input samples was improved, providing a stable data foundation for subsequent identification and control;

[0052] 2. By constructing a parallel combined dynamic neural network, different sub-networks can approximate the nonlinear dynamic characteristics under different local working conditions, thereby improving the modeling capability of complex liquid level systems.

[0053] 3. Through the online parameter / weight update mechanism, the identification model can be adaptively adjusted according to changes in on-site working conditions, thereby enhancing the system's anti-interference capability and adaptability;

[0054] 4. By introducing the identification model into the indirect adaptive controller, online correction of control parameters and real-time optimization of control quantities are achieved, thereby improving the level tracking accuracy, dynamic response speed and steady-state control effect.

[0055] 5. This invention is applicable to the liquid level control scenario of the dynamic wave scrubber in the flue gas acid production purification section, and has good engineering application value. Attached Figure Description

[0056] Figure 1 A schematic diagram of the liquid level control device in a dynamic wave scrubber.

[0057] Figure 2 This is a flowchart of the adaptive liquid level control method of the present invention;

[0058] Figure 3 This is a schematic diagram of a parallel combinational dynamic neural network structure;

[0059] Figure 4 Flowchart for online weight update and parameter estimation;

[0060] Figure 5 This is a block diagram of an indirect adaptive control structure;

[0061] Figure 6 This is a block diagram of the overall system structure of the present invention. Detailed Implementation

[0062] The technical solution of the present invention will be further described in detail below with reference to the accompanying drawings. It should be understood that the following embodiments are only used to explain the present invention and are not intended to limit the scope of protection of the present invention.

[0063] Example 1: Liquid Level Control Object of Dynamic Wave Washer

[0064] like Figure 1As shown, the liquid level control objects of the dynamic wave scrubber include inlet pipe 1, large-diameter nozzle 2, absorption zone 3, external liquid inlet 4, purified flue gas outlet 5, dewatering layer 6, return liquid level zone 7, return channel 8, and circulation pump 9. The flue gas to be purified enters through inlet pipe 1, and undergoes preliminary washing by fully contacting the circulating liquid in the large-diameter nozzle 2 and absorption zone 3. The liquid-containing gas then enters the main cavity and passes through the dewatering layer 6 to remove droplets. The purified flue gas is then discharged from outlet 5. Return liquid accumulates in level zone 7 and, under the action of circulation pump 9, participates in circulation again through external liquid inlet 4 and return channel 8. The liquid level height in level zone 7 is the main control object of this invention.

[0065] Example 2: Flowchart of Adaptive Liquid Level Control Method

[0066] like Figure 2 As shown, the method of the present invention first performs data acquisition for the liquid level control process, and then sequentially performs outlier removal, missing value interpolation, and minimum-maximum normalization to obtain standardized data; on this basis, a nonlinear liquid level model of the dynamic wave scrubber is established; subsequently, a parallel combined dynamic neural network is used for online identification, and the network parameters are corrected through online weight update and parameter estimation; then, based on the updated identification model, an indirect adaptive control law is calculated, the control quantity is output and applied to the actuator to adjust the liquid level of the dynamic wave scrubber.

[0067] In actual operation, the data acquisition in step one can be completed according to a fixed sampling period; the preprocessing results in step two serve as the direct input for steps three and four; the estimated liquid level and identification error output in step four further serve as the basis for parameter updates in step five and the solution of the control law in step six, thus forming a continuous technical chain of "data preprocessing - model identification - parameter update - control output".

[0068] Example 3: Parallel Combinatorial Dynamic Neural Network Identification

[0069] like Figure 3 As shown, the parallel combined dynamic neural network consists of two or more parallel sub-networks. The upper branch sub-network can use Sigmoid activation units to approximate a class of local operating conditions, while the lower branch sub-network can use another set of activation units to approximate another class of local operating conditions. The outputs of each sub-network are weighted and synthesized by a summing node to form the liquid level estimate output. This structure can improve the model's ability to express the time-varying and nonlinear characteristics of the liquid level in a dynamic wave scrubber.

[0070] Furthermore, the network receives normalized input data and normalized output data at each sampling time to generate an estimated output. The estimated output is consistent with the actual liquid level. The difference constitutes the identification error. This will serve as an important basis for the next online weight update.

[0071] Example 4: Online Weight Update and Parameter Estimation

[0072] like Figure 4 As shown, in the online weight update and parameter estimation process, normalized input data and normalized output data are fed into the dynamic neural network identifier, and the identifier outputs the estimated value. Actual output The error is fed into the error calculation node and compared with the estimated value to obtain the identification error. The identification error is then input into the online weight update and parameter estimation module to obtain the updated parameters / weights. This information is then fed back to the dynamic neural network identifier, which continues to perform identification at the next sampling time.

[0073] In this implementation process, parameter updates can be achieved through gradient descent, least squares, or other adaptive algorithms that meet stability requirements. As long as the network parameters can be corrected online based on the current identification error, they can fall within the protection scope of this invention.

[0074] Example 5: Indirect Adaptive Control

[0075] like Figure 5 As shown, this invention employs an indirect adaptive control structure. Controller Receive set liquid level input Based on the model parameter estimates provided by the online estimation module or target parameter Calculate control parameters Then the controller outputs the control quantity. Acting on the controlled object The controlled object outputs liquid level. Then, on the one hand, feedback is sent to the online estimation module to perform model correction, and on the other hand, feedback is sent to the controller to form a control closed loop.

[0076] In this invention, indirect adaptive control is chosen instead of direct adaptive control, mainly because the liquid level system of a dynamic wave scrubber is difficult to accurately represent over a long period of time using a fixed parameter model. Indirect adaptive control can estimate the object parameters first and then update the controller parameters accordingly, making it more suitable for time-varying working conditions in industrial settings.

[0077] Example 6: System Overall Structure

[0078] like Figure 6As shown, the system of the present invention may include a data acquisition unit, a data preprocessing unit, a nonlinear liquid level model building unit, a parallel combined dynamic neural network identification unit, an online parameter / weight update unit, an indirect adaptive control unit, an actuator unit, a liquid level control object for a dynamic wave scrubber, and a liquid level detection / feedback unit. Specifically, the data acquisition unit acquires liquid level-related process data; the data preprocessing unit performs outlier removal, missing value interpolation, and normalization; the nonlinear liquid level model building unit constructs a dynamic model of the controlled object; the identification unit and the parameter update unit form an online learning closed loop; the indirect adaptive control unit calculates the control quantity based on the set liquid level and feedback results; the actuator unit controls the pump / valve action based on the control quantity; and the liquid level detection / feedback unit transmits the detection results back to the data acquisition unit and the control unit.

[0079] Example 7: Complete Control Process

[0080] After the system is put into operation, the current liquid level and historical liquid level sequence of the dynamic wave scrubber are first acquired through the data acquisition unit; then, the data preprocessing unit removes abnormal sampled values, fills in missing sampled points, and completes normalization; next, the nonlinear liquid level model building unit and the parallel combined dynamic neural network identification unit are used to construct the liquid level estimation model under the current operating conditions; then, the online parameter / weight update unit continuously corrects the network parameters according to the identification error; based on this, the indirect adaptive control unit calculates the control quantity by combining the set liquid level, the current estimation model, and the feedback liquid level. Ultimately, the actuator unit drives the circulating pump or valve to regulate the circulating fluid flow rate, replenishment volume, or discharge volume, thereby achieving closed-loop stable control of the liquid level.

[0081] Through the above embodiments, the present invention can maintain high liquid level tracking accuracy and good control stability under industrial site conditions where the purified liquid level fluctuates greatly, the operating conditions change frequently, or the disturbance is obvious.

Claims

1. A method for controlling the liquid level in flue gas acid production and purification based on a dynamic neural network, characterized in that, Includes the following steps: Step (1): Collect liquid level control process data of the power wave washer. The liquid level control process data includes at least the liquid level sampling value, historical liquid level sequence and control input sequence related to liquid level adjustment. Step (2) Perform data preprocessing on the liquid level control process data. The data preprocessing includes outlier removal, missing value interpolation, and min-max normalization in sequence to obtain normalized time series data. Step (3): Based on the normalized time series data, establish a nonlinear liquid level model of the power wave washer liquid level system, and map the liquid level state variables and control inputs into the dynamic equations of the controlled object; Step (4): Based on the nonlinear liquid level model, a parallel combined dynamic neural network is constructed to identify the liquid level system of the power wave washer online and output the liquid level estimate. Step (5): Compare the estimated liquid level with the actual liquid level output to obtain the identification error, and update the parameters and weights of the parallel combined dynamic neural network online based on the identification error; Step (6): The online updated parallel combined dynamic neural network is introduced into the indirect adaptive controller as the controlled object model, and the control quantity is calculated in combination with the set liquid level. Step (7) outputs the control quantity to the actuator to control the operation of the circulating pump and / or valve, thereby adjusting the liquid level of the power wave scrubber, and sends the latest liquid level feedback result back to steps (1) and (5) to form a closed-loop iterative control.

2. The method for controlling the liquid level of flue gas acidification purification based on a dynamic neural network according to claim 1, characterized in that, Outlier removal in step (2) is based on the three-standard-deviation principle. First, the sample mean is calculated. and sample standard deviation When the observed value satisfies If the observed value is an outlier, it will be removed.

3. The method for controlling the liquid level of flue gas acidification purification based on a dynamic neural network according to claim 2, characterized in that, In step (2), missing value interpolation is performed by interpolating the missing sample points with adjacent valid sample points. In step (2), min-max normalization is performed by mapping the original data to... The interval is used to improve the convergence speed of model training and enhance data consistency under different operating conditions.

4. The method for controlling the liquid level of flue gas acid production purification based on a dynamic neural network according to claim 3, characterized in that, The nonlinear liquid level model in step (3) is expressed as follows: ,in Represents the liquid level state vector. Indicates control input, This represents the nonlinear vector field of the liquid level system in a dynamic wave scrubber.

5. The method for controlling the liquid level of flue gas acid production purification based on a dynamic neural network according to claim 4, characterized in that, The parallel combined dynamic neural network in step (4) includes at least two parallel sub-networks. Each parallel sub-network approximates different local dynamic characteristics of the liquid level system of the dynamic wave scrubber. The outputs of each parallel sub-network are combined to form the liquid level estimate. The estimation model of the parallel combined dynamic neural network satisfies: ； in, For a stable matrix, , This is the weight matrix from the input layer to the hidden layer. , This is the weight matrix from the hidden layer to the output layer. and This represents the sigmoid activation function.

6. The method for controlling the liquid level of flue gas acidification purification based on a dynamic neural network according to claim 5, characterized in that, The identification error in step (5) is expressed as: To ensure the stability of online updates, the Riccati equation is introduced. ； in, The matrix is ​​a positive definite solution matrix. It is a positive definite matrix. It is a positive definite matrix.

7. The method for controlling the liquid level of flue gas acid production purification based on a dynamic neural network according to claim 6, characterized in that, The online parameter / weight update in step (5) includes weight updates between the hidden layer and the output layer, as well as weight updates between the input layer and the hidden layer. The online weight update law includes: ； ； ； ； Among them, the identification error is , ,and The gain matrix is ​​constructed, and the learning rate of the network can be adjusted by modifying the parameters. To achieve dynamic adjustment, Let P be a positive definite matrix, and let P represent the positive definite solution matrix of the Riccati equation.

8. The method for controlling the liquid level of flue gas acid production purification based on a dynamic neural network according to claim 7, characterized in that, In step (6), the indirect adaptive controller is based on the reference state corresponding to the set liquid level. and current liquid level status Calculate tracking error The control quantity is then obtained according to the following control law: ； Where the number of hidden layer nodes l in the network is greater than or equal to the system state dimension n, and the matrix If a row satisfies the full rank condition, then its Moore–Penrose generalized inverse exists.

9. A flue gas acidification purification liquid level control method system based on a dynamic neural network for implementing the method as described in any one of claims 1-8, characterized in that, include: The data acquisition unit is used to collect data on the liquid level control process of the dynamic wave scrubber. The data preprocessing unit is used to perform outlier removal, missing value interpolation, and min-max normalization on the liquid level control process data. The nonlinear liquid level model establishment unit is used to establish a nonlinear liquid level model of the power wave scrubber liquid level system. A parallel combined dynamic neural network identification unit is used to perform online identification of the liquid level system of the power wave washer based on the nonlinear liquid level model to obtain the liquid level estimate. An online parameter / weight update unit is used to update the parameters and weights of the parallel combined dynamic neural network online based on the identification error between the liquid level estimate and the actual liquid level output. An indirect adaptive control unit is used to introduce the updated parallel combined dynamic neural network as the controlled object model into the controller, and calculate the control quantity in combination with the set liquid level. An actuator unit is used to control the operation of the circulating pump and / or valves according to the control quantity to adjust the liquid level of the power wave scrubber; The liquid level detection / feedback unit is used to collect the latest liquid level feedback results and transmit the feedback results back to the data acquisition unit, the online parameter / weight update unit, and the indirect adaptive control unit.

10. The system according to claim 9, characterized in that, The parallel combined dynamic neural network identification unit includes at least two parallel sub-networks, each corresponding to the dynamic identification of liquid level under different local working conditions; the actuator unit includes one or more of a circulating pump and a valve; and the liquid level detection / feedback unit includes a liquid level sensor or a liquid level transmitter.

Images