Intelligent control method and system for Claus sulfur recovery device

Abstract

The invention provides an intelligent control method and system for a Claus sulfur recovery device, and the method comprises the steps: calling a pre-established decoupling model, calculating and separating out independent and dynamic contribution values of methanol flash vapor drop and a liquid nitrogen wash tail gas peak value on the total air demand by using real-time data in an initial disturbance event record, outputting time sequence data influenced by two groups of independent disturbances; according to the output time sequence data influenced by the two groups of independent disturbances, through a superposition algorithm and in combination with respective time sequence characteristics, the net variation of the total air demand in a future preset period of time is calculated, and a predicted net air demand change track is obtained; and according to the predicted net air demand change track and in combination with the current system state, a target air flow adjustment scheme is calculated by applying an optimization control algorithm, the predicted air demand change is compensated in advance, and an obtained air flow adjustment instruction sequence is issued to an execution mechanism of the air supply system to be executed.

Description

Technical Field

[0001] The present 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 Art

[0002] Background: Sulfur recovery technology, as an important pillar in the field of petrochemicals and environmental protection, is directly related to the efficiency and environmental compliance of industrial waste gas treatment. Its core lies in the efficient conversion of sulfur-containing components into controllable products through the combustion process, thereby achieving resource utilization and emission compliance. Claus incinerators have attracted much attention due to their wide application in hydrogen sulfide treatment. Their stable operation and efficient conversion capabilities have a decisive impact on the economy and sustainability of the overall process. However, existing sulfur recovery control methods often show limitations when facing complex disturbances. Traditional control strategies mostly rely on static air-fuel ratio settings or simple feedback adjustments, which are difficult to adapt to multi-variable coupling and timing changes under dynamic conditions, resulting in fluctuations in sulfur conversion efficiency and even system instability. Especially in specific concurrent disturbance scenarios, such as the step-down of methanol flash gas flow and the superposition of liquid nitrogen tail gas peaks, the air demand shows an opposite trend. Conventional methods are difficult to quickly decouple and accurately respond, exposing their shortcomings in predictability and adaptability. Violent fluctuations in oxygen content can easily lead to incomplete combustion of sulfur species, and furnace temperature fluctuations interfere with the reaction equilibrium. The combined effect of the two makes it difficult to maintain a stable conversion rate of sulfur to SO2. Therefore, how to construct a predictive control strategy to decouple the coupling effects of oxygen content oscillation and furnace temperature fluctuation for the disturbance scenario of the concurrent step-down of methanol flash gas and the peak of liquid nitrogen washing tail gas has become the key issue focused on in this study. Summary of the invention

[0003] The present invention provides an intelligent control method for a Claus sulfur recovery unit, which mainly includes: Real-time monitor the flow rate of methanol flash vapor and the flow rate of tail gas from liquid nitrogen wash. Identify the disturbance scenario where the step-down of methanol flash vapor flow rate coincides with the peak of tail gas from liquid nitrogen wash through preset thresholds and logic. If this scenario is identified, collect the real-time data at this moment and within a preset time period afterwards to form an initial disturbance event record; call the pre-established decoupling model, use the real-time data in the initial disturbance event record to calculate and separate the contribution values of the decrease in methanol flash vapor and the peak of tail gas from liquid nitrogen wash to the total air demand respectively, and output two sets of time series data of independent disturbance impacts; according to the two sets of output time series data of independent disturbance impacts, calculate the net change in total air demand within a preset future time period through a superposition algorithm combined with their respective time series characteristics to obtain the predicted net air demand change trajectory; based on the predicted net air demand change trajectory and combined with the current system state, use an optimal control algorithm to calculate the target air flow adjustment plan, compensate in advance for the predicted air demand change, and issue the obtained air flow adjustment instruction sequence to the actuator of the air supply system for execution; collect the actual oxygen content, furnace temperature and products after the adjustment in real time, output them as a data stream containing control execution records and real-time performance feedback, compare the predicted net air demand trajectory with the set control target, evaluate the actual effect of the target air flow adjustment plan. If there is a significant deviation between the actual effect and the expectation, update the decoupling model parameters and / or the weights of the optimal control algorithm online, and output the adjusted model parameter set.

[0004] Furthermore, the methanol flash vapor flow rate and the LNG wash tail gas flow rate are monitored in real time. The disturbance scenario where the step - down of the methanol flash vapor flow rate coincides with the peak value of the LNG wash tail gas is identified through a pre - set threshold and logic. If this scenario is identified, the real - time data at this moment and within a subsequent pre - set time period of this scenario is collected to form an initial disturbance event record, including: continuously collecting the methanol flash vapor pressure measurement value and the LNG wash tail gas pressure measurement value by the pressure sensing monitoring unit according to a pre - set sampling period, smoothing the measurement values through a moving average filter to obtain pressure - processed values, and using a dual - threshold comparator to determine the pressure mutation marker for the pressure - processed values. The methanol flash vapor flow rate measurement value and the LNG wash tail gas flow rate measurement value are collected in real time by the gas flow sensing monitoring unit, the flow rate measurement values within a pre - set time period are recorded in the built - in memory of the sensing monitoring unit, and a random forest algorithm is used to detect the mutation points of the flow rate measurement values to obtain the moment points of the step - down of the methanol flash vapor flow rate and the moment points of the peak value of the LNG wash tail gas flow rate. The methanol flash vapor temperature measurement value and the LNG wash tail gas temperature measurement value are collected according to the temperature sensing monitoring unit, the temperature measurement values are decomposed by wavelet transform to obtain temperature characteristic values, and a recursive neural network is used to extract the temperature change feature vectors from the temperature characteristic values. A multi - dimensional feature matrix is constructed for the pressure mutation marker, the flow step - down moment points, the flow peak moment points, and the temperature change feature vectors, and a feature fuser is used to perform time - series alignment on the feature matrix. A scene identification data set is established based on the multi - dimensional feature matrix, the scene identification data set contains the pressure mutation marker, the flow mutation moment points, and the temperature feature vectors, a time - window constraint with a fixed duration is imposed on the data set, and disturbance event data records are generated based on the data set.

[0005] Further, call the pre-established decoupling model, use the real-time data in the initial perturbation event record, calculate and separate the contribution values of the methanol flash vapor decline and the peak value of the tail gas of the liquid nitrogen wash to the total air demand respectively, and output the time series data of two groups of independent perturbation effects, including: according to the initial perturbation event record, normalize the methanol flash vapor flow data and the liquid nitrogen wash tail gas flow data to obtain the standardized flow data, and use a sliding median filter to denoise the standardized flow data to obtain the smoothed flow data. Perform orthogonal transformation on the smoothed flow data by using principal component analysis to obtain the eigenvalue matrix and the eigenvector matrix, sort the eigenvectors according to the eigenvalue size, and select the first two eigenvectors to construct an orthogonal basis space. Map the smoothed flow data to the orthogonal basis space through a projection operator to obtain the methanol flash vapor load coefficient and the liquid nitrogen wash tail gas load coefficient, and use spline interpolation to smooth the load coefficient to obtain the load time series. Use a recurrent neural network to extract the dynamic features in the load time series, and 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 the dynamic feature vector. Calculate the correlation degree between the components of the dynamic feature vector through a mutual information entropy calculator, construct an independence evaluation matrix, and perform diagonalization processing on the independence evaluation matrix to obtain the independent contribution degree. Weight and combine the load time series according to the independent contribution degree to obtain the independent influence time series of the methanol flash vapor on the total air demand and the independent influence time series of the liquid nitrogen wash tail gas on the total air demand.

[0006] Furthermore, dynamically intercept the real-time data stream of initial perturbation events to generate a time-aligned input matrix, and input it into a decoupling model with a dual-channel feature extraction structure based on a gated recurrent unit, which respectively outputs the dynamic influence coefficient of the methanol component and the dynamic influence coefficient of the liquid nitrogen component. Generate an independent contribution sequence of the decrease in methanol flash vapor according to the product operation of the dynamic influence coefficient of the methanol component and the total air demand. Generate an independent contribution sequence of the peak value of the liquid nitrogen wash tail gas according to the product operation of the dynamic influence coefficient of the liquid nitrogen component and the total air demand. Perform timestamp alignment processing on the independent contribution sequences respectively, and output two sets of impact factor time series data with the same time resolution, including: replacing outliers in the initial perturbation event record by the median through the data preprocessing unit, sampling the event record at a fixed time interval to obtain the original sampling data, and using a sliding window to perform real-time interception on the original sampling data to obtain the time series intercepted data. Align the time series intercepted data according to the time marker aligner to obtain an aligned data matrix, and use the normalization calculation unit to perform zero-mean normalization processing on the aligned data matrix to obtain a standardized data matrix. Use a gated recurrent unit to perform dual-channel feature extraction on the standardized data matrix. The number of hidden layer neurons of the gated recurrent unit is the same as the dimension of the standardized data matrix. Extract the methanol component feature vector through the left channel and extract the liquid nitrogen component feature vector through the right channel. Perform linear transformation on the methanol component feature vector and the liquid nitrogen component feature vector according to the feature conversion unit, and use normalization processing to obtain the dynamic influence coefficient of the methanol component and the dynamic influence coefficient of the liquid nitrogen component. Obtain the total air demand from the process parameter database through the air demand calculation unit, and use a multiplier 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. Interpolate the product results according to a cubic spline interpolator to obtain an independent contribution sequence of methanol flash vapor and an independent contribution sequence of liquid nitrogen wash tail gas with consistent time resolution.

[0007] Further, based on the two sets of time series data affected by independent perturbations output, the net change in the total air demand within a preset future time period is calculated through a superposition algorithm in combination with their respective time series characteristics, obtaining a predicted net air demand change trajectory, including: extracting the methanol flash vapor perturbation feature and the liquid nitrogen wash tail gas perturbation feature from the independent perturbation influence sequence through a time series feature extractor, where the perturbation feature includes the occurrence time point, the duration period, and the change rate value, and using a long short-term memory network to perform feature recognition on the perturbation feature to obtain a perturbation feature vector. Based on the perturbation feature vector, a weight calculator calculates the methanol flash vapor sequence weight and the liquid nitrogen wash tail gas sequence weight, and a sequence superposition calculator performs a weighted superposition operation on the two sets of independent influence sequences, and a Butterworth low-pass filter is used to filter the superposed sequence to obtain a net influence sequence. For the perturbation feature vector, an autoregressive moving average is used to construct a time series prediction basis function, and the net influence sequence is extrapolated outside the sample within a preset time window to obtain a total air demand change amount prediction sequence. The total air demand change amount prediction sequence is subjected to wavelet decomposition by a trend extractor to obtain a trend component sequence and a fluctuation component sequence, and an exponential smoothing algorithm is used to smooth the trend component sequence to obtain a smoothed trend sequence. The autocorrelation function of the fluctuation component sequence is calculated by a delay estimator, the delay compensation amount is determined according to the autocorrelation peak position, and a phase compensator is used to perform delay correction on the smoothed trend sequence to obtain a total air demand net change prediction trajectory.

[0008] Furthermore, based on the predicted net air demand change trajectory and combined with the current system state, an optimized control algorithm is used to calculate the target air flow adjustment plan, compensating in advance for the predicted air demand changes. The obtained air flow adjustment command sequence is sent to the actuators of the air supply system for execution, including: obtaining the temperature acquisition sequence from the real-time temperature data of the Claus furnace collected by the multi-point temperature sensing and monitoring unit, obtaining the oxygen content acquisition sequence from the tail gas oxygen content data collected by the oxygen content detection unit, and using a Kalman filter to perform noise cancellation processing on the temperature acquisition sequence and the oxygen content acquisition sequence to obtain the state monitoring sequence. A data alignment unit is used to perform time mark alignment on the state monitoring sequence, and wavelet transform is used to perform multi-scale decomposition on the state monitoring sequence to obtain the feature coefficient sequence. The feature coefficient sequence and the predicted net air demand change trajectory are subjected to time series superposition to obtain the fusion feature sequence. Feature extraction is performed on the fusion feature sequence through the deep reinforcement learning unit, and the deep reinforcement learning unit includes a state space, an action space, and a reward function, and outputs an air flow adjustment parameter matrix. According to the constraint optimization unit, adjustment constraint conditions are extracted from the air flow adjustment parameter matrix, and the constraint conditions include a flow upper limit value, a flow lower limit value, and a change rate limit value. A dynamic programming algorithm is used to calculate the optimal adjustment sequence that satisfies the constraint conditions. The optimal adjustment sequence is time sampled through the sequence decomposition unit, and the adjustment time points and corresponding target flow values are extracted to generate an air flow adjustment command sequence. An instruction conversion unit is used to perform digital quantity conversion on the air flow adjustment command sequence to generate a standard control instruction for the actuator, and the standard control instruction is sent to the air supply actuator through the industrial bus interface.

[0009] Further, an air flow regulation coefficient is generated based on the predicted net air demand change trajectory and the current system state. The air flow regulation coefficient and the oxygen content gradient are fused, and a target flow adjustment time series is output, including: extracting change amplitude data and change rate data from the predicted net air demand change trajectory through a change analysis unit, and smoothing the change amplitude data and the change rate data using an exponentially weighted moving average calculator, where the weight coefficient of the calculator decays based on the data time interval, to obtain a smoothed change curve. The temperature data, pressure data, and gas component data are standardized by a system parameter processing unit to obtain a state parameter vector, and a feature mapping is performed on the state parameter vector and the smoothed change curve by a support vector regression calculation unit to obtain an initial flow regulation coefficient. A fixed-time window difference operation is performed on the oxygen content sampling data by a gradient calculation unit, and the data obtained from the difference operation is segmented using an adaptive threshold segmenter based on analysis of variance to obtain an oxygen content gradient curve. A weight allocation unit calculates the combined weight of the initial flow regulation coefficient and the oxygen content gradient curve based on the state parameter vector, and a weighted superposition operation is performed on the regulation coefficient and the gradient curve to obtain a comprehensive regulation parameter. A feedback compensation unit calculates a compensation coefficient based on the historical regulation effect, and dynamically compensates the comprehensive regulation parameter to obtain a corrected regulation parameter. According to a timing planning unit, the corrected regulation parameter is discretized according to a preset sampling period to generate an adjustment time series including adjustment times and target flows.

[0010] Further, verify the target flow adjustment timing sequence after compensating for the actuator response delay against the furnace temperature change rate constraint conditions. If the furnace temperature change rate exceeds the preset range, recalculate the target flow adjustment timing sequence. Split the verified timing sequence into air flow adjustment instructions, including: obtaining actuator response delay data through the response delay measurement unit, statistically analyzing the response delay data using a moving average calculator, extracting features from the delay data sequence based on a long short-term memory network, with the network input being the response delay time sequence and the output being the delay feature vector. Collect the furnace temperature data sequence through the temperature detection unit, perform a time difference operation on the furnace temperature data sequence using a difference calculator, and normalize the difference result to obtain the furnace temperature change rate curve. Compare the furnace temperature change rate curve with the preset upper threshold and lower threshold through a threshold judge. If it exceeds the preset threshold range, output a recalculation flag. Receive the recalculation flag by the constraint optimization calculator and regenerate the target flow adjustment curve based on the temperature constraint conditions and the delay feature vector, where the constraint conditions include the upper and lower limit values of the temperature change rate. Dynamically fit the target flow adjustment curve using the recursive least squares method to obtain a smooth adjustment curve, and determine whether the smooth adjustment curve meets the temperature constraint conditions through a constraint checker. Perform time discretization on the smooth adjustment curve that meets the constraint conditions by the instruction generation unit, and quantify the discrete points according to the preset flow levels using a flow classifier to generate an air flow adjustment instruction sequence.

[0011] Furthermore, the actual oxygen content, furnace temperature, and products after the execution of the adjustment are collected in real time, and output as a data stream containing control execution records and real-time performance feedback. The predicted net air demand trajectory is compared 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 decoupling model parameters and / or the weights of the optimization control algorithm are updated online, and the output is an adjusted set of model parameters, including: the original oxygen content data, furnace temperature original data, and sulfur dioxide concentration original data are collected in real time through a multi-parameter collector, and the data preprocessing unit is used to perform denoising and smoothing processing on the original data to obtain a processed data sequence, and the time series characteristics in the processed data sequence are extracted according to the long short-term memory network to obtain a real-time performance index vector. The control target calculation unit obtains the oxygen content target range, furnace temperature target range, and sulfur dioxide concentration target range from the process database, and the performance index evaluator calculates the deviation degree between the real-time performance index vector and the target range. The trajectory comparison calculator calculates the fitting degree between the real-time performance index vector and the predicted net air demand change trajectory, and the fitting degree calculation is based on the Euclidean distance metric. The difference detector compares the fitting degree value with the preset threshold for judgment. The Kalman filter is used to dynamically track the deviation degree and the fitting degree value to generate a deviation feature vector, and the threshold judge is used to judge the range of the deviation feature vector. If the deviation feature vector exceeds the preset range, the parameter optimization unit calculates the decoupling matrix correction amount based on the deviation feature vector, and the decoupling matrix is used to separate the influence contributions of methanol flash vapor and liquid nitrogen wash tail gas. The gradient optimization unit iteratively updates the decoupling matrix, and the backpropagation algorithm is used to optimize and adjust the control weight matrix, and the parameter integration unit generates an updated parameter set.

[0012] The present invention provides an intelligent control system for a Claus sulfur recovery unit, mainly including: a real-time monitoring and scenario recognition module, which is used to monitor the flow rate of methanol flash vapor and the flow rate of tail gas from liquid nitrogen wash in real time, and identify a disturbance scenario where the stepwise decrease in the flow rate of methanol flash vapor and the peak value of the tail gas from liquid nitrogen wash occur simultaneously through a preset threshold and logic. If this scenario is identified, the real-time data at this moment and within a preset period of time thereafter in this scenario is collected to form an initial disturbance event record; a decoupling model calculation module, which is used to call a pre-established decoupling model, and use the real-time data in the initial disturbance event record to calculate and separate the contribution values of the decrease in methanol flash vapor and the peak value of the tail gas from liquid nitrogen wash to the total air demand independently and dynamically, and output two sets of time series data of independent disturbance effects; a net air demand prediction module, which is used to calculate the net change in the total air demand within a preset period of time in the future based on the two sets of output time series data of independent disturbance effects, through a superposition algorithm and combined with their respective time series characteristics, to obtain the predicted net air demand change trajectory; an optimal control calculation module, which is used to calculate a target air flow adjustment plan according to the predicted net air demand change trajectory and combined with the current system state, using an optimal control algorithm to compensate in advance for the predicted air demand change, and send the obtained air flow adjustment instruction sequence to the actuator of the air supply system for execution; an effect evaluation and model update module, which is used to collect the actual oxygen content, furnace temperature, and products in real time after the adjustment is executed, and output a data stream including control execution records and real-time performance feedback, compare the predicted net air demand trajectory with the set control target, evaluate the actual effect of the target air flow adjustment plan, and if there is a significant deviation between the actual effect and the expectation, update the decoupling model parameters and / or the weights of the optimal control algorithm online, and output an adjusted set of model parameters. The technical solution provided by the embodiments of the present invention may include the following beneficial effects: The present invention discloses an intelligent control method for a Claus sulfur recovery unit. This method identifies and records a specific disturbance scenario by monitoring the flow rates of methanol flash vapor and tail gas from liquid nitrogen wash in real time. It uses a decoupling model to separate the independent effects of the two gases on air demand, predicts future air demand changes. Based on the prediction results and the system state, it calculates and executes an optimal air flow adjustment plan. At the same time, the present invention collects the performance indicators after the adjustment in real time, evaluates the control effect, and updates the model parameters and algorithm weights online. This method can effectively respond to complex disturbance scenarios, compensate for air demand changes in advance, achieve precise control of the Claus sulfur recovery unit, improve the system response speed and accuracy, and thus optimize the production process and product quality. BRIEF DESCRIPTION OF THE DRAWINGS

[0013] Figure 1 It is a flowchart of an intelligent control method for a Claus sulfur recovery unit of the present invention. DETAILED DESCRIPTION OF THE EMBODIMENTS

[0014] To enable those skilled in the art to better understand the technical solutions in this specification, the following will clearly and completely describe the technical solutions in the embodiments of this specification in conjunction with the accompanying drawings in the embodiments of this specification. Obviously, the described embodiments are only a part of the embodiments of this specification, rather than all the embodiments. Based on the embodiments in this specification, all other embodiments obtained by those of ordinary skill in the art without creative efforts shall fall within the scope of protection of this specification.

[0015] As Figure 1 , a specific intelligent control method for a Claus sulfur recovery unit in this embodiment may specifically include: S101. During the operation of the Claus sulfur recovery unit, use sensors to continuously monitor the flow rate of methanol flash gas and the flow rate of tail gas from liquid nitrogen washing. Identify the disturbance scenario where the step - down of the methanol flash gas flow rate coincides with the peak value of the tail gas flow rate from liquid nitrogen washing through a preset threshold logic. If this scenario is detected, lock the current state and collect relevant real - time data to generate an initial disturbance event record.

[0016] S1011. In the embodiment of the present invention, continuously collect the pressure values of the methanol flash gas and the tail gas from liquid nitrogen washing through a pressure sensor, obtain real - time pressure data with a sampling period of 100 milliseconds, apply a moving average filter to the collected pressure values, set the window length to 5 sampling points, generate a smoothed pressure processing value, and then use a dual - threshold comparator to judge pressure mutations. 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 periods, it is marked as a pressure mutation, and the mutation time point is recorded.

[0017] S1012. At the same time, use flow sensors to continuously monitor the flow rate of methanol flash gas and the flow rate of tail gas from liquid nitrogen washing, record the flow data within a 60 - second time window, and use the random forest algorithm to detect mutation points of the flow values. This algorithm is constructed based on 50 decision trees, with a maximum depth of 6 layers for each tree. Extract features such as flow mean, standard deviation, and peak factor, detect the step - down moment when the flow rate of methanol flash gas drops from 25 cubic meters per hour to below 10 cubic meters per hour, and the peak moment when the flow rate of tail gas from liquid nitrogen washing rises to 35 cubic meters per hour. Further collect the temperature data of the two gases through a temperature sensor, perform three - layer wavelet transform decomposition on the temperature values, select the db4 wavelet basis function to extract high - frequency and low - frequency coefficients, and then use a long short - term memory network with 128 hidden - layer neurons to extract temperature features and generate a temperature change feature vector.

[0018] S1013. After obtaining the pressure mutation marker, the flow mutation moment, and the temperature feature vector, construct a multi-dimensional feature matrix. Each row represents the state features at a time point and contains 12-dimensional data, including the pressure mutation marker, the methanol flash vapor flow rate value, the liquid nitrogen wash tail gas flow rate value, and the four-dimensional temperature feature vectors of the two gases. Align the matrix in time series through a feature fuser, apply a fixed time window constraint of 180 seconds, and set the step size to 30 seconds to generate a perturbation event data record containing the scenario identifier.

[0019] In the embodiment of the present invention, the flow rate changes of the methanol flash vapor and the liquid nitrogen wash tail gas have a specific pattern. For example, in a certain chemical plant scenario, when the pressure of the methanol flash vapor drops from 2.2 MPa to 0.6 MPa, the pressure mutation marker is triggered. At the same time, the flow rate drops from 23 cubic meters per hour to 8 cubic meters per hour, and the liquid nitrogen wash tail gas flow rate rises to 32 cubic meters per hour. The temperature data shows that the temperature of the methanol flash vapor drops rapidly by 15 to 25 degrees Celsius. The system generates a perturbation event record based on these features and accurately marks the abnormal time point.

[0020] It can be understood that the embodiment of the present invention does not overly limit the sensor type and algorithm parameters, which can be adjusted by technicians according to the actual scenario to ensure the accuracy of perturbation scenario recognition.

[0021] S102. Invoke the pre-trained decoupling model, and use the real-time data in the initial perturbation event record to analyze the independent dynamic effects of the methanol flash vapor flow rate decrease and the liquid nitrogen wash tail gas flow rate peak on the total air demand, and generate two sets of time series data to characterize their respective contribution values.

[0022] In the embodiment of the present invention, the decoupling process first extracts the methanol flash vapor flow rate and the liquid nitrogen wash tail gas flow rate data from the initial perturbation event record. These data are collected in real time by sensors and usually contain flow rate fluctuations and noise. To ensure the analysis accuracy, the system preprocesses the original flow rate data. The data is mapped to the standardized interval of 0 to 1 through a normalization calculation unit to eliminate the dimension difference. 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 wash tail gas flow rate ranges from 15 to 25 cubic meters per hour. After normalization, it is convenient for unified processing. Subsequently, a sliding median filter is used to denoise the standardized data, and the filter window length is set to 7 sampling points to effectively smooth the mutation noise in the sensor measurement, such as the abnormal jump in the flow rate caused by equipment jitter.

[0023] Furthermore, the principal component analysis method is used to perform orthogonal transformation on the smoothed flow data, generating an eigenvalue matrix and an eigenvector matrix. The first two principal components are selected to construct an orthogonal basis space by sorting the eigenvalues, and the smoothed flow data is projected onto this space. The load coefficients of the methanol flash vapor and the tail gas of liquid nitrogen wash are calculated respectively, and then a continuous load time series is generated through spline interpolation smoothing. In the embodiment of the present invention, the principal component analysis aims to capture the main directions of data variation. Practical applications show that the first principal component explains approximately 75% of the variance, and the second principal component contributes 20%. Together, they cover more than 95% of the data information. After projection, the load coefficient of the methanol flash vapor on the first principal component is approximately 0.85, and the load coefficient of the tail gas of liquid nitrogen wash on the second principal component is approximately 0.78, reflecting the orthogonal characteristics of the influence of the two gases. Cubic polynomials are used for spline interpolation to ensure that the time resolution of the load sequence is increased to 1 second, retaining the details of dynamic changes.

[0024] Next, the system inputs the load time series into a recurrent neural network for dynamic feature extraction. The network consists of an input layer, two hidden layers, and an output layer. Each hidden layer is configured with 128 neurons. A 16-dimensional dynamic feature vector is extracted through forward propagation. Subsequently, a mutual information entropy calculator is used to evaluate the correlation between the components of the feature vector, constructing an independence evaluation matrix and performing diagonalization processing to obtain the independent contribution degrees of the methanol flash vapor and the tail gas of liquid nitrogen wash. Then, the load sequences are weighted and combined according to the contribution degrees, and two independent influence time series are output. In an actual scenario, mutual information entropy analysis shows that the correlation between the first 8 dimensions of the feature vector of the methanol flash vapor and the last 8 dimensions of the feature vector of the tail gas of liquid nitrogen wash is relatively low, and the average mutual information value is less than 0.15, indicating strong independence after decoupling. For example, when the flow rate of the methanol flash vapor decreases by 25%, its independent influence on air demand decreases by 18% within 5 minutes, while when the flow rate of the tail gas of liquid nitrogen wash increases by 30%, the air demand increases by 12%, and the superimposed error of the two is less than 3%, verifying the accuracy of the method.

[0025] In addition, in the embodiment of the present invention, to further improve the decoupling accuracy, the system can dynamically intercept the real-time disturbance event data stream, generate a time-aligned input matrix, and input it into a dual-channel decoupling model based on gated recurrent units. This model extracts the features of the methanol flash vapor through the left channel and the features of the tail gas of liquid nitrogen wash through the right channel, and outputs dynamic influence coefficients respectively. The number of hidden layer neurons of the gated recurrent unit matches the dimension of the input matrix. For example, when the input is standardized data of 60 time points, 60 neurons are configured, and the information flow is controlled through an update gate and a reset gate to capture the temporal characteristics of the flow rate changes.

[0026] After generating the dynamic influence coefficients, the system calculates the product of the dynamic influence coefficient of the methanol component and the total air demand through a multiplier to obtain the independent contribution sequence of the decrease in methanol flash vapor, and at the same time calculates the product of the dynamic influence coefficient of the liquid nitrogen component and the total air demand to obtain the independent contribution sequence of the peak value of the liquid nitrogen wash tail gas. And through cubic spline interpolation, the time resolution of the two sequences is ensured to be consistent, and unified time series data is output. In the chemical industry scenario, the reference value of the total air demand is 100 cubic meters per hour. When the methanol flash vapor flow rate drops from 26 cubic meters per hour to 12 cubic meters per hour, its dynamic influence coefficient is -0.35, and the contribution value is -35 cubic meters per hour. When the liquid nitrogen wash tail gas flow rate rises from 18 cubic meters per hour to 32 cubic meters per hour, the influence coefficient is 0.28, and the contribution value is 28 cubic meters per hour. After interpolation processing, the sequence resolution reaches 100 milliseconds, clearly showing the dynamic differences in the rapid establishment of methanol influence and the gradual formation of liquid nitrogen influence.

[0027] In the embodiment of the present invention, the data preprocessing link further 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 failure and exceeds the normal range, the 5-point median filter is used to replace the outlier. The time series truncation uses a 60-second sliding window with a step size of 10 seconds. The time deviation is aligned through linear interpolation, and the maximum deviation is controlled within 50 milliseconds. The feature extraction results of the dual-channel model show that the 3rd to 5th dimensions in the methanol component feature vector reflect the flow rate change rate, and the 7th to 9th dimensions in the liquid nitrogen component feature vector highlight the peak duration, enhancing the pertinence of decoupling.

[0028] It can be understood that the embodiment of the present invention does not strictly limit the length of the filtering window or the number of neural network layers. Those skilled in the art can adjust the parameters according to actual needs to adapt to the disturbance characteristics of different production scenarios.

[0029] S103. According to the output two sets of independent disturbance influence time series data of methanol flash vapor and liquid nitrogen wash tail gas, combined with their respective time series characteristics, calculate the net change amount of the total air demand in the future for a period of time through the superposition algorithm, and generate the predicted total air demand change trajectory.

[0030] In the embodiments of the present invention, the system first uses a temporal feature extractor to extract key features from the independent influence sequences of methanol flash vapor and tail gas from liquid nitrogen wash, including the perturbation occurrence time, duration, and change rate. These features reflect the dynamic patterns of the influence of the two gases on air demand. For example, in chemical production, the flow rate of methanol flash vapor may suddenly drop from 25 cubic meters per hour to 10 cubic meters per hour, with a change rate of approximately -0.3 cubic meters per hour per second and a duration between 120 and 180 seconds, while the flow rate of the tail gas from liquid nitrogen wash slowly rises from 18 cubic meters per hour to 32 cubic meters per hour, with a change rate of approximately 0.1 cubic meters per hour per second and a duration of up to 300 to 360 seconds. The extracted features are identified through a long short-term memory network to generate perturbation feature vectors.

[0031] Furthermore, the system uses a weight calculator to determine the weights of the methanol flash vapor sequence and the tail gas from liquid nitrogen wash sequence based on the perturbation feature vectors, and performs a weighted superposition operation on the two independent influence sequences through a sequence superposition calculator. Subsequently, a Butterworth low-pass filter is used to smooth the superposition result to generate a net influence sequence to characterize the comprehensive change of the total air demand. In the embodiments of the present invention, the long short-term memory network is configured with two hidden layers, each layer containing 64 neurons, screening key temporal information through forget gates and input gates, and the output feature vector has a dimension of 16. The weight calculation is based on the mutability and persistence scores of the feature vectors. For example, the mutability feature score of methanol flash vapor is relatively high at 0.85, and the persistence score is relatively low at 0.35, while the mutability score of the tail gas from liquid nitrogen wash is only 0.25 and the persistence score is as high as 0.92. Based on this, the weight of the methanol sequence is calculated as 0.6, and the weight of the liquid nitrogen sequence is 0.4. After superposition, the cut-off frequency of the filter is set to 0.1 Hz, effectively filtering out high-frequency noise to ensure the smoothness and credibility of the net influence sequence.

[0032] On this basis, the system constructs a time series prediction basis function through an autoregressive moving average filter, extrapolates the net impact sequence outside the sample within a 480-second prediction window to generate a predicted sequence of the total air demand change, and uses the wavelet decomposition method to split it into a trend component sequence and a fluctuation component sequence. Then, through delay correction, the final predicted trajectory of the net change in the total air demand is generated. In practical applications, the autoregressive moving average filter sets the autoregressive order to 4 and the moving average order to 2, and uses the historical data of the past 10 minutes to train the basis function. The prediction results show that the sudden decrease in the methanol flash vapor results in a decrease in air demand of 15 cubic meters per hour in the first 180 seconds, while the peak value of the tail gas from the liquid nitrogen wash increases by 12 cubic meters per hour between 240 and 360 seconds. The wavelet decomposition adopts a three-layer decomposition structure and selects the db6 wavelet basis function. The trend component reflects the long-term change trend, and the fluctuation component captures the characteristics of short-term disturbances. The delay estimator determines the delay compensation amount by calculating the autocorrelation function of the fluctuation component sequence and 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.

[0033] In the embodiment of the present invention, the time characteristics of the disturbances of the two gases are fully considered in the prediction process. The sudden decrease in the methanol flash vapor usually quickly affects the air demand, while the gradual increase in the tail gas from the liquid nitrogen wash shows a longer response period. Through weighted superposition and filtering, the system can smoothly integrate the two effects and avoid prediction errors caused by time series misalignment. In a case of a chemical plant, the predicted trajectory shows that the negative impact of the methanol flash vapor appears 120 seconds in advance, while the positive impact of the tail gas from the liquid nitrogen wash is delayed by 60 seconds. The corrected trajectory has a coincidence degree with the actual demand change of more than 95%.

[0034] In addition, to enhance the robustness of the prediction, the system applies the exponential smoothing algorithm to the trend component sequence with a smoothing coefficient set to 0.3, which not only retains the main change trend but also effectively suppresses the interference of short-term fluctuations. In actual operation, this method enables the air demand prediction to reflect the disturbance impact in advance. It can be understood that the embodiment of the present invention does not impose fixed restrictions on the prediction window length or filtering parameters, and those skilled in the art can flexibly adjust them according to specific process requirements to optimize the prediction accuracy.

[0035] S104. According to the predicted trajectory of the net air demand change and in combination with the current system state of the Claus sulfur recovery unit, generate a target air flow adjustment plan through an optimization control algorithm, compensate for the air demand change in advance, and send the adjustment instruction to the actuator.

[0036] In the embodiments of the present invention, first, according to the predicted trajectory of the net air demand change, combined with the real-time furnace temperature and oxygen content data, an air flow rate adjustment scheme is calculated. The furnace temperature is collected by multi-point temperature sensors distributed at different positions of the reaction furnace, and is maintained at 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, and the normal range is between 2% and 3%, with 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, the data alignment unit uses linear interpolation to unify the two groups of data to a 100-millisecond time resolution. The wavelet transform is decomposed into 3 layers based on the db4 basis function to extract low-frequency trend and high-frequency fluctuation characteristics, forming a characteristic coefficient sequence, which is then superimposed with the predicted trajectory to generate a fusion characteristic sequence.

[0037] S1041. Further, the deep reinforcement learning unit extracts features from the fusion feature sequence to generate an air flow rate adjustment parameter matrix, and calculates the optimal adjustment sequence based on the dynamic programming algorithm, and finally converts it into an executable control instruction. In the embodiments of the present invention, the state space of the deep reinforcement learning unit contains 16 dimensions, such as furnace temperature, oxygen content, and predicted demand change. The action space defines the air flow rate adjustment range as -20% to +20%. The reward function comprehensively considers temperature deviation, oxygen content stability, and adjustment cost, and optimizes the strategy through multiple rounds of trial and error. The output adjustment parameter matrix contains information on the adjustment timing and amplitude. The constraint optimization unit extracts the conditions of the flow rate upper limit of 120 cubic meters per hour, the lower limit of 80 cubic meters per hour, and the change rate not exceeding 2 cubic meters per hour per second. The dynamic programming algorithm aims to minimize the deviation and generates a stepped adjustment sequence. For example, within 5 minutes when the predicted air demand decreases by 15 cubic meters per hour, it is adjusted in stages: 5 cubic meters per hour is reduced in the first minute, 6 cubic meters per hour is reduced in the next two minutes, and 4 cubic meters per hour is reduced in the last two minutes. The instruction conversion unit samples the sequence into 30 control points at 10-second intervals and sends them to the air supply actuator through the Modbus bus in the form of a 4-20 mA standard signal.

[0038] In addition, the air flow regulation coefficient and the oxygen content gradient are integrated to optimize the adjustment scheme. The change amplitude and rate data are extracted from the predicted trajectory. For example, the air demand drops from 100 cubic meters per hour to 85 cubic meters per hour, and the change rate is about -0.05 cubic meters per hour per second. The exponential weighted moving average calculator is used for smoothing, with an initial weight of 0.2 and a 50% decay every 60 seconds, to generate a smoothed change curve. The state parameter vector is composed of normalized temperature, pressure, and gas component data. Support vector regression maps features with a Gaussian kernel function, and the calculated initial regulation coefficient ranges from 0.85 to 1.15. The oxygen content gradient is obtained through a 60-second window difference operation. For example, the oxygen content rises from 2.5% to 3.2%, with a rate of 0.006% per second. The adaptive threshold segmenter divides the stable and fluctuating segments based on a variance of 0.15. The weight allocation unit dynamically adjusts the combined weights of the coefficient and the gradient according to the state. When the temperature is 1100 degrees Celsius, the weight of the regulation coefficient is 0.7, and the weight of the gradient is 0.3. When the temperature is too high, up to 1150 degrees Celsius, the weight of the gradient rises to 0.6. After weighted superposition, the comprehensive regulation parameter is corrected by historical feedback compensation and finally discretized into an adjustment time series at 5-second intervals to ensure a smooth transition.

[0039] S1042. After the instruction is generated, the system checks the actuator response delay and the furnace temperature change rate constraint to ensure the feasibility of the adjustment scheme. If it exceeds the range, it is re-optimized and the final instruction sequence is generated. In the embodiment of the present invention, the actuator response delay is obtained through real-time measurement, and the typical value is between 2 and 5 seconds. After smoothing with a 20-second sliding average window, the input long short-term memory network extracts an 8-dimensional delay feature vector. The furnace temperature change rate is calculated by a 5-second difference. For example, when the furnace temperature drops from 1150 degrees Celsius to 1080 degrees Celsius, and the change rate of -14 degrees Celsius per minute exceeds the threshold of ±10 degrees Celsius per minute, a recalculation is triggered. The constraint optimization regenerates the adjustment curve based on the delay feature and the temperature constraint, and the recursive least squares method is smoothed with a forgetting factor of 0.95. After verification, the change rate is controlled within ±9 degrees Celsius per minute. The final instruction sequence is quantized into 5 levels from 90 to 110 cubic meters per hour, including 36 control points, to ensure that the fluctuations of the furnace temperature and the oxygen content are respectively stable within ±10 degrees Celsius and ±0.3%.

[0040] In the embodiment of the present invention, this multi-level optimization and verification mechanism effectively improves the response speed and stability of air supply. Technicians can adjust the parameters according to the actual working conditions to further optimize the control effect.

[0041] S105. While sending the air flow adjustment instruction sequence to the actuator of the air supply system of the Claus sulfur recovery unit, collect the data of the adjusted oxygen content, furnace temperature, and sulfur dioxide concentration in real time, generate a data stream containing execution records and performance feedback, evaluate the adjustment effect by comparing with the predicted net air demand trajectory and control target, and online optimize the decoupling model and control algorithm parameters when the effect deviation is large to improve the accuracy and response speed of subsequent control.

[0042] In the embodiment of the present invention, the system monitors key process parameters in real time through a multi-parameter collector. Among them, the oxygen content data is collected at a period of 100 milliseconds, with a range of 2% to 3% and accompanied by noise of ±0.2%; the furnace temperature data is collected at a period of 200 milliseconds, with a range of 1000 to 1200 degrees Celsius and noise of ±5 degrees Celsius; the sulfur dioxide concentration data is collected at a period of 500 milliseconds, with a range of 8% to 12% and noise of ±0.5%. To ensure data quality, the data preprocessing unit uses a 5-point moving average method to denoise and smooth the original data, generating a stable processed data sequence. Subsequently, the long short-term memory network extracts temporal features from these sequences. The network is configured with a two-layer structure, with 64 neurons in each layer. The input is the processed data of 60 time points, and the output is 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 change trends of the parameters.

[0043] Furthermore, obtain the control target interval from the process database, that is, the oxygen content is 2.3% to 2.7%, the furnace temperature is 1080 to 1150 degrees Celsius, and the sulfur dioxide concentration is 9% to 11%. Calculate the deviation degree between the real-time performance index vector and the target interval through the performance index evaluator, and use the trajectory comparison calculator to measure the fitting degree based on the Euclidean distance to judge the coincidence degree between the actual effect and the predicted trajectory. In the embodiment of the present invention, the performance evaluation uses the Mahalanobis distance to quantify the deviation, and the fitting degree threshold is set to 0.9. If the fitting degree is lower than the threshold, the Kalman filter dynamically tracks the deviation degree and the fitting degree, and generates a deviation feature vector through prediction and correction. For example, in a certain disturbance, the oxygen content rises to 3.1%, the furnace temperature drops to 1050 degrees Celsius, and the sulfur dioxide concentration reaches 11.5%, and the fitting degree is only 0.82, indicating that there is an obvious deviation between the actual response and the prediction.

[0044] 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 the gradient optimization and backpropagation algorithms to generate an updated parameter set to optimize subsequent control. In the embodiment of the present invention, the decoupling matrix has a 4×4 structure and is used to separate the effects of methanol flash vapor and liquid nitrogen wash tail gas. The correction amount calculation shows that the influence weight of methanol flash vapor on the oxygen content needs to be increased from 0.65 to 0.78, the influence on the furnace temperature is reduced from 0.55 to 0.45, while the influence of liquid nitrogen wash tail gas on the sulfur dioxide concentration is increased from 0.35 to 0.52. The gradient optimization sets the learning rate to 0.01. After 50 iterations, the weight matrix converges. The new parameter set reduces the deviation by about 35% and shortens the response time by 40% in the next similar perturbation, significantly improving the control robustness.

[0045] In addition, during actual operation, the system ensures process stability through multi-parameter real-time monitoring and feedback optimization. For example, when the air flow rate is adjusted, if the oxygen content exceeds the target upper limit by 0.4%, the furnace temperature is lower than the lower limit by 30 degrees Celsius, and the sulfur dioxide concentration exceeds the standard by 0.5%, the deviation feature vector reflects the problem of response lag. The optimized parameter set adjusts the estimation accuracy of the decoupling model's contribution to the perturbation, making subsequent adjustments closer to actual requirements. It can be understood that the embodiment of the present invention does not strictly limit the sampling period or the number of network layers, and those skilled in the art can adjust according to specific scenarios to further improve the effect.

[0046] The present invention provides an intelligent control system for a Claus sulfur recovery device, which mainly includes: a real-time monitoring and scenario recognition module for real-time monitoring of the methanol flash vapor flow rate and the liquid nitrogen wash tail gas flow rate, and identifying a disturbance scenario where the methanol flash vapor flow rate undergoes a stepwise decrease and the liquid nitrogen wash tail gas peak occur simultaneously through a preset threshold and logic. If this scenario is identified, the real-time data at this moment and within a subsequent preset time period of this scenario is collected to form an initial disturbance event record; a decoupling model calculation module for calling a pre-established decoupling model and using the real-time data in the initial disturbance event record to calculate and separate the contribution values of the methanol flash vapor decrease and the liquid nitrogen wash tail gas peak to the total air demand independently and dynamically, and outputting two sets of time series data of independent disturbance impacts; a net air demand prediction module for calculating the net change in the total air demand within a future preset time period based on the two sets of output time series data of independent disturbance impacts through a superposition algorithm and combining their respective time series characteristics, and obtaining the predicted net air demand change trajectory; an optimal control calculation module for calculating a target air flow adjustment plan according to the predicted net air demand change trajectory and combining the current system state, using an optimal control algorithm to compensate in advance for the predicted air demand change, and sending the obtained air flow adjustment instruction sequence to the actuator of the air supply system for execution; an effect evaluation and model update module for real-time collecting the actual oxygen content, furnace temperature, and products after the adjustment is executed, outputting a data stream including control execution records and real-time performance feedback, comparing the predicted net air demand trajectory with the set control target, evaluating the actual effect of the target air flow adjustment plan, and if there is a significant deviation between the actual effect and the expectation, online updating the decoupling model parameters and / or the weights of the optimal control algorithm, and outputting an adjusted set of model parameters. The above description is only a preferred embodiment of the present application and an explanation of the applied technical principles. Those skilled in the art should understand that the scope of the invention involved in the present application is not limited to the technical solution formed by the specific combination of the above technical features, but should also cover other technical solutions formed by any combination of the above technical features or their equivalent features without departing from the concept of the present application. For example, a technical solution formed by mutually replacing the above features with (but not limited to) technical features with similar functions disclosed in the present application.

Claims

1. An intelligent control method for a Claus sulfur recovery unit, characterized in that, The method includes: Real-time monitoring of the methanol flash vapor flow rate and the tail gas flow rate of liquid nitrogen washing. Through a preset threshold and logic, a disturbance scenario where the stepwise decrease of the methanol flash vapor flow rate coincides with the peak value of the tail gas of liquid nitrogen washing is identified. If this scenario is identified, real-time data at this moment and within a preset time period afterwards in this scenario is collected to form an initial disturbance event record. Call a pre-established decoupling model, use the real-time data in the initial disturbance event record, calculate and separate the contribution values of the decrease in methanol flash vapor and the peak value of the tail gas of liquid nitrogen washing to the total air demand independently and dynamically, and output time series data of two sets of independent disturbance effects. According to the output time series data of the two sets of independent disturbance effects, through a superposition algorithm and combined with their respective time series characteristics, calculate the net change in the total air demand within a preset future time period to obtain the predicted net air demand change trajectory. Based on the predicted net air demand change trajectory and combined with the current system state, use an optimal control algorithm to calculate the target air flow adjustment plan, compensate in advance for the predicted air demand change, and send the obtained air flow adjustment instruction sequence to the actuator of the air supply system for execution.

2. The method according to claim 1, characterized in that, The real-time monitoring of the methanol flash vapor flow rate and the tail gas flow rate of liquid nitrogen washing, through a preset threshold and logic, identifies a disturbance scenario where the stepwise decrease of the methanol flash vapor flow rate coincides with the peak value of the tail gas of liquid nitrogen washing. If this scenario is identified, real-time data at this moment and within a preset time period afterwards in this scenario is collected to form an initial disturbance event record, including: Collect the methanol flash vapor pressure measurement value and the tail gas pressure measurement value of liquid nitrogen washing through a pressure sensing monitoring unit, and perform a moving average filtering process on the pressure measurement value to obtain a pressure processed value and a pressure mutation mark; According to the pressure processed value, perform mutation point detection on the methanol flash vapor flow rate measurement value and the tail gas flow rate measurement value of liquid nitrogen washing to obtain the moment point of the flow rate stepwise decrease and the moment point of the flow rate peak value; For the flow rate moment point, perform wavelet transform decomposition on the methanol flash vapor temperature measurement value and the tail gas temperature measurement value of liquid nitrogen washing to obtain temperature characteristic values, and use a recurrent neural network to perform feature extraction on the temperature characteristic values to obtain a temperature change feature vector; Generate a disturbance event data record according to the pressure mutation mark, the flow rate moment point, and the temperature change feature vector.

3. The method according to claim 1, characterized in that The call to the pre-established decoupling model, using the real-time data in the initial disturbance event record, calculates and separates the contribution values of the decrease in methanol flash vapor and the peak value of the tail gas of liquid nitrogen washing to the total air demand independently and dynamically, and outputs time series data of two sets of independent disturbance effects, including: Perform normalization processing on the methanol flash vapor flow rate data and the tail gas flow rate data of liquid nitrogen washing according to the initial disturbance event record to obtain normalized flow rate data, and perform denoising processing on the normalized flow rate data through a sliding median filter to obtain smooth flow rate data; Perform orthogonal transformation on the smooth flow rate data using principal component analysis to obtain an eigenvalue matrix and an eigenvector matrix, and sort the eigenvectors according to the eigenvalue size and select the first two eigenvectors to construct an orthogonal basis space; The smoothed flow rate data is mapped to the orthogonal basis space through a projection operator to obtain the methanol flash vapor load coefficient and the LNG wash tail gas load coefficient, and spline interpolation is used for smoothing to obtain the load time series; A recurrent neural network is used to extract the dynamic features in the load time series to obtain a dynamic feature vector, and the mutual information entropy calculator is used to calculate the correlation degree between the components of the dynamic feature vector to construct an independence evaluation matrix, and the independence evaluation matrix is diagonalized to obtain the independent contribution degree; According to the independent contribution degree, the load time series is weighted and combined to obtain the independent influence time series of methanol flash vapor on the total air demand and the independent influence time series of LNG wash tail gas on the total air demand.

4. The method according to claim 1, characterized in that The pre-established decoupling model is called, and the real-time data in the initial disturbance event record is used to calculate and separate the contribution values of the methanol flash vapor drop and the LNG wash tail gas peak to the independent and dynamic influence of the total air demand respectively, and the time series data of two groups of independent disturbance influences are output, including: dynamically intercepting the real-time initial disturbance event data stream to generate a time-aligned input matrix, inputting the decoupling model including a dual-channel feature extraction structure based on a gated recurrent unit, and respectively outputting the dynamic influence coefficient of the methanol component and the dynamic influence coefficient of the LNG component; generating an independent contribution sequence of the methanol flash vapor drop according to the product operation of the dynamic influence coefficient of the methanol component and the total air demand; generating an independent contribution sequence of the LNG wash tail gas peak according to the product operation of the dynamic influence coefficient of the LNG component and the total air demand; respectively performing timestamp alignment processing on the independent contribution sequences to output two groups of influence factor time series data with the same time resolution.

5. The method according to claim 1, wherein According to the output time series data of two groups of independent disturbance influences, the net change amount of the total air demand within a preset period of time in the future is calculated through a superposition algorithm and combined with their respective time series characteristics to obtain the predicted net air demand change trajectory, including: A time series feature extractor is used to extract the methanol flash vapor disturbance feature and the LNG wash tail gas disturbance feature from the independent disturbance influence sequence, and the disturbance feature vector is obtained through recognition by a long short-term memory network according to the disturbance feature; For the disturbance feature vector, the methanol flash vapor sequence weight and the LNG wash tail gas sequence weight are calculated through a weight calculator, and the two groups of independent influence sequences are subjected to weighted superposition operation and filtering processing to obtain a net influence sequence; According to the net influence sequence, an autoregressive moving average is used to construct a time series prediction basis function, and out-of-sample extension is performed within a preset time window to obtain a total air demand change amount prediction sequence; For the total air demand change amount prediction sequence, a trend component sequence and a fluctuation component sequence are obtained through wavelet decomposition, the time delay compensation amount is determined according to the autocorrelation function of the fluctuation component sequence, and the trend component sequence is corrected for time delay to obtain the total air demand net change prediction trajectory.

6. The method according to claim 1, characterized in that, Based on the predicted change trajectory of the net air demand and combined with the current system state, an optimized control algorithm is used to calculate the target air flow adjustment plan, compensating in advance for the predicted change in air demand, and the obtained air flow adjustment instruction sequence is sent to the actuator of the air supply system for execution, including: Using a Kalman filter to perform noise elimination 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 a state monitoring sequence; Performing time stamp alignment according to the state monitoring sequence, performing multi-scale decomposition on the state monitoring sequence through wavelet transform to obtain a feature coefficient sequence, and performing time series superposition on the feature coefficient sequence and the predicted change trajectory of the net air demand to obtain a fusion feature sequence; Performing feature extraction through a deep reinforcement learning unit according to the fusion feature sequence to obtain an air flow adjustment parameter matrix; Extracting adjustment constraint conditions according to the air flow adjustment parameter matrix, generating an air flow adjustment instruction sequence, and sending it to the actuator of the air supply system for execution.

7. The method according to claim 6, characterized in that, It also includes: Generating an air flow adjustment coefficient according to the predicted change trajectory of the net air demand and the current system state, fusing the air flow adjustment coefficient and the oxygen content gradient, and outputting a target flow adjustment time series sequence; Verifying the target flow adjustment time series sequence after compensating for the actuator response delay. If the furnace temperature change rate exceeds the preset range, recalculate the target flow adjustment time series sequence, and split the time series sequence that passes the verification into air flow adjustment instructions.

8. The method according to claim 1, characterized in that The method also includes: real-time collecting the actual oxygen content, furnace temperature, and products after the adjustment is executed, outputting a data stream including control execution records and real-time performance feedback, comparing the predicted net air demand trajectory with the set control target, and evaluating the actual effect of the target air flow adjustment plan; if there is a significant deviation between the actual effect and the expectation, online update the decoupling model parameters and / or the weights of the optimized control algorithm, and output an adjusted model parameter set.

9. The method according to claim 8, characterized in that, The real-time collection of the actual oxygen content, furnace temperature, and products after the adjustment is executed, outputting a data stream including control execution records and real-time performance feedback, comparing the predicted net air demand trajectory with the set control target, and evaluating the actual effect of the target air flow adjustment plan; If there is a significant deviation between the actual effect and the expectation, online update the decoupling model parameters and / or the weights of the optimized control algorithm, and output an adjusted model parameter set, including: Obtaining the original oxygen content data, furnace temperature original data, and sulfur dioxide concentration original data, performing denoising and smoothing processing through a data preprocessing unit to obtain a processed data sequence, and extracting time series features from the processed data sequence according to a long short-term memory network to obtain a real-time performance index vector; Obtaining the oxygen content target interval, furnace temperature target interval, and sulfur dioxide concentration target interval from the process database, and calculating the deviation degree between the real-time performance index vector and the target interval through a performance index evaluator; Calculating the fitting degree between the real-time performance index vector and the predicted change trajectory of the net air demand through a trajectory comparison calculator to obtain a fitting degree value, and judging whether the fitting degree value exceeds a preset threshold through a difference detector; If the deviation degree exceeds the preset range, the decoupling matrix correction amount is calculated by the parameter optimization unit based on the deviation feature vector, and the control weight matrix is optimized and adjusted according to the decoupling matrix correction amount to obtain an updated parameter set.

10. An intelligent control system for a Claus sulfur recovery unit, characterized in that, The system includes: a real-time monitoring and scenario recognition module, which is used to monitor the methanol flash vapor flow rate and the liquid nitrogen wash tail gas flow rate in real time, identify the disturbance scenario where the methanol flash vapor flow rate suddenly drops and the liquid nitrogen wash tail gas peak occur simultaneously through a preset threshold and logic, and if this scenario is identified, collect the real-time data at this moment and in the subsequent preset time period of this scenario to form an initial disturbance event record; a decoupling model calculation module, which is used to call a pre-established decoupling model, use the real-time data in the initial disturbance event record to calculate and separate the respective independent and dynamic influence contributions of the methanol flash vapor drop and the liquid nitrogen wash tail gas peak to the total air demand, and output two sets of time series data of independent disturbance influences; a net air demand prediction module, which is used to calculate the net change amount of the total air demand in the future preset time period according to the output two sets of time series data of independent disturbance influences, through a superposition algorithm and combined with their respective time series characteristics, to obtain the predicted net air demand change trajectory; an optimal control calculation module, which is used to calculate the target air flow adjustment plan according to the predicted net air demand change trajectory and combined with the current system state, use the optimal control algorithm to compensate in advance for the predicted air demand change, and send the obtained air flow adjustment instruction sequence to the actuator of the air supply system for execution; an effect evaluation and model update module, which is used to collect the actual oxygen content, furnace temperature and products in real time after the adjustment is executed, output a data stream including control execution records and real-time performance feedback, compare the predicted net air demand trajectory with the set control target, evaluate the actual effect of the target air flow adjustment plan, and if there is a significant deviation between the actual effect and the expectation, update the decoupling model parameters and / or the optimal control algorithm weights online, and output an adjusted model parameter set.

Images