Model prediction-based ammonium sulfate production process control method

Abstract

The invention provides an ammonium sulfate production process control method based on model prediction, which comprises the following steps: acquiring real-time data of flue gas flow, sulfur dioxide concentration, slurry PH value and ammonia water concentration, analyzing dynamic change characteristics of the flue gas flow, the sulfur dioxide concentration, the slurry PH value and the ammonia water concentration under different coal types and load working conditions, and constructing a flue gas component mutation characteristic library; transmitting an ammonia water adding amount and / or material transfer pump rotating speed control instruction to a desulfurization tower control system, executing dynamic adjustment of an absorption circulating tank, updating real-time monitoring data, comparing output prediction results, calculating deviation, and generating an optimized future prediction sequence; and judging whether the pH value and the ammonium sulfate concentration of the absorption circulating tank can continuously meet the preset upper and lower limit range conditions in the future preset time or not by comparing the prediction result sequences before and after optimization, if so, completing a dynamic adjustment period, and if not, recalculating the adjustment value of the ammonia water addition amount and / or the rotating speed of the material transfer pump.

Description

Technical Field

[0001] The present invention relates to the field of information technology, and particularly to a control method for ammonium sulfate production process based on model prediction. Background Art

[0002] Under the background of the global energy structure transformation and increasingly strict environmental protection requirements, the efficiency and stability of the flue gas desulfurization system in coal-fired power plants are crucial. Sulfur oxides (SOx) are one of the main air pollutants, and efficient desulfurization is a key link in realizing clean energy production. However, the desulfurization system in thermal power plants is a complex dynamic process, and there is a strong non-linear coupling relationship among the flue gas flow rate, sulfur dioxide concentration, slurry pH value, and ammonia water concentration. These factors jointly affect the desulfurization efficiency and the quality of the by-product ammonium sulfate. Specifically, the fluctuations in the flue gas flow rate and sulfur dioxide concentration directly affect the pH value of the slurry in the absorption circulation tank, and the pH value is the key to controlling the ammonia water addition amount to maintain the desulfurization reaction balance. At the same time, the ammonia water concentration and the slurry pH value jointly determine the production rate and crystallization state of ammonium sulfate. Therefore, to achieve precise control of the desulfurization system, it is necessary to deeply understand and quantify the dynamic coupling relationship among these four key parameters. Existing technologies usually adopt control strategies based on empirical formulas or simple models, which are difficult to adapt to the rapid changes in flue gas composition and operating conditions. First, it is a great challenge to construct a multi-dimensional feature library that can comprehensively reflect the mutation characteristics of flue gas composition, which requires in-depth analysis of historical data under different coal types and load conditions. Second, how to use non-linear regression algorithms to establish a multi-variable model that can accurately describe the dynamic coupling relationship among the flue gas flow rate, sulfur dioxide concentration, slurry pH value, and ammonia water concentration is also a technical problem to be solved urgently. This model not only needs to consider the correlation among the parameters, but also can reflect the data change characteristics under different operating conditions. In addition, combining the historical data of the pH value and ammonium sulfate concentration in the absorption circulation tank to determine the input variables of the model and construct an effective training data set is crucial for achieving accurate prediction. Finally, based on this model, how to accurately predict the change trends of the pH value and ammonium sulfate concentration in the absorption circulation tank in the future period, and dynamically adjust the ammonia water addition amount and the rotation speed of the transfer pump according to the prediction results to ensure the continuous and stable operation of the system, and at the same time enable the model to have the ability of adaptive learning, continuously optimize the parameters through real-time data, and improve the prediction accuracy, is the core technical challenge for realizing intelligent and efficient desulfurization control. These technological advancements will provide strong support for environmental protection and clean energy production. Summary of the Invention

[0003] The present invention provides a control method for ammonium sulfate production process based on model prediction, mainly including: Obtain real-time data of flue gas flow rate, sulfur dioxide concentration, slurry pH value, and ammonia water concentration. Analyze the dynamic change characteristics of flue gas flow rate, sulfur dioxide concentration, slurry pH value, and ammonia water concentration under different coal types and load conditions, and construct a mutation feature library of flue gas components; According to the mutation feature library of flue gas components, use the non-linear regression algorithm to analyze the correlation between flue gas flow rate and sulfur dioxide concentration, the correlation between flue gas flow rate and slurry pH value, the correlation between sulfur dioxide concentration and slurry pH value, and the correlation between ammonia water concentration and slurry pH value. Conduct multivariate dynamic coupling analysis of the four monitoring indicators of flue gas flow rate, sulfur dioxide concentration, slurry pH value, and ammonia water concentration, and output the change trends of flue gas flow rate, sulfur dioxide concentration, slurry pH value, and ammonia water concentration in a future period predicted based on the current monitoring indicators; Within the preset prediction time window, based on the flue gas flow rate, sulfur dioxide concentration, slurry pH value, and ammonia water concentration in the current and preset time periods, predict the change trends of the pH value and ammonium sulfate concentration in the future absorption circulation tank, generate a prediction sequence, and generate a control instruction for the ammonia water addition amount and / or a control instruction for the transfer pump speed according to the prediction sequence; Transmit the control instruction for the ammonia water addition amount and / or the transfer pump speed to the desulfurization tower control system, execute the dynamic adjustment of the absorption circulation tank, and at the same time update the real-time monitoring data, compare the output prediction results, calculate the deviation, and generate an optimized future prediction sequence; By comparing the prediction result sequences before and after optimization, judge whether the pH value and ammonium sulfate concentration in the absorption circulation tank can continuously meet the preset upper and lower limit range conditions within the future preset time. If it is satisfied, complete a dynamic adjustment cycle; if not, recalculate the adjustment values of the ammonia water addition amount and / or the transfer pump speed.

[0004] Further, obtain the real-time data of flue gas flow rate, sulfur dioxide concentration, slurry pH value, and ammonia water concentration. Analyze the dynamic change characteristics of flue gas flow rate, sulfur dioxide concentration, slurry pH value, and ammonia water concentration under different coal types and load conditions, and construct a flue gas composition mutation feature library, including: for coal type identification, obtain flue gas concentration data from the flue gas emission port and coal quality data from the coal quality data acquisition device. Use a multi-layer sampler to sample the flue gas concentration data once per second and the coal quality data once per hour. Establish a corresponding relationship dataset between flue gas concentration and coal quality data through a regression algorithm. Generate standard condition reference data based on the corresponding relationship dataset between flue gas concentration and coal quality data. Use a multi-point sensor array to real-time collect flue gas flow rate data, sulfur dioxide concentration data, ammonia water concentration data, and slurry pH value data, and compare with the standard condition reference data to generate real-time condition deviation data. Obtain real-time flue gas flow velocity data through a flue gas flow velocity monitoring device, collect unit load data from the operating conditions, and establish a curve of flue gas composition and load change according to the real-time condition deviation data, and calculate the flue gas composition change trend data. Set a flue gas concentration mutation threshold according to the flue gas composition change trend data. When the mutation threshold is triggered, record the ammonia water concentration data and slurry pH value data, and use a recursive neural network to generate optimized desulfurization process parameter data. Use a multi-dimensional data analysis method to classify the optimized desulfurization process parameter data, establish a flue gas composition mutation feature database under different coal types and different load intervals, and record the process parameter data at the moment of mutation. Calculate the adjustment parameters of ammonia water concentration and slurry pH value according to the flue gas composition mutation feature database, and establish a process parameter adjustment rule dataset through a data hierarchical storage method.

[0005] Furthermore, according to the flue gas composition mutation feature library, a non-linear regression algorithm is used to analyze the correlations between flue gas flow rate and sulfur dioxide concentration, between flue gas flow rate and slurry pH value, between sulfur dioxide concentration and slurry pH value, and between ammonia water concentration and slurry pH value, and to conduct a multi-variable dynamic coupling analysis of four monitoring indicators, namely flue gas flow rate, sulfur dioxide concentration, slurry pH value, and ammonia water concentration, and output the changing trends of flue gas flow rate, sulfur dioxide concentration, slurry pH value, and ammonia water concentration within a certain period of time in the future predicted based on the current monitoring indicators, including: obtaining historical sampling data from the flue gas composition mutation feature database, converting the four indicators of flue gas flow rate, sulfur dioxide concentration, slurry pH value, and ammonia water concentration to the range from zero to one through maximum-minimum normalization processing, and calculating the initial correlation degree data between the four indicators through the Pearson correlation coefficient. Setting a 10-minute sliding time window according to the initial correlation degree data, with the window sliding interval set to 1 minute, and calculating the quadratic polynomial fitting functions for pairwise pairing of the four indicators using the least squares method within each time window to generate the data on the mutual influence strength between the indicators. Establishing a fourth-order state variable equation set based on the data on the mutual influence strength between the indicators, where each variable in the equation set represents a monitoring indicator and the variable coefficient matrix represents the influence weights between the indicators, and solving the parameter values of the equation set through the recursive least squares method. Constructing a four-dimensional state space using the solved parameter values of the equation set, calculating the dynamic coupling trajectories of the four indicators in the state space, and generating a set of non-linear mapping functions between the indicators. Establishing a dynamic coupling response curve based on the set of non-linear mapping functions, where the abscissa in the curve represents the time series and the ordinate represents the normalized values of the four indicators, and calculating the changing trends of the indicators through the curve slope. Using the Kalman filter to predict the values of the four indicators within the next 30 minutes based on the dynamic coupling response curve, calculating the root mean square error between the predicted values and the measured values through a sliding prediction window, and generating the prediction accuracy evaluation data. Using the prediction accuracy evaluation data to iteratively optimize the parameters of the Kalman filter and constructing a predicted value correction function to achieve the dynamic prediction of the changing trends of the four indicators within the future time window.

[0006] Furthermore, extract the data during the period of sudden change in flue gas flow rate, match it with the feature vectors in the flue gas composition mutation feature library to generate a dynamic weight matrix, adjust the input parameters of the non-linear regression algorithm, calculate the response lag coefficient of the slurry pH value, and perform an associated mapping with the ammonia water concentration adjustment rate to generate the dynamic association parameters between the flue gas flow rate and the sulfur dioxide concentration. Input these parameters into the multi-variable dynamic coupling model to output the predicted trend data of the future flue gas flow rate, sulfur dioxide concentration, slurry pH value, and ammonia water concentration, including: obtain the flue gas flow rate data from the on-line flue gas monitoring device at a sampling interval of 10 seconds, calculate the change rate of the flue gas flow rate through a 5-minute sliding time window. If the change rate exceeds the preset upper and lower limit intervals of the reference value, extract the four-dimensional feature vectors at the corresponding moment from the flue gas composition mutation feature library. The four-dimensional feature vectors include the flue gas flow rate, sulfur dioxide concentration, slurry pH value, and ammonia water concentration. Establish a 4×4 dynamic weight matrix based on the four-dimensional feature vectors. Each row of the matrix represents the influence weight of a single index on other indexes. Collect the slurry pH value data from the pH value on-line monitoring device at an interval of 5 seconds, and use the non-linear regression algorithm to calculate the time displacement between the pH value curve and the flue gas flow rate curve to obtain the response lag coefficient of the pH value. Obtain the ammonia water concentration data at an interval of 30 seconds from the ammonia water concentration automatic detection device according to the response lag coefficient of the pH value, calculate the change amount of the ammonia water concentration per unit time to obtain the adjustment rate curve, and use the feedforward neural network to establish the dynamic mapping function between the flue gas flow rate and the sulfur dioxide concentration. Input the dynamic mapping function into the non-linear regressor to generate the dynamic association parameters between the flue gas flow rate and the sulfur dioxide concentration. The dynamic association parameters characterize the change response relationship between the two. Construct the flue gas flow rate change curve within a 30-minute prediction time window based on the dynamic weight matrix, the response lag coefficient of the pH value, and the dynamic association parameters, and calculate the predicted trend curves of the three indexes of sulfur dioxide concentration, slurry pH value, and ammonia water concentration through matrix mapping. The time resolution of the predicted trend curves is 1 minute.

[0007] Further, obtain the historical change data of the pH value and the historical change data of the ammonium sulfate concentration in the absorption circulation tank, analyze the correlation between the pH value and the ammonium sulfate concentration in the absorption circulation tank, and form a model training data set to optimize the multivariable dynamic coupling model, including: obtaining the historical pH value data from the online pH value monitoring device of the absorption circulation tank at 5-second intervals, screening the historical data using a 30-minute sliding time window, setting the sliding interval to 1 minute, filtering the noise signal in the monitoring data through a 5-point median filter, and generating a time series change curve of the pH value. Collecting the historical concentration data from the online ammonium sulfate analyzer at 1-minute intervals, downsampling the pH value time series change curve according to the data sampling timestamp, and supplementing the missing points in the ammonium sulfate concentration data using the three-point Lagrange interpolation method. Calculating the Pearson correlation coefficient between the pH value and the ammonium sulfate concentration using a 15-minute calculation window based on the time-aligned data, generating a time series curve of the correlation coefficient through the sliding calculation window, and establishing a mapping function between the pH value and the ammonium sulfate concentration using the support vector regression algorithm. Setting a threshold for the change rate of the correlation coefficient according to the time series curve of the correlation coefficient, extracting the time points of the correlation coefficient mutation, and recording the pH value data and the ammonium sulfate concentration data for 15 minutes before and after the mutation point. Using a deep neural network to extract features from the data before and after the mutation point, constructing a dynamic response feature vector of the pH value and the ammonium sulfate concentration, and generating an optimized training data set for the multivariable dynamic coupling model. Adjusting the coupling coefficient between the pH value and the ammonium sulfate concentration in the multivariable dynamic coupling model according to the optimized training data set, and iteratively optimizing the model parameters using the backpropagation algorithm.

[0008] Further, within a preset prediction time window, based on the flue gas flow rate, sulfur dioxide concentration, slurry pH value, and ammonia water concentration in the current and preset time periods, predict the future change trends of the pH value and ammonium sulfate concentration in the absorption circulation tank, and generate a prediction sequence. Generate an ammonia water addition amount control instruction and / or a transfer pump speed control instruction according to the prediction sequence, including: collect four real-time data of flue gas flow rate, sulfur dioxide concentration, slurry pH value, and ammonia water concentration at 10-second intervals from a multi-point sensor array arranged at the inlet, middle, and outlet of the absorption tower, and calculate the change trend data of the pH value and ammonium sulfate concentration in the absorption circulation tank within a 30-minute prediction time window through a multi-variable dynamic coupling model. Calculate the time displacement of the pH value and ammonium sulfate concentration using the support vector regression algorithm based on the change trend data, and update the prediction sequence in real time through a 5-minute sliding prediction window, and calculate the maximum fluctuation amplitude of the pH value and ammonium sulfate concentration. Establish a mapping function between the pH value and the ammonia water addition amount based on the maximum fluctuation amplitude, set a stable control threshold for the pH value within the range of 5.5 to 6.5, and predict the ammonia water addition amount adjustment parameter using a recursive neural network. Calculate the change amount of the transfer pump flow rate per unit time according to the ammonia water addition amount adjustment parameter, and generate the corresponding relationship parameter between the reference speed and the adjusted speed of the transfer pump using a proportional-integral algorithm. Calculate the target speed of the transfer pump according to the corresponding relationship parameter, generate a speed control instruction within the preset minimum speed and maximum speed range, and transmit it to the on-site controller through an industrial Ethernet bus. Collect the deviation data between the actual speed and the set speed of the transfer pump from the on-site controller, dynamically correct the speed control instruction using a feedback correction algorithm, and establish a speed control protection limit under abnormal conditions.

[0009] Further, if the predicted sequence shows that the pH value of the future absorption circulation tank will be lower than the preset lower threshold, calculate the difference from the lower threshold to obtain the predicted deviation of the pH value. Considering the influence of the sulfite / sulfate ratio in the slurry on the pH buffering capacity, calculate the additional ammonia water addition required, and generate a control instruction to increase the ammonia water addition, including: extracting the predicted pH value data of the absorption circulation tank at 1-minute intervals from the predicted sequence, calculating the difference from the preset lower pH threshold of 5.5 through a comparison operator, smoothing the difference sequence using a Kalman filter with a measurement noise variance of 0.1 to obtain the predicted pH deviation sequence. According to the predicted pH deviation sequence, call the variable coefficient matrix representing the correlation between ammonia water concentration, flue gas flow rate, sulfur dioxide concentration, and slurry pH value in the 4×4 multi-variable dynamic coupling model, and use the support vector regression algorithm to calculate the influence weights of the flue gas flow rate change and sulfur dioxide concentration change on the pH value change. Obtain the sulfite and sulfate concentration data from the desulfurization slurry sampling device every 5 minutes, generate a buffering capacity index reflecting the acid-base neutralization capacity of the slurry through a recursive calculator, and use a non-linear mapping function to calculate the ammonia water supplement required for a unit pH value change. Calculate the theoretical supplement amount based on the product of the unit supplement amount and the predicted pH deviation, generate a theoretical supplement amount correction parameter in combination with the current flue gas conditions, and calculate the total actual ammonia water supplement required. Read the addition amount reference curve corresponding to different pH value ranges from the pre-established ammonia water addition amount database, generate an addition amount increment value based on the total actual supplement amount, and generate an ammonia water addition increase instruction through a digital signal processor. Use a feedback checker to online monitor the pH value change trend after the execution of the addition amount increase instruction, and trigger the adaptive adjustment mechanism of the addition amount control parameter when the pH value recovery rate is lower than the expected value.

[0010] Further, if the predicted sequence shows that the ammonium sulfate concentration in the future absorption circulation tank will be higher than the preset upper threshold, calculate the difference between the predicted value and the preset upper threshold to obtain the predicted deviation of the ammonium sulfate concentration. Combine the influence of the ammonium sulfate crystal particle size distribution on supersaturation, calculate the rotational speed of the transfer pump that needs to be increased, and generate a control instruction to increase the rotational speed of the transfer pump, including: Extract the predicted data of the ammonium sulfate concentration in the circulation tank for the next 30 minutes from the predicted sequence at 1-minute intervals, use the support vector regression algorithm to predict the concentration change trend, and calculate the difference from 30% of the preset upper threshold of the ammonium sulfate concentration through a comparison operator to obtain the predicted deviation of the ammonium sulfate concentration. Extract the three historical data of the transfer pump rotational speed, slurry density, and slurry temperature in the last 24 hours from the historical operation database, use the long short-term memory network to establish a 4×4-dimensional parameter influence weight matrix, and calculate the influence coefficients of the three parameters on the ammonium sulfate concentration change rate. Collect the ammonium sulfate crystal particle size distribution data from the particle size detector every 5 minutes, measure the crystal particle size by the laser diffraction method, and use a recursive calculator to generate the average crystal particle size and standard deviation. Calculate the solution supersaturation index based on the average crystal particle size and standard deviation, and the supersaturation index characterizes the dissolution equilibrium state of the crystal in the slurry. Calculate the reference rotational speed of the transfer pump within the range of 900 to 1500 revolutions per minute according to the supersaturation index and the parameter influence weight matrix, and use the proportional integral algorithm with an integral time of 60 seconds to generate the rotational speed adjustment amount. Send a rotational speed increase instruction to the transfer pump controller through a digital signal processor, and use an online feedback checker to monitor the rotational speed response curve in real time. When the rotational speed adjustment shows overshoot or lag, trigger the adaptive correction mechanism.

[0011] Further, the ammonia water addition amount and / or the control instruction of the transfer pump speed are transmitted to the desulfurization tower control system to perform dynamic adjustment of the absorption circulation tank. Meanwhile, the real-time monitoring data is updated, the predicted results of the comparison output are compared, the deviation is calculated, and an optimized future prediction sequence is generated, including: encoding and packing the ammonia water addition amount control instruction and the transfer pump speed control instruction by adopting the Modbus-TCP communication protocol through the industrial Ethernet bus, and transmitting them to the desulfurization tower controller in the 16-bit integer format, and monitoring the signal transmission status of the acquisition devices arranged at the inlet, middle part and outlet of the absorption tower. Judging the execution situation of the control instruction according to the execution status data returned by the acquisition device, collecting the pH value, ammonium sulfate concentration, slurry density and temperature data in the circulation tank at 5-second intervals through the sensor array, and comparing the multi-variable dynamic coupling prediction sequence to calculate the deviation value. Performing gradient descent operation on the predicted deviation value by using an adaptive calculator with an initial learning rate of 0.01, and stopping the optimization when the deviation value of 5 consecutive iterations is less than the preset threshold, and generating optimized weight parameters. Establishing a 4×4 dimensional model parameter matrix according to the optimized weight parameters, and online updating the parameter matrix by using a recurrent neural network to generate corrected coupling parameter data. Calculating the prediction sequence for the next 30 minutes by using a time window with a length of 10 minutes and a sliding interval of 1 minute, and judging the accuracy of the prediction sequence through the root mean square error index. Dynamically adjusting the length of the prediction time window according to the root mean square error index, and starting the local cache mechanism to save historical data when the network communication is interrupted, and performing data synchronization after the communication is restored.

[0012] Further, compare the adjusted absorption efficiency and the pressure fluctuation value in the tank with the predicted result sequence to generate an error distribution matrix. Iteratively optimize the weight parameters of the dynamic coupling relationship in the multivariable dynamic coupling model based on the error distribution matrix. Generate the predicted result sequence for the next cycle according to the optimized weight parameters of the dynamic coupling relationship. The predicted result sequence includes the ammonia water demand trend and the speed matching interval at multiple future time points, including: obtaining the absorption efficiency data and the pressure fluctuation data in the tank from the on-line acquisition device at 10-second intervals, preprocessing the data using a Kalman filter with a measurement noise variance of 0.1, calculating the error sequence between the predicted value and the measured value through mathematical statistics methods, and generating a 4×4 error distribution matrix. Iteratively optimize the weight parameters in the multivariable dynamic coupling model according to the error distribution matrix using the gradient descent algorithm. Stop the optimization when the error value of 5 consecutive iterations is less than 0.05, and generate the optimized weight parameter set. Construct a 4×4 dynamic coupling matrix according to the optimized weight parameter set in four dimensions of absorption efficiency, pressure fluctuation, ammonia water concentration, and speed. Calculate the coupling coefficient between parameters through a recurrent neural network. Dynamically update the coupling coefficient using a 10-minute sliding time window to generate the response relationship curve between the ammonia water concentration and the transfer pump speed, and calculate the response time constant. Calculate the ammonia water demand trend within the next 30 minutes using an adaptive predictor according to the response time constant, and generate the speed matching data within the range of 900 to 1500 revolutions per minute of the transfer pump. Record the prediction sequence data at 1-minute intervals through a data hierarchical memory, evaluate the prediction accuracy using the root mean square error index, and mark the data with a predicted value deviation exceeding the preset threshold.

[0013] Further, by comparing the predicted result sequences before and after optimization, it is determined whether the pH value and ammonium sulfate concentration in the absorption circulation tank can continuously meet the preset upper and lower limit range conditions within a preset future time. If they meet, a dynamic adjustment cycle is completed. If not, the adjustment values of the ammonia addition amount and / or the transfer pump speed are recalculated, including: extracting the predicted sequence data before and after optimization from the prediction result memory at 1-minute intervals, calculating the root mean square error of the two predicted sequences using a 5-minute sliding time window, and constructing a trend prediction curve of the pH value and ammonium sulfate concentration through a support vector regression algorithm. According to the prediction curve, calculate the deviation value of the pH value from the preset range of 5.5 to 6.5 within the next 30 minutes, and the deviation value of the ammonium sulfate concentration from the preset range of 15% to 30%. Use a dual-threshold comparator to generate a parameter overlimit flag. Calculate the adjustment cycle according to the parameter overlimit flag using the preset response time parameter. If the adjustment cycle is less than 30 minutes, use a recurrent neural network to recalculate the adjustment parameters of the ammonia addition amount and the transfer pump speed. Generate a corrected value of the ammonia addition amount and a corrected value of the transfer pump speed according to the adjustment parameters, calculate new control parameters using a proportional-integral algorithm, and generate a control instruction sequence updated once per minute. Transmit the control instruction sequence to the actuator using a fieldbus protocol, obtain the execution status data from the actuator feedback interface, and judge the instruction execution result through an online checker. If the execution status data shows an abnormal change trend of the pH value or ammonium sulfate concentration, trigger an emergency protection mechanism and limit the control parameters within the preset safe operation range.

[0014] The technical solutions provided by the embodiments of the present invention may include the following beneficial effects: The present invention discloses a method for controlling the ammonium sulfate production process based on model prediction. This method obtains the real-time data of flue gas flow, sulfur dioxide concentration, slurry pH value, and ammonia concentration, analyzes the dynamic change characteristics of these indicators under different working conditions, and constructs a multi-dimensional flue gas component mutation feature library. A model reflecting the dynamic coupling mechanism of the four monitoring indicators is established using a non-linear regression algorithm to predict the change trends of each indicator in the future for a period of time. Based on the prediction results, control instructions for the ammonia addition amount and the transfer pump speed are generated to achieve the dynamic adjustment of the absorption circulation tank. At the same time, the present invention adopts an online optimization strategy, continuously optimizes the model parameters by comparing the model output with the actual monitoring data, and improves the prediction accuracy. This method can effectively improve the control accuracy and efficiency of the desulfurization system, reduce energy consumption, and achieve a double improvement in environmental protection and economic benefits. Description of the Drawings

[0015] Figure 1 It is a flowchart of a method for controlling the ammonium sulfate production process based on model prediction of the present invention. Detailed Embodiments

[0016] Next, the technical solutions in the embodiments of the present invention will be clearly and detailedly described in conjunction with the accompanying drawings in the embodiments of the present invention. The described embodiments are only a part of the embodiments of the present invention.

[0017] As Figure 1 , a method for controlling the ammonium sulfate production process based on model prediction in this embodiment may specifically include: S101 Collect data on flue gas flow rate, sulfur dioxide concentration, slurry pH value, and ammonia concentration in real time, analyze the dynamic change characteristics under different coal types and load conditions, and establish a multi-dimensional flue gas component mutation feature library. The desulfurization system has a control function based on model prediction, which can be integrated into the system software or implemented by an independent controller, and the specific implementation form is determined according to the actual scenario. After the control function is started, the system collects data on flue gas flow rate, sulfur dioxide concentration, slurry pH value, and ammonia concentration in real time through a multi-point sensor array, and analyzes the change rules of these parameters under different working conditions.

[0018] S1011 Use a multi-layer sampler to collect flue gas concentration data and coal quality data, establish a corresponding relationship data set between the two, generate standard working condition reference data, and compare the real-time working condition data through a multi-point sensor array to calculate the deviation value to determine the change trend of flue gas components.

[0019] Specifically, the multi-layer sampler arranges sampling points at positions such as the boiler outlet, the inlet of the economizer, and the outlet of the air preheater, collects flue gas concentration data once per second, and simultaneously collects coal quality data once per hour. Analyze the correlation between flue gas concentration and coal quality through a regression algorithm to establish a corresponding relationship data set. Based on this, generate standard working condition reference data. For example, when a 600 MW unit operates at 80% to 100% load, the flue gas flow rate is about 1.8 million cubic meters per hour, the sulfur dioxide concentration is 700 mg per cubic meter, and the slurry pH value is stable at 5.8. Compare the real-time collected flue gas flow rate, sulfur dioxide concentration, and slurry pH value data with the reference data to generate real-time working condition deviation data, which reflects the deviation of the current operating state from the standard.

[0020] S1012 Construct a curve of flue gas components and load changes according to the real-time working condition deviation data and unit load data, set a mutation threshold for flue gas concentration, record ammonia concentration and slurry pH value data when the threshold is triggered, use a recurrent neural network to generate optimized data for desulfurization process parameters, and establish a mutation feature library.

[0021] Obtain flow rate data through a flue gas flow rate monitoring device, and combine it with the load data in the operating conditions to establish a curve of flue gas composition versus load change. For example, when the load drops from 100% to 80%, the flue gas flow rate drops from 25 meters per second to 20 meters per second, and the sulfur dioxide concentration drops by about 15%. Set a mutation threshold according to the change trend. For example, a record is triggered when the sulfur dioxide concentration deviation exceeds 50 milligrams per cubic meter. Input the ammonia concentration and slurry pH value data during the recording period into a recurrent neural network to generate optimization parameters. Based on multi-dimensional data analysis, classify the optimization parameters by coal type and load range to construct a mutation feature library. For example, for bituminous coal with a sulfur content of 2%, the sulfur dioxide concentration is 800 milligrams per cubic meter at 90% load and drops to 650 milligrams per cubic meter at 80% load, and these data form process parameter adjustment rules.

[0022] The multi-layer sampler sets sampling points at different heights, with 4 sampling holes arranged at each point, and the sampling frequency is 1 Hz to ensure data synchronization. Taking a certain operating condition as an example, when the load drops from 100% to 85%, the sulfur dioxide concentration drops from 850 milligrams per cubic meter to 680 milligrams per cubic meter, and the system automatically adjusts the ammonia concentration from 20% to 17% and the slurry pH value from 5.8 to 5.5 to achieve dynamic optimization. For operating conditions with frequent mutations, the sampling frequency can be increased to 2 Hz to ensure data integrity.

[0023] The mutation feature library established through the above steps can reflect the flue gas composition characteristics under different coal types and loads, providing data support for subsequent model prediction.

[0024] S102 According to the multi-dimensional flue gas composition mutation feature library, use the non-linear regression algorithm to analyze the correlations between flue gas flow rate and sulfur dioxide concentration, flue gas flow rate and slurry pH value, sulfur dioxide concentration and slurry pH value, and ammonia concentration and slurry pH value, establish a multi-variable dynamic coupling model reflecting the dynamic coupling mechanism of flue gas flow rate, sulfur dioxide concentration, slurry pH value, and ammonia concentration, and predict the change trends of the four indicators in the future period through this model.

[0025] The multi-variable dynamic coupling model aims to quantitatively describe the interaction relationship between flue gas flow rate, sulfur dioxide concentration, slurry pH value, and ammonia concentration in a mathematical way. To this end, first extract historical data from the flue gas composition mutation feature library, including real-time monitoring values of flue gas flow rate, sulfur dioxide concentration, slurry pH value, and ammonia concentration. These data are processed by maximum-minimum normalization and mapped to the interval from 0 to 1 to eliminate the dimension difference for subsequent analysis. For example, at the rated load of a 600 MW unit, the flue gas flow rate of 1.8 million cubic meters per hour can be normalized to 0.85, the sulfur dioxide concentration of 850 milligrams per cubic meter can be normalized to 0.75, the slurry pH value of 5.8 can be normalized to 0.6, and the ammonia concentration of 20% can be normalized to 0.7.

[0026] S1021 Set a sliding time window, process the data of the standardized flue gas flow rate, sulfur dioxide concentration, slurry pH value, and ammonia water concentration. Use the least squares method to calculate the quadratic polynomial fitting functions for pairwise pairing of each index, generate the influence intensity data between the indexes, and construct a fourth-order state variable equation set based on this.

[0027] Specifically, the sliding time window is set to 10 minutes, and the sliding interval is 1 minute. Within each window, pairwise pairing analysis is performed on the four indexes. By fitting the quadratic polynomial function using the least squares method, the non-linear relationship between each index is obtained. For example, when the flue gas flow rate decreases from 1.8 million cubic meters per hour to 1.6 million cubic meters per hour, the sulfur dioxide concentration decreases from 850 mg per cubic meter to 750 mg per cubic meter. The fitting function shows that for every 0.1 million cubic meters per hour reduction in the flue gas flow rate, the sulfur dioxide concentration decreases by approximately 45 mg per cubic meter. Based on these fitting results, the influence intensity data between the indexes is generated, and then a fourth-order state variable equation set is constructed, where each variable represents an index, and the coefficient matrix of the equation set reflects the dynamic influence weights between the indexes. The parameters of the equation set are solved by the recursive least squares method to obtain stable coefficient values for subsequent model construction.

[0028] S1022 Based on the parameters of the solved fourth-order state variable equation set, construct a four-dimensional state space model, calculate the dynamic coupling trajectory of the four indexes, and analyze the trajectory characteristics through a Kalman filter to predict the change trends of the flue gas flow rate, sulfur dioxide concentration, slurry pH value, and ammonia water concentration within the next 30 minutes.

[0029] Using the solved parameters, the system constructs a four-dimensional state space model, where each dimension corresponds to the standardized value of an index. The dynamic coupling trajectory in the four-dimensional space reflects the non-linear change law between the indexes. For example, when the slurry pH value rises from 5.8 to 6.2, the ammonia water concentration drops from 20% to 17%, and the sulfur dioxide concentration may fluctuate by 50 mg per cubic meter, and this fluctuation stabilizes within 5 to 8 minutes. The system generates a set of non-linear mapping functions based on the trajectory to form a dynamic coupling response curve, and the slope of the curve is used to calculate the index change trend. Then, a Kalman filter is used to filter the response curve to predict the index values in the next 30 minutes. In a certain working condition where the load drops from 100% to 90%, the prediction results show that the flue gas flow rate will drop from 1.65 million cubic meters per hour to 1.55 million cubic meters per hour within 15 minutes, the sulfur dioxide concentration will drop from 780 mg per cubic meter to 720 mg per cubic meter, the slurry pH value will fluctuate between 5.8 and 6.0, and the ammonia water concentration needs to be adjusted from 18% to 16%. To ensure the prediction accuracy, the system calculates the root mean square error between the predicted value and the measured value through a sliding prediction window, and controls the error within 3% under stable working conditions and optimizes it to within 5% under dynamic working conditions.

[0030] Extract the flue gas flow mutation period data from the flue gas composition mutation feature library, match it with the feature vectors in the feature library, generate a dynamic weight matrix, optimize the parameters of the non-linear regression algorithm, calculate the response lag coefficient of the slurry pH value and correlate it with the ammonia water concentration adjustment rate, generate the dynamic correlation parameters between the flue gas flow and the sulfur dioxide concentration, and input them into the multi-variable dynamic coupling model to predict the future trend.

[0031] Collect the flue gas flow data at 10-second intervals through an on-line monitoring device, and calculate the flow change rate with a 5-minute sliding time window. If the change rate exceeds the preset reference value, such as the upper limit of 0.5% per minute under normal operating conditions, extract the four-dimensional feature vectors corresponding to the period from the feature library, including the flue gas flow, sulfur dioxide concentration, slurry pH value, and ammonia water concentration. Based on this, construct a 4×4 dynamic weight matrix, where each row represents the influence weight of a single index on other indexes. For example, in a certain mutation, the flue gas flow drops from 1.8 million cubic meters per hour to 1.5 million cubic meters per hour, and the change rate reaches 3.3% per minute. The matrix shows that the influence weight of the flue gas flow on the sulfur dioxide concentration is 0.85, and on the slurry pH value is 0.35. Then, the system collects the slurry pH value data at 5-second intervals, uses the non-linear regression algorithm to calculate the time displacement between the pH value curve and the flue gas flow curve, and obtains a pH value response lag coefficient of about 90 seconds, which reflects the inertial characteristics of the desulfurization system. At the same time, collect the ammonia water concentration data at 30-second intervals, calculate the adjustment rate curve, establish a dynamic mapping function between the flue gas flow and the sulfur dioxide concentration through a feedforward neural network, and generate dynamic correlation parameters to characterize the response relationship between the two. Finally, combining the weight matrix, lag coefficient, and correlation parameters, the system predicts the index trend curve for the next 30 minutes with a time resolution of 1 minute, and dynamically adjusts the parameters through deviation evaluation.

[0032] S1023 Obtain the historical data of the pH value and ammonium sulfate concentration in the absorption circulation tank, analyze their correlation, generate a training data set, and optimize the parameters of the multi-variable dynamic coupling model through support vector regression and deep neural network to improve the prediction accuracy.

[0033] Historical data is collected from the pH monitoring device at 5 - second intervals. The data is screened through a 30 - minute sliding time window and a 5 - point median filter is used to remove noise, generating a time - series change curve of pH. Data is collected from the ammonium sulfate concentration analyzer at 1 - minute intervals. The timestamps are aligned and missing values are supplemented through downsampling and three - point Lagrange interpolation. For example, when the ammonium sulfate concentration drops from 25% to 22% and 1 - minute data is missing, the interpolated value is estimated to be 23.2%. The Pearson correlation coefficient between the two is calculated using a 15 - minute calculation window, which normally remains at - 0.85 under normal operating conditions and may drop below - 0.4 during mutations. The system establishes a mapping function between pH and ammonium sulfate concentration through the support vector regression algorithm. When the correlation coefficient mutates, data for 15 minutes before and after is extracted, and a deep neural network is used to extract dynamic response feature vectors, including the change rate and response time difference. Based on this, model parameters are optimized. For example, the coupling coefficient between pH and ammonium sulfate concentration is adjusted from 0.75 to 0.82, reducing the prediction error from 7% to 4% and improving the model's adaptability to changes in operating conditions.

[0034] The multi - variable dynamic coupling model realizes the dynamic prediction of key indicators of the desulfurization system through the above steps. The model construction and optimization process do not overly limit the specific algorithm implementation details, which can be adjusted by technicians according to actual needs to adapt to the operating characteristics of different units.

[0035] S103 Using the multi - variable dynamic coupling model, within a preset 30 - minute prediction time window, input the current and historical collected data of flue gas flow rate, sulfur dioxide concentration, slurry pH value, and ammonia water concentration, predict the future change trends of the pH value and ammonium sulfate concentration in the absorption circulation tank, generate a prediction sequence, and generate a control instruction for the ammonia water addition amount or the rotational speed of the transfer pump according to the prediction result.

[0036] Through a multi-point sensor array arranged at the inlet, middle, and outlet of the absorption tower, data on flue gas flow rate, sulfur dioxide concentration, slurry pH value, and ammonia water concentration are collected in real-time at 10-second intervals. These data are input into a multivariable dynamic coupling model to generate prediction sequences for the pH value and ammonium sulfate concentration in the absorption circulation tank within the next 30 minutes. To ensure prediction accuracy, the system uses the support vector regression algorithm to analyze the time-shift relationship between the pH value and ammonium sulfate concentration, and calculates the maximum fluctuation amplitude of both through a 5-minute sliding prediction window. For example, in a 600-megawatt unit, when the load drops from 100% to 80%, the flue gas flow rate drops from 1.8 million cubic meters per hour to 1.5 million cubic meters per hour, and the model predicts that the pH value will drop from 5.8 to 5.5 within 15 minutes, and the ammonium sulfate concentration will rise from 25% to 28%. Based on the maximum fluctuation amplitude, the system establishes a mapping function between the pH value and the ammonia water addition amount, sets the target pH value control range to 5.5 to 6.5, and calculates the adjustment parameter for the ammonia water addition amount through a recurrent neural network. If the predicted pH value drops to 5.4, the model calculates that the ammonia water addition amount needs to be increased by 2.5 cubic meters per hour. Then, the system calculates the change in the flow rate of the transfer pump according to the adjustment parameter, determines the target speed using the proportional-integral algorithm, generates a control command within the range of 900 to 1500 revolutions per minute, and transmits it to the field controller through the industrial Ethernet. The field controller feeds back the deviation between the actual speed and the set value. If the deviation exceeds 50 revolutions per minute, the system dynamically adjusts the command through the feedback correction algorithm to ensure a smooth transition of the speed.

[0037] S1031 If the prediction sequence shows that the pH value of the absorption circulation tank will be lower than the preset lower limit threshold of 5.5, calculate the deviation between the predicted value and the lower limit, and determine the required ammonia water addition amount and generate a control command in combination with the correlation of flue gas parameters and the slurry buffering characteristics.

[0038] Extract pH value data from the prediction sequence at 1-minute intervals, smooth the difference from the lower threshold of 5.5 through a Kalman filter, and generate a pH value prediction deviation sequence. For example, if the predicted pH value drops from 5.8 to 5.3, the deviation is 0.2. Call the 4×4 correlation matrix in the multivariable dynamic coupling model and use the support vector regression algorithm to calculate the influence weights of flue gas flow and sulfur dioxide concentration on the pH value. Under a certain working condition, when the flue gas flow increases to 2 million cubic meters per hour and the sulfur dioxide concentration rises to 850 milligrams per cubic meter, the weights are 0.6 and 0.8 respectively, resulting in a 0.3 drop in the pH value. The system collects data on sulfite and sulfate concentrations every 5 minutes through a slurry sampling device and calculates the buffering capacity index. If the sulfite concentration is 15 millimoles per liter and the sulfate concentration is 25 millimoles per liter, the index is 0.75, indicating that the slurry has a strong buffering capacity. Based on this, calculate that 2.5 cubic meters per hour of ammonia water needs to be supplemented for a unit change in pH value, and combined with the deviation of 0.2, the theoretical supplement amount is 0.5 cubic meters per hour. Considering the correction parameter of 1.2 for high-load working conditions, finally determine the actual supplement amount to be 0.6 cubic meters per hour. Extract the reference curve from the ammonia water addition amount database and generate an increment instruction. If the pH value recovery rate is lower than expected, for example, only 0.1 unit per hour instead of 0.15, the system adaptively adjusts to 0.7 cubic meters per hour to ensure that the pH value is stable at 5.7.

[0039] If the prediction sequence shows that the ammonium sulfate concentration will be higher than the preset upper threshold by 30%, calculate the prediction deviation, analyze the relationship between the transfer pump speed and the concentration and the influence of crystal particle size in combination with historical data, determine the increase in speed and generate a control instruction.

[0040] Extract ammonium sulfate concentration data from the prediction sequence at 1-minute intervals, predict the trend through the support vector regression algorithm and calculate the deviation from the 30% upper limit. For example, if the predicted concentration rises to 32%, the deviation is 2%. The system extracts the transfer pump speed, slurry density, and temperature from the historical data of the most recent 24 hours, constructs a 4×4 weight matrix using the long short-term memory network, and calculates the influence coefficient. If the speed increases from 1200 to 1350 revolutions per minute, the density is 1.2 kilograms per liter, and the temperature is 45 degrees Celsius, the concentration decrease rate is about 0.8% per hour, and the weights are 0.6, 0.3, and 0.1 respectively. Collect crystal particle size data through a laser diffraction particle size analyzer every 5 minutes. If the average particle size is 250 micrometers, the standard deviation is 35 micrometers, and the supersaturation index is 1.15, it indicates mild supersaturation. Combining the deviation and the weights, calculate that the speed needs to be increased to 1380 revolutions per minute. Generate an adjustment signal using a proportional-integral algorithm with an integral time of 60 seconds to avoid sudden impacts. If the speed response lags by 20 seconds, the system adaptively corrects the adjustment parameters to ensure a smooth reach of the target value.

[0041] Through multi-point sensor arrays and dynamic model prediction, the system can monitor the desulfurization process in real time and generate precise control instructions. The dynamic adjustment of pH value and ammonium sulfate concentration not only improves the desulfurization efficiency but also optimizes resource utilization to meet the requirements of different working conditions.

[0042] S104 Transmit the generated ammonia addition amount and transfer pump speed control instructions to the desulfurization tower control system through the industrial Ethernet bus, perform dynamic adjustment of the absorption circulation tank, and at the same time collect real-time monitoring data and compare it with the prediction results of the multi-variable dynamic coupling model, calculate the deviation and use the gradient descent method to optimize the model parameters online, and generate a more accurate future prediction sequence to improve the control effect.

[0043] Adopt the Modbus-TCP communication protocol, encode the ammonia addition amount and transfer pump speed control instructions into a 16-bit integer format. For example, encode the ammonia addition amount of 2.5 cubic meters per hour as 25000, and the transfer pump speed of 1350 revolutions per minute as 13500, and efficiently transmit them to the desulfurization tower controller through the industrial Ethernet bus. To ensure transmission reliability, the system monitors the network status through the heartbeat packet mechanism. If the communication delay exceeds 100 milliseconds, an alarm is triggered. Collect the pH value, ammonium sulfate concentration, slurry density, and temperature data in the circulation tank from the sensor arrays at the inlet, middle, and outlet of the absorption tower at 5-second intervals, and calculate the deviation compared with the model prediction sequence. For example, the predicted pH value drops from 5.8 to 5.3 within 30 minutes, while the measured value is 5.4, and the deviation is 0.1. The system uses the gradient descent method with an initial learning rate of 0.01 to iteratively optimize the deviation, and stops when the deviation is less than 0.05 for 5 consecutive iterations, generating optimized weight parameters. Based on this, construct a 4×4 model parameter matrix, update the coupling parameters through a recurrent neural network, and calculate the root mean square error of the future 30-minute prediction sequence. If the error indicates that the prediction window needs to be adjusted, the system dynamically extends it to 12 minutes, and enables local caching to save the data of the last 4 hours when the communication is interrupted, and synchronously uploads it after recovery to ensure data continuity.

[0044] S1041 Compare the adjusted absorption efficiency and the pressure fluctuation value in the tank with the prediction sequence to generate an error distribution matrix, based on which iteratively optimize the weight parameters of the multi-variable dynamic coupling model, and generate the prediction result sequence for the next cycle, including the ammonia demand trend and the speed matching interval.

[0045] Absorption efficiency and pressure fluctuation data are acquired from the online acquisition device at 10 - second intervals, pre - processed through a Kalman filter with a measurement noise variance of 0.1 to generate a smooth trend curve. For example, under a certain working condition, the absorption efficiency drops from 95% to 92%, and the pressure fluctuation rises from 200 Pa to 350 Pa. A 4×4 error distribution matrix is generated by comparing with the predicted values, where the deviation of the absorption efficiency is 3% and the deviation of the pressure fluctuation is 75 Pa. The gradient - descent algorithm is used to iteratively optimize the weight parameters with a step size of 0.01. After multiple rounds of calculation, when the error is continuously lower than 0.05 for 5 times, an optimized set of weight parameters is generated. Based on this, a dynamic coupling matrix is constructed, and the coupling coefficient is calculated through a recurrent neural network. For example, the coefficient between the absorption efficiency and the ammonia - water concentration is 0.85, and the coefficient with the rotational speed is 0.65. The coefficient is updated using a 10 - minute sliding time window to generate a response relationship curve, and the future 30 - minute ammonia - water demand trend and the rotational - speed matching interval are predicted according to the response time constant. If the predicted ammonia - water demand increases by 0.8 cubic meters per hour, the rotational speed needs to increase from 1200 to 1350 revolutions per minute. The data is stored at 1 - minute intervals, and if the deviation exceeds 1%, it is marked as abnormal for subsequent optimization.

[0046] Through real - time data update and model optimization, the system can dynamically adapt to the changes in the desulfurization process to ensure the accuracy of the prediction sequence. During a certain operation, the control instruction adjusts the ammonia - water addition amount to 2.5 cubic meters per hour, and the rotational speed rises to 1350 revolutions per minute. The measured pH value recovers from 5.4 to 5.7, and the ammonium sulfate concentration drops from 28% to 26%, which basically coincides with the predicted trend, and the root - mean - square error is 1.8%, indicating that the control effect is significantly better than the traditional method. Moreover, through the cache and synchronization mechanism, it can effectively cope with network anomalies, providing a reliable guarantee for the stable operation of the process.

[0047] S105 By comparing the predicted result sequences before and after optimization, analyze whether the pH value and ammonium sulfate concentration in the absorption circulation tank can continuously meet the preset range conditions within the next 30 minutes, that is, the pH value is between 5.5 and 6.5, and the ammonium sulfate concentration is between 15% and 30%. If it meets the conditions, the dynamic adjustment cycle is completed; if it exceeds the range, recalculate the adjustment values of the ammonia - water addition amount and the rotational speed of the transfer pump to ensure the stable operation of the system.

[0048] Extract the predicted sequence data before and after optimization from the prediction result memory at 1-minute intervals, and calculate the root mean square error of the two groups of sequences using a 5-minute sliding time window. For example, in a certain adjustment, the error drops from 0.2 to 0.15, indicating a significant optimization effect. Based on this error value, a support vector regression algorithm is used to construct prediction curves for pH value and ammonium sulfate concentration. This algorithm maps to a high-dimensional space through a kernel function, analyzes the non-linear relationship between historical data and current inputs, and generates smooth prediction results that closely follow the actual trend. The prediction curves show the change trajectories of pH value and ammonium sulfate concentration within the next 30 minutes. For example, the pH value gradually decreases from 5.8 to 5.3, and the ammonium sulfate concentration increases from 25% to 32%. The system compares the predicted values with the preset ranges, calculates the deviation parameters, such as a pH value deviation of 0.2 and an ammonium sulfate concentration deviation of 2%, and determines whether it exceeds the limit through a dual-threshold comparator. If it exceeds the limit, an over-limit flag is generated. The over-limit flag triggers the calculation of the adjustment period, and the adjustment duration is determined based on the preset response time parameter. For example, for 12 minutes, when it is less than the 30-minute threshold, the system starts the parameter correction process.

[0049] S1051 If the prediction result shows that the parameter exceeds the limit, the system uses a recurrent neural network to recalculate the correction values of the ammonia addition amount and the transfer pump speed according to the deviation parameters, and transmits them to the actuator for adjustment through the fieldbus protocol. At the same time, the execution effect is verified to ensure that the control target is achieved.

[0050] When the predicted pH value is lower than 5.5 or the ammonium sulfate concentration is higher than 30%, the recurrent neural network calculates the adjustment value based on the current input data such as flue gas flow rate and sulfur dioxide concentration and the deviation parameters. For example, under a certain working condition, the predicted pH value drops to 5.3, and the ammonia addition amount needs to be increased to 3.2 cubic meters per hour. The ammonium sulfate concentration rises to 32%, and the transfer pump speed needs to be increased from 1200 revolutions per minute to 1380 revolutions per minute. The recurrent neural network iteratively calculates through multiple layers of neurons, combines historical adjustment experience, and outputs accurate correction values. The system uses a proportional-integral algorithm to convert the correction values into a control instruction sequence updated every minute to ensure a smooth adjustment process and avoid sudden impacts. The instructions are transmitted to the actuator through the Modbus-TCP protocol, and the actuator feeds back the status data at 1-second intervals. For example, when the actual speed reaches 1350 revolutions per minute, fluctuations occur due to an increase in the slurry viscosity. After the on-line checker detects the deviation, an adjustment is triggered. The emergency protection mechanism limits the speed to the safe value of 1350 revolutions per minute, and at the same time finely tunes the ammonia addition amount to 3.0 cubic meters per hour to ensure stable operation of the equipment. After adjustment, the pH value rises back to 5.8, and the ammonium sulfate concentration drops to 28%, meeting the control target.

[0051] Through the comparison of the predicted sequences before and after optimization and dynamic adjustment, the system realizes the precise control of the parameters of the absorption circulation tank. During a certain operation, the predicted pH value before optimization dropped to 5.4, and after optimization, it was adjusted to 5.6, with the measured value being 5.7, and the error was reduced to 0.1, indicating a significant improvement in the model accuracy. The prediction of the ammonium sulfate concentration was corrected from 31% to 29%, and the measured value was 28.5%, also showing higher reliability. After the adjustment cycle is completed, if the parameters remain stable, it enters the monitoring state; if they are still out of limits, the calculation and execution process are repeated to ensure the efficiency and stability of the desulfurization process.

[0052] As described above, the above embodiments are only used to illustrate the technical solutions of the present application and are not intended to limit them; although the present application has been described in detail with reference to the foregoing embodiments, those of ordinary skill in the art should understand that: they can still modify the technical solutions recorded in the foregoing embodiments, or perform equivalent replacements on some of the technical features; and these modifications or replacements do not cause the essence of the corresponding technical solutions to deviate from the spirit and scope of the technical solutions of the various embodiments of the present application.

Claims

1. A method for controlling the ammonium sulfate production process based on model prediction, characterized in that, The method includes: Obtaining real-time data of flue gas flow rate, sulfur dioxide concentration, slurry pH value, and ammonia water concentration, analyzing the dynamic change characteristics of flue gas flow rate, sulfur dioxide concentration, slurry pH value, and ammonia water concentration under different coal types and load conditions, and constructing a flue gas component mutation characteristic library; According to the flue gas component mutation characteristic library, using a non-linear regression algorithm, by analyzing the correlation between flue gas flow rate and sulfur dioxide concentration, the correlation between flue gas flow rate and slurry pH value, the correlation between sulfur dioxide concentration and slurry pH value, and the correlation between ammonia water concentration and slurry pH value, predicting the change trends of flue gas flow rate, sulfur dioxide concentration, slurry pH value, and ammonia water concentration in a future period of time; Within a preset prediction time window, according to the flue gas flow rate, sulfur dioxide concentration, slurry pH value, and ammonia water concentration in the current and preset time periods, predicting the change trends of the pH value and ammonium sulfate concentration in the future absorption circulation tank, generating a prediction sequence, and generating an ammonia water addition amount control instruction and / or a transfer pump speed control instruction according to the prediction sequence; Transmitting the ammonia water addition amount and / or transfer pump speed control instruction to the desulfurization tower control system, performing dynamic adjustment of the absorption circulation tank, and at the same time comparing the prediction result and the real-time monitoring data, and generating an optimized prediction sequence according to the calculated deviation; By comparing the prediction result sequences before and after optimization, judging whether the pH value and ammonium sulfate concentration of the absorption circulation tank can continuously meet the preset upper and lower limit range conditions within a future preset time. If they meet, a dynamic adjustment cycle is completed. If they do not meet, the adjustment values of the ammonia water addition amount and / or the transfer pump speed are recalculated.

2. The method according to claim 1, wherein The obtaining of real-time data of flue gas flow rate, sulfur dioxide concentration, slurry pH value, and ammonia water concentration, analyzing the dynamic change characteristics of flue gas flow rate, sulfur dioxide concentration, slurry pH value, and ammonia water concentration under different coal types and load conditions, and constructing a flue gas component mutation characteristic library includes: Using a multi-layer sampler to obtain flue gas concentration data and coal quality data, and establishing a corresponding relationship data set; Generating standard condition reference data according to the corresponding relationship data set, collecting flue gas flow rate data, sulfur dioxide concentration data, and slurry pH value data through a multi-point sensor array, and comparing with the standard condition reference data to obtain real-time condition deviation data; Establishing a flue gas component and load change curve according to the real-time condition deviation data and unit load data, and obtaining flue gas component change trend data; Setting a flue gas concentration mutation threshold according to the flue gas component change trend data, recording ammonia water concentration data and slurry pH value data when the mutation threshold is triggered, and generating desulfurization process parameter optimization data using a recursive neural network; Classifying the desulfurization process parameter optimization data using a multi-dimensional data analysis method, and establishing a flue gas component mutation characteristic database under different coal types and different load intervals.

3. The method according to claim 1, wherein According to the flue gas component mutation characteristic library, using a non-linear regression algorithm, by analyzing the correlation between flue gas flow rate and sulfur dioxide concentration, the correlation between flue gas flow rate and slurry pH value, the correlation between sulfur dioxide concentration and slurry pH value, and the correlation between ammonia water concentration and slurry pH value, predicting the change trends of flue gas flow rate, sulfur dioxide concentration, slurry pH value, and ammonia water concentration in a future period of time includes: Obtain the data of four indicators, namely flue gas flow rate, sulfur dioxide concentration, slurry pH value, and ammonia water concentration, from the flue gas composition database, and obtain the standardized indicator data through maximum-minimum normalization processing; Set a sliding time window for the standardized indicator data, and use the least squares method to calculate the quadratic polynomial fitting functions of the four indicators paired pairwise, and obtain the influence intensity data between the indicators; Establish a fourth-order state variable equation set according to the influence intensity data between the indicators, solve the parameter values of the equation set by the recursive least squares method, and construct a four-dimensional state space model; Use the four-dimensional state space model to calculate the dynamic coupling trajectory of the four indicators, and analyze the dynamic coupling trajectory through a Kalman filter to obtain the predicted values of the changes in the four indicators within the future time window.

4. The method according to claim 3, wherein It further includes: Extract the data of the flue gas flow rate mutation period, match it with the eigenvectors in the flue gas composition mutation feature library, generate a dynamic weight matrix, adjust the input parameters of the nonlinear regression algorithm, calculate the pH value response lag coefficient of the slurry, and associate and map it with the ammonia water concentration adjustment rate to generate the dynamic correlation parameters between the flue gas flow rate and the sulfur dioxide concentration, and input them into the multivariable dynamic coupling model to output the predicted trend data of the future flue gas flow rate, sulfur dioxide concentration, slurry pH value, and ammonia water concentration, specifically including: Obtain the flue gas flow rate data from the on-line flue gas monitoring device according to the sampling interval, calculate the change rate of the flue gas flow rate through a sliding time window, and extract the four-dimensional eigenvector according to the change rate exceeding the preset reference value interval; Establish a dynamic weight matrix according to the four-dimensional eigenvector, and use the nonlinear regression algorithm to calculate the time displacement between the pH value curve and the flue gas flow rate curve to obtain the pH value response lag coefficient; Obtain the ammonia water concentration data according to the pH value response lag coefficient, and establish a dynamic mapping function between the flue gas flow rate and the sulfur dioxide concentration by using a feedforward neural network to obtain the dynamic correlation parameters; Construct the flue gas flow rate change curve within the prediction time window according to the dynamic weight matrix, the pH value response lag coefficient, and the dynamic correlation parameters, and calculate the predicted trend curves of the three indicators of sulfur dioxide concentration, slurry pH value, and ammonia water concentration through matrix mapping.

5. The method according to claim 4, wherein It further includes: Obtain the historical change data of the pH value and the historical change data of the ammonium sulfate concentration in the absorption circulation tank, analyze the correlation between the pH value and the ammonium sulfate concentration in the absorption circulation tank, and form a model training data set to optimize the multivariable dynamic coupling model, specifically including: Obtain the historical data collected by the pH value monitoring device of the absorption circulation tank, and the historical data is screened through a sliding time window and then passed through a median filter to obtain the pH value time series change curve; According to the sampling timestamps of the pH value time series change curve and the historical data of the ammonium sulfate concentration, obtain the time-aligned data sequence through downsampling and Lagrange interpolation method; Calculate the correlation coefficient of the time-aligned data sequence using a calculation window, and establish a mapping function between the pH value and the ammonium sulfate concentration through a support vector regression algorithm; If the change rate of the correlation coefficient exceeds a preset threshold, extract the data sequences before and after the mutation point. After the data sequences are subjected to deep neural network feature extraction, dynamic response feature vectors are obtained, and the parameters of the dynamic coupling model are optimized by the backpropagation algorithm.

6. The method according to claim 5, wherein It also includes: If the prediction sequence shows that the pH value of the future absorption circulation tank will be lower than the preset lower threshold, calculate the difference from the lower threshold to obtain the predicted deviation of the pH value; combine the influence of the sulfite / sulfate ratio in the slurry on the pH buffering capacity, calculate the additional ammonia water addition amount required, and generate a control instruction to increase the ammonia water addition amount, specifically including: Extract pH value data from the pH value prediction sequence at a preset time interval, smooth the difference between the pH value data and the preset pH threshold through a Kalman filter to obtain a pH value prediction deviation sequence; according to the pH value prediction deviation sequence, use the support vector regression algorithm to calculate the influence weight of flue gas parameters on the pH value; obtain concentration data through a slurry sampling device, use a recursive calculator to generate a buffering capacity index, calculate the ammonia water supplement amount required for a unit pH value change according to the buffering capacity index; obtain the theoretical supplement amount according to the product of the unit supplement amount and the pH value prediction deviation, correct the theoretical supplement amount through flue gas operating parameters, calculate the actual total ammonia water supplement amount, and generate a control instruction to increase the ammonia water addition amount.

7. The method according to claim 6, wherein It also includes: If the prediction sequence shows that the ammonium sulfate concentration in the future absorption circulation tank will be higher than the preset upper threshold, calculate the difference between the predicted value and the preset upper threshold to obtain the predicted deviation of the ammonium sulfate concentration, and combine the influence of the ammonium sulfate crystal particle size distribution on the supersaturation, calculate the required increase in the transfer pump speed, and generate a control instruction to increase the transfer pump speed, specifically including: Use the support vector regression algorithm to analyze the ammonium sulfate concentration prediction sequence in the circulation tank, and calculate the predicted deviation of the ammonium sulfate concentration through a comparison operator; Establish a long short-term memory network model according to the predicted deviation of the ammonium sulfate concentration, and obtain the influence coefficient of the ammonium sulfate concentration change rate through weight matrix operation; Obtain ammonium sulfate crystal particle size distribution data through a laser diffraction particle size analyzer, and calculate the crystal supersaturation index through a recursive calculator; Calculate the reference speed of the transfer pump according to the crystal supersaturation index and the concentration change rate influence coefficient, and generate a control instruction to increase the transfer pump speed.

8. The method according to claim 7, wherein Transmit the control instruction for the ammonia water addition amount and / or the transfer pump speed to the desulfurization tower control system, perform dynamic adjustment of the absorption circulation tank, and at the same time compare the prediction results and real-time monitoring data, and generate an optimized prediction sequence according to the calculated deviation, including: Receive the control instruction transmitted by the industrial Ethernet bus, obtain the pH value, ammonium sulfate concentration, slurry density and temperature data in the circulation tank from the acquisition device according to the control instruction, compare with the prediction sequence to obtain the prediction deviation value, and the prediction deviation value generates weight parameters through gradient descent operation by an adaptive calculator; Establish a model parameter matrix according to the weight parameters, and after the model parameter matrix is updated by a recursive neural network, generate an optimized future prediction sequence.

9. The method according to claim 8, wherein It also includes: Compare the adjusted absorption efficiency and the pressure fluctuation value in the tank with the predicted result sequence to generate an error distribution matrix; Iteratively optimize the weight parameters of the dynamic coupling relationship in the multivariable dynamic coupling model based on the error distribution matrix; Generate the predicted result sequence for the next cycle according to the optimized weight parameters of the dynamic coupling relationship. The predicted result sequence includes the ammonia water demand trend and the speed matching interval at multiple future time points, specifically including: Obtain the absorption efficiency data and pressure fluctuation data collected by the online acquisition device at preset time intervals, and preprocess them through a Kalman filter to obtain an error distribution matrix; Iteratively optimize the weight parameters in the dynamic coupling model according to the error distribution matrix using the gradient descent algorithm. If the continuously iterated error value is less than the preset threshold, generate an optimized set of weight parameters; Construct a dynamic coupling matrix for the optimized set of weight parameters, and calculate the coupling coefficient through a recursive neural network; Update the coupling coefficient using a sliding time window to obtain a response relationship curve, calculate the ammonia water demand trend according to the response time constant, and generate speed matching data.

10. The method according to claim 1, wherein By comparing the predicted result sequences before and after optimization, judge whether the pH value and ammonium sulfate concentration in the absorption circulation tank can continuously meet the preset upper and lower limit range conditions within a preset future time. If satisfied, complete a dynamic adjustment cycle. If not satisfied, recalculate the adjustment values of the ammonia water addition amount and / or the transfer pump speed, including: Calculate the root mean square error value for the prediction sequence using a sliding time window, and construct a prediction curve for the pH value and ammonium sulfate concentration according to the root mean square error value through the support vector regression algorithm; Calculate the deviation parameters for the pH value and ammonium sulfate concentration according to the prediction curve, and compare the deviation parameters through a double threshold comparator to obtain a parameter overrun mark; If the adjustment cycle corresponding to the parameter overrun mark is less than the preset time, use a recursive neural network to calculate the correction value of the ammonia water addition amount and the correction value of the transfer pump speed according to the deviation parameters; Transmit the correction value to the actuator through the fieldbus protocol, obtain the execution status data from the actuator feedback interface, and judge the instruction execution result through the online checker.

Images