Intelligent control method and system for a claus sulfur recovery plant

Abstract

The application provides a kind of intelligent control method and system of Claus sulfur recovery unit, comprising: calling the decoupling model established in advance, using the real-time data in initial disturbance event record, calculate and separate out the contribution value of methanol flash gas descent and liquid nitrogen tail gas peak respectively on the independent and dynamic influence of total air demand, output two groups of time series data of independent disturbance influence;According to the output of two groups of time series data of independent disturbance influence, by superposition algorithm and combining respective time sequence characteristics, the net change amount of total air demand in future preset period is calculated, and the predicted net air demand change trajectory is obtained;According to the predicted net air demand change trajectory, combined with the current system state, the target air flow adjustment scheme is calculated by using optimization control algorithm, the predicted air demand change is compensated in advance, and the obtained air flow regulation instruction sequence is issued to the actuator of air supply system for execution.

Description

Technical Field

[0001] This invention relates to the field of information technology, and in particular to an intelligent control method and system for a Claus sulfur recovery device. Background Technology

[0002] Background: Sulfur recovery technology, a crucial pillar in the petrochemical and environmental protection sectors, directly impacts the efficiency of industrial waste gas treatment and environmental compliance. Its core lies in the efficient conversion of sulfur-containing components into controllable products through combustion, thereby achieving resource utilization and emission compliance. Claus incinerators have garnered significant attention due to their widespread application in hydrogen sulfide treatment, and their stable operation and efficient conversion capabilities have a decisive impact on the overall economics and sustainability of the process. However, existing sulfur recovery control methods often exhibit limitations when facing complex disturbances. Traditional control strategies rely heavily on static air-fuel ratio settings or simple feedback adjustments, making it difficult to adapt to multi-variable coupling and temporal changes under dynamic operating conditions, leading to fluctuations in sulfur conversion efficiency and even system instability. Especially in specific concurrent disturbance scenarios, such as the superposition of a step decrease in methanol flash vapor flow and a peak in liquid nitrogen scrubbing tail gas, air demand exhibits opposite trends, making it difficult for conventional methods to quickly decouple and respond accurately, revealing their shortcomings in predictability and adaptability. Severe fluctuations in oxygen content easily lead to incomplete combustion of sulfur species, while furnace temperature fluctuations disrupt reaction equilibrium; the combined effect of these two factors makes it difficult to maintain a stable sulfur-to-SO2 conversion rate. Therefore, the key issue this study focuses on is how to construct a predictive control strategy to decouple the coupling effects of oxygen content oscillation and furnace temperature fluctuation in the disturbance scenario of simultaneous methanol flash vapor step decrease and liquid nitrogen wash tail gas peak. Summary of the Invention

[0003] This invention provides an intelligent control method for a Claus sulfur recovery device, mainly comprising:

[0004] Real-time monitoring of methanol flash vapor flow and liquid nitrogen wash tail gas flow; identification of disturbance scenarios where methanol flash vapor flow drops sharply and liquid nitrogen wash tail gas peaks occur simultaneously through pre-set thresholds and logic; if such a scenario is identified, real-time data of the scenario at that moment and within the subsequent preset time period are collected to form an initial disturbance event record.

[0005] By calling the pre-established decoupling model and using real-time data from the initial disturbance event record, the independent and dynamic contributions of the methanol flash vapor decline and liquid nitrogen wash tail gas peak to the total air demand are calculated and separated, and the time series data of the two independent disturbance effects are output.

[0006] Based on the time series data of the two independent disturbances, the net change in total air demand within a preset period is calculated by superposition algorithm and combined with their respective time series characteristics, thus obtaining the predicted net air demand change trajectory.

[0007] Based on the predicted trajectory of changes in net air demand and combined with the current system status, the target air flow adjustment scheme is calculated using an optimized control algorithm to compensate for the predicted changes in air demand in advance. The resulting air flow adjustment command sequence is then sent to the actuators of the air supply system for execution.

[0008] The system collects and executes the actual oxygen content, furnace temperature, and products after adjustment in real time. The output is a data stream containing control execution records and real-time performance feedback. It compares the predicted clean air demand trajectory with the set control target to evaluate the actual effect of the target air flow adjustment scheme. If the actual effect deviates significantly from the expectation, it updates the decoupling model parameters and / or optimizes the control algorithm weights online, and outputs the adjusted model parameter set.

[0009] Furthermore, the methanol flash vapor flow rate and liquid nitrogen wash tail gas flow rate are monitored in real time. A disturbance scenario involving a step decrease in methanol flash vapor flow rate and a peak in liquid nitrogen wash tail gas flow rate is identified using pre-set thresholds and logic. If such a scenario is identified, real-time data for that moment and subsequent preset time periods are collected to form an initial disturbance event record. This includes: continuously collecting methanol flash vapor pressure measurements and liquid nitrogen wash tail gas pressure measurements at a preset sampling period using a pressure sensing monitoring unit; smoothing the measurements using a moving average filter to obtain a pressure processing value; and using a dual threshold comparator to determine a pressure change marker for the pressure processing value. The methanol flash vapor flow rate and liquid nitrogen wash tail gas flow rate measurements are also collected in real time using a gas flow sensing monitoring unit. The flow rate measurements within a preset time period are recorded in the built-in memory of the sensing monitoring unit. A random forest algorithm is used to detect abrupt change points in the flow rate measurements to obtain the moment of the step decrease in methanol flash vapor flow rate and the moment of the peak in liquid nitrogen wash tail gas flow rate. Based on the temperature measurements of methanol flash vapor and liquid nitrogen wash tail gas collected by the temperature sensing and monitoring unit, wavelet transform decomposition is performed on the temperature measurements to obtain temperature feature values. A recurrent neural network is then used to extract features from these temperature feature values ​​to obtain a temperature change feature vector. A multidimensional feature matrix is ​​constructed for the pressure mutation marker, the flow step drop time point, the flow peak time point, and the temperature change feature vector. A feature fusion processor is used to perform time-series alignment on the feature matrix. A scene identification dataset is established based on the multidimensional feature matrix. This scene identification dataset includes pressure mutation markers, flow mutation time points, and temperature feature vectors. A fixed-duration time window constraint is applied to the dataset, and disturbance event data records are generated based on this dataset.

[0010] Furthermore, a pre-established decoupling model is invoked, and real-time data from the initial disturbance event records is used to calculate and separate the independent and dynamic contributions of the methanol flash vapor decrease and the liquid nitrogen wash tail gas peak to the total air demand. Two sets of time-series data on independent disturbance effects are output, including: normalizing the methanol flash vapor flow rate data and liquid nitrogen wash tail gas flow rate data according to the initial disturbance event records to obtain standardized flow rate data; using a sliding median filter to denoise the standardized flow rate data to obtain smoothed flow rate data; performing orthogonal transformation on the smoothed flow rate data using principal component analysis to obtain eigenvalue matrices and eigenvector matrices; sorting the eigenvectors according to the eigenvalue magnitude; and selecting the top two eigenvectors to construct an orthogonal basis space; mapping the smoothed flow rate data to the orthogonal basis space using a projection operator to obtain the methanol flash vapor load coefficient and the liquid nitrogen wash tail gas load coefficient; and smoothing the load coefficients using spline interpolation to obtain the load time series. A recurrent neural network is used to extract dynamic features from the load time series. The recurrent neural network includes an input layer, a hidden layer, and an output layer; the input is the load time series, and the output is a dynamic feature vector. The correlation between the components of the dynamic feature vector is calculated using a mutual information entropy calculator, and an independence evaluation matrix is ​​constructed. This matrix is ​​then diagonalized to obtain the independent contribution. Based on these independent contribution values, the load time series are weighted and combined to obtain the independent impact time series of methanol flash vapor on total air demand and the independent impact time series of liquid nitrogen scrubbing exhaust gas on total air demand.

[0011] Furthermore, the real-time initial disturbance event data stream is dynamically intercepted to generate a time-aligned input matrix. The input includes a decoupled model containing a dual-channel feature extraction structure based on a gated loop unit, and outputs the dynamic influence coefficients of methanol and liquid nitrogen components, respectively. Based on the product of the methanol component dynamic influence coefficient and the total air demand, an independent contribution sequence for the decrease in methanol flash vapor is generated. Similarly, based on the product of the liquid nitrogen component dynamic influence coefficient and the total air demand, an independent contribution sequence for the peak value of liquid nitrogen scrubbing exhaust gas is generated. The independent contribution sequences are then timestamped to align them, outputting two sets of time series data of influence factors with the same time resolution. This includes: replacing outliers in the initial disturbance event records with medians using a data preprocessing unit; sampling the event records at fixed time intervals to obtain raw sampled data; and using a sliding window to intercept the raw sampled data in real time to obtain time-series intercepted data. The time-series intercepted data is then aligned using a time stamp aligner to obtain an aligned data matrix, and a normalization calculation unit performs zero-mean normalization on the aligned data matrix to obtain a standardized data matrix. A gated recurrent unit (GRU) is used to perform dual-channel feature extraction on the standardized data matrix. The number of hidden layer neurons in the GRU is the same as the dimension of the standardized data matrix. The methanol component feature vector is extracted through the left channel, and the liquid nitrogen component feature vector is extracted through the right channel. A feature transformation unit performs a linear transformation on the methanol and liquid nitrogen component feature vectors, and normalization is applied to obtain the dynamic influence coefficients of the methanol and liquid nitrogen components. The total air demand is obtained from the process parameter database by the air demand calculation unit. A multiplier is used to calculate the product of the dynamic influence coefficient of the methanol component and the total air demand, and the product of the dynamic influence coefficient of the liquid nitrogen component and the total air demand, respectively. A cubic spline interpolator is used to interpolate the product results to obtain independent contribution sequences of methanol flash vapor and liquid nitrogen wash tail gas with consistent time resolution.

[0012] Furthermore, based on the time-series data of the two sets of independent disturbances, the net change in total air demand within a preset future period is calculated using an overlay algorithm combined with their respective time-series characteristics, resulting in a predicted net air demand change trajectory. This includes: extracting methanol flash vapor disturbance features and liquid nitrogen wash tail gas disturbance features from the independent disturbance impact sequences using a time-series feature extractor. These disturbance features include the occurrence time, duration, and rate of change. A long short-term memory network is used to identify the disturbance features, obtaining a disturbance feature vector. A weight calculator is used to calculate the weights of the methanol flash vapor sequence and the liquid nitrogen wash tail gas sequence based on the disturbance feature vector. A sequence overlay calculator is used to perform a weighted overlay operation on the two sets of independent impact sequences. A Butterworth low-pass filter is used to filter the overlay sequence to obtain the net impact sequence. For the disturbance feature vector, an autoregressive moving average is used to construct a time-series prediction basis function. The net impact sequence is then extended out-of-sample within a preset time window to obtain a predicted sequence for the total air demand change. The total air demand change prediction sequence is decomposed using a trend extractor to obtain a trend component sequence and a fluctuation component sequence. An exponential smoothing algorithm is then used to smooth the trend component sequence to obtain a smoothed trend sequence. The autocorrelation function of the fluctuation component sequence is calculated using a delay estimator, and the delay compensation amount is determined based on the autocorrelation peak position. A phase compensator is then used to perform delay correction on the smoothed trend sequence to obtain the predicted trajectory of the net change in total air demand.

[0013] Furthermore, based on the predicted trajectory of net air demand changes and combined with the current system status, an optimized control algorithm is used to calculate the target airflow adjustment scheme to compensate for the predicted changes in air demand in advance. The obtained airflow adjustment command sequence is then sent to the actuators of the air supply system for execution. This includes: obtaining a temperature acquisition sequence by collecting real-time temperature data from the Claus furnace through a multi-point temperature sensing monitoring unit; obtaining an oxygen content acquisition sequence by collecting exhaust gas oxygen content data through an oxygen content detection unit; and using a Kalman filter to perform noise reduction processing on the temperature acquisition sequence and oxygen content acquisition sequence to obtain a status monitoring sequence. A data alignment unit is used to align the status monitoring sequence with time stamps, and wavelet transform is used to perform multi-scale decomposition of the status monitoring sequence to obtain a feature coefficient sequence. The feature coefficient sequence is then time-series superimposed with the predicted net air demand change trajectory to obtain a fused feature sequence. A deep reinforcement learning unit is used to extract features from the fused feature sequence. The deep reinforcement learning unit includes a state space, an action space, and a reward function, and outputs an airflow adjustment parameter matrix. The constraint optimization unit extracts adjustment constraints from the airflow adjustment parameter matrix. These constraints include an upper limit for flow, a lower limit for flow, and a rate of change limit. A dynamic programming algorithm is used to calculate the optimal adjustment sequence that satisfies these constraints. The sequence decomposition unit samples the optimal adjustment sequence over time, extracting the adjustment time points and corresponding target flow values ​​to generate an airflow adjustment command sequence. The command conversion unit converts the airflow adjustment command sequence into digital signals, generating standard control commands for the actuators. These standard control commands are then sent to the air supply actuators via an industrial bus interface.

[0014] Furthermore, an airflow adjustment coefficient is generated based on the predicted net air demand change trajectory and the current system state. This coefficient is then fused with the oxygen content gradient to output a target flow adjustment time series. This includes: extracting amplitude and rate of change data from the predicted net air demand change trajectory using a change analysis unit; smoothing the amplitude and rate of change data using an exponentially weighted moving average calculator, where the calculator's weight coefficients decay based on data time intervals to obtain a smoothed change curve; standardizing temperature, pressure, and gas component data using a system parameter processing unit to obtain a state parameter vector; performing feature mapping between the state parameter vector and the smoothed change curve using a support vector regression calculation unit to obtain an initial flow adjustment coefficient; performing fixed-time-window differential operations on the oxygen content sampling data using a gradient calculation unit; segmenting the differential data using an adaptive threshold segmenter based on variance analysis to obtain an oxygen content gradient curve; and calculating the combined weight of the initial flow adjustment coefficient and the oxygen content gradient curve using a weighted superposition operation based on the state parameter vector to obtain a comprehensive adjustment parameter. The feedback compensation unit calculates compensation coefficients based on historical adjustment effects, and dynamically compensates the comprehensive adjustment parameters to obtain corrected adjustment parameters. The time-series planning unit then discretizes the corrected adjustment parameters according to a preset sampling period to generate an adjustment time-series sequence containing the adjustment time and target flow rate.

[0015] Furthermore, the target flow adjustment timing sequence after actuator response delay compensation is verified against the furnace temperature change rate constraint. If the furnace temperature change rate exceeds a preset range, the target flow adjustment timing sequence is recalculated. The verified timing sequence is then split into air flow adjustment commands, including: acquiring actuator response delay data through a response delay measurement unit; performing statistical analysis on the response delay data using a moving average calculator; and extracting features from the delay data sequence using a long short-term memory network (LSTM), where the network input is the response delay time sequence and the output is a delay feature vector. A furnace temperature data sequence is collected by a temperature detection unit, and a time difference operation is performed on the furnace temperature data sequence using a difference calculator. The difference result is then normalized to obtain a furnace temperature change rate curve. A threshold judge compares the furnace temperature change rate curve with preset upper and lower threshold values. If it exceeds the preset threshold range, a recalculation flag is output. A constraint optimization calculator receives the recalculation flag and regenerates the target flow adjustment curve based on the temperature constraint and the delay feature vector. The constraint includes an upper and lower limit value for the temperature change rate. The target flow rate adjustment curve is dynamically fitted using the recursive least squares method to obtain a smooth adjustment curve. A constraint checker then determines whether the smooth adjustment curve meets temperature constraints. The instruction generation unit performs time discretization on the smooth adjustment curve that meets the constraints, and a flow rate classifier quantizes the discrete points according to preset flow rates to generate an air flow rate adjustment instruction sequence.

[0016] Furthermore, the actual oxygen content, furnace temperature, and products after real-time acquisition and adjustment are analyzed, and the output is a data stream containing control execution records and real-time performance feedback. This data stream is compared with the predicted net air demand trajectory and the set control target to evaluate the actual effect of the target airflow adjustment scheme. If the actual effect deviates significantly from the expectation, the decoupled model parameters and / or control algorithm weights are updated online, and the output is an adjusted model parameter set. This set includes: real-time acquisition of raw oxygen content, furnace temperature, and sulfur dioxide concentration data via a multi-parameter acquisition device; denoising and smoothing of the raw data using a data preprocessing unit to obtain a processed data sequence; and extraction of temporal features from the processed data sequence using a Long Short-Term Memory (LSTM) network to obtain a real-time performance index vector. A control target calculation unit obtains the target ranges for oxygen content, furnace temperature, and sulfur dioxide concentration from the process database, and a performance index evaluator calculates the deviation between the real-time performance index vector and the target range. A trajectory comparison calculator calculates the goodness of fit between the real-time performance index vector and the predicted net air demand change trajectory. The goodness of fit calculation is based on Euclidean distance, and a difference detector compares the goodness of fit value with a preset threshold. A Kalman filter is used to dynamically track the deviation and fit values, generating a deviation feature vector. A threshold detector then judges the range of this deviation feature vector. If the deviation feature vector exceeds a preset range, a parameter optimization unit calculates a decoupling matrix correction based on the deviation feature vector. This decoupling matrix is ​​used to separate the influence contributions of methanol flash vapor and liquid nitrogen wash tail gas. The decoupling matrix is ​​iteratively updated by a gradient optimization unit, and the control weight matrix is ​​optimized and adjusted using a backpropagation algorithm. An updated parameter set is then generated by a parameter integration unit.

[0017] This invention provides an intelligent control system for a Claus sulfur recovery device, mainly comprising:

[0018] The real-time monitoring and scene recognition module is used to monitor the methanol flash vapor flow rate and liquid nitrogen wash tail gas flow rate in real time. It identifies disturbance scenarios where the methanol flash vapor flow rate drops sharply and the liquid nitrogen wash tail gas flow rate peaks simultaneously through preset thresholds and logic. If such a scenario is identified, real-time data of the scenario at that moment and within the subsequent preset time period are collected to form an initial disturbance event record.

[0019] The decoupling model calculation module is used to call a pre-established decoupling model, utilize real-time data from the initial disturbance event record, calculate and separate the independent and dynamic contributions of the methanol flash vapor decline and liquid nitrogen scrubbing tail gas peak to total air demand, and output two sets of time series data of independent disturbance effects, including:

[0020] Based on the initial disturbance event records, the methanol flash vapor flow data and liquid nitrogen wash tail gas flow data are normalized to obtain standardized flow data. The standardized flow data is then denoised using a sliding median filter to obtain smooth flow data.

[0021] Principal component analysis is used to perform orthogonal transformation on the smoothed flow data to obtain eigenvalue matrix and eigenvector matrix. The eigenvectors are sorted according to the magnitude of the eigenvalues, and the first two eigenvectors are selected to construct an orthogonal basis space.

[0022] The smoothed flow data is mapped to the orthogonal basis space by the projection operator to obtain the methanol flash vapor load coefficient and the liquid nitrogen wash tail gas load coefficient. Spline interpolation is used for smoothing to obtain the load time series.

[0023] A recurrent neural network is used to extract dynamic features from the load time series to obtain a dynamic feature vector. The correlation between the components of the dynamic feature vector is calculated using a mutual information entropy calculator to construct an independence evaluation matrix. The independence evaluation matrix is ​​then diagonalized to obtain the independent contribution.

[0024] The load time series are weighted and combined according to the independent contribution to obtain the independent impact time series of methanol flash vapor on total air demand and the independent impact time series of liquid nitrogen scrubbing tail gas on total air demand.

[0025] The net air demand forecasting module is used to calculate the net change in total air demand over a future preset period based on the time series data of two independent disturbances, using an overlay algorithm and combining their respective time series characteristics, and thus obtain the predicted net air demand change trajectory.

[0026] The optimization control calculation module is used to calculate the target air flow adjustment scheme based on the predicted trajectory of changes in net air demand and the current system status, using an optimization control algorithm to compensate for the predicted changes in air demand in advance, and to send the obtained air flow adjustment command sequence to the actuator of the air supply system for execution.

[0027] The effect evaluation and model update module is used to collect the actual oxygen content, furnace temperature and products after the adjustment in real time. The output is a data stream containing control execution records and real-time performance feedback. It compares the predicted net air demand trajectory with the set control target to evaluate the actual effect of the target air flow adjustment scheme. If there is a significant deviation between the actual effect and the expectation, the decoupled model parameters and / or the control algorithm weights are updated online, and the output is the adjusted model parameter set.

[0028] The technical solutions provided by the embodiments of the present invention may include the following beneficial effects:

[0029] This invention discloses an intelligent control method for a Claus sulfur recovery unit. The method identifies and records specific disturbance scenarios by real-time monitoring of methanol flash vapor and liquid nitrogen scrubbing tail gas flow rates. A decoupling model is used to separate the independent effects of the two gases on air demand, predicting future changes in air demand. Based on the prediction results and system status, the optimal air flow adjustment scheme is calculated and executed. Simultaneously, this invention collects adjusted performance indicators in real time, evaluates the control effect, and updates model parameters and algorithm weights online. This method can effectively cope with complex disturbance scenarios, compensate for changes in air demand in advance, achieve precise control of the Claus sulfur recovery unit, improve system response speed and accuracy, and thus optimize the production process and product quality. Attached Figure Description

[0030] Figure 1 This is a flowchart of an intelligent control method for a Claus sulfur recovery device according to the present invention. Detailed Implementation

[0031] To enable those skilled in the art to better understand the technical solutions in this specification, the technical solutions in the embodiments of this specification will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of this specification, and not all embodiments. Based on the embodiments in this specification, all other embodiments obtained by those skilled in the art without creative effort should fall within the scope of protection of this specification.

[0032] like Figure 1 The intelligent control method for a Claus sulfur recovery device in this embodiment may specifically include:

[0033] S101. During the operation of the Claus sulfur recovery unit, the methanol flash vapor flow rate and liquid nitrogen wash tail gas flow rate are monitored in real time using sensors. The disturbance scenario of a step drop in methanol flash vapor flow rate and a peak in liquid nitrogen wash tail gas flow rate is identified by a preset threshold logic. If the scenario is detected, the current state is locked and relevant real-time data is collected to generate an initial disturbance event record.

[0034] S1011. In this embodiment of the invention, the pressure values ​​of methanol flash vapor and liquid nitrogen wash tail gas are continuously collected by a pressure sensor. Real-time pressure data is obtained with a sampling period of 100 milliseconds. The collected pressure values ​​are subjected to moving average filtering with a window length of 5 sampling points to generate smoothed pressure values. Then, a dual threshold comparator is used to determine pressure mutations, where the high threshold is set to 2.5 MPa and the low threshold is set to 0.8 MPa. When the pressure exceeds the threshold range for three consecutive cycles, it is marked as a pressure mutation, and the mutation time point is recorded.

[0035] S1012. Simultaneously, the flow rate of methanol flash vapor and liquid nitrogen wash tail gas are monitored in real time using flow sensors. Flow data within a 60-second time window is recorded. A random forest algorithm is used to detect abrupt changes in flow values. This algorithm is based on 50 decision trees, each with a maximum depth of 6 layers. It extracts features such as flow mean, standard deviation, and peak factor. It detects the step drop in methanol flash vapor flow rate from 25 cubic meters per hour to below 10 cubic meters per hour, and the peak value of liquid nitrogen wash tail gas flow rate rising to 35 cubic meters per hour. Furthermore, temperature data of the two gases are collected using temperature sensors. The temperature values ​​are decomposed using a three-layer wavelet transform. The db4 wavelet basis function is used to extract high-frequency and low-frequency coefficients. Then, a long short-term memory network containing 128 hidden layer neurons is used to extract temperature features and generate a temperature change feature vector.

[0036] S1013. After obtaining the pressure mutation marker, flow mutation time, and temperature feature vector, a multi-dimensional feature matrix is ​​constructed. Each row represents the state feature at a time point and contains 12 dimensions of data, including the pressure mutation marker, methanol flash vapor flow rate, liquid nitrogen wash tail gas flow rate, and four-dimensional temperature feature vectors of the two gases. The matrix is ​​time-series aligned by a feature fusion unit, and a fixed time window constraint of 180 seconds is applied with a step size of 30 seconds to generate disturbance event data records containing scene identifiers.

[0037] In this embodiment of the invention, the flow rate changes of methanol flash vapor and liquid nitrogen scrubbing tail gas exhibit specific patterns. For example, in a chemical plant scenario, when the methanol flash vapor pressure drops from 2.2 MPa to 0.6 MPa, a pressure surge flag is triggered. Simultaneously, the flow rate decreases from 23 cubic meters per hour to 8 cubic meters per hour, while the liquid nitrogen scrubbing tail gas flow rate increases to 32 cubic meters per hour. Temperature data reflects a rapid drop in methanol flash vapor temperature of 15 to 25 degrees Celsius. Based on these characteristics, the system generates disturbance event records, accurately marking abnormal time points.

[0038] It is understood that the embodiments of the present invention do not impose too many restrictions on sensor types and algorithm parameters, which can be adjusted by technicians according to the actual scenario to ensure the accuracy of disturbance scene recognition.

[0039] S102. Call the pre-trained decoupling model and use real-time data from the initial disturbance event record to analyze the independent dynamic impact of the decrease in methanol flash vapor flow rate and the peak flow rate of liquid nitrogen wash tail gas on total air demand, generating two sets of time series data to characterize their respective contribution values.

[0040] In this embodiment of the invention, the decoupling process first extracts methanol flash vapor flow rate and liquid nitrogen scrubbing tail gas flow rate data from the initial disturbance event record. These data are acquired in real time by sensors and typically contain flow fluctuations and noise. To ensure analytical accuracy, the system preprocesses the raw flow data by mapping it to a standardized range of 0 to 1 using a normalization calculation unit to eliminate dimensional differences. For example, in chemical production, the methanol flash vapor flow rate may vary between 20 and 30 cubic meters per hour, while the liquid nitrogen scrubbing tail gas flow rate ranges from 15 to 25 cubic meters per hour; normalization facilitates unified processing. Subsequently, a sliding median filter is used to denoise the standardized data, with the filter window length set to 7 sampling points, effectively smoothing abrupt noise in sensor measurements, such as abnormal flow rate jumps caused by equipment vibration.

[0041] Furthermore, principal component analysis (PCA) is used to orthogonally transform the smoothed flow data, generating eigenvalue and eigenvector matrices. The first two principal components are selected by eigenvalue ranking to construct an orthogonal basis space, and the smoothed flow data is projected onto this space. Loading coefficients for methanol flash vapor and liquid nitrogen scrubbing tail gas are calculated separately, and then spline interpolation is used for smoothing to generate a continuous load time series. In this embodiment, PCA aims to capture the main direction of data variation. Practical application shows that the first principal component explains approximately 75% of the variance, and the second principal component contributes 20%, together covering more than 95% of the data information. After projection, the loading coefficient of methanol flash vapor on the first principal component is approximately 0.85, and the loading coefficient of liquid nitrogen scrubbing tail gas on the second principal component is approximately 0.78, reflecting the orthogonal characteristics of the influence of the two gases. The spline interpolation uses a cubic polynomial to ensure that the time resolution of the load series is improved to 1 second, preserving the details of dynamic changes.

[0042] Next, the system inputs the load time series into a recurrent neural network for dynamic feature extraction. This network consists of an input layer, two hidden layers, and an output layer. Each hidden layer has 128 neurons. A 16-dimensional dynamic feature vector is extracted through forward propagation. Then, a mutual information entropy calculator is used to evaluate the correlation between the components of the feature vector, constructing an independence evaluation matrix and diagonalizing it to obtain the independent contributions of methanol flash vapor and liquid nitrogen scrubbing tail gas. The load sequences are then weighted and combined according to the contribution values ​​to output two sets of independent impact time series. In a real-world scenario, mutual information entropy analysis shows that the first 8 dimensions of the methanol flash vapor feature vector and the last 8 dimensions of the liquid nitrogen scrubbing tail gas feature vector have low correlation, with an average mutual information value of less than 0.15, indicating strong independence after decoupling. For example, when the methanol flash vapor flow rate decreases by 25%, its independent impact on air demand decreases by 18% within 5 minutes, while when the liquid nitrogen scrubbing tail gas flow rate increases by 30%, air demand increases by 12%. The superposition error between the two is less than 3%, verifying the accuracy of the method.

[0043] Furthermore, in this embodiment of the invention, to further improve decoupling accuracy, the system can dynamically extract real-time disturbance event data streams, generate time-aligned input matrices, and input them into a dual-channel decoupling model based on a gated recurrent unit. This model extracts methanol flash vapor features through the left channel and liquid nitrogen wash tail gas features through the right channel, outputting dynamic influence coefficients respectively. The number of hidden layer neurons in the gated recurrent unit matches the dimension of the input matrix; for example, when the input is standardized data at 60 time points, 60 neurons are configured. Information flow is controlled through update and reset gates to capture the temporal characteristics of flow rate changes.

[0044] After generating the dynamic influence coefficients, the system uses a multiplier to calculate the product of the methanol component's dynamic influence coefficient and the total air demand, obtaining the independent contribution sequence of the methanol flash vapor decrease. Simultaneously, it calculates the product of the liquid nitrogen component's dynamic influence coefficient and the total air demand, obtaining the independent contribution sequence of the liquid nitrogen scrubbing tail gas peak. Cubic spline interpolation ensures consistent time resolution for both sequences, outputting unified time series data. In the chemical industry scenario, the baseline total air demand is 100 cubic meters per hour. When the methanol flash vapor flow rate decreases from 26 cubic meters per hour to 12 cubic meters per hour, its dynamic influence coefficient is -0.35, with a contribution of -35 cubic meters per hour. Conversely, when the liquid nitrogen scrubbing tail gas flow rate increases from 18 cubic meters per hour to 32 cubic meters per hour, the influence coefficient is 0.28, with a contribution of 28 cubic meters per hour. After interpolation, the sequence resolution reaches 100 milliseconds, clearly demonstrating the dynamic difference between the rapid establishment of the methanol influence and the gradual formation of the liquid nitrogen influence.

[0045] In this embodiment of the invention, the data preprocessing stage also includes outlier detection and replacement. For example, when the methanol flash vapor flow rate jumps to 150 cubic meters per hour due to a sensor malfunction, exceeding the normal range, a 5-point median filter is used to replace the outlier. A 60-second sliding window with a 10-second step size is used for time series extraction, and time deviations are aligned using linear interpolation, with the maximum deviation controlled within 50 milliseconds. Feature extraction results from the dual-channel model show that dimensions 3 to 5 of the methanol component feature vector reflect the flow rate change rate, while dimensions 7 to 9 of the liquid nitrogen component feature vector highlight the peak duration, enhancing the targeted nature of the decoupling.

[0046] It is understood that the embodiments of the present invention do not strictly limit the length of the filtering window or the number of neural network layers. Technicians can adjust the parameters according to actual needs to adapt to the disturbance characteristics of different production scenarios.

[0047] S103. Based on the two sets of independent disturbance impact time series data of methanol flash vapor and liquid nitrogen wash tail gas, and combined with their respective time series characteristics, calculate the net change in total air demand over a future period using an overlay algorithm, and generate the predicted trajectory of total air demand change.

[0048] In this embodiment of the invention, the system first utilizes a temporal feature extractor to extract key features from the independent impact sequences of methanol flash vapor and liquid nitrogen scrubbing tail gas, including the time of disturbance occurrence, duration, and rate of change. These features reflect the dynamic patterns of the impact of the two gases on air demand. For example, in chemical production, the methanol flash vapor flow rate may suddenly drop from 25 cubic meters per hour to 10 cubic meters per hour, with a rate of change of approximately -0.3 cubic meters per hour per second, lasting between 120 and 180 seconds, while the liquid nitrogen scrubbing tail gas flow rate slowly increases from 18 cubic meters per hour to 32 cubic meters per hour, with a rate of change of approximately 0.1 cubic meters per hour per second, lasting up to 300 to 360 seconds. The extracted features are identified through a long short-term memory network to generate a disturbance feature vector.

[0049] Furthermore, the system employs a weight calculator to determine the weights of the methanol flash vapor sequence and the liquid nitrogen scrubbing tail gas sequence based on the perturbation feature vector. A sequence superposition calculator then performs a weighted superposition operation on the two independent impact sequences. Subsequently, a Butterworth low-pass filter is used to smooth the superposition result, generating a net impact sequence to characterize the comprehensive change in total air demand. In this embodiment, the Long Short-Term Memory (LSTM) network is configured with two hidden layers, each containing 64 neurons. Key temporal information is filtered through forget gates and input gates, resulting in a 16-dimensional output feature vector. Weight calculation is based on the mutability and persistence scores of the feature vectors. For example, methanol flash vapor has a high mutability score of 0.85 and a low persistence score of 0.35, while liquid nitrogen scrubbing tail gas has a mutability score of only 0.25 and a persistence score as high as 0.92. Based on this, the weight of the methanol sequence is calculated to be 0.6, and the weight of the liquid nitrogen sequence is 0.4. After superposition, the filter cutoff frequency is set to 0.1 Hz to effectively filter out high-frequency noise and ensure the smoothness and reliability of the net impact sequence.

[0050] Based on this, the system constructs a time-series prediction basis function using an autoregressive moving average, extends the net impact sequence out of the sample within a 480-second prediction window, generates a prediction sequence for the total air demand change, and uses wavelet decomposition to separate it into trend and fluctuation components. Finally, delay correction is applied to generate the final predicted trajectory for the net change in total air demand. In practical applications, the autoregressive moving average is set to an autoregressive order of 4 and the moving average order to 2. The basis function is trained using historical data from the past 10 minutes. The prediction results show that the decrease in methanol flash vapor leads to a 15 cubic meter per hour reduction in air demand in the first 180 seconds, while the peak value of liquid nitrogen scrubbing gas increases by 12 cubic meters per hour between 240 and 360 seconds. The wavelet decomposition uses a three-level structure, selecting the db6 wavelet basis function. The trend component reflects long-term trends, while the fluctuation component captures short-term disturbance characteristics. The delay estimator determines the delay compensation amount by calculating the autocorrelation function of the fluctuation component sequence. It finds that the autocorrelation coefficient reaches 0.75 at 180 seconds, indicating that the disturbance period is about 3 minutes. The phase compensator corrects the trend sequence accordingly to ensure that the predicted trajectory is consistent with the actual response.

[0051] In this embodiment of the invention, the prediction process fully considers the temporal differences between the two gas disturbances. The sudden drop in methanol flash vapor typically affects air demand rapidly, while the gradual rise in liquid nitrogen scrubbing exhaust gas exhibits a longer response period. Through weighted superposition and filtering, the system can smoothly integrate the two influences, avoiding prediction errors caused by timing misalignment. In a chemical plant case, the predicted trajectory shows that the negative impact of methanol flash vapor appears 120 seconds earlier, while the positive impact of liquid nitrogen scrubbing exhaust gas is delayed by 60 seconds. The corrected trajectory shows a greater than 95% agreement with actual demand changes.

[0052] Furthermore, to enhance the robustness of the forecast, the system applies an exponential smoothing algorithm to the trend component sequence, with a smoothing coefficient set to 0.3. This preserves the main trend while effectively suppressing short-term fluctuations. In actual operation, this method allows air demand forecasts to reflect the impact of disturbances in advance. It is understood that this embodiment of the invention does not impose fixed limitations on the forecast window length or filtering parameters; technicians can flexibly adjust them according to specific process requirements to optimize forecast accuracy.

[0053] S104. Based on the predicted trajectory of net air demand changes and combined with the current system status of the Claus sulfur recovery unit, a target air flow adjustment scheme is generated through optimized control algorithm to compensate for changes in air demand in advance, and the adjustment command is sent to the actuator.

[0054] In this embodiment of the invention, an airflow regulation scheme is first calculated based on the predicted trajectory of net air demand changes, combined with real-time furnace temperature and oxygen content data. The furnace temperature is collected by multiple temperature sensors distributed across different locations within the reactor, maintaining a range of 1050 to 1150 degrees Celsius during normal operation, with a sampling period of 100 milliseconds. The tail gas oxygen content is obtained by the outlet detection unit, with a normal range of 2% to 3%, and a sampling period of 200 milliseconds. To eliminate measurement noise, a Kalman filter is used to process the temperature and oxygen content data sequences. Through state prediction and measurement update iterations, the temperature noise is reduced from ±5 degrees Celsius to ±2 degrees Celsius, and the oxygen content noise is reduced from ±0.2% to ±0.1%. Subsequently, a data alignment unit uses linear interpolation to unify the two sets of data to a 100-millisecond time resolution. Wavelet transform is performed based on the db4 basis function for a three-level decomposition, extracting low-frequency trends and high-frequency fluctuation features to form a feature coefficient sequence, which is then superimposed with the predicted trajectory to generate a fused feature sequence.

[0055] S1041. Further, the deep reinforcement learning unit extracts features from the fused feature sequence to generate an air flow regulation parameter matrix, and calculates the optimal regulation sequence based on a dynamic programming algorithm, ultimately transforming it into executable control commands. In this embodiment of the invention, the state space of the deep reinforcement learning unit contains 16 dimensions, such as furnace temperature, oxygen content, and predicted demand changes. The action space defines the air flow regulation range as -20% to +20%. The reward function integrates temperature deviation, oxygen content stability, and regulation cost, and optimizes the strategy through multiple rounds of trial and error. The output regulation parameter matrix contains adjustment timing and magnitude information. The constraint optimization unit extracts the conditions of an upper limit of 120 cubic meters per hour, a lower limit of 80 cubic meters per hour, and a change rate not exceeding 2 cubic meters per hour per second. The dynamic programming algorithm generates a stepped regulation sequence with the goal of minimizing the deviation. For example, within 5 minutes of a predicted reduction of 15 cubic meters per hour in air demand, the adjustment is carried out in stages: a reduction of 5 cubic meters per hour in the first minute, a reduction of 6 cubic meters per hour in the next two minutes, and a reduction of 4 cubic meters per hour in the last two minutes. The command conversion unit samples the sequence into 30 control points at 10-second intervals and sends them to the air supply actuator via the Modbus bus as a 4-20 mA standard signal.

[0056] Furthermore, an optimization scheme integrating airflow regulation coefficients and oxygen content gradients is employed. The amplitude and rate of change data are extracted from the predicted trajectory; for example, if air demand decreases from 100 cubic meters per hour to 85 cubic meters per hour, the rate of change is approximately -0.05 cubic meters per hour per second. An exponentially weighted moving average calculator is used for smoothing, with an initial weight of 0.2, decaying by 50% every 60 seconds to generate a smooth change curve. The state parameter vector consists of standardized temperature, pressure, and gas composition data. Support vector regression maps the features using a Gaussian kernel function, calculating the initial regulation coefficient ranging from 0.85 to 1.15. The oxygen content gradient is obtained through a 60-second window difference operation; for example, if oxygen content increases from 2.5% to 3.2% at a rate of 0.006% per second, an adaptive threshold segmenter divides the system into stable and fluctuating segments based on a variance of 0.15. The weighting unit dynamically adjusts the combined weights of the coefficients and gradients based on the state. At a temperature of 1100 degrees Celsius, the adjustment coefficient weight is 0.7 and the gradient weight is 0.3. When the temperature rises to 1150 degrees Celsius, the gradient weight increases to 0.6. After weighted superposition, the comprehensive adjustment parameters are compensated and corrected by historical feedback, and finally discretized into an adjustment time sequence with 5-second intervals to ensure a smooth transition.

[0057] S1042. After the command is generated, the system verifies the actuator response delay and furnace temperature change rate constraints to ensure the feasibility of the adjustment scheme. If the constraints exceed the range, the system re-optimizes and generates the final command sequence. In this embodiment of the invention, the actuator response delay is obtained through real-time measurement, with a typical value of 2 to 5 seconds. After smoothing with a 20-second moving average window, the delay feature vector is extracted by a long short-term memory network. The furnace temperature change rate is calculated using a 5-second difference. For example, if the furnace temperature drops from 1150 degrees Celsius to 1080 degrees Celsius, and the change rate of -14 degrees Celsius per minute exceeds the ±10 degrees Celsius per minute threshold, a recalculation is triggered. Constraint optimization regenerates the adjustment curve based on the delay features and temperature constraints, and smooths it with a 0.95 forgetting factor using recursive least squares. After verification, the change rate is controlled within ±9 degrees Celsius per minute. The final command sequence is quantized into five levels from 90 to 110 cubic meters per hour, containing 36 control points, ensuring that the furnace temperature and oxygen content fluctuations are stable within ±10 degrees Celsius and ±0.3%, respectively.

[0058] In this embodiment of the invention, this multi-level optimization and verification mechanism effectively improves the response speed and stability of the air supply, and technicians can adjust the parameters according to the actual working conditions to further optimize the control effect.

[0059] S105. While sending the air flow adjustment command sequence to the actuator of the air supply system of the Claus sulfur recovery unit, the system collects the adjusted oxygen content, furnace temperature and sulfur dioxide concentration data in real time, generates a data stream containing execution records and performance feedback, evaluates the adjustment effect by comparing it with the predicted clean air demand trajectory and control target, and optimizes the decoupling model and control algorithm parameters online when the effect deviation is large, so as to improve the accuracy and response speed of subsequent control.

[0060] In this embodiment of the invention, the system monitors key process parameters in real time using a multi-parameter data acquisition unit. Oxygen content data is acquired at 100-millisecond intervals, ranging from 2% to 3%, with ±0.2% noise; furnace temperature data is acquired at 200-millisecond intervals, ranging from 1000 to 1200 degrees Celsius, with ±5 degrees Celsius noise; and sulfur dioxide concentration data is acquired at 500-millisecond intervals, ranging from 8% to 12%, with ±0.5% noise. To ensure data quality, the data preprocessing unit uses a 5-point moving average method to denoise and smooth the raw data, generating a stable processed data sequence. Subsequently, a Long Short-Term Memory (LSTM) network extracts temporal features from these sequences. This network has a two-layer structure with 64 neurons per layer, taking 60 time points of processed data as input and outputting a 12-dimensional real-time performance index vector, including 4-dimensional oxygen content features, 4-dimensional furnace temperature features, and 4-dimensional sulfur dioxide concentration features, capturing the dynamic trends of the parameters.

[0061] Furthermore, the target control range is obtained from the process database, namely, oxygen content of 2.3% to 2.7%, furnace temperature of 1080 to 1150 degrees Celsius, and sulfur dioxide concentration of 9% to 11%. The deviation of the real-time performance index vector from the target range is calculated using a performance index evaluator, and the goodness of fit is measured using a trajectory comparison calculator based on Euclidean distance to determine the degree of agreement between the actual effect and the predicted trajectory. In this embodiment of the invention, the performance evaluation uses Mahalanobis distance to quantify the deviation, and the goodness of fit threshold is set to 0.9. If the goodness of fit is lower than the threshold, a Kalman filter dynamically tracks the deviation and the goodness of fit, generating a deviation feature vector through prediction and correction. For example, in a certain perturbation, the oxygen content rises to 3.1%, the furnace temperature drops to 1050 degrees Celsius, and the sulfur dioxide concentration reaches 11.5%, with a goodness of fit of only 0.82, indicating a significant deviation between the actual response and the prediction.

[0062] When the deviation exceeds the preset range, the parameter optimization unit calculates the decoupling matrix correction amount based on the deviation feature vector, and adjusts the control weight matrix using gradient optimization and backpropagation algorithms to generate an updated parameter set to optimize subsequent control. In this embodiment, the decoupling matrix is ​​a 4×4 structure used to separate the effects of methanol flash vapor and liquid nitrogen scrubbing tail gas. The correction amount calculation shows that the weight of methanol flash vapor's influence on oxygen content needs to be increased from 0.65 to 0.78, and its influence on furnace temperature needs to be reduced from 0.55 to 0.45, while the influence of liquid nitrogen scrubbing tail gas on sulfur dioxide concentration needs to be increased from 0.35 to 0.52. The gradient optimization is set with a learning rate of 0.01. After 50 iterations, the weight matrix converges, and the new parameter set reduces the deviation by approximately 35% and shortens the response time by 40% in the next similar disturbance, significantly improving control robustness.

[0063] Furthermore, in actual operation, the system ensures process stability through real-time monitoring and feedback optimization of multiple parameters. For example, when the air flow rate is adjusted, and the oxygen content exceeds the target upper limit by 0.4%, the furnace temperature is below the lower limit by 30 degrees Celsius, and the sulfur dioxide concentration exceeds the standard by 0.5%, the deviation feature vector reflects the response lag problem. The optimized parameter set adjusts the estimation accuracy of the decoupled model's contribution to disturbances, making subsequent adjustments closer to actual needs. It is understood that the embodiments of this invention do not strictly limit the sampling period or the number of network layers; technicians can adjust them according to specific scenarios to further improve the effect.

[0064] This invention provides an intelligent control system for a Claus sulfur recovery device, mainly comprising:

[0065] The real-time monitoring and scene recognition module is used to monitor the methanol flash vapor flow rate and liquid nitrogen wash tail gas flow rate in real time. It identifies disturbance scenarios where the methanol flash vapor flow rate drops sharply and the liquid nitrogen wash tail gas flow rate peaks simultaneously through preset thresholds and logic. If such a scenario is identified, real-time data of the scenario at that moment and within the subsequent preset time period are collected to form an initial disturbance event record.

[0066] The decoupling model calculation module is used to call the pre-established decoupling model, use the real-time data in the initial disturbance event record to calculate and separate the independent and dynamic contributions of the methanol flash vapor drop and the liquid nitrogen wash tail gas peak to the total air demand, and output the time series data of the two independent disturbance effects.

[0067] The net air demand forecasting module is used to calculate the net change in total air demand over a future preset period based on the time series data of two independent disturbances, using an overlay algorithm and combining their respective time series characteristics, and thus obtain the predicted net air demand change trajectory.

[0068] The optimization control calculation module is used to calculate the target air flow adjustment scheme based on the predicted trajectory of changes in net air demand and the current system status, using an optimization control algorithm to compensate for the predicted changes in air demand in advance, and to send the obtained air flow adjustment command sequence to the actuator of the air supply system for execution.

[0069] The effect evaluation and model update module is used to collect the actual oxygen content, furnace temperature and products after the adjustment in real time. The output is a data stream containing control execution records and real-time performance feedback. It compares the predicted net air demand trajectory with the set control target to evaluate the actual effect of the target air flow adjustment scheme. If there is a significant deviation between the actual effect and the expectation, the decoupled model parameters and / or the control algorithm weights are updated online, and the output is the adjusted model parameter set.

[0070] The above description is merely a preferred embodiment of this application and an explanation of the technical principles employed. Those skilled in the art should understand that the scope of the invention involved in this application is not limited to technical solutions formed by specific combinations of the above-described technical features, but should also cover other technical solutions formed by arbitrary combinations of the above-described technical features or their equivalents without departing from the concept of this application. For example, technical solutions formed by substituting the above features with (but not limited to) technical features with similar functions disclosed in this application.

Claims

1. A smart control method for a Claus sulfur recovery device, characterized in that, The method includes: Real-time monitoring of methanol flash vapor flow and liquid nitrogen wash tail gas flow; identification of disturbance scenarios where methanol flash vapor flow drops sharply and liquid nitrogen wash tail gas peaks occur simultaneously through pre-set thresholds and logic; if such a scenario is identified, real-time data of the scenario at that moment and within the subsequent preset time period are collected to form an initial disturbance event record. By invoking a pre-established decoupling model and utilizing real-time data from the initial disturbance event records, the independent and dynamic contributions of the methanol flash vapor decline and liquid nitrogen scrubbing tail gas peak to total air demand are calculated and separated. Two sets of time-series data on the independent disturbance effects are output, including: Based on the initial disturbance event records, the methanol flash vapor flow data and liquid nitrogen wash tail gas flow data are normalized to obtain standardized flow data. The standardized flow data is then denoised using a sliding median filter to obtain smooth flow data. Principal component analysis is used to perform orthogonal transformation on the smoothed flow data to obtain eigenvalue matrix and eigenvector matrix. The eigenvectors are sorted according to the magnitude of the eigenvalues, and the first two eigenvectors are selected to construct an orthogonal basis space. The smoothed flow data is mapped to the orthogonal basis space by the projection operator to obtain the methanol flash vapor load coefficient and the liquid nitrogen wash tail gas load coefficient. Spline interpolation is used for smoothing to obtain the load time series. A recurrent neural network is used to extract dynamic features from the load time series to obtain a dynamic feature vector. The correlation between the components of the dynamic feature vector is calculated using a mutual information entropy calculator to construct an independence evaluation matrix. The independence evaluation matrix is ​​then diagonalized to obtain the independent contribution. The load time series are weighted and combined according to the independent contribution to obtain the independent impact time series of methanol flash vapor on total air demand and the independent impact time series of liquid nitrogen scrubbing tail gas on total air demand. Based on the time series data of the two independent disturbances, the net change in total air demand within a preset period is calculated by superposition algorithm and combined with their respective time series characteristics, thus obtaining the predicted net air demand change trajectory. Based on the predicted trajectory of changes in net air demand and combined with the current system status, the target airflow adjustment scheme is calculated using an optimized control algorithm to compensate for the predicted changes in air demand in advance. The resulting airflow adjustment command sequence is then sent to the actuators of the air supply system for execution.

2. The method according to claim 1, characterized in that, The real-time monitoring of methanol flash vapor flow rate and liquid nitrogen scrubbing tail gas flow rate identifies disturbance scenarios where a step decrease in methanol flash vapor flow rate and a peak in liquid nitrogen scrubbing tail gas flow rate occur simultaneously through pre-set thresholds and logic. If such a scenario is identified, real-time data for that moment and subsequent preset time periods are collected to form an initial disturbance event record, including: The pressure measurement values ​​of methanol flash vapor and liquid nitrogen wash tail gas are collected by the pressure sensing and monitoring unit. The pressure measurement values ​​are then processed by moving average filtering to obtain the pressure processing value and pressure change marker. Based on the pressure processing value, abrupt change point detection was performed on the methanol flash vapor flow rate measurement value and the liquid nitrogen wash tail gas flow rate measurement value to obtain the flow rate step drop time point and the flow rate peak time point. For the specified flow rate time point, wavelet transform decomposition is performed on the measured values ​​of methanol flash vapor temperature and liquid nitrogen wash tail gas temperature to obtain temperature feature values. A recurrent neural network is then used to extract features from the temperature feature values ​​to obtain a temperature change feature vector. Disturbance event data records are generated based on the pressure mutation marker, the flow rate time point, and the temperature change feature vector.

3. The method according to claim 1, characterized in that, The process involves calling a pre-established decoupling model and using real-time data from the initial disturbance event record to calculate and separate the independent and dynamic contributions of the methanol flash vapor decrease and the liquid nitrogen scrubbing tail gas peak to the total air demand. It outputs two sets of time-series data on independent disturbance effects, including: dynamically extracting the real-time initial disturbance event data stream to generate a time-aligned input matrix; inputting the decoupling model containing a dual-channel feature extraction structure based on a gated cyclic unit; and outputting the dynamic influence coefficients of the methanol and liquid nitrogen components, respectively. Based on the product of the methanol component dynamic influence coefficient and the total air demand, an independent contribution sequence for the methanol flash vapor decrease is generated; based on the product of the liquid nitrogen component dynamic influence coefficient and the total air demand, an independent contribution sequence for the liquid nitrogen scrubbing tail gas peak is generated; and the independent contribution sequences are timestamped to output two sets of time-series data on influence factors with the same time resolution.

4. The method according to claim 1, characterized in that, The process involves using time-series data of two independent disturbances, combined with an overlay algorithm and their respective time-series characteristics, to calculate the net change in total air demand over a predetermined period, thus obtaining the predicted net air demand change trajectory, including: A temporal feature extractor is used to extract methanol flash vapor perturbation features and liquid nitrogen wash tail gas perturbation features from independent perturbation impact sequences. Based on the perturbation features, a long short-term memory network is used to identify the perturbation feature vector. For the aforementioned perturbation feature vector, the weights of the methanol flash vapor sequence and the liquid nitrogen wash tail gas sequence are calculated using a weight calculator. The net influence sequence is obtained by weighted superposition and filtering of the two independent influence sequences. Based on the net impact sequence, an autoregressive moving average is used to construct a time series prediction basis function, and an out-of-sample extension is performed within a preset time window to obtain a prediction sequence of total air demand change. For the total air demand change prediction sequence, wavelet decomposition is used to obtain the trend component sequence and the fluctuation component sequence. The time delay compensation amount is determined according to the autocorrelation function of the fluctuation component sequence, and the time delay correction is performed on the trend component sequence to obtain the total air demand net change prediction trajectory.

5. The method according to claim 1, characterized in that, Based on the predicted trajectory of net air demand changes and combined with the current system status, an optimized control algorithm is used to calculate the target airflow adjustment scheme to compensate for the predicted changes in air demand in advance. The resulting airflow adjustment command sequence is then sent to the actuators of the air supply system for execution, including: A Kalman filter is used to perform noise removal processing on the temperature data sequence collected by the temperature sensing and monitoring unit and the oxygen content data sequence collected by the oxygen content detection unit to obtain the state monitoring sequence. The state monitoring sequence is time-stamped and aligned. The state monitoring sequence is then decomposed into a multi-scale feature coefficient sequence by wavelet transform. The feature coefficient sequence is then time-series superimposed with the predicted net air demand change trajectory to obtain a fused feature sequence. Based on the fused feature sequence, feature extraction is performed using a deep reinforcement learning unit to obtain the airflow regulation parameter matrix; Based on the airflow regulation parameter matrix, the regulation constraints are extracted, an airflow regulation command sequence is generated, and the command is sent to the actuator of the air supply system for execution.

6. The method according to claim 5, characterized in that, Also includes: Based on the predicted trajectory of net air demand change and the current system status, an air flow adjustment coefficient is generated. The air flow adjustment coefficient is then integrated with the oxygen content gradient to output a target flow adjustment time series. After verifying the actuator response delay compensation, the target flow adjustment timing sequence is checked. If the furnace temperature change rate exceeds the preset range, the target flow adjustment timing sequence is recalculated, and the verified timing sequence is split into air flow adjustment commands.

7. The method according to claim 1, characterized in that, The method further includes: real-time acquisition of the actual oxygen content, furnace temperature, and products after adjustment, outputting a data stream containing control execution records and real-time performance feedback, comparing the predicted clean air demand trajectory with the set control target, and evaluating the actual effect of the target air flow adjustment scheme; if the actual effect deviates significantly from the expectation, updating the decoupling model parameters and / or optimizing the control algorithm weights online, and outputting an adjusted model parameter set.

8. The method according to claim 7, characterized in that, The real-time acquisition and execution of the adjusted actual oxygen content, furnace temperature and products are output as a data stream containing control execution records and real-time performance feedback. The predicted clean air demand trajectory is compared with the set control target to evaluate the actual effect of the target air flow adjustment scheme. If the actual results deviate significantly from the expected results, the decoupled model parameters and / or the control algorithm weights are updated online, and the output is an adjusted set of model parameters, including: The raw data of oxygen content, furnace temperature and sulfur dioxide concentration are acquired, and the data preprocessing unit performs noise reduction and smoothing to obtain a processed data sequence. The time-series features are extracted from the processed data sequence according to the long short-term memory network to obtain a real-time performance index vector. Obtain the target ranges for oxygen content, furnace temperature, and sulfur dioxide concentration from the process database, and calculate the deviation of the real-time performance index vector from the target ranges using the performance index evaluator. The fitting degree value is obtained by using the trajectory comparison calculator to calculate the fitting degree between the real-time performance index vector and the predicted net air demand change trajectory. The fitting degree value is then determined by the difference detector to see if the fitting degree value exceeds a preset threshold. If the deviation exceeds the preset range, the parameter optimization unit calculates the decoupling matrix correction amount based on the deviation feature vector, and optimizes and adjusts the control weight matrix according to the decoupling matrix correction amount to obtain the updated parameter set.

9. An intelligent control system for a Claus sulfur recovery device, characterized in that, The system includes: The real-time monitoring and scene recognition module is used to monitor the methanol flash vapor flow rate and liquid nitrogen wash tail gas flow rate in real time. It identifies disturbance scenarios where the methanol flash vapor flow rate drops sharply and the liquid nitrogen wash tail gas flow rate peaks simultaneously through preset thresholds and logic. If such a scenario is identified, real-time data of the scenario at that moment and within the subsequent preset time period are collected to form an initial disturbance event record. The decoupling model calculation module is used to call a pre-established decoupling model, utilize real-time data from the initial disturbance event record, calculate and separate the independent and dynamic contributions of the methanol flash vapor decline and liquid nitrogen scrubbing tail gas peak to total air demand, and output two sets of time series data of independent disturbance effects, including: Based on the initial disturbance event records, the methanol flash vapor flow data and liquid nitrogen wash tail gas flow data are normalized to obtain standardized flow data. The standardized flow data is then denoised using a sliding median filter to obtain smooth flow data. Principal component analysis is used to perform orthogonal transformation on the smoothed flow data to obtain eigenvalue matrix and eigenvector matrix. The eigenvectors are sorted according to the magnitude of the eigenvalues, and the first two eigenvectors are selected to construct an orthogonal basis space. The smoothed flow data is mapped to the orthogonal basis space by the projection operator to obtain the methanol flash vapor load coefficient and the liquid nitrogen wash tail gas load coefficient. Spline interpolation is used for smoothing to obtain the load time series. A recurrent neural network is used to extract dynamic features from the load time series to obtain a dynamic feature vector. The correlation between the components of the dynamic feature vector is calculated using a mutual information entropy calculator to construct an independence evaluation matrix. The independence evaluation matrix is ​​then diagonalized to obtain the independent contribution. The load time series are weighted and combined according to the independent contribution to obtain the independent impact time series of methanol flash vapor on total air demand and the independent impact time series of liquid nitrogen scrubbing tail gas on total air demand. The net air demand forecasting module is used to calculate the net change in total air demand over a future preset period based on the time series data of two independent disturbances, using an overlay algorithm and combining their respective time series characteristics, and thus obtain the predicted net air demand change trajectory. The optimized control calculation module is used to calculate the target air flow adjustment scheme based on the predicted trajectory of changes in net air demand and the current system status, using an optimized control algorithm to compensate for the predicted changes in air demand in advance, and to send the obtained air flow adjustment command sequence to the actuator of the air supply system for execution. The effect evaluation and model update module is used to collect the actual oxygen content, furnace temperature and products after the adjustment in real time. The output is a data stream containing control execution records and real-time performance feedback. It compares the predicted net air demand trajectory with the set control target to evaluate the actual effect of the target air flow adjustment scheme. If there is a significant deviation between the actual effect and the expectation, the decoupled model parameters and / or the control algorithm weights are updated online, and the output is the adjusted model parameter set.

Images