An energy-saving control method for reflux heating in a distillation column

Abstract

This invention relates to the field of process optimization and control technology, and provides an energy-saving control method for reflux heating in distillation columns. By denoising, aligning, and completing various sensor data such as column temperature, flow rate, and pressure, standardized process parameters are generated. Combined with edge computing units, multiple real-time operating condition features are extracted, dynamically constructing a process semantic graph and achieving millisecond-level recalibration of variable correlation strength, eliminating priority ambiguity issues in multivariate modeling. Based on model prediction calibration output, the reflux temperature setpoint is automatically adjusted, driving the process actuator in a closed loop, thereby achieving adaptive trade-off control between product quality and energy consumption targets. This invention significantly improves the accuracy of distillation column operating condition identification, response speed, and energy consumption optimization capabilities.

Description

Technical Field

[0001] This invention relates to the field of process optimization and control technology, and in particular to an energy-saving control method for reflux heating in a distillation column. Background Technology

[0002] Currently, traditional methods for optimizing and controlling energy consumption in distillation columns generally employ static regression models or fixed neural network structures. These models treat key process parameters such as column top temperature, reflux temperature, and reboiler vapor pressure as having equal weights or weights assigned based on experience under a single operating condition. Through multivariate statistical analysis or machine learning training, a global relationship model between these parameters is obtained, and the operating conditions within the column are optimized accordingly. While these methods can achieve a certain level of predictive accuracy and energy savings under ideal or single-condition operating conditions, the limitations of traditional solutions become increasingly apparent as the demands for high-purity separation and extreme energy consumption constraints in fine chemical industries such as lithium battery material recovery and specialty solvent refining become increasingly prominent. First, fixed-weighted variable coupling models struggle to handle dynamic and ever-changing industrial conditions. Minor fluctuations in feed composition, unpredictable changes in equipment status (such as reboiler scaling), and seasonal shifts in environmental conditions (temperature, humidity, air pressure, etc.) significantly impact the thermodynamic balance and energy distribution within the tower. Existing models cannot dynamically adjust the weight distribution among variables based on actual operating conditions, leading to slow responses to sudden or evolving disturbances. This significantly reduces the accuracy of model predictions and the robustness of the control system. Second, current mainstream modeling methods rely heavily on big data, often implying the influence of variables within the model's black-box structure, lacking transparency and interpretability in distinguishing variable priorities. This not only reduces industrial users' trust in the model but also creates knowledge barriers for later maintenance and process optimization. In practice, the actual contributions of parameters such as column top temperature, reflux liquid temperature, and reboiler steam pressure vary significantly depending on operating conditions, but it is usually difficult to trace the reasons and processes for variable weight adjustments. Furthermore, existing multivariate models are mostly linear or quasi-linear frameworks, making it difficult to accurately characterize complex physical mechanisms and operational semantics such as non-equilibrium, limited heat transfer, and equipment degradation. For example, under conditions such as localized reboiler scaling or a sudden increase in feed moisture content, the control priority of certain variables should be temporarily increased or decreased; otherwise, separation efficiency may decrease or unnecessary energy consumption may increase. Due to the lack of a mechanism-based dynamic variable priority allocation strategy, traditional methods struggle to achieve online adjustment of variable priorities without model reconstruction. Moreover, once the weights between variables are determined, subsequent changes often require manual intervention and offline retraining, failing to meet the real-time and adaptive requirements of production sites. Especially in extreme application scenarios involving precision separation, high purity requirements, and energy sensitivity, delayed model updates directly impact product quality and energy costs. Therefore, a novel modeling framework is needed, possessing the capabilities of process knowledge induction, operational semantic expression, and equipment state adaptation. This framework can map key parameters of the distillation column into structured process nodes, integrate process property databases with actual operating condition labeling information, and shift variable weight allocation from static to dynamic migration in real time according to operating conditions. This enables transparent dynamic adjustment of variable priorities and traceability of physical mechanisms, improving the model's prediction accuracy and control stability under different disturbances and evolving conditions. This provides robust data and decision support for reflux temperature setting, energy consumption optimization, and separation accuracy assurance, better meeting the dual demands of energy saving and quality in high-end distillation processes. Summary of the Invention

[0003] In order to solve the above-mentioned technical problems, the present invention provides an energy-saving control method for reflux heating in a distillation column.

[0004] The technical solution of this invention is implemented as follows: an energy-saving control method for reflux heating in a distillation column, comprising: S1: Acquire multi-source heterogeneous sensor data during the operation of the distillation column, and preprocess the multi-source heterogeneous sensor data to generate a standardized process parameter sequence; S2: Based on the standardized process parameter sequence, extract the operating condition feature tags, identify the real-time operating condition status according to the operating condition feature tags, and generate a dynamic operating condition tag stream; S3: Based on the boiling point-liquid phase composition-pressure relationship in the NMP physical property database, construct a multi-layer process semantic map containing a static topology layer, a dynamic weighting layer, and a semantic constraint layer; S4: Input the dynamic working condition label stream into the multi-layer process semantic graph, match the preset trigger conditions to locate the associated edge nodes, recalibrate the correlation strength between variables in the dynamic weight layer according to the preset migration rules, and generate a dynamic variable priority weight matrix. S5: Based on the dynamic variable priority weight matrix, calculate the coupling relationship between key nodes to generate the output value of the temperature-efficiency-energy consumption correlation prediction model; S6: Compare the output value of the temperature-efficiency-energy consumption correlation prediction model with the preset NMP purity target threshold, calculate the separation efficiency deviation and energy consumption redundancy, and generate the reflux temperature setpoint correction factor. S7: Using the reflux temperature setpoint correction factor, combined with the dynamic variable priority weight matrix, calculate the reflux temperature dynamic setting trajectory, and combined with temperature closed-loop error correction, generate a drive command to drive the closed-loop controller to perform reflux temperature adjustment action. S8: Monitor the actual column top temperature and distillate color online spectral characteristic values ​​after the reflux temperature adjustment action is executed. If a control response lag or a decrease in prediction accuracy is detected, trigger the weight transfer protocol to update the dynamic variable priority weight matrix and complete the online adaptive optimization closed loop of the model.

[0005] The present invention provides an energy-saving control method for reflux heating in a distillation column, which has the following beneficial effects: (1) In view of the problem that traditional fixed weight models are difficult to cope with dynamic and changing industrial field conditions and the lack of transparent differentiation of variable priorities in the black box of the model, this invention constructs a multi-layer process semantic graph and incorporates six types of core nodes with clear thermodynamic meaning and control semantics into a unified knowledge expression framework, realizing the structural deconstruction and formal modeling of the internal mechanism of the system. In the static topology layer, this invention explicitly defines causal chains such as "changes in feed flow rate cause fluctuations in vapor-liquid load, which in turn affect the reflux liquid temperature response". Combined with the edge weight mechanism based on the working condition label activation in the dynamic weight layer, the system can adaptively adjust the influence distribution between key variables under different operating conditions. For example, when the feed moisture content is detected to increase, the influence weight of feed flow rate on reflux liquid temperature is automatically increased, while the traditional judgment basis of column top temperature on separation efficiency is weakened, so as to more accurately reflect the actual mass transfer behavior under high moisture conditions. This design significantly improves the working condition identification accuracy and prediction consistency of the distillation column under non-standard working conditions, and effectively overcomes the defects of traditional models such as slow response and inaccuracy under disturbance environment. (2) To address the problems of traditional methods lacking interpretable adjustment paths, variable priorities not being dynamically divided online, and the need for manual offline retraining for weight changes, which cannot meet the real-time requirements of the field, this invention introduces a millisecond-level weight transfer protocol driven by lightweight chemical condition labels. This protocol enables real-time knowledge injection and parameter recalibration without retraining the model. Once the status labels such as "mid-stage reboiler scaling" and "sudden increase in ambient humidity" output by edge-side sensors and analysis modules are matched to the preset trigger conditions in the semantic graph, the weight transfer logic of the local subgraph is immediately initiated. Only the predefined intensity correction strategy is executed for key related edges, avoiding the computational burden and stability risks caused by global reconstruction. The entire process is fully embedded in the DCS control system and executed locally without relying on cloud computing power. This meets the stringent requirements of industrial sites for response speed and operational reliability. All weight changes originate from traceable process rules and expert experience. Each adjustment can be attributed to specific physical state or operational constraint changes, providing a complete audit trail and compliance transparency. This breaks through the technical bottleneck of the limited application of black-box models in high-risk chemical scenarios. (3) To address the challenges of accurately characterizing complex physical mechanisms such as non-equilibrium heat transfer and equipment degradation using linear or quasi-linear frameworks, and the difficulty in achieving synergistic optimization of quality and energy consumption under the dual constraints of extreme energy consumption and high-purity separation in lithium battery recycling and special solvent refining, this invention integrates ternary phase equilibrium relationships, thermosensitive degradation thresholds, and color stability indicators unique to the lithium battery recycling field from the NMP property database. It embeds strong prior knowledge boundaries in the semantic constraint layer to ensure that control decisions are always within the safe and feasible domain, while also considering the consistency requirements of product quality. This mechanism-data-objective triple-coupling architecture design enables the closed-loop controller not only to respond to current disturbances but also to predict potential deviation trends, guiding the reflux temperature setpoint towards the optimal trajectory in advance, thereby achieving synergistic optimization of energy consumption, efficiency, and purity. Attached Figure Description

[0006] Figure 1 A flowchart of an energy-saving control method for reflux heating in a distillation column according to the present invention; Figure 2 This is a sub-flowchart of an energy-saving control method for reflux heating in a distillation column according to the present invention. Figure 3 This is another sub-flowchart of an energy-saving control method for reflux heating in a distillation column according to the present invention. Detailed Implementation

[0007] Embodiments of the present invention are described in detail below, examples of which are shown in the accompanying drawings, wherein the same or similar reference numerals denote the same or similar elements or elements having the same or similar functions throughout. The embodiments described below with reference to the accompanying drawings are exemplary and are only used to explain the present invention, and should not be construed as limiting the present invention.

[0008] The following disclosure provides many different embodiments or examples for implementing different structures of the invention. To simplify the disclosure, specific examples of components and arrangements are described below. Of course, these are merely examples and are not intended to limit the invention. Furthermore, reference numerals and / or letters may be repeated in different examples; such repetition is for simplification and clarity and does not in itself indicate a relationship between the various embodiments and / or arrangements discussed. like Figure 1 As shown, this invention provides an energy-saving control method for reflux heating in a distillation column, specifically including: S1: Acquire multi-source heterogeneous sensor data during the operation of the distillation column, including column top temperature, reflux liquid temperature, reboiler vapor pressure, feed flow rate, column bottom liquid level and condenser outlet temperature, and perform noise reduction and time sequence alignment preprocessing on the multi-source heterogeneous sensor data to generate a standardized process parameter sequence. S2: Based on the standardized process parameter sequence, use the edge computing unit to extract lightweight chemical condition feature tags, identify real-time operating conditions including sudden changes in feed moisture content, reboiler scaling stage identification and environmental humidity jumps, and generate dynamic operating condition tag stream. S3: Based on the boiling point-liquid phase composition-pressure relationship in the NMP physical property database, construct a process semantic graph containing a static topology layer, a dynamic weight layer, and a semantic constraint layer. The static topology layer defines the inherent causal relationship between nodes, and the dynamic weight layer binds the type of operating condition label that can be activated, generating a multi-layer process semantic graph structure. S4: Input the dynamic working condition label stream into the multi-layer process semantic structure, match the preset trigger conditions to locate the associated edge nodes, recalibrate the correlation strength between variables in the dynamic weight layer according to the preset migration rules, and generate a dynamic variable priority weight matrix. S5: Based on the dynamic variable priority weight matrix, the coupling relationship between phase balance sensitive nodes and energy transfer dominant nodes is solved to eliminate the ambiguity of variable priority in multivariate coupling modeling and generate the output value of the temperature-efficiency-energy consumption correlation prediction model. S6: Compare the output value of the temperature-efficiency-energy consumption correlation prediction model with the target threshold of the product quality correlation node, calculate the separation efficiency deviation and energy consumption redundancy, and generate the reflux temperature setpoint correction factor. S7: Using the reflux temperature setpoint correction factor, combined with the dynamic variable priority weight matrix, calculate the reflux temperature dynamic setting trajectory, and combined with temperature closed-loop error correction, generate a drive command to drive the closed-loop controller to perform reflux temperature adjustment action. S8: Monitor the actual column top temperature and distillate color online spectral characteristic values ​​after the reflux temperature adjustment action is executed. If a control response lag or a decrease in prediction accuracy is detected, trigger the weight transfer protocol to update the dynamic variable priority weight matrix and complete the online adaptive optimization closed loop of the model.

[0009] Step S1: Acquire multi-source heterogeneous sensor data during the operation of the distillation column, including column top temperature, reflux liquid temperature, reboiler vapor pressure, feed flow rate, column bottom liquid level, and condenser outlet temperature. Perform noise reduction and time-series alignment preprocessing on the multi-source heterogeneous sensor data to generate a standardized process parameter sequence. Specifically, this includes: S1.1: Acquire the raw data stream from multiple heterogeneous sensors during the operation of the distillation column, including the analog voltage signal output by the top temperature sensor, the digital communication message from the reflux liquid temperature transmitter, the pulse frequency signal from the reboiler steam pressure gauge, the vortex pulse sequence from the feed flow meter, the radar echo data packet from the bottom level gauge, and the thermocouple millivolt signal from the condenser outlet temperature. Perform protocol parsing and unit normalization on the raw data stream from multiple heterogeneous sensors to generate a set of initial process parameters with uniform dimensions. The 4-20mA analog current signal output by the temperature sensor at the top of the tower is collected and quantized into a 12-bit digital value through a high-precision A / D conversion module. Based on the linear mapping relationship, the digital value is converted into a Celsius value to eliminate zero-point drift error in the analog transmission process. The digital communication messages sent by the reflux liquid temperature transmitter based on the Modbus RTU protocol are analyzed, the original temperature code value in the register address is extracted, and floating-point calculation is performed in combination with the transmitter range coefficient to obtain a real-time reading of the reflux liquid temperature with high resolution. Receive the pulse frequency signal output from the reboiler steam pressure gauge, use a high-frequency counter to count the number of pulses within a fixed time window, and calculate the instantaneous steam pressure value based on the pressure-frequency characteristic curve to ensure rapid response and capture of dynamic pressure changes; Read the vortex pulse sequence generated by the feed flow meter, perform Schmitt trigger shaping on the pulse signal to eliminate glitch interference, accumulate the number of effective pulses per unit time and multiply by the instrument coefficient to convert it into volumetric flow rate data of standard cubic meters per hour. The radar echo data packets emitted by the tower bottom level gauge are demodulated, and the peak frequency in the echo spectrum is extracted by fast Fourier transform. Combined with the correction formula for the propagation speed of electromagnetic waves in the NMP steam environment, the accurate tower bottom level height is calculated by inversion. The microvolt-level millivolt signal generated by the thermocouple at the condenser outlet is amplified, and a cold junction temperature compensation algorithm is executed to eliminate the influence of ambient temperature fluctuations. The accurate condensate temperature measurement value is obtained by looking up the table and interpolating according to the NIST standard calibration table. Establish a unified data standardization mapping model to map the physical quantities of different dimensions and precisions to the [0,1] interval or standard engineering units, forming a set of initial process parameters with consistent structure. By using multi-protocol parallel parsing and physical dimension standardization processing, the multi-source heterogeneous raw signals from the previous step are transformed into a set of initial process parameters with a unified format and unified dimensions, realizing the transformation of the underlying data into a standardized information flow, and providing a high-quality data foundation for subsequent denoising and timing alignment. For example, the tower top temperature sensor outputs a 12mA current, which is converted to a value of 2048 by a 12-bit A / D converter, corresponding to 120.5℃; Modbus message parsing yields a reflux liquid temperature of 118.2℃; the steam pressure pulse count rate is 1500Hz, equivalent to 0.65MPa; the vortex pulse accumulation is 3000 pulses / minute, equivalent to a feed flow rate of 5.2m³ / h; the radar spectrum peak corresponds to a liquid level of 8.5 meters; and the thermocouple potential, after compensation of 15.3mV, corresponds to a temperature of 95.1℃. All data is normalized to form an initial parameter vector containing 6 dimensions, and the data refresh rate is synchronized to 100ms to ensure consistency in subsequent processing. S1.2: Based on the initial set of process parameters, wavelet transform denoising algorithm is used to denoise the column top temperature signal, reflux liquid temperature signal and reboiler steam pressure signal containing high-frequency random noise, so as to eliminate abnormal fluctuations caused by sensor electromagnetic interference and generate a denoised pure process parameter sequence. S1.3: Based on the timestamp information of each signal in the denoised pure process parameter sequence, a linear interpolation resampling algorithm is used to perform time axis calibration on the feed flow rate data, tower bottom liquid level data and condenser outlet temperature data that are phase misaligned due to transmission delay, so as to eliminate the time lag effect between multiple variables and generate a time-aligned process parameter synchronization sequence. S1.4: The sliding window statistical test method is used to fill missing values ​​and remove outliers in the time-aligned process parameter synchronization sequence. Neighbor mean repair is performed on data breakpoints in scenarios such as sudden changes in feed flow or excessive reboiler steam pressure to ensure data continuity and generate a clean process parameter sequence that passes integrity verification. S1.5: Based on the clean process parameter sequence that has passed the integrity verification, perform data structuring and encapsulation operations to map the column top temperature, reflux liquid temperature, reboiler steam pressure, feed flow rate, column bottom liquid level and condenser outlet temperature into array objects with a fixed format, so as to form a standardized process parameter sequence that can be directly called by the edge computing unit, wherein the edge computing unit is used to perform millisecond-level calculations of lightweight chemical condition feature extraction and weight transfer protocol.

[0010] Step S2: Based on the standardized process parameter sequence, the edge computing unit extracts lightweight chemical condition feature tags to identify real-time operating conditions, including sudden changes in feed moisture content, reboiler scaling stage indicators, and abrupt changes in ambient humidity, generating a dynamic operating condition tag stream. Specifically, this includes: S2.1: Obtain the feed flow rate time series data and the tower top temperature time series data from the standardized process parameter sequence, and use the sliding window wavelet transform algorithm to perform multi-scale decomposition processing on the feed flow rate time series data and the tower top temperature time series data to separate high-frequency noise components and low-frequency trend components, and generate a denoised feature dataset containing feed fluctuation feature vectors and temperature response hysteresis feature vectors. S2.2: Based on the feed fluctuation feature vector in the denoised feature dataset, call the preset infrared spectral soft measurement mapping model to perform moisture content inversion calculation, map the feed fluctuation feature vector into a real-time feed moisture content estimate, and perform a comparison logic operation between the real-time feed moisture content estimate and the preset moisture content change threshold range to generate a feed moisture content change initial judgment flag bit that characterizes the abnormal state of the feed components; Receive the feed fluctuation feature vector from the denoised feature dataset generated in step S2.1. This vector contains the low-frequency trend component and the high-frequency residual component after wavelet decomposition, and serves as the input basis for moisture content inversion. The infrared spectral soft measurement mapping model pre-placed in the edge computing unit is invoked. This model is based on the near-infrared absorption spectral characteristics of the NMP-water binary mixture and uses a multivariate partial least squares regression algorithm to process the feed fluctuation feature vector. Extract the absorbance values ​​at specific wavelengths corresponding to the stretching vibration frequency band of the OH bonds in water molecules from the feature vector, construct a spectral response matrix, and calculate the estimated real-time feed moisture content using the following formula: in, This is a real-time estimated value for the feed moisture content. This is the preprocessed spectral response matrix. For the calibration set reference concentration vector, This represents the matrix transpose operation; The calculated real-time feed moisture content estimate is compared with the preset moisture content mutation threshold range. This threshold range is determined based on the historical data statistics of the NMP distillation process's sensitivity to feed moisture, and is divided into normal fluctuation range, slight abnormality range and severe mutation range. If the real-time feed moisture content estimate falls into a severe abrupt change zone and the duration exceeds the set time window threshold, then an abnormal state of the feed components is determined to have occurred. Generate a preliminary flag bit for abrupt changes in feed moisture content to characterize the abnormal state of feed components. This flag bit includes an anomaly level code, a trigger timestamp, and a confidence score, which is used to drive the weight transfer of the process semantic graph. Through the above-mentioned infrared spectroscopy soft measurement inversion and threshold comparison logic, the abstract feed fluctuation characteristics are transformed into concrete operating condition status indicators, realizing millisecond-level accurate identification of sudden changes in feed moisture content, and providing reliable status input for dynamically adjusting the priority of model variables; For example, the sampling wavelength range of the infrared spectroscopy soft measurement model is set to 1300nm to 1600nm, and the characteristic absorption peak of water molecules at 1450nm is selected as the key variable. When the absorbance value at 1450nm in the feed flow fluctuation feature vector suddenly increases from 0.15 to 0.28, it is substituted into the partial least squares regression model to calculate the estimated real-time feed moisture content as 1.2wt%. The preset upper limit of the moisture content change threshold is 0.8wt%, and the duration threshold is 5 seconds. Since 1.2wt% exceeds 0.8wt% and the duration is 8 seconds, the logic judgment module generates a preliminary judgment flag for the feed moisture content change, marked as "Level_High", with a confidence score of 0.95. This flag immediately triggers the enhancement operation for the "feed flow" node weight in subsequent steps, significantly improving the model's response speed and control accuracy to moisture disturbance conditions. S2.3: Extract the temperature response hysteresis feature vector and reboiler steam pressure time series data from the denoised feature dataset, construct a trend slope analysis algorithm based on the heat transfer coefficient decay rate, perform coupling correlation calculation on the temperature response hysteresis feature vector and the reboiler steam pressure time series data, obtain the actual heat transfer efficiency decline rate index of the reboiler, classify the scaling degree level according to the cumulative change amplitude of the actual heat transfer efficiency decline rate index of the reboiler, and generate a reboiler scaling stage identification code that characterizes the aging state of the equipment; The temperature response hysteresis feature vector and reboiler steam pressure time series data are extracted from the denoised feature dataset and used as the input benchmark for heat transfer efficiency analysis. A time-window-based sliding calculation mechanism is constructed, and a historical data window of fixed length is set to capture the dynamic trend of the heat transfer process; The least squares method was used to linearly fit the steam pressure data and the tower bottom temperature response data within the window, and the instantaneous slope of the heat transfer temperature difference and heat load was calculated. The actual heat transfer coefficient attenuation rate of the reboiler is calculated using the following formula: in, The heat transfer coefficient decay rate, The current heat load, As the baseline heat load, For heat exchange area, The logarithmic mean temperature difference Runtime; Perform cumulative integration on the calculated attenuation rate sequence to generate a fouling accumulation index that characterizes the degree of equipment aging.

[0011] The scaling accumulation index is compared with the preset scaling threshold ranges for light, medium, and heavy scaling to make a logical judgment: When the index is in the first threshold range, it is marked as the initial scaling state; when it is in the second threshold range, it is marked as the intermediate scaling state; when it exceeds the third threshold, it is marked as the severe scaling state. Based on the judgment results, corresponding reboiler scaling stage identification codes are generated, including codes C1, C2 and C3. Through the trend slope analysis and threshold comparison processing based on the heat transfer coefficient decay rate, the multi-source sensor data is transformed into the reboiler scaling stage identification code that characterizes the aging state of the equipment, so as to realize the quantitative classification and real-time perception of the equipment health status. For example, the sliding window length is set to 30 minutes, and the sampling frequency is 1 second. Reboiler steam pressure data and column bottom temperature data are collected to calculate the current heat load. 500kW, base heat load 520kW, heat exchange area The area is 10 square meters, the logarithmic mean temperature difference ΔT is 20 K, and the operating time t is 3600 seconds. Substituting these values ​​into the formula, the heat transfer coefficient decay rate is calculated. -2.78×10 -4 kW / (m²·K·s). The cumulative integral of the degradation rate over the past 24 hours yields a scaling accumulation index of 0.5. Preset thresholds for mild scaling are 0.3, moderate scaling is 0.8, and severe scaling is 1.5. Since 0.3 < 0.5 < 0.8, it is determined to be a mid-stage scaling condition, generating the identifier code C2. This code is directly used to activate the dynamic weight layer in the subsequent process semantic graph, significantly improving the model's adaptability to conditions of decreased heat transfer efficiency. S2.4: Access the raw relative humidity data transmitted by the external meteorological station API, perform a first-order difference operation on the raw relative humidity data to obtain the environmental humidity change rate sequence, dynamically match the environmental humidity change rate sequence with the environmental humidity jump judgment threshold, and when the environmental humidity change rate sequence is detected to exceed the preset jump threshold, generate an environmental humidity jump event marker that represents the external environmental disturbance. Access the raw relative humidity data transmitted from the external weather station API, establish a lightweight data communication channel based on the HTTP / HTTPS protocol, and periodically obtain JSON format data packets containing timestamps and relative humidity values; The received raw environmental relative humidity data is validated, and abnormal records such as null values, non-numeric characters and those exceeding the physical limit of 0%-100% are removed to generate a cleaned environmental humidity time series. A sliding window mechanism is used to extract environmental humidity data from the current moment and the previous N sampling points, constructing a local time series segment to support dynamic rate of change calculation. First-order difference operations are performed on the environmental humidity time series, and the rate of change of environmental humidity is calculated using the following formula: in, The rate of change of ambient humidity at the current moment. The measured relative humidity at the current time t. This is the measured value of the ambient relative humidity at the previous sampling time t-1; The calculated sequence of environmental humidity change rate is dynamically matched and compared with the preset environmental humidity jump judgment threshold, which is set according to the sensitive range of environmental temperature and humidity based on the heat exchange efficiency of the distillation column condenser. When the absolute value of the rate of change of ambient humidity at M consecutive sampling points exceeds a preset jump threshold, an external environmental disturbance event is determined to have occurred. An ambient humidity jump event marker characterizing the external environmental disturbance is generated, which includes an event type code, a trigger time timestamp, and a confidence score. Through the above differential operation and threshold matching processing, the continuous environmental humidity analog quantity is transformed into discrete event state identifiers, realizing the rapid identification and quantification of the impact of external meteorological disturbances, and providing real-time trigger signals for the dynamic adjustment of the condenser node weights in the subsequent process semantic graph. For example, the sampling period is set to 10 seconds, the sliding window size N is 6, and the number of consecutive judgment points M is 3. The preset threshold for judging sudden changes in ambient humidity is 2% / min. At a certain moment, the ambient humidity sequence collected is [60.1%, 60.3%, 60.5%, 61.2%, 62.5%, 64.0%]. The change rates of the last three intervals are calculated to be 0.7%, 1.3%, and 1.5%, which are converted to minute change rates of 4.2% / min, 7.8% / min, and 9.0% / min, respectively. Since the three consecutive points exceed the threshold of 2% / min, the system immediately generates an ambient humidity jump event marker, with the marker type 'EXT_HUMIDITY_SURGE' and the confidence level set to 0.95. This marker is immediately input into step S2.5, triggering an increase in the weight of the condenser outlet temperature node in the semantic graph, effectively compensating for the fluctuation in condensation efficiency caused by the sudden increase in ambient humidity, and significantly improving the robustness of the reflux temperature control. S2.5: Gather the initial judgment flag of the feed moisture content change, the identification code of the reboiler scaling stage, and the environmental humidity jump event flag, and use the timestamp alignment mechanism to encapsulate the above heterogeneous state identifiers into a structured data frame with a unified format. Add the operating condition type enumeration value and confidence score field to generate a dynamic operating condition label stream that can be used to drive the execution of the multi-layer process semantic graph weight migration protocol.

[0012] like Figure 2 As shown, step S3 involves constructing a process semantic graph containing a static topology layer, a dynamic weighting layer, and a semantic constraint layer based on the boiling point-liquid phase composition-pressure relationship in the NMP physical property database. The static topology layer defines the inherent causal relationships between nodes, and the dynamic weighting layer binds activatable operating condition label types, generating a multi-layered process semantic graph structure. Specifically, this includes: S3.1: Based on the boiling point-liquid phase composition-pressure ternary relationship data in the NMP physical property database, physical entity mapping is performed on phase equilibrium sensitive nodes, energy transfer dominant nodes, disturbance response hysteresis nodes, equipment status characterization nodes, product quality related nodes, and operation constraint boundary nodes to generate a set of spectral nodes with clear thermodynamic meaning. Call the NMP property database interface to read the boiling point-liquid phase composition-pressure ternary phase equilibrium data table under standard atmospheric pressure and vacuum conditions, and analyze the azeotropic composition point and dew point curve slope characteristics of NMP and water under different pressures. Based on the principle of phase equilibrium thermodynamics, the tower top temperature sensor and the reflux liquid temperature transmitter are mapped as phase equilibrium sensitive nodes, and their physical properties are defined as the characteristic quantity of the mass transfer driving force at the gas-liquid interface, and associated with the dew point temperature calculation function. Extract the opening signal of the reboiler steam pressure regulating valve and the pulse count value of the feed mass flow meter, map them as energy transfer-dominant nodes, and establish the energy coupling relationship between the release of latent heat of steam and the absorption of sensible heat of material based on the enthalpy balance equation. Collect radar wave flight time data of the liquid level in the tower bottom and the return water temperature signal of the cooling water at the condenser outlet, construct a disturbance response hysteresis node, and quantify the delayed response characteristics of liquid level fluctuations to changes in the liquid holdup in the tower through a first-order inertial link model. By integrating the reboiler heat exchanger tube wall temperature difference thermocouple array data and the harmonic spectrum analysis results of the reflux pump motor current, a device status characterization node is generated. The heat transfer coefficient decay model is used to correlate the tube wall fouling thermal resistance with the decreasing trend of pump unit mechanical efficiency. By integrating the characteristic absorption peak intensity data of the distillate from the online UV-Vis spectrometer with the NMP purity estimate output by the soft measurement model, a product quality correlation node is established, and dual quality index constraints of colorimetric Lab* spatial coordinates and purity at the ppm level are set. Read the safety valve set pressure value, the ultimate pressure threshold of the tower top vacuum system and the critical temperature data of NMP thermal decomposition, delineate the operational constraint boundary type node, and define the legal value range of each variable in the spectrum and the alarm trigger boundary. Through the above multi-dimensional physical entity mapping process, discrete sensing signals are transformed into a set of graph nodes with clear thermodynamic meaning and control semantics, realizing a structured dimensionality upgrade from raw data to process mechanism knowledge, and providing standardized node input for the subsequent construction of static topology layer; For example, for an NMP recovery distillation column, vapor-liquid balance data at 10 kPa absolute pressure was read to determine the column top temperature node threshold as 85°C. The reboiler vapor pressure of 0.4 MPa was mapped to the energy node baseline value. When the harmonic distortion rate of the reflux pump current exceeded 5%, the equipment status node was marked as "slight wear". Finally, a set of 12 core nodes in 6 categories was generated. Node attributes included name, unit, physical meaning, and initial confidence level, ensuring that the basic data for the map construction had clear physical interpretability. S3.2: Utilize the causal association rules in historical high-quality operating data to perform logical connection processing on the inherent causal relationships between nodes in the graph node set, so as to construct a static topology layer structure that defines fixed transmission paths between nodes; Obtain the set of graph nodes with clear thermodynamic meaning generated by S3.1, including six types of core nodes such as phase equilibrium sensitive type and energy transfer dominant type, as the entity basis for constructing the static topology layer; We retrieved historical high-quality operating databases of distillation columns, screened multivariate time-series data segments under steady-state conditions, and used the Granger causality test algorithm to conduct bidirectional causality analysis on the time-series data between each node. The Granger causality F-statistic of variable X on variable Y is calculated using the following formula: in, For the sum of squared residuals of the constrained model, For the sum of squared residuals of the unconstrained model, The lag order is... The sample length is used to quantify the predictive contribution of the predecessor node to the successor node. The calculated F-statistic is compared with the preset significance level threshold. If the F-value exceeds the threshold and the P-value is less than 0.05, it is determined that there is a statistically significant one-way causal relationship between the two nodes. By combining the mechanism rules in the NMP physical property knowledge graph, physical consistency verification is performed on statistically significant causal edges to eliminate pseudo-causal connections that violate the second law of thermodynamics or the law of conservation of mass. Retain the causal edges that pass the verification, assign them directional attributes to define the signal transmission path, and form a directed acyclic graph structure from feed flow rate to column top temperature and from vapor pressure to reflux liquid temperature. The resulting directed graph structure is subjected to connectivity testing to ensure that all key control nodes are in the main connected components, thus eliminating control blind spots caused by isolated nodes. By using the above processing method, discrete process nodes are transformed into static topological layer structures with fixed transmission paths, realizing a structured representation of the inherent causal relationship of the distillation process and providing stable skeleton support for dynamic weight transfer. For example, 5000 sets of steady-state operating data with NMP recovery rates greater than 99.8% over the past three months were selected. The Granger causality test lag order k was set to 3, and the significance level α to 0.05. Calculations showed that the F-statistic for reboiler steam pressure to top temperature was 12.45, with a p-value of 0.0003, indicating a strong causal relationship; while the F-statistic for bottom liquid level to feed flow rate was 1.2, with a p-value of 0.31, indicating no direct causal relationship, and this edge was removed. Finally, a static topology layer containing 18 nodes and 42 directed edges was constructed, ensuring the physical interpretability and logical rigor of the model structure. S3.3: Based on the triggering characteristics of real-time operating conditions such as sudden changes in feed moisture content, reboiler scaling stage indicators, and sudden changes in ambient humidity, the connection edges in the static topology layer structure are assigned initial confidence weights and bound to activatable operating condition label types to generate a dynamic weight layer structure that supports dynamic recalibration. Obtain the node connection relationships defined in the static topology layer and the baseline values ​​of causal correlation strength in historical high-quality running data as the initialization input for the dynamic weight layer; For each directed edge in the static topology layer, the initial confidence weight is calculated based on the Pearson correlation coefficient between the source node and the target node in the historical data, and the basic coupling strength mapping between variables is established. in, Let be the initial confidence weights from node i to node j. and Let be the observation values ​​of node i and node j at the k-th sampling time, respectively, and n be the total number of historical samples. Let i be the arithmetic mean of all sample data from source node i. Let be the arithmetic mean of all sample data for target node j; The three types of operating condition labels—abrupt change in feed moisture content, reboiler scaling stage identifier, and sudden change in ambient humidity—are defined as trigger condition indexes for the dynamic weight layer, and a mapping table between the labels and specific edge sets is established. For the feed moisture content abrupt change label, bind two key paths 'feed flow rate - top temperature' and 'reflux liquid temperature - separation efficiency', and set the weight adjustment range in the high sensitivity response mode; For the reboiler scaling stage identification, the 'steam pressure-heat transfer coefficient' and 'tower liquid level-energy consumption' paths are linked, and a thermal resistance compensation weight coefficient that increases with the degree of scaling is set. For the ambient humidity jump label, bind the 'condenser outlet temperature - reflux liquid temperature' path and set the heat exchange efficiency correction weight factor under ambient temperature disturbance; Add an activation status flag and a default dormant weight value to each bound connection edge to ensure that the model maintains the stability of the static topology under non-triggered conditions; Through the above processing method, the static causal relationship is transformed into a dynamic weight structure with working condition perception capability, so as to achieve the expected technical effect of automatic recalibration of variable priority according to real-time working conditions. For example, during the steady-state operation of the NMP distillation column, the initial confidence weight of the edge from 'feed flow rate' to 'top temperature' is calculated to be 0.85. When the edge computing unit detects a 'sudden change in feed moisture content' label, the system retrieves the mapping table and locks this edge and its associated 'reflux temperature-separation efficiency' edge. According to preset rules, the weight of the 'feed flow rate-top temperature' edge is dynamically adjusted to 0.95 to enhance the contribution of feed fluctuations to the prediction of the top state; at the same time, the weight of the 'reflux temperature-separation efficiency' edge is reduced from 0.70 to 0.60 to suppress the prediction deviation of separation efficiency caused by component changes. After the adjustment, the standard deviation of the model's prediction residual for the top temperature decreases from 0.3℃ to 0.15℃, significantly improving the control accuracy under complex operating conditions. S3.4: Based on NMP thermal stability constraints, product quality constraints, and equipment safe operation constraints, the node value range and edge weight change interval in the dynamic weight layer structure are subjected to boundary constraint embedding processing to generate a semantic constraint layer structure containing hard operation restrictions. Obtain the initial confidence weights and associated edge definitions of each node in the dynamic weight layer, and use them as input objects for boundary constraint embedding processing; Read the critical temperature threshold data for thermosensitive degradation from the NMP property database and use it as an NMP thermal stability constraint to determine the absolute safe upper limit of the column top temperature and reflux liquid temperature in the phase equilibrium sensitive node. Based on the stability requirements of NMP product color in lithium battery recycling process, the allowable fluctuation range of online spectral characteristic values ​​of distillate color is extracted, used as product quality constraint, and mapped to the output constraint range of product quality correlation node; The pressure setting parameters of safety valves and the lower limit of vacuum at the top of the tower in the distributed control system are retrieved as pressure-temperature coupled safety boundaries for the operational constraint boundary nodes, which are then used as safety constraints for equipment operation. A boundary constraint operator based on a penalty function is constructed to transform the above hard constraints into a system of mathematical inequalities, which is used to limit the feasible region of edge weight changes in the dynamic weight layer. The boundary constraint function for the range of node values ​​is calculated using the following formula: in, Let be the constraint penalty value for node i. For real-time predicted values ​​of nodes, and These are the lower and upper limits determined by the node based on its physical properties and safety specifications. Conditions for crossing the boundary; Apply the boundary constraint function to each associated edge of the dynamic weight layer. When the weight migration causes the node prediction value to approach the boundary, automatically decay the weight change magnitude coefficient of the corresponding edge. Nonlinear truncation is applied to the range of edge weight changes to ensure that the redistribution of variable priorities will not exceed the physical limits set by the semantic constraint layer when the extreme working condition label is triggered. Generate a semantic constraint layer structure that includes hard constraints on node values ​​and soft constraints on edge weight changes, and complete the legality verification and encapsulation of the dynamic weight layer; By using boundary constraint embedding, the dynamic weight layer structure of the previous step is transformed into a semantic constraint layer structure with safety and robustness, thereby achieving the expected technical effect of ensuring process safety during the model adaptive adjustment process. For example, for the NMP distillation column top temperature node, the thermosensitive degradation threshold is set to 205℃, and the safety valve tripping pressure corresponds to a temperature of 210℃. When the reboiler scaling stage flag triggers weight migration, causing an increase in the weight of 'steam pressure' on 'top temperature', the semantic constraint layer monitors the predicted top temperature value in real time. If the predicted value reaches 202℃, the constraint operator is activated, linearly reducing the coefficient of the weight change of the associated edge from 1.0 to 0.2 to prevent the temperature from further rising into the degradation zone. Simultaneously, for the distillate color node, the b-value fluctuation range in the Lab color space is set to ±0.5. If the 'feed moisture content mutation' label causes the separation efficiency weight adjustment, potentially leading to color exceeding the standard, the constraint layer forcibly locks the relevant weights unchanged until the operating conditions stabilize. The final output semantic constraint layer ensures that, under all dynamic adjustment scenarios, key process parameters remain within a safe and compliant operating window. S3.5: Integrate the static topology layer structure, dynamic weight layer structure, and semantic constraint layer structure, and perform multi-layer data fusion and consistency verification processing to generate a multi-layer process semantic graph structure with complete process mechanism knowledge and adaptive potential.

[0013] like Figure 3 As shown, step S4 involves: inputting the dynamic operating condition label stream into the multi-layer process semantic structure, matching preset trigger conditions to locate associated edge nodes, recalibrating the correlation strength between variables in the dynamic weight layer according to preset migration rules, and generating a dynamic variable priority weight matrix. Specifically, this includes: S4.1: Obtain the dynamic operating condition tag stream output by the edge computing unit, and use the hash index algorithm to quickly retrieve the feed moisture content change tag, reboiler scaling stage identifier and ambient humidity jump tag in the dynamic operating condition tag stream to generate an active operating condition tag set with a unique identifier, so as to establish the current process state input benchmark that needs to be responded to. S4.2: Based on the set of activated operating condition labels, traverse the dynamic weight layer in the multi-layer process semantic graph, and use the semantic matching engine to compare each label in the set of activated operating condition labels with the activated operating condition label types bound in the dynamic weight layer. Generate a list of target edge nodes containing the correlation edges to be adjusted between phase balance sensitive nodes and energy transfer dominant nodes, so as to lock the specific variable coupling path affected by the current operating condition. Receive the set of activation condition tags generated by S4.1. This set contains key disturbance identifiers of the current distillation column operating status, such as feed moisture content change tags, reboiler scaling stage identifiers, or ambient humidity jump tags. The semantic matching engine is invoked to load the dynamic weight layer data structure in the multi-layer process semantic graph. This structure stores metadata about all node connection edges and their bound active working condition label types. Iterate through each tag element in the set of active working condition tags, extract its working condition type enumeration value and confidence score, and use them as the primary key index for semantic retrieval. In the dynamic weight layer, a breadth-first search is performed, and the extracted working condition type enumeration values ​​are matched with the trigger conditions bound to each connection edge for exact string matching and fuzzy similarity calculation. When it is detected that the binding label type of a connection edge is consistent with the currently active label, the connection edge is marked as an associated edge to be adjusted, and the physical attributes of its start node and end node are recorded. The focus is on selecting connection edges where the starting node is a phase equilibrium sensitive node and the ending node is an energy transfer dominant node. Such edges directly reflect the coupling strength between the temperature field and the energy field. The selected edges to be adjusted and their corresponding source and target node IDs are encapsulated into a target edge node list. Each element in the list contains the edge ID, initial weight value, and the name of the associated physical variable. By using semantic matching and topology filtering, the activation labels from the previous step are transformed into a list of target edge nodes containing specific coupling paths, enabling precise locking of key variable relationships affected by the current working conditions and providing clear operational objects for subsequent weight recalibration. S4.3: Based on the target edge node list, call the weight migration rule library pre-set in the semantic constraint layer, use the rule inference engine to read the preset migration rules that match the target edge node list, and generate the weight change direction instruction and weight change magnitude coefficient for each related edge to be adjusted, so as to clarify the specific quantitative strategy for the priority adjustment of each variable. Receive a target edge node list, which contains the identifiers of the associated edges to be adjusted between phase balance sensitive nodes and energy transfer dominant nodes affected by the current operating condition label; Call the weighted migration rule library pre-set in the semantic constraint layer, use the rule inference engine to traverse and search the target edge node list, and match the preset migration rule entries corresponding to the current active working condition label type; Read the matched preset migration rules, parse the weight change direction instructions defined therein, and determine whether to perform an enhancement operation or a weakening operation on the confidence weight of each associated edge to be adjusted, so as to reflect the dynamic change trend of the coupling strength between variables. Based on the severity level parameters of the working conditions bound in the rules, the coefficient of weight change amplitude is calculated. This coefficient represents the quantitative step size of variable priority adjustment under the current disturbance intensity, ensuring that the weight adjustment is both responsive and avoids oscillation. The weight change direction instruction is combined with the weight change magnitude coefficient to generate a specific quantization strategy data packet for each associated edge to be adjusted, which explicitly specifies the control parameters for the transition from the initial confidence weight to the new weight; Through rule-based reasoning and parameter quantization, abstract working condition labels are transformed into specific weight adjustment instructions and amplitude coefficients, enabling precise quantification of priority adjustment strategies for each variable and providing a direct basis for subsequent millisecond-level weight recalibration. For example, when the label "Sudden Change in Feed Moisture Content" is detected and the level is "High", the rule inference engine matches rule ID_R01. This rule specifies: strengthen the association weight of "feed flow rate" with "reflux temperature" by +1; weaken the association weight of "top temperature" with "separation efficiency" by -1. Based on the ratio of the moisture content deviation of 0.5wt% to the threshold of 0.2wt%, the amplitude coefficient is calculated to be 0.3. A strategy data package is generated: the weight increment of edge E_1 (feed flow rate → reflux temperature) is +0.3, and the weight decrement of edge E_2 (top temperature → separation efficiency) is -0.3. This process ensures that under high moisture content conditions, the model focuses more on the direct impact of feed fluctuations on temperature, significantly improving the predictive model's adaptability to component changes. S4.4: Based on the weight change direction instruction and weight change magnitude coefficient, perform linear interpolation recalibration on the initial confidence weights corresponding to the target edge node list in the dynamic weight layer to generate updated inter-variable correlation strength values, so as to complete the millisecond-level dynamic redistribution of variable priorities without reconstructing the model structure. Obtain the weight change direction instruction and weight change magnitude coefficient generated in step S4.3, and combine them with the current initial confidence weight value corresponding to the target edge node list in the dynamic weight layer to construct a linear interpolation recalibration calculation model; For each associated edge in the target edge node list, read the trigger status of its bound working condition label, determine whether the edge is in the active migration range, and if it is in the active state, extract the preset weight migration step size parameter. A piecewise linear interpolation algorithm is adopted to map the coefficient of weight change amplitude to a normalization adjustment factor in the range of 0 to 1. The starting point and the ending point of the interpolation calculation are determined according to the weight change direction instruction. The starting point is the initial confidence weight at the current time, and the ending point is the boundary weight extreme value defined by the semantic constraint layer. Perform a single-step weight update calculation using the following formula: in, This represents the numerical value of the correlation strength between variables after recalibration. The current initial confidence weight, This is the coefficient representing the magnitude of the weight change. This is the direction correction factor (1 for positive migration, -1 for negative migration). The boundary weight extreme values ​​set for the semantic constraint layer; Saturation limiting is applied to the calculated updated association strength values ​​to ensure that the values ​​strictly fall within the [0, 1] confidence interval defined by the semantic constraint layer, preventing model divergence caused by weight overflow due to drastic operating condition fluctuations. Traverse all associated edges in the target edge node list and perform the above linear interpolation recalibration operation in parallel to generate a temporary weight set containing the values ​​of the correlation strength between all updated variables. By using linear interpolation recalibration, the weight adjustment strategy of the previous step is transformed into specific numerical values ​​of the correlation strength between variables that conform to physical constraints. This enables millisecond-level dynamic redistribution of variable priorities without reconstructing the model topology, significantly improving the model's adaptive response speed to complex conditions such as sudden changes in feed composition and equipment aging. For example, when a mid-stage reboiler fouling label is detected, the current initial confidence weight for the "steam pressure-heat transfer efficiency" correlation edge is... The boundary weight extreme value is 0.65, set by the semantic constraint layer. The coefficient for the magnitude of weight change is 0.85. The direction correction factor is 0.2. The value is 1. Substituting into the formula, we get... = 0.65 + 0.2 * 1 * (0.85 - 0.65) = 0.69. This value does not exceed the [0, 1] interval and is directly used as the updated correlation strength. Meanwhile, for the "tower top temperature - separation efficiency" correlation edge, due to the decrease in sensitivity caused by scaling, the following settings are applied: β is 0.4, and β is -1. It is 0.1, currently. The value is 0.7, and the calculated value is... =0.7+0.1*(-1)*(0.4 - 0.7)=0.73. The system synchronously updates the weights of these two edges, so that the model automatically reduces its dependence on the top temperature of the tower and increases its attention to the steam pressure in subsequent calculations, effectively controlling the prediction bias; S4.5: Summarize all updated correlation strength values ​​between variables, and use a matrix encapsulation algorithm to map the updated correlation strength values ​​between variables into a two-dimensional array structure to generate a dynamic variable priority weight matrix, so as to provide weight input parameters with real-time operating condition adaptability for the subsequent solution of the temperature-efficiency-energy consumption correlation prediction model.

[0014] Step S5: Based on the dynamic variable priority weight matrix, the coupling relationship between phase equilibrium sensitive nodes and energy transfer dominant nodes is calculated to eliminate variable priority ambiguity in multivariate coupling modeling and generate the output value of the temperature-efficiency-energy consumption correlation prediction model. Specifically, this includes: S5.1: Obtain the priority weight matrix of dynamic variables and the definition of inherent causal relationships in the static topology layer. Use the graph traversal algorithm to extract the effective connection edges between phase balance sensitive nodes and energy transfer dominant nodes, and generate a coupling relationship subgraph structure containing the initial association strength. Obtain the priority weight matrix of dynamic variables and the definition of inherent causal relationships in the static topology layer as the initial input data for graph traversal; A depth-first search algorithm is used to perform node reachability analysis on the multi-layer process semantic graph. Starting from a phase balance sensitive node, path tracing is performed along the directed edges defined by the static topology layer. During the traversal, valid connection edges directly connected to energy transfer dominant nodes are identified and filtered out, while indirect redundant paths via disturbance response lag nodes or device status characterization nodes are eliminated. Extract the baseline association strength coefficient of each valid connection edge in the static topology layer, and construct the original coupling relationship set containing the source node identifier, the target node identifier and the baseline weight; Map the real-time weight values ​​in the dynamic variable priority weight matrix to the corresponding edge nodes in the original set of coupling relationships, replacing the original static benchmark weights; The mapped weight data is normalized to ensure that the sum of the weights of all outgoing edges from the same source node satisfies the probability distribution constraint. Generate a coupling subgraph structure containing the updated association strength, which retains only the direct causal links between phase balance sensitive nodes and energy transfer dominant nodes; By using graph traversal and weight mapping, the dynamic weight matrix generated in the previous step is transformed into a physically interpretable coupled relationship subgraph structure data, realizing the explicit and structured expression of variable priority in multivariate coupled modeling, and eliminating the technical defects of ambiguous variable contribution in traditional black box models. For example, in the case of a sudden change in feed moisture content in an NMP distillation column, the dynamic variable priority weight matrix shows that the weight of 'feed flow rate' to 'reflux temperature' increases from 0.3 to 0.65. The system performs a depth-first search starting with 'top temperature' and identifies two valid paths: 'top temperature → reflux temperature' and 'top temperature → reboiler vapor pressure'. The baseline association strengths of these two edges in the static topology layer are extracted to be 0.8 and 0.5, respectively. The corresponding real-time weights of 0.92 and 0.78 in the dynamic weight matrix are mapped to these two edges. All outgoing edge weights of the 'top temperature' node are normalized, resulting in a normalization factor of 1.7. The corrected association strength of 'top temperature → reflux temperature' is 0.54, and the association strength of 'top temperature → reboiler vapor pressure' is 0.46. The resulting coupling subgraph clearly demonstrates the dominant control effect of the column top temperature on the reflux liquid temperature under the current operating conditions, significantly improving the model's prediction accuracy for water content fluctuations and ensuring the accuracy of subsequent control command generation. S5.2: Based on the aforementioned coupling relationship subgraph structure, read the recalibrated correlation strength values ​​between variables in the dynamic weight layer, and use the weighted adjacency matrix construction method to map the discrete weights into a continuous coefficient matrix to generate a dynamic coupling coefficient matrix that characterizes the interaction strength of variables under the current working condition. Read the node set and dynamic variable priority weight matrix in the coupling relationship subgraph structure, and extract all effective connection edges between phase balance sensitive nodes and energy transfer dominant nodes and their corresponding recalibrated association strength values. An N-row, N-column zero matrix is ​​constructed as the initial carrier of the dynamic coupling coefficient matrix, where N is the total number of key process variables involved in the solution in the graph, and the matrix row and column indices correspond to the variable numbers in the standardized process parameter sequences such as column top temperature, reflux liquid temperature, and reboiler steam pressure. Traverse the list of target edge nodes, obtain the source node index i and target node index j for each connection edge, and recalibrate the association strength value after linear interpolation in the dynamic weight layer. Mapped to the i-th row and j-th column position of the dynamic coupling coefficient matrix; For variable pairs with bidirectional coupling, positive influence coefficients and negative feedback coefficients are filled in according to the causal direction definition in the semantic constraint layer to ensure matrix asymmetry and accurately reflect the thermodynamic hysteresis characteristics of the distillation process. The discrete weights are normalized to be continuous using the following formula to eliminate the interference of dimensional differences on the calculation of coupling strength: in, Let be the element in the i-th row and j-th column of the dynamic coupling coefficient matrix. The value represents the recalibrated association strength, and N represents the total number of active valid connection edges under the current operating condition. The value representing the association strength between source node i and the k-th target node; Sparsity optimization is performed on the generated dynamic coupling coefficient matrix, setting weak coupling coefficients below a preset confidence threshold to zero, preserving the core variable interaction paths that significantly affect the system state, and reducing the computational complexity of subsequent matrix operations; By constructing a weighted adjacency matrix, discrete variable priority weights are transformed into a continuous dynamic coupling coefficient matrix, thereby achieving a quantitative representation of the interaction strength of variables under the current working condition and eliminating priority ambiguity in multivariate coupling modeling. S5.3: Using the dynamic coupling coefficient matrix and combining the real-time observations in the standardized process parameter sequence, perform multivariate linear weighted fusion operation to quantify the contribution of each node to the system state and generate an intermediate-state coupling feature vector that reflects the real-time thermodynamic state. Read the real-time observations from the dynamic coupling coefficient matrix and the standardized process parameter sequence, and construct a multivariate linear weighted fusion operation model to quantify the node contribution. The real-time observations of the phase equilibrium sensitive node and the energy transfer dominant node are assembled into a state vector X, which includes key process parameters such as column top temperature, reflux liquid temperature, reboiler steam pressure and feed flow rate. Calling the dynamic coupling coefficient matrix The matrix elements Characterize the correlation strength between the j-th input variable and the i-th system state variable, ensuring that the matrix dimension matches the state vector; Perform matrix multiplication to calculate the intermediate state coupling eigenvector H, which is mathematically expressed as: in, To reflect the intermediate state coupling characteristic vector of real-time thermodynamic state, This is the dynamic coupling coefficient matrix. This is a standardized process parameter state vector; The physical meaning of the calculated intermediate state coupling feature vector H is analyzed, the thermodynamic contribution index corresponding to each component is extracted, and the combination of key variables that dominate the current separation process is identified. By performing multivariate linear weighted fusion operations, the dynamic coupling coefficient matrix generated in the previous step is transformed into an intermediate coupling feature vector that reflects the real-time thermodynamic state, thereby eliminating the ambiguity of variable priority in multivariate coupling modeling and achieving accurate quantification of the contribution of each node to the system state. S5.4: Based on the intermediate state coupling feature vector, call the boiling point-composition-pressure ternary relationship rule library embedded in the semantic constraint layer to perform nonlinear compensation correction on the prediction deviation of the phase equilibrium sensitive node, and generate the corrected coupling feature vector after physical property constraint calibration. Receive the intermediate coupling feature vector generated by S5.3, which contains preliminary predicted values ​​of column top temperature, reflux liquid temperature and vapor pressure calculated based on dynamic weights; The pre-defined NMP boiling point-composition-pressure ternary relationship rule library in the semantic constraint layer is analyzed to extract the theoretical gas-liquid equilibrium curve parameters and thermodynamic deviation correction coefficients under the current operating pressure. Using the pressure node values ​​in the intermediate-state coupling feature vector as index keys, the corresponding saturated vapor pressure data and dew point temperature benchmark values ​​are retrieved from the rule base. Calculate the residual between the preliminary predicted temperature and the theoretical dew point temperature to identify the phase equilibrium deviation caused by non-ideal solution behavior or local mass transfer resistance; A nonlinear compensation function is constructed, and a correction term from the Margules equation is introduced to quantify the effect of component activity coefficients on the boiling point shift. The corrected temperature deviation is calculated using the following formula: in, Here, k is the temperature compensation value, and k is the thermodynamic correction constant. For the product of pressure division, For total pressure, The activity coefficient is estimated in real time; The calculated temperature compensation value is superimposed on the temperature component of the intermediate state coupling eigenvector to eliminate the prediction blind zone of the linear model in the strongly nonlinear phase transition region. The temperature value after synchronous verification is checked to see if it exceeds the NMP thermal degradation threshold. If it does, it is truncated according to the boundary constraints of the semantic constraint layer. By integrating the temperature component after physical property calibration with the unaffected efficiency and energy consumption components, the corrected coupled feature vector is reconstructed. By using a nonlinear compensation process that embeds property rules, the linear approximation result of the previous step is transformed into a corrected coupled feature vector that conforms to the true laws of thermodynamics, thereby achieving a dual improvement in model prediction accuracy and physical interpretability. For example, when the intermediate-state coupling feature vector shows a tower top pressure of 12.5 kPa and a preliminary predicted temperature of 185.2 °C, the system retrieves the theoretical dew point of the NMP-water system at this pressure as 184.8 °C from the rule base. The detected feed moisture content label is 0.9 wt%, and the activity coefficient is calculated. It is 1.05. Substituting into the formula, we get... The value is 0.35℃. Adding 0.35℃ to 185.2℃ yields a corrected temperature of 185.55℃. This temperature is verified to be within the 200℃ degradation threshold. The output is a feature vector containing this corrected value, significantly eliminating the negative bias in temperature prediction under high moisture content conditions. S5.5: Based on the modified coupled feature vector, calculate the comprehensive correlation index between the tower top temperature, separation efficiency and reboiler steam consumption through the regression mapping function, eliminate the influence of variable priority ambiguity, and finally generate the output value of the temperature-efficiency-energy consumption correlation prediction model. Receive the corrected coupling feature vector after physical property constraint calibration, which contains the weighted state information of phase balance sensitive nodes and energy transfer dominant nodes under the current operating conditions; A nonlinear mapping function based on partial least squares regression is constructed, with the modified coupled feature vector as the input of independent variables and the column top temperature, separation efficiency and reboiler steam consumption as the target set of dependent variables. The kernel trick is used to perform high-dimensional mapping of the linear regression space, and the radial basis kernel function is introduced to handle the nonlinear coupling effect between variables, thereby enhancing the model's ability to fit the thermosensitive degradation characteristics of NMP. Define a regression loss function, which includes a prediction error term and a regularization penalty term. Iteratively solve the regression coefficient matrix using the gradient descent method to ensure the generalization stability of the model in the multidimensional feature space. in, This is the predicted output vector for the temperature-efficiency-energy consumption correlation. This is the corrected coupled feature vector. For kernel mapping functions, This is the regression coefficient matrix. This is residual noise; The predicted value of the tower top temperature is calculated in real time, and the physical rationality of the predicted value is verified based on the ternary relationship of boiling point-composition-pressure, and abnormal solutions that violate the laws of thermodynamics are eliminated. Simultaneously calculate the predicted value of separation efficiency, and combine it with the purity soft measurement data in the product quality-related nodes to evaluate the theoretical separation accuracy under the current reflux ratio; Estimate the reboiler steam consumption, integrate the weights of steam pressure and heat transfer coefficient based on the law of conservation of energy, and quantify the unit energy consumption index under the current operating conditions. The above three prediction indicators are encapsulated into structured data objects to form the output values ​​of the temperature-efficiency-energy consumption correlation prediction model; By using regression mapping and nonlinear compensation processing, the modified coupled feature vector from the previous step is transformed into multi-objective prediction data with physical interpretability, thereby achieving the expected technical effect of eliminating variable priority ambiguity and accurately predicting the system's energy efficiency status. For example, the corrected coupled feature vector is set to 6 dimensions, including normalized parameters such as column top temperature, reflux liquid temperature, and steam pressure. The radial basis function kernel width is selected as 0.5, and the regularization coefficient as 0.01. The current-time feature vector is input and, after kernel mapping, a high-dimensional feature space representation is obtained. Through 20 iterations of gradient descent, the regression coefficient matrix converges. The calculated predicted column top temperature is 85.3℃, the predicted separation efficiency is 99.92%, and the predicted reboiler steam consumption is 1.25 tons / hour. These output values ​​show minimal deviation from the measured data from the online analyzer, significantly improving the model's prediction accuracy under fluctuating feed moisture content conditions and providing a reliable basis for subsequent control command generation.

[0015] Step S6: Compare the output value of the temperature-efficiency-energy consumption correlation prediction model with the target threshold of the product quality correlation node, calculate the separation efficiency deviation and energy consumption redundancy, and generate a reflux temperature setpoint correction factor. Specifically, this includes: S6.1: Obtain the real-time separation efficiency prediction data and real-time energy consumption prediction data from the output values ​​of the temperature-efficiency-energy consumption correlation prediction model, and read the preset NMP purity target threshold and unit output steam consumption benchmark threshold in the product quality correlation node. Use the multivariate difference calculation algorithm to calculate the deviation between the real-time separation efficiency prediction data and the NMP purity target threshold, and at the same time calculate the redundancy between the real-time energy consumption prediction data and the unit output steam consumption benchmark threshold to generate a working condition deviation feature vector containing the separation efficiency deviation value and the energy consumption redundancy value. Receive the output value of the temperature-efficiency-energy consumption correlation prediction model from step S5, and extract the real-time separation efficiency prediction data and real-time energy consumption prediction data as the calculation benchmark. The system reads the preset NMP purity target threshold from the product quality-related nodes. This threshold is set to 99.95wt% based on the national standard GB / T 30684-2014 for lithium-ion grade NMP solvent and the customer's specific color requirements. It also reads the unit output steam consumption benchmark threshold, which is determined based on the average steam consumption under historical best operating conditions and is set to 1.2 tons of steam / ton of NMP. Using a multivariate difference calculation algorithm, the real-time separation efficiency prediction data is converted into an equivalent NMP purity estimate, and the algebraic difference between the estimate and the NMP purity target threshold is calculated to generate a separation efficiency deviation value. If the predicted purity is lower than the target threshold, the deviation value is negative, indicating insufficient separation accuracy; otherwise, it is positive, indicating excessive separation redundancy. Simultaneously, energy redundancy calculation is performed by comparing real-time energy consumption prediction data with the benchmark threshold for steam consumption per unit output, calculating the difference between the two, and generating an energy redundancy value. A positive value indicates that the current predicted energy consumption is higher than the benchmark, and there is room for energy saving; a negative value indicates that the energy consumption is better than the benchmark, but it is necessary to verify whether separation efficiency has been sacrificed. A feature vector of operating condition deviation is constructed, encapsulating the separation efficiency deviation value and energy consumption redundancy value into a two-dimensional vector structure. This vector is standardized to eliminate dimensional differences, ensuring that the subsequent fuzzy logic reasoning module can accurately identify whether the current operating condition tends to "maintain purity" or "reduce energy consumption." Through the above multivariate difference calculation and vector encapsulation processing, the model prediction results of the previous step are transformed into a condition deviation feature vector containing the separation efficiency deviation value and the energy consumption redundancy value, realizing the quantitative characterization of the current operating state of the distillation column relative to the ideal target state, and providing accurate input for subsequent dynamic weight allocation. For example, when the estimated NMP purity corresponding to the real-time separation efficiency prediction data is 99.92 wt%, and the target NMP purity threshold is 99.95 wt%, the separation efficiency deviation is calculated to be -0.03 wt%. Meanwhile, if the real-time energy consumption prediction data is 1.25 tons of steam / ton of NMP, and the benchmark threshold for steam consumption per unit output is 1.2 tons of steam / ton of NMP, then the energy consumption redundancy is 0.05 tons of steam / ton of NMP. The system encapsulates [-0.03, 0.05] into a working condition deviation feature vector. This vector indicates that the current working condition has a risk of not meeting purity standards and that energy consumption is too high. Subsequent control strategies will prioritize improving separation accuracy and appropriately relax energy-saving constraints, thereby significantly improving product qualification rate and gradually optimizing energy consumption indicators. S6.2: Based on the separation efficiency deviation value and energy consumption redundancy value in the operating condition deviation feature vector, the preset thermodynamic constraint rule library is called, and the separation accuracy priority coefficient and energy saving potential priority coefficient are constructed using the fuzzy logic reasoning mechanism. The separation efficiency deviation value is mapped to the separation accuracy priority coefficient, and the energy consumption redundancy value is mapped to the energy saving potential priority coefficient. A dynamic weight allocation parameter set is generated to characterize the control target trade-off relationship under the current operating condition. S6.3: The historical response trajectory data of the phase equilibrium sensitive node is weighted and corrected by using the dynamic weight allocation parameter set. Combined with the reboiler heat transfer lag time constant, the rolling time domain optimization algorithm based on model predictive control is executed to calculate the optimal reflux liquid temperature change rate under the premise of satisfying the NMP purity target threshold constraint, and generate the initial reflux temperature setpoint adjustment sequence with feedforward compensation characteristics. Obtain the dynamic weight allocation parameter set and the historical response trajectory data of the phase balance sensitive node, and map the separation accuracy priority coefficient and the energy saving potential priority coefficient to the weighting factor of each sampling point in the historical temperature sequence, respectively. A weighted moving average filter is applied to the historical reflux temperature time series data. The weight distribution within the filter window is adjusted according to the priority coefficient under the current operating conditions to suppress non-critical fluctuations and highlight the trend of controlled variables, thereby generating a smoothed baseline temperature trajectory. Read the reboiler heat transfer hysteresis time constant, construct a rolling optimization window that includes the control time domain length and control step size, and input the smoothed reference temperature trajectory as the initial guess solution into the model predictive controller. An objective function is established with the goal of minimizing separation efficiency deviation and energy consumption redundancy. The NMP purity target threshold is introduced as a hard constraint. The optimal reflux liquid temperature change rate sequence in the future control time domain is solved using a quadratic programming algorithm. in, To optimize the objective function value, For separation accuracy weighting coefficients, This is the energy consumption penalty weighting coefficient. To predict the separation efficiency at step size k, The target threshold for NMP purity. This represents the rate of change of reflux liquid temperature. Extract the temperature change rate corresponding to the first control step in the optimization solution, and combine it with the feedforward compensation to correct the disturbance caused by the sudden change in feed flow rate, thereby generating an initial reflux temperature setpoint adjustment sequence with anti-interference capability. By using the rolling time-domain optimization and feedforward compensation mechanism of model predictive control, the dynamic weight allocation parameters are transformed into specific temperature control command sequences, thereby achieving refined control that balances separation accuracy and energy-saving goals. For example, the control time domain length N is set to 10, the control step size is 1 minute, and the separation accuracy weighting coefficient is... The energy consumption penalty weighting coefficient is 0.8. The initial setpoint is 0.2. When a sudden change in feed moisture content is detected, resulting in a separation efficiency deviation of 0.5%, the system retrieves historical response trajectory data and applies a weighted moving average filter to generate a baseline temperature trajectory. Based on the reboiler heat transfer lag time constant of 15 minutes, a rolling optimization window is constructed. A quadratic programming problem is solved to obtain the optimal reflux liquid temperature change rate sequence for the next 10 minutes, with an initial change rate of +0.3℃ / min. Combined with a feedforward compensation of 0.1℃, an initial reflux temperature setpoint adjustment sequence is generated, with an initial adjustment of +0.4℃. This sequence effectively offsets feed disturbances, causing the predicted separation efficiency to converge to the target threshold range within three control cycles, while steam consumption remains below the baseline threshold, significantly improving the response speed and steady-state accuracy of the control system. S6.4: Perform safety boundary verification on the initial reflux temperature setpoint adjustment sequence, compare it with the temperature limit corresponding to the lower limit of the column top vacuum and the safety valve set pressure in the operation constraint boundary node. If the initial reflux temperature setpoint adjustment sequence exceeds the safety boundary, perform truncation and amplitude limiting operation to generate a feasible domain sequence of reflux temperature setpoints that conforms to the process safety specifications. S6.5: Based on the instantaneous difference between the feasible domain sequence of the reflux temperature setpoint and the current actual reflux liquid temperature, the proportional-integral-derivative (PID) control algorithm is used to calculate the final control increment. This control increment is encapsulated as a reflux temperature setpoint correction factor and output to the closed-loop controller to drive the coordinated action of the reboiler steam regulating valve opening and the reflux pump frequency, thus completing the closed-loop transformation from model prediction deviation to specific execution instructions.

[0016] Step S7: Using the reflux temperature setpoint correction factor and the dynamic variable priority weight matrix, the dynamic setpoint trajectory of the reflux temperature is calculated, and combined with the temperature closed-loop error correction, a drive command is generated to drive the closed-loop controller to execute the reflux temperature adjustment action. Specifically, this includes: S7.1: Obtain the reflux temperature setpoint correction factor and the priority weight matrix of the dynamic variables generated in the previous steps, and use the multi-input multi-output model predictive control algorithm to solve the coupling deviation between the phase balance sensitive node and the energy transfer dominant node, and generate the reflux temperature dynamic setpoint trajectory containing the feedforward compensation. The reflux temperature setpoint correction factor output from step S6 and the dynamic variable priority weight matrix generated from step S4 are used as the initial boundary conditions for multi-input multi-output model predictive control. Real-time state vectors of phase equilibrium sensitive nodes and energy transfer dominant nodes are extracted from the process semantic graph, and a multi-dimensional state space model including column top temperature, reflux liquid temperature, reboiler steam pressure and feed flow rate is constructed. Based on the dynamic variable priority weight matrix, the system matrix in the state space model is weighted and corrected to quantify the coupling strength and nonlinear disturbance terms among the variables under the current working condition. A rolling time-domain optimization algorithm is adopted, with the prediction time domain and control time domain set, and the optimal solution of the future control sequence is solved by minimizing the separation efficiency deviation and energy consumption redundancy as the objective function. The minimum value of the objective function J is calculated using the following formula: in, To predict the system output value at step i in the time domain, Set values ​​for the reference trajectory. To control the increment, To control the weighting coefficients, To predict the length of the time domain; A feedforward compensation mechanism is introduced to map the fluctuations in feed flow rate and the sudden changes in moisture content into measurable disturbance vectors, which are then superimposed into the model prediction equation to offset the hysteresis effect. The optimal return flow temperature adjustment sequence in the future control time domain is obtained by solving the problem. The first control variable is extracted as the dynamic setpoint at the current moment, forming a dynamic setpoint trajectory of return flow temperature including feedforward compensation. By using a multi-input multi-output model predictive control algorithm, the static correction factor is transformed into a dynamic temperature trajectory with operating condition adaptive capability, enabling the distillation process to respond quickly and track accurately to complex disturbances. For example, setting the prediction time domain 10, controlling the time domain The weighting factor is 3. The weighting factor for feed flow rate on reflux temperature in the dynamic variable priority weight matrix is ​​adjusted from 0.3 to 0.6 when a sudden change in feed moisture content is detected. The system reads the current top temperature of 120.5℃ and the reflux temperature of 85.2℃, and substitutes them into the corrected state-space model. The sum of squares of the deviations between the predicted output for the next 10 steps and the reference trajectory of 120.0℃ is calculated, while the rate of change of the control increment is constrained to not exceed 2℃ / min. After solving the quadratic programming problem, the first step control increment is obtained as +0.8℃, and the dynamic setting trajectory of reflux temperature is generated as [86.0, 86.5, 86.8...]℃. This trajectory effectively offsets the lag rise in top temperature caused by feed fluctuations, keeping the top temperature stable within the set value ±0.2℃ range, significantly improving the robustness and energy-saving effect of the control system. S7.2: Based on the dynamic setting trajectory of the reflux temperature and the real-time collected reflux liquid temperature feedback value, the temperature closed-loop error and its rate of change are calculated using an adaptive fuzzy PID control strategy, and the reboiler heat load demand increment signal and the reflux cooling capacity demand increment signal are generated. The system receives the dynamic reflux temperature setting trajectory generated by S7.1 and the real-time acquired reflux liquid temperature feedback value, and calculates the temperature closed-loop error at the current moment. A differential algorithm is used to perform differentiation on the temperature closed-loop error within a continuous sampling period to obtain the temperature error change rate, characterizing the system response trend. An adaptive fuzzy PID controller is constructed, using the temperature closed-loop error and its rate of change as dual input variables, mapped to a preset fuzzy domain. Based on the large hysteresis and nonlinear characteristics of the NMP distillation process, a fuzzy rule base is invoked to perform logical inference, dynamically adjusting the proportional, integral, and derivative gain coefficients. A defuzzification algorithm is used to convert the gain correction output from fuzzy inference into precise control parameters, and the proportional coefficient, integral time constant, and derivative time constant of the PID controller are updated in real time. Based on the updated control parameters, proportional-integral-derivative (PID) calculations are performed on the temperature closed-loop error to generate a basic heat load adjustment signal. A feedforward compensation mechanism is introduced, combined with the correction factor of the reflux temperature setpoint output by S6, to superimpose and correct the basic heat load adjustment signal, eliminating the steady-state error caused by model prediction bias. The corrected heat load signal is decomposed into the heat load demand increment on the reboiler steam side and the reflux cooling capacity demand increment on the condenser side through the energy balance equation; Through the above adaptive fuzzy PID control strategy and feedforward compensation processing, the temperature trajectory tracking error is transformed into a dynamic and adaptive incremental signal of reboiler heat load demand and incremental signal of reflux cooling demand, thereby achieving precise decoupled control of the distillation column heat input and significantly improving the response speed and anti-interference capability of the temperature control system. For example, the temperature sampling period is set to 1 second, and the fuzzy domain is divided into seven levels: {negative large, negative medium, negative small, zero, positive small, positive medium, positive large}. When the detected temperature closed-loop error is -0.5℃ and the error change rate is -0.1℃ / s, the fuzzy inference engine matching rule "if the error is negative small and the change rate is negative small, then the proportional gain increases medium and the integral time decreases small" is applied, resulting in an output proportional coefficient correction of +0.2 and an integral time correction of -5s. After defuzzification, the PID proportional coefficient is adjusted from 2.0 to 2.2, and the integral time is adjusted from 30s to 25s. Combined with a feedforward correction factor of 0.8kW, the calculated reboiler heat load demand increment is 1.5kW, and the reflux cooling demand increment is -0.3kW. This control action stabilizes the reflux liquid temperature within 120s to the set value ±0.1℃ range, reducing overshoot by 40% and shortening the settling time by 25% compared to traditional fixed-parameter PID control. S7.3: Perform a nonlinear mapping transformation of steam flow rate and pressure on the incremental signal of reboiler heat load demand, and combine it with the real-time monitoring value of reboiler steam pressure to correct the valve characteristic curve and generate the target opening command of reboiler steam regulating valve. Receive the reboiler heat load demand increment signal output from step S7.2 and use it as the reference input for target steam flow rate regulation; The system calls the saturated steam property lookup table pre-set in the edge computing unit, and uses a linear interpolation algorithm to invert the corresponding saturated steam density and specific enthalpy value based on the current real-time monitored reboiler steam pressure value. Based on the laws of mass conservation and energy conservation, a nonlinear mapping model from steam flow rate to valve opening is constructed to convert heat load demand into theoretical volumetric flow rate command; Considering the inherent nonlinearity of the flow characteristic curve and the dead zone effect of the control valve, a valve positioner feedback correction mechanism is introduced to read the current actual valve opening position signal; A piecewise linearization compensation algorithm is adopted to reversely correct the excessive gain in the small opening region and the saturation characteristics in the full opening region, and generate preliminary opening compensation coefficients. By combining the pipeline resistance coefficient and real-time data of the pressure difference before and after the valve, the Bernoulli equation of fluid mechanics is used to correct the deviation between the theoretical flow rate and the actual flow capacity. The flow demand value after differential pressure correction is mapped to the final valve stroke percentage command, and a slew rate limit is applied to prevent excessive valve action that could cause water hammer. Through the above-mentioned nonlinear mapping transformation of steam flow-pressure and valve characteristic curve correction, the incremental signal of heat load demand in the previous step is transformed into a target opening command of reboiler steam regulating valve with physical executable capability, thereby achieving high-precision heat energy supply control. For example, when the reboiler steam pressure monitoring value is 0.45 MPa, the saturated steam density is found to be 2.55 kg / m³. If the increase in heat load demand corresponds to an increase in steam mass flow rate of 50 kg / h, the theoretical increase in volumetric flow rate is 0.0196 m³ / h. For a DN50 equal percentage control valve, at the current opening of 30%, the flow coefficient Kv is 12.5. After differential pressure correction, it is calculated that the opening needs to be increased by 2.3%. The system applies a 0.5% / s action rate limit to this increment, generating a smooth target opening command and sending it to the actuator, effectively eliminating temperature oscillations caused by valve nonlinearity, and significantly improving the stability and accuracy of the reboiler's heating response. S7.4: Based on the incremental signal of the reflux cooling demand and the real-time monitoring value of the condenser outlet temperature, the required head and flow matching point are calculated using the pump set variable frequency speed regulation characteristic model, and a reflux pump frequency target adjustment command is generated. The system receives the incremental signal of reflux cooling demand generated by S7.2 and the real-time monitoring value of condenser outlet temperature, which are used as input variables for the pump set variable frequency speed regulation characteristic model. The system calls up the pre-set centrifugal pump hydraulic characteristic curve database, retrieves the corresponding head-flow (HQ) curve equation coefficients and efficiency-flow (η-Q) fitting parameters according to the current return pump model identifier, and constructs the basis of the pump set static hydraulic model. Based on the real-time monitoring value of the condenser outlet temperature, the fluid density and dynamic viscosity correction coefficients at the current temperature are obtained by querying the NMP property database. The Reynolds number is corrected for the HQ curve under standard operating conditions to generate a dynamic head characteristic curve that adapts to real-time properties. The incremental signal of reflux cooling demand is converted into a target volume flow rate setpoint. Combined with the distillation column top pressure data, the pipeline system resistance coefficient is calculated, and a system resistance characteristic equation including static head and friction loss is constructed. The intersection of the dynamic head characteristic curve and the system resistance characteristic equation is solved by Newton's iteration method. The coordinates of the operating point that meets the target flow requirement and is in the high-efficiency zone are determined, and the theoretical shaft power and the required net positive suction head value corresponding to the operating point are extracted. Based on the extracted theoretical shaft power, combined with the motor's rated power factor and the inverter's transmission efficiency, the target operating frequency is calculated using the following formula: Where f is the target operating frequency, As the reference power frequency, This is the current theoretical shaft power. This is the power compensation term corresponding to the increase in cooling demand. The shaft power is the reference operating condition. The calculated target operating frequency is subjected to slope limiting processing to prevent water hammer effect in the pipeline caused by frequency abrupt changes, and a smooth transition return pump frequency target adjustment command is generated. By solving the variable frequency speed regulation characteristic model of the pump group and smoothing the frequency, the cooling demand of the previous step is transformed into specific frequency commands of the actuator, so as to achieve precise matching and energy-saving optimization control of the return liquid flow rate. S7.5: Integrate the target opening command of the reboiler steam regulating valve and the target adjustment command of the reflux pump frequency, apply the actuator action rate limit and safety interlock logic verification, generate the optimized distillation column operation control command and send it to the underlying distributed control system to execute the reflux temperature adjustment action.

[0017] Step S8: Monitor the actual column top temperature and distillate color online spectral characteristic values ​​after the reflux temperature adjustment action. If a control response lag or a decrease in prediction accuracy is detected, trigger the weight transfer protocol to update the dynamic variable priority weight matrix, completing the online adaptive optimization closed loop of the model. Specifically, this includes: S8.1: Acquire the real-time sensor raw signal after the reflux temperature adjustment action is performed, and use the multi-channel synchronous acquisition module to perform high-frequency sampling on the actual column top temperature time series data and the online spectral characteristic value of distillate color to generate the actual operating status raw dataset containing timestamps; S8.2: Based on the original dataset of the actual operating status, the sliding window filtering algorithm is used to denoise the actual column top temperature time series data and extract the trend term. At the same time, the principal component analysis algorithm is used to reduce the dimension of the online spectral feature value of the distillate color and reconstruct it to generate a standardized feature vector of the actual operating status. S8.3: Compare the standardized actual operating state feature vector with the output value of the temperature-efficiency-energy consumption correlation prediction model point by point, calculate the residual sequence between the actual column top temperature and the predicted temperature, and the deviation of the actual distillate color online spectral feature value from the target threshold, and generate a set of control performance evaluation indicators. S8.4: Based on the control performance evaluation index set, apply dynamic threshold judgment logic to detect whether the convergence rate of the residual sequence is lower than the preset response speed lower limit or whether the deviation amplitude exceeds the allowable error range. If either condition is met, it is determined to be a control response lag or prediction accuracy decrease event, and a weight migration trigger signal is generated. S8.5: In response to the weight migration trigger signal, call the preset weight migration protocol to read the currently matched working condition label type, recalculate the association strength coefficient of the relevant edge nodes in the dynamic weight layer according to the migration rules, update the dynamic variable priority weight matrix, and complete the online adaptive optimization closed loop of the model.

[0018] The technical solution of the present invention has been described above with reference to the preferred embodiments shown in the accompanying drawings. However, it will be readily understood by those skilled in the art that the scope of protection of the present invention is obviously not limited to these specific embodiments. Without departing from the principles of the present invention, those skilled in the art can make equivalent changes or substitutions to the relevant technical features, and the technical solutions after these changes or substitutions will all fall within the scope of protection of the present invention.

[0019] The above description is merely a preferred embodiment of the present invention and is not intended to limit the invention. Various modifications and variations can be made to the present invention by those skilled in the art. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and rules of the present invention should be included within the scope of protection of the present invention.

Claims

1. An energy-saving control method for reflux heating in a distillation column, characterized in that, Includes the following steps: S1: Acquire multi-source heterogeneous sensor data during the operation of the distillation column, and preprocess the multi-source heterogeneous sensor data to generate a standardized process parameter sequence; S2: Based on the standardized process parameter sequence, extract the operating condition feature tags, identify the real-time operating condition status according to the operating condition feature tags, and generate a dynamic operating condition tag stream; S3: Based on the boiling point-liquid phase composition-pressure relationship in the NMP physical property database, construct a multi-layer process semantic map containing a static topology layer, a dynamic weighting layer, and a semantic constraint layer; S4: Input the dynamic working condition label stream into the multi-layer process semantic graph, match the preset trigger conditions to locate the associated edge nodes, recalibrate the correlation strength between variables in the dynamic weight layer according to the preset migration rules, and generate a dynamic variable priority weight matrix. S5: Based on the dynamic variable priority weight matrix, calculate the coupling relationship between key nodes to generate the output value of the temperature-efficiency-energy consumption correlation prediction model; S6: Compare the output value of the temperature-efficiency-energy consumption correlation prediction model with the preset NMP purity target threshold, calculate the separation efficiency deviation and energy consumption redundancy, and generate the reflux temperature setpoint correction factor. S7: Using the reflux temperature setpoint correction factor and the dynamic variable priority weight matrix, calculate the dynamic reflux temperature setpoint trajectory, and combine it with the temperature closed-loop error correction to generate a drive command to drive the closed-loop controller to perform the reflux temperature adjustment action.

2. The energy-saving control method for reflux heating in a distillation column according to claim 1, characterized in that, Following S7, the following also includes: S8: Monitor the actual top temperature of the tower after the reflux temperature adjustment action is performed. If a control response lag or a decrease in prediction accuracy is detected, trigger the weight migration protocol to update the dynamic variable priority weight matrix and complete the online adaptive optimization closed loop of the model.

3. The energy-saving control method for reflux heating in a distillation column according to claim 1, characterized in that, The multi-source heterogeneous sensor data includes column top temperature, reflux liquid temperature, reboiler vapor pressure, feed flow rate, column bottom liquid level, and condenser outlet temperature.

4. The energy-saving control method for reflux heating in a distillation column according to claim 1, characterized in that, The dynamic operating condition tag stream includes a feed moisture content change flag, a reboiler scaling stage identifier, and an ambient humidity jump event tag.

5. The energy-saving control method for reflux heating in a distillation column according to claim 1, characterized in that, S3 specifically includes: Based on the boiling point-composition-pressure ternary relationship data in the NMP physical property database, physical entity mapping is performed on phase equilibrium sensitive nodes, energy transfer dominant nodes, disturbance response hysteresis nodes, equipment status characterization nodes, product quality related nodes, and operation constraint boundary nodes to generate a set of graph nodes. By utilizing the causal association rules in the historical high-quality operation data of the distillation column, the inherent causal relationships between the nodes in the graph node set are logically connected to construct a static topology layer structure that defines fixed conduction paths between nodes; Based on the triggering characteristics of real-time operating conditions, initial confidence weights are assigned to the connecting edges in the static topology layer structure and active operating condition label types are bound to form a dynamic weight layer structure. Based on NMP thermal stability constraints, product quality constraints, and equipment safe operation constraints, the node value range and edge weight change interval in the dynamic weight layer structure are subjected to boundary constraint embedding processing to form a semantic constraint layer structure. By integrating the static topology layer structure, dynamic weight layer structure, and semantic constraint layer structure, multi-layer data fusion and consistency verification are performed to generate a multi-layer process semantic graph structure.

6. The energy-saving control method for reflux heating in a distillation column according to claim 1, characterized in that, S4 specifically includes: Obtain the dynamic operating condition tag stream generated in S2, quickly retrieve the real-time operating condition status in the dynamic operating condition tag stream, and generate an active operating condition tag set. Based on the set of activated operating condition labels, the dynamic weight layer in the multi-layer process semantic graph constructed by S3 is traversed, and each label in the set of activated operating condition labels is compared with the type of activated operating condition labels bound in the dynamic weight layer to generate a list of target edge nodes. Based on the target edge node list, the weight migration rule library in the semantic constraint layer of the multi-layer process semantic graph is called, the preset migration rule matching the target edge node list is read, and the weight change direction instruction and weight change magnitude coefficient for each associated edge to be adjusted are generated. Based on the weight change direction instruction and the weight change magnitude coefficient, perform linear interpolation recalibration operation on the initial confidence weights corresponding to the target edge node list in the dynamic weight layer to generate updated inter-variable correlation strength values. Summarize all updated correlation strength values ​​between variables, map these values ​​into a two-dimensional array structure, and generate a dynamic variable priority weight matrix.

7. The energy-saving control method for reflux heating in a distillation column according to claim 1, characterized in that, S5 specifically includes: Obtain the dynamic variable priority weight matrix generated by S4 and the inherent causal relationship definition in the static topology layer, extract the effective connection edges between phase balance sensitive nodes and energy transfer dominant nodes, and generate a coupling relationship subgraph structure; Based on the aforementioned coupling relationship subgraph structure, the recalibrated correlation strength values ​​between variables in the dynamic weight layer are read, and the discrete weights are mapped to a continuous coefficient matrix using the weighted adjacency matrix construction method to generate a dynamic coupling coefficient matrix. Using the dynamic coupling coefficient matrix and combining it with the real-time observations in the standardized process parameter sequence generated by S1, a linear weighted fusion operation is performed to generate an intermediate coupling feature vector. Based on the intermediate state coupling feature vector, the boiling point-composition-pressure ternary relationship rule library embedded in the semantic constraint layer is invoked to perform nonlinear compensation correction on the prediction deviation of the phase equilibrium sensitive node, and generate the corrected coupling feature vector. Based on the modified coupled feature vector, the comprehensive correlation index between the column top temperature, separation efficiency and reboiler steam consumption is calculated, and the output value of the temperature-efficiency-energy consumption correlation prediction model is generated.

8. The energy-saving control method for reflux heating in a distillation column according to claim 7, characterized in that, The method for generating the coupling relationship subgraph structure is as follows: The dynamic variable priority weight matrix and the inherent causal relationship definition in the static topology layer are obtained as the initial input data for graph traversal. Node reachability analysis is performed on the multi-layer process semantic graph. Starting from the phase balance sensitive node, path tracing is performed along the directed edges defined in the static topology layer. During the traversal, effective connection edges directly connected to energy transfer dominant nodes are identified and filtered out, while indirect redundant paths via disturbance response lag nodes or equipment state characterization nodes are eliminated. The baseline correlation strength coefficient of each effective connection edge in the static topology layer is extracted to construct an original coupling relationship set containing source node identifier, target node identifier, and baseline weight. The real-time weight values ​​in the dynamic variable priority weight matrix are mapped to the corresponding edge nodes in the original coupling relationship set, replacing the original static benchmark weights. The mapped weight data is normalized to generate a coupling subgraph structure containing the updated association strength.

Images