A dynamic data analysis and decision system for coke oven flue gas SDS desulfurization reagent

Abstract

This invention discloses a dynamic data analysis and decision-making system for SDS desulfurization agents in coke oven flue gas, relating to the fields of air pollutant control and automatic control technology, including a data acquisition module. This invention objectively and quantitatively identifies the inherent reaction delay time of the system from historical data and constructs a concentration prediction model. Simultaneously, it uses a multiple linear regression correction method to correct the predicted values ​​of the outlet concentration in real time, further eliminating the cumulative error of the model. By calculating and updating the current average emission value and the remaining allowable emission key intermediate variables in real time, the feeding management algorithm integrates historical trends, real-time predictions, and emission budgets, combining the majority judgment rule with the dynamically calculated emission capacity and the corrected predicted value to intelligently decide the timing of feeding start and stop. This mechanism can minimize unnecessary agent additions while strictly ensuring that the hourly average emission concentration meets the standard.

Description

Technical Field

[0001] This invention relates to the fields of air pollutant control and automatic control technology, and in particular to a dynamic data analysis and decision-making system for SDS desulfurization agents in coke oven flue gas. Background Technology

[0002] Coke ovens have high production intensity, and their flue gas is characterized by low temperature and large fluctuations. Furthermore, SO2 emission concentrations must be strictly controlled at 30 mg / m³. 3 Currently, although the sodium bicarbonate dry desulfurization (SDS) process can effectively achieve desulfurization, the following prominent problems still exist: First, the cost of sodium bicarbonate raw material is high, and it is prone to caking and moisture absorption during actual storage and use, requiring continuous monitoring and intervention from operators; second, the task of monitoring production data is arduous, the labor intensity of operators is high, and it is difficult to keep track of the feed rate and emission concentration in real time, which can easily lead to overdosing and waste of reagents; third, the control of SO2 concentration at the outlet lacks a systematic and adaptive strategy, mainly relying on the experience of operators for adjustment, resulting in insufficient control accuracy and stability.

[0003] In the existing technology, some studies have attempted to introduce model predictive control to improve the automation level of the desulfurization process, but there are still obvious limitations: on the one hand, the prediction models used often have limited prediction accuracy and are difficult to adapt to the complex fluctuations of coke oven flue gas conditions; on the other hand, most schemes only focus on concentration control itself, without fully considering the balance between reagent reaction efficiency and economic cost, and lack dynamic optimization of the dosing process. To address the aforementioned technical shortcomings, a solution is proposed. Summary of the Invention

[0004] This invention adopts the following technical solution: a dynamic data analysis and decision-making system for SDS desulfurization agents in coke oven flue gas, including a data acquisition module, a concentration monitoring module, a dosing decision-making module, a safety monitoring module, and an anomaly handling module. The data acquisition module is used to acquire and integrate flue gas parameter data at the inlet and outlet of the desulfurization system in real time, and to acquire the pre-grinding mill current, unloading frequency, sequential control program running status and running time, and to determine the reaction delay time during desulfurization. The concentration monitoring module, based on the collected continuous historical operating data and combined with the time series training dataset, establishes a flue gas prediction model through key time series characteristic analysis, so as to output a predicted sequence of outlet SO concentration for future time steps. The dosing decision module is used to perform dynamic dosing decision analysis based on the real-time judgment of the outlet concentration trend over a historical period and the predicted outlet SO2 concentration data. Based on the preset emission limit rules, it dynamically generates decision instructions for dosing implementation. The safety monitoring module is used to receive decision instructions from the dynamic decision and control module, execute automated sequential start and sequential stop procedures, control the operation of the feeding equipment, provide a graphical operation interface to display the operating data, outlet SO2 concentration trend, system alarm information, and receive user function start / stop, parameter setting and sequential reset operation instructions. The exception handling module monitors the running status of sequential start and stop in real time, and enables the follow-up start and follow-up stop functions to deal with the incorrect operation of sequential start and stop. If feeding is in progress during backflushing and exchange, the sequential stop function should be activated to stop feeding.

[0005] Furthermore, the pre-grinding mill current, unloading frequency, sequential control program operation status and runtime are obtained, and the reaction delay time during desulfurization is determined. The specific process is as follows: By deploying continuous emission monitoring systems at the inlet and outlet of the desulfurization system, the emissions in the flue gas are collected simultaneously. concentration, Content, humidity, pressure, flow rate, smoke temperature, and flow rate parameters; Equipment operation data: Key status parameters of the reagent preparation and dosing system are acquired in real time through PLC or equipment sensors, including at least the mill operating current, the operating frequency of the discharge valve, and the start / stop status and duration of the sequential control program; Add a unified high-precision timestamp to all collected real-time data streams to ensure that process data and equipment status data across systems are strictly aligned in the time dimension. Store the time-aligned multivariate time series data in a structured historical database to form a long-term running dataset that covers the complete operating conditions and can be used for model training and post-analysis. Based on the aforementioned historical operational dataset, the exit is calculated. The cross-correlation function between concentration changes and feeding actions, or the application of system identification algorithms, can be used to quantitatively analyze and determine the relationship between changes in feeding and the outlet. The number of inherent reaction delay steps between the concentration and the stable response is denoted as the key parameter u, where u step refers to the time window parameter, which refers to the chemical reaction delay time of the SDS desulfurization system. Furthermore, based on the collected continuous historical operational data and combined with the time series training dataset, a concentration prediction model is established through key time series characteristic analysis. The specific process is as follows: S200. Based on the stored continuous historical operating data, extract a complete sample sequence containing the multi-source time-series variables and the corresponding outlet SO2 concentration values ​​at that time. Preprocess the dataset, including data cleaning, outlier removal, and normalization, and divide it into a training set, a validation set, and a test set according to a preset ratio to form a time-series training dataset for model learning. S201. Construct a deep learning concentration prediction model based on an improved long short-term memory network, including a feature extraction layer: use a one-dimensional convolutional neural network to extract local features from the input multi-dimensional time series, and perform adaptive weight adjustment on the extracted features; Multi-layer LSTM units are used to capture long-term temporal dependencies, and a multi-head attention mechanism is introduced to perform secondary filtering and fusion of the hidden states of the LSTM. Output layer: Maps high-level features to output values ​​for the next u time steps using a fully connected network. Concentration prediction sequence.

[0006] Furthermore, it is used to determine the export concentration trend over historical periods in real time, combined with predicted exports. The concentration data is used for dynamic dosing decision analysis, and the specific process is as follows: Based on real-time and historical exits Concentration data, calculate the actual average emissions for the current hour. and the permissible average emissions over the remaining time period. ; Based on the established trend judgment rules, analyze the changing trend of the outlet concentration within the past u-step; When the trend meets the first preset condition, the flue gas prediction model module is invoked to obtain the outlet value for the next u-step. Initial concentration predictions are generated, and subsequent predictions are corrected online using the actual measurements that follow. Based on the corrected prediction results, combined with the emission intensity limit and the second preset condition, feeding start command and feeding stop command are generated.

[0007] Furthermore, based on preset emission limit rules, decision instructions for material feeding are dynamically generated, and the judgment logic for stopping material feeding includes: Using the reaction delay time u as the analysis window length, the number of time steps in which the measured concentration of SO2 at the outlet does not exceed the preset feeding stop threshold D in the most recent u consecutive sampling moments is counted in real time. If the number of steps exceeds half the window length, the current emission trend is determined to meet the first trigger condition for stopping the judgment; When the first triggering condition is met, the concentration prediction model is invoked to obtain the initial prediction sequence of the outlet SO2 concentration for the next u time steps, and the prediction sequence is rolled online using the latest collected real concentration data to generate the corrected future concentration prediction curve. This includes determining the export within the complete future U-step forecast period. The state in which the predicted concentration value is not lower than the predicted I value at the corresponding time. Determine whether the cumulative duration of continuous feeding since the start of this feeding has exceeded the preset maximum safe operating time T; After two judgments, if either condition is met, a feeding stop command is immediately generated and sent to the safety monitoring unit to terminate the drug delivery.

[0008] Furthermore, the logic for determining the feeding start-up conditions includes: Step 1: Calculate the actual cumulative emissions from the start time t1 to the current time tn for the current hour, and obtain the actual average emissions for the current hour. and the permissible average emissions over the remaining time period. ; Step 2: Using the reaction delay time u as the analysis window length, count the number of time steps in the past u consecutive sampling times where the measured concentration of SO2 at the outlet is not lower than I; If the number of time steps exceeds half the window length, the system emission trend is determined to meet the first trigger condition. When the first trigger condition is met, the prediction model is called to obtain the initial prediction sequence of the outlet SO2 concentration for the next u time steps, and the prediction sequence is rolled online using the latest collected real concentration data to generate the corrected future concentration prediction curve. Step 3: Based on the corrected prediction curve, calculate the predicted I value for each future time point, and determine the exit point within the complete future u-step prediction period. The state in which the predicted concentration value is not lower than the predicted I value at the corresponding time. If step three verification passes, then further check for unplanned and sudden process exchange and backflush signals within the preset process look-ahead time window. The system generates a material feeding start command and sends it to the safety monitoring unit only when the trend triggering condition, the prediction persistence condition, and the process safety condition are met simultaneously.

[0009] Furthermore, the concentration monitoring module also includes online correction of subsequent predicted values ​​using actual measured values. The specific method is multiple linear regression correction, and the specific process is as follows: Let the initial prediction sequence be Obtain the sequence of the first m actual values ​​after prediction. Where m is a fixed number of steps less than u, the prediction error sequence of the first m steps is calculated. ,in m=15; For the k-th step The predicted value, its correction value The calculation formula is: in, , Let be the weighting coefficient of the i-th error term. In the range of 0.2 to 0.05, β is an attenuation factor that ranges from 0.8 to 0.95.

[0010] In summary, due to the adoption of the above technical solution, the beneficial effects of the present invention are: This coke oven flue gas SDS desulfurization agent dynamic data analysis and decision-making system objectively and quantitatively identifies the inherent reaction delay time of the system from historical data and constructs a concentration prediction model. Simultaneously, it uses a multiple linear regression correction method to correct the predicted values ​​of the outlet concentration in real time, further eliminating the cumulative error of the model. By calculating and updating the current average emission value and the remaining allowable emission key intermediate variables in real time, it dynamically constructs an emission budget management system that conforms to the national hourly average standard. The feeding management algorithm integrates historical trends, real-time predictions, and emission budgets, combining the majority judgment rule with the dynamically calculated emission capacity and the corrected predicted value to form a standardized and reproducible intelligent feeding start-up and shutdown decision logic. This intelligently determines the timing of feeding start-up and shutdown. This mechanism can minimize unnecessary agent addition while strictly ensuring that the hourly average emission concentration meets the standard, avoiding the conservative waste common in manual operation, and significantly improving the reaction efficiency and utilization rate of sodium bicarbonate. Attached Figure Description

[0011] Figure 1 This is a schematic diagram of the system steps in this invention; Figure 2 This is a schematic diagram of the deep learning model structure in an embodiment of the present invention; Figure 3 This is a flowchart of the dynamic data analysis and control system for coke oven flue gas SDS desulfurization agents based on improved LSTM in an embodiment of the present invention; Figure 4 This is a flowchart of the training process for the dynamic data analysis and control system model of coke oven flue gas SDS desulfurization agent based on the improved LSTM in an embodiment of the present invention. Detailed Implementation

[0012] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention. Example 1:

[0013] like Figure 1-4As shown, a dynamic data analysis and decision-making system for SDS desulfurization agents in coke oven flue gas includes a data acquisition module, a concentration monitoring module, a dosing decision module, a safety monitoring module, and an anomaly handling module. The data acquisition module is used to acquire and integrate flue gas parameter data at the inlet and outlet of the desulfurization system in real time, and to acquire the pre-grinding mill current, unloading frequency, sequential control program running status and running time, and to determine the reaction delay time during desulfurization. The concentration monitoring module, based on the collected continuous historical operating data and combined with the time series training dataset, establishes a flue gas prediction model through key time series characteristic analysis, so as to output a predicted sequence of outlet SO concentration for future time steps. The dosing decision module is used to perform dynamic dosing decision analysis based on the real-time judgment of the outlet concentration trend over a historical period and the predicted outlet SO2 concentration data. Based on the preset emission limit rules, it dynamically generates decision instructions for dosing implementation. The safety monitoring module is used to receive decision instructions from the dynamic decision and control module, execute automated sequential start and sequential stop procedures, control the operation of the feeding equipment, provide a graphical operation interface to display the operating data, outlet SO2 concentration trend, system alarm information, and receive user function start / stop, parameter setting and sequential reset operation instructions. The exception handling module monitors the running status of sequential start and stop in real time, and enables the follow-up start and follow-up stop functions to deal with the incorrect operation of sequential start and stop. If feeding is in progress during backflushing and exchange, the sequential stop function should be activated to stop feeding.

[0014] The pre-grinding mill current, unloading frequency, sequential control program operation status and runtime are obtained, and the reaction delay time during desulfurization is determined. The specific process is as follows: By deploying continuous emission monitoring systems at the inlet and outlet of the desulfurization system, the emissions in the flue gas are collected simultaneously. concentration, Content, humidity, pressure, flow rate, smoke temperature, and flow rate parameters; Equipment operation data: Key status parameters of the reagent preparation and dosing system are acquired in real time through PLC or equipment sensors, including at least the mill operating current, the operating frequency of the discharge valve, and the start / stop status and duration of the sequential control program; Add a unified high-precision timestamp to all collected real-time data streams to ensure that process data and equipment status data across systems are strictly aligned in the time dimension. Store the time-aligned multivariate time series data in a structured historical database to form a long-term running dataset that covers the complete operating conditions and can be used for model training and post-analysis. Based on the aforementioned historical operational dataset, the exit is calculated. The cross-correlation function between concentration changes and feeding actions, or the application of system identification algorithms, can be used to quantitatively analyze and determine the relationship between changes in feeding and the outlet. The number of inherent reaction delay steps between the concentration and the stable response is denoted as the key parameter u, where u step refers to the time window parameter, which refers to the chemical reaction delay time of the SDS desulfurization system. Based on the collected continuous historical operational data and combined with the time series training dataset, a concentration prediction model is established through key time series characteristic analysis. The specific process is as follows: S200. Based on the stored continuous historical operation data, extract the multi-source time-series variables and their corresponding exit times. The dataset contains a complete sample sequence of concentration values. Preprocessing of this dataset includes data cleaning, outlier removal, and normalization. It is then divided into training, validation, and test sets according to a predetermined ratio to form a time-series training dataset for model learning. S201. Construct a deep learning concentration prediction model based on an improved long short-term memory network, including a feature extraction layer: use a one-dimensional convolutional neural network to extract local features from the input multi-dimensional time series, and perform adaptive weight adjustment on the extracted features; Multi-layer LSTM units are used to capture long-term temporal dependencies, and a multi-head attention mechanism is introduced to perform secondary filtering and fusion of the hidden states of the LSTM. Output layer: Maps high-level features to output values ​​for the next u time steps using a fully connected network. Concentration prediction sequence.

[0015] Used to determine the export concentration trend over a historical period in real time, combined with predicted exports. The concentration data is used for dynamic dosing decision analysis, and the specific process is as follows: Based on real-time and historical exits Concentration data, calculate the actual average emissions for the current hour. and the permissible average emissions over the remaining time period. ; Based on the established trend judgment rules, analyze the changing trend of the outlet concentration within the past u-step; When the trend meets the first preset condition, the flue gas prediction model module is invoked to obtain the outlet value for the next u-step. Initial concentration predictions are generated, and subsequent predictions are corrected online using the actual measurements that follow. Based on the corrected prediction results, combined with the emission intensity limit and the second preset condition, feeding start command and feeding stop command are generated.

[0016] Based on preset emission limit rules, decision instructions for feeding are dynamically generated. The judgment logic for feeding stop instructions includes: Using the reaction delay time u as the analysis window length, the output is statistically analyzed in real time over the most recent u consecutive sampling times. The number of time steps during which the measured concentration does not exceed the preset feeding stop threshold D; If the number of steps exceeds half the window length, the current emission trend is determined to meet the first trigger condition for stopping the judgment; When the first triggering condition is met, the concentration prediction model is invoked to obtain the initial prediction sequence of the outlet SO2 concentration for the next u time steps, and the prediction sequence is rolled online using the latest collected real concentration data to generate the corrected future concentration prediction curve. This includes determining the export within the complete future U-step forecast period. The state in which the predicted concentration value is not lower than the predicted I value at the corresponding time. Determine whether the cumulative duration of continuous feeding since the start of this feeding has exceeded the preset maximum safe operating time T; After two judgments, if either condition is met, a feeding stop command is immediately generated and sent to the safety monitoring unit to terminate the drug delivery.

[0017] The logic for determining the start-up conditions includes: Step 1: Calculate the actual cumulative emissions from the start time t0 to the current time tn for the current hour, and obtain the actual average emissions for the current hour. and the permissible average emissions over the remaining time period. ; Step 2: Using the reaction delay time u as the analysis window length, statistically analyze the measured SO2 concentration at the outlet over the past u consecutive sampling times, where the concentration is not lower than... Time steps; If the number of time steps exceeds half the window length, the system emission trend is determined to meet the first trigger condition. When the first trigger condition is met, the prediction model is called to obtain the initial prediction sequence of the outlet SO2 concentration for the next u time steps, and the prediction sequence is rolled online using the latest collected real concentration data to generate the corrected future concentration prediction curve. Step 3: Based on the corrected prediction curve, calculate the predicted I value for each future time point, and determine the exit point within the complete future u-step prediction period. The state in which the predicted concentration value is not lower than the predicted I value at the corresponding time. If step three verification passes, then further check for unplanned and sudden process exchange and backflush signals within the preset process look-ahead time window. The system generates a material feeding start command and sends it to the safety monitoring unit only when the trend triggering condition, the prediction persistence condition, and the process safety condition are met simultaneously. The system operates on a 1-second cycle, reading the flue gas at the CEMS inlet and outlet each cycle. concentration, The system records parameters such as quantity, humidity, pressure, flow rate, flue gas temperature and flow rate, as well as mill current and discharge valve frequency. It also records the current feeding status and current time variables, updates intermediate variables every cycle, and calculates the maximum emission I based on the set one-hour average. And the maximum average emission to Ileft within the remaining time, since the national standard stipulates that the average emission per hour shall not exceed 30 mg / m3, so let the theoretical maximum emission per hour be I, then we can obtain, where t is the length of one hour, calculated as 3600 seconds, lmax is the maximum emission per second, which is 30, in this embodiment it is set to 28, then I is 100800, the specific real-time calculation formula is: Current hourly cumulative average emissions : in, The output at the i-th sampling time concentration, The duration of this concentration value (in seconds). The current time is the time elapsed since the start of this hour; Average emissions allowed for the remaining time : For hourly total emissions limits, This represents the current actual cumulative emissions; Feeding start-up trend condition: Feeding stops when the feeding duration exceeds the set maximum time (10 minutes in this embodiment). The past u-step outlet concentration trend is judged in real time using a set value counting method, with the formula: where I{} is an exponential function, 1 when the condition is true and 0 otherwise, u is the delay step size, Dset is the set lower limit value, and prediction starts when N is true. After receiving the corrected prediction value, Imean and Ileft are updated. Then, feeding stops when N is true at future u-step times.

[0018] To initiate trend counting, if If this is triggered, the material feeding process will begin, where 1{·} is an indicator function. This represents the SO2 concentration at the outlet; it is set to 1 if the condition is true, and 0 otherwise. like If this occurs, it will trigger, where D is the preset lower limit threshold for stopping material feeding. To stop trend counting.

[0019] The concentration monitoring module also includes online correction of subsequent predicted values ​​using actual measured values. The specific method is multiple linear regression correction, and the process is as follows: Let the initial prediction sequence be Obtain the sequence of the first m actual values ​​after prediction. Where m is a fixed number of steps less than u, the prediction error sequence of the first m steps is calculated. ,in m=15; For the k-th step The predicted value, its correction value The calculation formula is: in, , Let be the weighting coefficient of the i-th error term. In the range of 0.2 to 0.05, β is an attenuation factor ranging from 0.8 to 0.95; During material feeding, the mill current value is recorded. If more than 10 of the last 15 values ​​are abnormal, an alarm is triggered. The calculation formula is as follows: ; Where I{·} is an exponential function, which is 1 when the condition is true and 0 otherwise. Let j be the mill current. This refers to the normal operating current range of the mill. For alarm output; The system reads the trigger values ​​of each relevant device through communication with the PLC to determine whether the feeding command is actually executed. When backflushing or switching occurs, feeding will stop if feeding has started; if there is a problem with the mill current during feeding, an alarm will pop up.

[0020] The size of the interval and threshold is set to facilitate comparison. The size of the threshold depends on the amount of sample data and the number of bases set by those skilled in the art for each set of sample data; as long as it does not affect the ratio between the parameter and the quantized value.

[0021] The above formulas are all dimensionless calculations. The formulas are derived from software simulations based on a large amount of collected data to obtain the most recent real-world results. The preset parameters in the formulas are set by those skilled in the art according to the actual situation. In the two embodiments provided in this application, it should be understood that the disclosed system can be implemented in other ways; for example, the device embodiments described above are merely illustrative, and the division of modules is merely a logical functional division. In actual implementation, there may be other division methods, such as multiple modules or components can be combined or integrated into another system, or some features can be ignored or not executed; another point is that the mutual coupling or direct coupling or communication connection shown or discussed may be through some interfaces, and the indirect coupling or communication connection of devices or modules may be electrical, mechanical or other forms. The above description is only a preferred embodiment of the present invention, but the scope of protection of the present invention is not limited thereto. Any equivalent substitutions or modifications made by those skilled in the art within the scope of the technology disclosed in the present invention, based on the technical solution and inventive concept of the present invention, should be covered within the scope of protection of the present invention.

Claims

1. A dynamic data analysis and decision-making system for SDS desulfurization agents in coke oven flue gas, characterized in that, It includes a data acquisition module, a concentration monitoring module, a dosing decision module, a safety monitoring module, and an anomaly handling module. The data acquisition module is used to acquire and integrate flue gas parameter data at the inlet and outlet of the desulfurization system in real time, and to acquire the pre-grinding mill current, unloading frequency, sequential control program running status and running time, and to determine the reaction delay time during desulfurization. The concentration monitoring module, based on the collected continuous historical operating data and combined with the time series training dataset, establishes a flue gas prediction model through key time series characteristic analysis, so as to output a predicted sequence of outlet SO concentration for future time steps. The dosing decision module is used to perform dynamic dosing decision analysis based on the real-time judgment of the outlet concentration trend over a historical period and the predicted outlet SO2 concentration data. Based on the preset emission limit rules, it dynamically generates decision instructions for dosing implementation. The safety monitoring module is used to receive decision instructions from the dynamic decision and control module, execute automated sequential start and sequential stop procedures, control the operation of the feeding equipment, provide a graphical operation interface to display the operating data, outlet SO2 concentration trend, system alarm information, and receive user function start / stop, parameter setting and sequential reset operation instructions. The exception handling module monitors the running status of sequential start and stop in real time, and enables the follow-up start and follow-up stop functions to deal with the incorrect operation of sequential start and stop. If feeding is in progress during backflushing and exchange, the sequential stop function should be activated to stop feeding.

2. The dynamic data analysis and decision-making system for coke oven flue gas SDS desulfurization agents according to claim 1, characterized in that, The pre-grinding mill current, unloading frequency, sequential control program operation status and runtime are obtained, and the reaction delay time during desulfurization is determined. The specific process is as follows: By deploying continuous emission monitoring systems at the inlet and outlet of the desulfurization system, the emissions in the flue gas are collected simultaneously. concentration, Content, humidity, pressure, flow rate, smoke temperature, and flow rate parameters; Equipment operation data: Key status parameters of the reagent preparation and dosing system are acquired in real time through PLC or equipment sensors, including at least the mill operating current, the operating frequency of the discharge valve, and the start / stop status and duration of the sequential control program; Add a unified high-precision timestamp to all collected real-time data streams to ensure that process data and equipment status data across systems are strictly aligned in the time dimension. Store the time-aligned multivariate time series data in a structured historical database to form a long-term running dataset that covers the complete operating conditions and can be used for model training and post-analysis. Based on the aforementioned historical operational dataset, the exit is calculated. The cross-correlation function between concentration changes and feeding actions, or the application of system identification algorithms, can be used to quantitatively analyze and determine the relationship between changes in feeding and the outlet. The inherent reaction delay steps between the concentration and the stable response are denoted as the key parameter u, where u is the time window parameter referring to the chemical reaction delay time of the SDS desulfurization system.

3. The dynamic data analysis and decision-making system for SDS desulfurization agents in coke oven flue gas according to claim 1, characterized in that, Based on the collected continuous historical operational data and combined with the time series training dataset, a concentration prediction model is established through key time series characteristic analysis. The specific process is as follows: S200. Based on the stored continuous historical operating data, extract a complete sample sequence containing the multi-source time-series variables and the corresponding outlet SO2 concentration values ​​at that time. Preprocess the dataset, including data cleaning, outlier removal, and normalization, and divide it into a training set, a validation set, and a test set according to a preset ratio to form a time-series training dataset for model learning. S200. Construct a deep learning concentration prediction model based on an improved long short-term memory network. Includes a feature extraction layer: using a one-dimensional convolutional neural network to extract local features from the input multi-dimensional time series, and adaptively adjusting the weights of the extracted features; Multi-layer LSTM units are used to capture long-term temporal dependencies, and a multi-head attention mechanism is introduced to perform secondary filtering and fusion of the hidden states of the LSTM. Output layer: Maps high-level features to output values ​​for the next u time steps using a fully connected network. Concentration prediction sequence.

4. The dynamic data analysis and decision-making system for SDS desulfurization agents in coke oven flue gas according to claim 1, characterized in that, Used to determine the export concentration trend over a historical period in real time, combined with predicted exports. The concentration data is used for dynamic dosing decision analysis, and the specific process is as follows: Based on real-time and historical exits Concentration data, calculate the actual average emissions for the current hour. and the permissible average emissions over the remaining time period. ; Based on the established trend judgment rules, analyze the changing trend of the outlet concentration within the past u-step; When the trend meets the first preset condition, the flue gas prediction model module is invoked to obtain the outlet value for the next u-step. Initial concentration predictions are generated, and subsequent predictions are corrected online using the actual measurements that follow. Based on the corrected prediction results, combined with the emission intensity limit and the second preset condition, feeding start command and feeding stop command are generated.

5. The dynamic data analysis and decision-making system for coke oven flue gas SDS desulfurization agents according to claim 4, characterized in that, Based on preset emission limit rules, decision instructions for feeding are dynamically generated. The judgment logic for feeding stop instructions includes: Using the reaction delay time u as the analysis window length, the number of time steps in which the measured concentration of SO2 at the outlet does not exceed the preset feeding stop threshold D in the most recent u consecutive sampling moments is counted in real time. If the number of steps exceeds half the window length, the current emission trend is determined to meet the first trigger condition for stopping the judgment; When the first triggering condition is met, the concentration prediction model is invoked to obtain the initial prediction sequence of the outlet SO2 concentration for the next u time steps, and the prediction sequence is rolled online using the latest collected real concentration data to generate the corrected future concentration prediction curve. This includes determining the export within the complete future U-step forecast period. The state in which the predicted concentration value is not lower than the predicted I value at the corresponding time. Determine whether the cumulative duration of continuous feeding since the start of this feeding has exceeded the preset maximum safe operating time T; After two judgments, if either condition is met, a feeding stop command is immediately generated and sent to the safety monitoring unit to terminate the drug delivery.

6. The dynamic data analysis and decision-making system for SDS desulfurization agents in coke oven flue gas according to claim 1, characterized in that, The logic for determining the feeding start-up conditions includes: Step 1: Calculate the actual cumulative emissions from the start time t0 to the current time tn in the current hour, obtain the current average emission value Q, and calculate the maximum and average values ​​of emissions that can be emitted in the remaining time. Step 2: Using the reaction delay time u as the analysis window length, count the number of time steps in the past u consecutive sampling times where the measured concentration of SO2 at the outlet is not lower than I; If the number of time steps exceeds half the window length, the system emission trend is determined to meet the first trigger condition. When the first trigger condition is met, the prediction model is called to obtain the initial prediction sequence of the outlet SO2 concentration for the next u time steps, and the prediction sequence is rolled online using the latest collected real concentration data to generate the corrected future concentration prediction curve. Step 3: Based on the corrected prediction curve, calculate the predicted I value for each future time point, and determine the exit point within the complete future u-step prediction period. The state in which the predicted concentration value is not lower than the predicted I value at the corresponding time. If step three verification passes, then further check for unplanned and sudden process exchange and backflush signals within the preset process look-ahead time window. The system generates a material feeding start command and sends it to the safety monitoring unit only when the trend triggering condition, the prediction persistence condition, and the process safety condition are met simultaneously.

7. The dynamic data analysis and decision-making system for coke oven flue gas SDS desulfurization agents according to claim 1, characterized in that, The concentration monitoring module also includes online correction of subsequent predicted values ​​using actual measured values. The specific method is multiple linear regression correction, and the process is as follows: Let the initial prediction sequence be Obtain the sequence of the first m actual values ​​after prediction. Where m is a fixed number of steps less than u, the prediction error sequence of the first m steps is calculated. ,in m=15; For the k-th step The predicted value, its correction value The calculation formula is: in, , Let be the weighting coefficient of the i-th error term. In the range of 0.2 to 0.05, β is an attenuation factor that ranges from 0.8 to 0.95.

Images