Rectification column flooding risk prediction and control decision method for light hydrocarbon separation device

Abstract

This application provides a method for predicting and controlling flooding risks in distillation columns of light hydrocarbon separation units. It collects real-time operating process parameters for each process node within the distillation column's operating cycle, generating a full-cycle process operation characteristic profile of the distillation column that characterizes the unit's overall operating features based on these parameters. Based on this profile, it establishes a thermodynamic coupling relationship between endogenous state parameters and external excitation / disturbance parameters to construct a deviation evolution characteristic body. The deviation evolution characteristic body undergoes digital characterization processing to obtain a risk characteristic evolution carrier, which in turn generates a risk situation evolution characteristic identifier representing the spatial distribution of risk conditions. Finally, it performs hierarchical clustering mapping processing on the risk situation evolution characteristic identifier to obtain a risk efficiency level label representing the current risk severity, and generates a hierarchical flooding risk prediction and control decision-making scheme for flooding failures in light hydrocarbon separation units.

Description

Technical Field

[0001] This application relates to the fields of petrochemical and energy management technology, and more specifically, to a method for predicting and controlling flooding risks in distillation columns of light hydrocarbon separation units. Background Technology

[0002] In petrochemical production processes, the light hydrocarbon separation unit in the refining plant plays a crucial role in separating natural gas condensate or light hydrocarbon mixtures into products such as ethane, propane, butane, and light oil. The core equipment of this unit includes distillation columns such as ethane stripper, propane stripper, and butane stripper. Due to dynamic fluctuations in feed composition, processing scale, and system pressure, the operating conditions of these distillation columns are extremely complex. During the operation of these distillation units, flooding is one of the most serious process failures. Once it occurs, it directly leads to loss of tray efficiency, deterioration of product quality, and even triggers safety accidents such as overpressure alarms and interlocking shutdowns.

[0003] In existing monitoring schemes for distillation columns in light hydrocarbon separation units, a single-parameter alarm strategy based on fixed thresholds is typically employed. This scheme first deploys pressure and temperature sensors at the top and bottom of the distillation column, respectively. Then, a monitoring system collects discrete process data in real time, such as pressure drop and top temperature. Finally, the collected real-time data is compared one by one with pre-set safe operating thresholds. When a parameter exceeds the fixed threshold, an early warning signal is issued to alert the operator for intervention.

[0004] However, this monitoring scheme based on a single parameter and a fixed threshold has significant technical drawbacks. Flooding in a distillation column is a complex evolutionary process of imbalance in the dynamics of the gas-liquid two-phase fluid within the column, involving deep coupling of multiple variables such as pressure, temperature, flow rate, and liquid level. Traditional schemes only focus on the real-time values ​​of discrete parameters, lacking the feature extraction of the inherent thermodynamic correlations between multi-dimensional operating parameters. This results in a weak ability to perceive the trend of the operating condition evolving from a steady state to a flooding critical state. When a single parameter reaches the alarm threshold, the distillation column is often already in a substantial flooding state. The lag in early warning makes it difficult for operational intervention to effectively suppress the failure trend, leading to increased risks of unplanned shutdowns and increased production capacity losses. This approach cannot meet the needs of light hydrocarbon separation units for advanced prediction and graded control of flooding risks. Summary of the Invention

[0005] This application provides a method for predicting and controlling the risk of flooding in a distillation column of a light hydrocarbon separation unit, so as to at least alleviate the above-mentioned technical problems.

[0006] A method for predicting and controlling flooding risks in distillation columns of light hydrocarbon separation units, comprising: Step 1: Collect real-time operating process parameters corresponding to each process node during the operating cycle of the distillation column by a sensor monitoring array deployed at each process node of the light hydrocarbon separation unit, so as to generate a full-cycle process operation characteristic profile of the distillation column that characterizes the full-dimensional operating characteristics of the unit based on the real-time operating process parameters. Step 2: Based on the full-cycle process operation characteristic conformation of the distillation column, establish the thermodynamic coupling relationship between endogenous state parameters and external excitation disturbance parameters to construct a deviation evolution characteristic body. The deviation evolution characteristic body is used to characterize the degree of drift of the current operating condition of the distillation column relative to the steady-state equilibrium point. Step 3: Perform digital characterization processing on the deviation evolution feature body to obtain the risk feature evolution carrier and generate a risk situation evolution feature identifier body that represents the spatial distribution of risk situation. Step 4: Perform hierarchical clustering mapping on the risk situation evolution feature identifier to obtain a risk energy efficiency level label that represents the current risk severity and generate a hierarchical distillation column flooding risk prediction and control decision scheme for flooding failure of light hydrocarbon separation unit.

[0007] Technical advantages of the technical solution provided in this application First, traditional solutions can only collect discrete process data from a few locations such as the top and bottom of the column, failing to achieve a global characterization of the operating conditions of the entire distillation column across all process nodes and the entire operating cycle, resulting in a one-sided extraction of operating condition features. This solution addresses this by deploying sensor monitoring arrays at each process node of the distillation column to collect real-time operating process parameters corresponding to each process node within the distillation column's operating cycle. This generates a full-cycle process operating characteristic profile of the distillation column that represents the full-dimensional operating characteristics of the unit. Compared to traditional methods that only monitor a few discrete parameters, this approach covers the operating status of all process nodes of the distillation column, achieving a systematic characterization of the unit's full-cycle and multi-dimensional operating characteristics. This provides a more comprehensive operating condition data foundation for flooding risk analysis and alleviates the problem of significant risk analysis bias caused by insufficient monitoring dimensions and one-sided operating condition characterization in traditional solutions.

[0008] Secondly, traditional solutions focus only on the real-time values ​​of discrete parameters, lacking the ability to explore the inherent thermodynamic coupling relationships between multi-dimensional operating parameters. This prevents the capture of the risk evolution process from the fundamental level of gas-liquid two-phase fluid dynamic imbalance corresponding to flooding, and results in a weak perception of the trend of the operating condition evolving from steady state to flooding critical state. This solution, based on the generated full-cycle process operation characteristic conformation of the distillation column, establishes the thermodynamic coupling relationship between endogenous state parameters and external excitation disturbance parameters, constructing a deviation evolution feature body that characterizes the degree of drift of the current operating condition of the distillation column relative to the steady-state equilibrium point. Compared to the traditional method of simply comparing single-parameter values ​​with fixed thresholds, this approach can characterize the drift process of the distillation column's operating condition relative to the steady-state equilibrium point from the perspective of multi-parameter thermodynamic coupling, rather than simply judging whether a parameter exceeds its limit. This improves the perception of the trend of the operating condition evolving towards the flooding critical state and alleviates the problem of early warning lag in traditional solutions.

[0009] Third, traditional solutions only output a single over-limit alarm signal, failing to systematically and quantify the overall situation of liquid flooding risk, thus hindering subsequent refined management. This solution performs digital characterization processing on the deviation evolution feature body to obtain a risk feature evolution carrier, and generates a risk situation evolution feature identifier that can characterize the spatial distribution of risk situation. Compared to the traditional processing method that can only output a single alarm signal, this solution can transform the evolution process of operating condition drift into a quantifiable risk situation representation, comprehensively presenting the spatial distribution and evolution of liquid flooding risk, and providing a quantitative basis for subsequent risk classification and control decisions.

[0010] Fourth, traditional solutions lack a tiered control strategy that matches risk levels. When a single parameter reaches the alarm threshold, the distillation column is often already in a substantial flooding state. Operational intervention is insufficient to effectively suppress the failure trend, leading to high risks of unplanned shutdowns, capacity losses, and safety accidents. This approach fails to meet the needs of light hydrocarbon separation units for proactive prediction and tiered control of flooding risks. This solution performs tiered clustering mapping on the risk situation evolution characteristic identifier to obtain a risk efficiency level label that characterizes the current risk severity. Based on this, a tiered flooding risk prediction and control decision scheme is generated for distillation column flooding failures. Compared to traditional methods lacking tiered control capabilities, this approach can match different risk severity levels to output corresponding tiered control decisions. It can take appropriate intervention measures at different stages of flooding risk evolution, mitigating the problems of delayed intervention and mismatched intervention measures in traditional solutions, which result in high risks of unplanned shutdowns, capacity losses, and safety accidents. This approach better meets the actual needs of light hydrocarbon separation unit distillation columns for proactive prediction and tiered control of flooding risks. Attached Figure Description

[0011] Figure 1This is a flowchart illustrating a flood risk prediction and control decision-making method for a distillation column in a light hydrocarbon separation unit, as described in an embodiment of this application.

[0012] Figure 2 This is a structural diagram of a distillation column flooding risk prediction and control decision-making device for a light hydrocarbon separation unit, according to an embodiment of this application.

[0013] Figure 3 This is a structural diagram of an electronic device according to an embodiment of this application. Detailed Implementation

[0014] like Figure 1 As shown in the figure, a method for predicting and controlling flooding risk in a distillation column of a light hydrocarbon separation unit is provided in this application, comprising: Step 1: Collect real-time operating process parameters corresponding to each process node during the operating cycle of the distillation column by a sensor monitoring array deployed at each process node of the light hydrocarbon separation unit, so as to generate a full-cycle process operation characteristic profile of the distillation column that characterizes the full-dimensional operating characteristics of the unit based on the real-time operating process parameters. Step 2: Based on the full-cycle process operation characteristic conformation of the distillation column, establish the thermodynamic coupling relationship between endogenous state parameters and external excitation disturbance parameters to construct a deviation evolution characteristic body. The deviation evolution characteristic body is used to characterize the degree of drift of the current operating condition of the distillation column relative to the steady-state equilibrium point. Step 3: Perform digital characterization processing on the deviation evolution feature body to obtain the risk feature evolution carrier and generate a risk situation evolution feature identifier body that represents the spatial distribution of risk situation. Step 4: Perform hierarchical clustering mapping on the risk situation evolution feature identifier to obtain a risk energy efficiency level label that represents the current risk severity and generate a hierarchical distillation column flooding risk prediction and control decision scheme for flooding failure of light hydrocarbon separation unit.

[0015] Optionally, the step of generating a full-cycle process operation characteristic profile of the distillation column based on the real-time operating process parameters, which characterizes the full-dimensional operating features of the device, includes: All real-time operating process parameters collected are subjected to time-series synchronization calibration to generate digital characterization data covering the full range of operating conditions of the distillation column; The digital characterization data is analyzed for variables to generate a full-cycle process operation characteristic profile of the distillation column that characterizes the full-dimensional operation features of the device.

[0016] Preferably, during the timing synchronization calibration of all collected real-time operating process parameters, the sensor monitoring array deployed at each process node of the distillation column in the light hydrocarbon separation unit first collects the real-time operating process parameters. Based on the actual connection relationship and mass and heat transfer mechanism of the distillation column in the light hydrocarbon separation process, the sensor monitoring array not only covers the top pressure monitoring node, bottom pressure monitoring node, top temperature monitoring node, and bottom temperature monitoring node of the distillation column itself, but also extends to the upstream and downstream process pipeline nodes closely related to the operation of the distillation column, such as the feed flow monitoring node on the feed pipeline, the reflux flow monitoring node on the reflux pipeline, the bottom liquid level monitoring node on the bottom liquid phase pipeline, and the temperature and pressure monitoring nodes of the cold box heat exchange unit. Sensors at each process node continuously sense and convert physical quantities within the distillation column's operating cycle according to their respective preset sampling frequencies (e.g., pressure sensors at 100 sampling times per second, temperature sensors at 50 sampling times per second, and flow sensors at 80 sampling times per second), thereby outputting a series of initial operating process parameter data streams carrying timestamps.

[0017] Preferably, after acquiring the initial operating process parameter data streams with timestamps output from each process node, a timing synchronization calibration process is performed on all the aforementioned initial operating process parameter data streams. Because sensor monitoring arrays corresponding to different physical quantities have objective differences in hardware response characteristics, signal transmission link delays, and analog-to-digital conversion mechanisms, the initial operating process parameter data streams are not naturally aligned on the time axis. Directly using unsynchronized data for subsequent analysis will introduce characterization bias. The timing synchronization calibration process first performs periodic global timing calibration on the internal clock sources of each sensor monitoring array using a preset high-precision time synchronization protocol (e.g., Network Time Protocol or a GPS-based timing mechanism) to eliminate accumulated errors caused by clock drift. Subsequently, for the initial operating process parameter data streams received within each sampling period, interpolation and decimation operations are performed according to a predefined resampling time grid (e.g., setting a uniform ten-millisecond time reference interval), remapping the originally unevenly distributed initial operating process parameter data streams onto a unified time series coordinate axis. After time-series synchronization calibration, the original isolated and asynchronous initial operating process parameter data streams of each process node are integrated into a full operating dataset that is strictly synchronized in the time dimension. This full operating dataset is defined as digital characterization data covering the full-domain operating condition characteristics of the distillation column.

[0018] Preferably, after generating digital characterization data covering the full-domain operating characteristics of the distillation column, variable analysis processing is immediately performed on this digital characterization data. Variable analysis processing is not simply listing the numerical values ​​contained in the digital characterization data, but rather, based on the thermodynamic working mechanism and fluid dynamics characteristics of the distillation column in the light hydrocarbon separation unit, it classifies the physical properties of the parameters within the digital characterization data and defines their process roles. Specifically, during variable analysis processing, the operating process parameters reflecting the thermodynamic equilibrium state and interphase mass transfer state within the distillation column are categorized into a set of endogenous state parameters. For example, the column top pressure parameter after time-synchronous calibration is categorized as the real-time characterization value of column top pressure in the endogenous state parameters; the column bottom pressure parameter after time-synchronous calibration is categorized as the real-time characterization value of column bottom pressure in the endogenous state parameters; the column top temperature parameter after time-synchronous calibration is categorized as the real-time characterization value of column top temperature in the endogenous state parameters; and the column bottom temperature parameter after time-synchronous calibration is categorized as the real-time characterization value of column bottom temperature in the endogenous state parameters.

[0019] Preferably, the variable analysis process also categorizes the operating process parameters from the external boundary conditions of the distillation column, which can directly disturb the dynamic state of the gas-liquid two-phase fluid within the column, into another set of external excitation disturbance parameters. For example, the feed flow rate parameter corresponding to the feed line after time-synchronous calibration is categorized as the feed flow rate disturbance characterization value in the external excitation disturbance parameters; the reflux flow rate parameter corresponding to the reflux line after time-synchronous calibration is categorized as the reflux flow rate disturbance characterization value in the external excitation disturbance parameters; and the bottom liquid level parameter after time-synchronous calibration is categorized as the bottom liquid level disturbance characterization value in the external excitation disturbance parameters. Through the above classification operation, the variable analysis process decomposes the digital characterization data into two interacting parameter sets in terms of physical attributes: endogenous state parameters and external excitation disturbance parameters, thus clarifying the internal determinants and external driving factors of the evolution of the distillation column's operating state.

[0020] Preferably, based on the endogenous state parameters and external excitation disturbance parameters obtained through variable analysis, a full-cycle process operation characteristic configuration of the distillation column, representing the full-dimensional operational characteristics of the device, is further generated. This full-cycle process operation characteristic configuration of the distillation column is not merely a simple concatenation of endogenous state parameters and external excitation disturbance parameters over time, but rather a dynamic evolution trajectory description in a multi-dimensional state space. In the generation process of this full-cycle process operation characteristic configuration of the distillation column, firstly, using the timestamp as the index axis, all endogenous state parameters at the same moment (including real-time values ​​of top pressure, bottom pressure, top temperature, and bottom temperature) are constructed into an endogenous state vector, and simultaneously, all external excitation disturbance parameters at the same moment (including values ​​of feed flow rate disturbance, reflux flow rate disturbance, and bottom liquid level disturbance) are constructed into an external excitation vector. Subsequently, the endogenous state vector and external excitation vector corresponding to each sampling moment are sequentially arranged in the time series to form a characteristic curve that continuously maps the entire process state changes of the distillation column during its operating cycle. This characteristic curve is the full-cycle process operation characteristic configuration of the distillation column, which completely preserves the dynamic information of the entire process from stable operation to potential instability in terms of data format.

[0021] Preferably, unlike traditional methods that only use finite discrete parameters such as top and bottom pressures, top and bottom temperatures for threshold comparison, the full-cycle process operation characteristic profile of the distillation column generated in this application achieves a systematic characterization of the operating status of all process nodes of the distillation column. The full-cycle process operation characteristic profile of the distillation column in this application can cover more dimensional process parameters, such as temperature gradient parameters related to cold box heat exchange efficiency and power parameters related to compressor unit load rate, which can all be included as extended dimensions of endogenous state parameters or external excitation disturbance parameters. By generating the full-cycle process operation characteristic profile of the distillation column, a high-fidelity data foundation is provided for subsequently establishing the thermodynamic coupling relationship between endogenous state parameters and external excitation disturbance parameters. This allows the subsequent construction of the deviation evolution characteristic body to no longer rely on isolated judgments of single-parameter thresholds, but rather on the collaborative evolution analysis of multi-dimensional state vectors, thus laying a data and characteristic foundation for capturing the subtle trend of the distillation column evolving from steady state to flooding critical state earlier.

[0022] Optionally, the step of performing variable analysis on the digital characterization data to generate a full-cycle process operation characteristic profile of the distillation column representing the full-dimensional operational characteristics of the device includes: The digital characterization data is analyzed to classify and obtain the endogenous state parameters characterizing the top pressure, bottom pressure, top temperature, and bottom temperature of the column, as well as the external excitation disturbance parameters characterizing the feed flow rate, reflux flow rate, and bottom liquid level. Based on the endogenous state parameters and the external excitation disturbance parameters, a full-cycle process operation characteristic profile of the distillation column is generated, representing the full-dimensional operation characteristics of the device.

[0023] Preferably, during the process of performing variable analysis on the digital characterization data covering the overall operating conditions of the distillation column to classify and obtain the endogenous state parameters and the external excitation disturbance parameters, based on the inherent mechanism of the distillation column in the gas-liquid two-phase fluid dynamic equilibrium and thermodynamic mass transfer process of the light hydrocarbon separation unit, attribute decomposition processing is performed on the multi-dimensional operating process parameters contained in the digital characterization data covering the overall operating conditions of the distillation column, which have undergone time-series synchronous calibration. This attribute decomposition processing first distinguishes two sets of parameters that play different causal roles in the flooding evolution process. Specifically, those operating process parameters that directly reflect the state of the gas-liquid two-phase interface, the pressure gradient distribution, the temperature gradient distribution, and the interphase mass transfer driving force determined by them inside the distillation column are extracted from the digital characterization data covering the overall operating conditions of the distillation column and classified into a parameter set representing the thermodynamic state response characteristics inside the distillation column. This parameter set is the endogenous state parameters. Simultaneously, operating process parameters that exert external influences on the internal equilibrium state of the distillation column, such as those originating from the external process boundary and characterizing feed disturbances, reflux disturbances, and reboiler holdup disturbances, are extracted from the digital characterization data covering the full-domain operating conditions of the distillation column. These parameters are then categorized into another set of parameters representing the external driving factors of the distillation column; this set of parameters constitutes the external excitation disturbance parameters. This classification operation clarifies the physical boundary between internal state variables and external control variables in the evolution of the distillation column's operating state, providing a clear distinction of the objects of action for subsequently establishing the thermodynamic coupling relationship between the two.

[0024] Preferably, based on classifying the digital characterization data covering the full-domain operating conditions of the distillation column into the endogenous state parameters and the external excitation disturbance parameters, the specific parameter items included in the endogenous state parameters are further defined and named. Since the pressure and temperature inside the distillation column are direct thermodynamic state parameters determining the gas-liquid phase equilibrium and fluid dynamic stability, the column top pressure parameter, after time-series synchronous calibration, is categorized and named as the real-time characterization value of the column top pressure in the endogenous state parameters. This real-time characterization value of the column top pressure is used to quantitatively describe the instantaneous operating pressure value of the column top gas phase space at each synchronous sampling moment; the column bottom pressure parameter, after time-series synchronous calibration, is categorized and named as the real-time characterization value of the column bottom pressure in the endogenous state parameters. This real-time characterization value of the column bottom pressure is used to quantitatively describe the instantaneous operating pressure value of the column top gas phase space at each synchronous sampling moment. The system quantifies the instantaneous combined static pressure and operating pressure of the liquid phase region at the bottom of the column. The time-synchronized top temperature parameters are categorized and named as real-time top temperature characterization values ​​within the endogenous state parameters. These real-time top temperature characterization values ​​are used to quantitatively describe the instantaneous temperature of the distillate vapor phase at each synchronous sampling moment. Similarly, the time-synchronized bottom temperature parameters are categorized and named as real-time bottom temperature characterization values ​​within the endogenous state parameters. These real-time bottom temperature characterization values ​​are used to quantitatively describe the instantaneous temperature of the liquid phase mixture at each synchronous sampling moment. These real-time top pressure characterization values, real-time bottom pressure characterization values, real-time top temperature characterization values, and real-time bottom temperature characterization values ​​together constitute the core dimensions of the endogenous state parameters. These parameters typically exhibit coordinated, trend-like drift characteristics before flooding occurs in the distillation column, rather than isolated abrupt changes.

[0025] Preferably, at the same time, the specific parameter items included in the external excitation disturbance parameters are defined and named. Since fluctuations in the feed flow rate directly change the gas-liquid phase load distribution in the column, adjustments to the reflux flow rate directly affect the heat load of the top condenser system and the liquid flow rate in the column, and rises and falls in the bottom liquid level affect the heating stability of the reboiler and the liquid holdup in the column, these are all direct external disturbance sources that induce the distillation column operating conditions to deviate from the steady-state equilibrium point. Accordingly, the feed flow rate parameters corresponding to the feed lines that have undergone time-synchronous calibration are categorized and named as feed flow rate disturbance characterization values ​​in the external excitation disturbance parameters. These feed flow rate disturbance characterization values ​​are used to quantitatively describe the instantaneous mass flow rate or volumetric flow rate of the light hydrocarbon mixture entering the distillation column at each synchronous sampling moment. The reflux flow rate parameters corresponding to the reflux lines that have undergone time-synchronous calibration are categorized and named as reflux flow rate disturbance characterization values ​​in the external excitation disturbance parameters. These reflux flow rate disturbance characterization values ​​are used to quantitatively describe the instantaneous mass flow rate or volumetric flow rate of the reflux liquid returning from the top condenser to the distillation column at each synchronous sampling moment. The bottom liquid level parameters corresponding to the bottom liquid phase lines that have undergone time-synchronous calibration are categorized and named as bottom liquid level disturbance characterization values ​​in the external excitation disturbance parameters. These bottom liquid level disturbance characterization values ​​are used to quantitatively describe the instantaneous liquid phase medium accumulation height in the bottom of the column at each synchronous sampling moment. The aforementioned feed flow rate disturbance values, reflux flow rate disturbance values, and bottom liquid level disturbance values ​​together constitute the core dimensions of the external excitation disturbance parameters. Changes in these parameters will act as input excitations on the distillation column system, driving the endogenous state parameters to undergo corresponding responsive evolution.

[0026] Preferably, after obtaining the endogenous state parameters consisting of the real-time values ​​of the top pressure, bottom pressure, top temperature, and bottom temperature, and the external excitation disturbance parameters consisting of the values ​​of the feed flow disturbance, reflux flow disturbance, and bottom level disturbance through the above variable analysis, a full-cycle process operation characteristic configuration of the distillation column, representing the full-dimensional operation characteristics of the device, is generated based on the endogenous state parameters and the external excitation disturbance parameters. This generation process does not simply list or concatenate the two sets of parameters in time sequence, but rather uses a unified timestamp as an index to construct two feature vectors describing the operating state of the distillation column at each synchronous sampling moment. Specifically, at each sampling time point that has undergone time-series synchronization calibration, the real-time characterization values ​​of the column top pressure, column bottom pressure, column top temperature, and column bottom temperature corresponding to the same moment are combined to construct a multi-dimensional array structure describing the internal thermodynamic state of the distillation column at that moment. This multi-dimensional array structure is defined as an endogenous state vector. Each dimension component of the endogenous state vector corresponds to the instantaneous state of a core thermodynamic variable inside the distillation column at that moment.

[0027] Preferably, at the same sampling time point for constructing the endogenous state vector, using the same unified timestamp as an index, the feed flow rate disturbance characterization value, the reflux flow rate disturbance characterization value, and the column bottom level disturbance characterization value corresponding to the same moment are combined to construct a multidimensional array structure describing the external boundary condition disturbance acting on the distillation column at that moment. This multidimensional array structure is defined as an external excitation vector. Each dimension component of the external excitation vector corresponds to the instantaneous state of an external driving variable acting on the distillation column at that moment. Thus, at each synchronous sampling moment within the distillation column's operating cycle, a pair of mutually paired feature vectors is generated, namely, an endogenous state vector and an external excitation vector. This pair of vectors jointly locates the operating state point of the distillation column at that moment in the state space. The former reflects the internal state response of the system, and the latter reflects the externally applied excitation conditions. The coupled evolution trajectory of the two in the time series contains the complete dynamic information of the distillation column's operating condition changes.

[0028] Preferably, the endogenous state vectors and external excitation vectors generated at all synchronous sampling moments within the distillation column's operating cycle are then sequentially arranged and coupled for storage according to their respective timestamps. Mathematically, this operation forms a dynamic trajectory curve that continuously extends in the time dimension and continuously maps the entire process state transition of the distillation column in a multi-dimensional state space. This dynamic trajectory curve is defined as a characteristic curve, which is the full-cycle process operation characteristic configuration of the distillation column. This full-cycle process operation characteristic configuration not only completely preserves the dynamic changes of all endogenous state parameters throughout the entire process from stable operation to potential instability at the data level, but also simultaneously records the complete change history of the external excitation disturbance parameters driving this state evolution. Unlike traditional methods that only focus on the instantaneous values ​​of discrete parameters such as top pressure, bottom pressure, top temperature, and bottom temperature, the full-cycle process operation characteristic configuration of the distillation column couples the internal state of the distillation column with external stimuli in the form of structured vectors in a temporal sequence. This provides a multi-dimensional, high-fidelity feature base that preserves dynamic evolution information for establishing the thermodynamic coupling relationship between the endogenous state parameters and the external excitation disturbance parameters.

[0029] Preferably, by generating the full-cycle process operation characteristic conformation of the distillation column, the subsequent construction of the deviation evolution characteristic body can be based on the coordinated evolution trend of the endogenous state vector and the external excitation vector in the multi-dimensional state space, rather than relying on isolated comparisons of single parameters such as the real-time characterization values ​​of the column top pressure and the real-time characterization values ​​of the column bottom pressure with their respective preset fixed safety thresholds. This analysis method based on the evolution trajectory of multi-dimensional state vectors can perceive the slight drift of the distillation column's operating conditions from the macroscopic perspective of the gas-liquid two-phase fluid dynamics equilibrium, rather than waiting for a certain monitoring parameter to significantly exceed the limit before issuing a warning. Therefore, the generation of the full-cycle process operation characteristic conformation of the distillation column provides structured data support for capturing the nascent risk of flooding in the early stages of distillation column flooding risk, i.e., when the operating conditions are still in the weak trend stage of evolving from steady state to flooding critical state, thus creating conditions for taking intervention measures earlier to suppress the further development of the failure trend.

[0030] Optionally, a thermodynamic coupling relationship between endogenous state parameters and external excitation disturbance parameters is established based on the full-cycle process operation characteristic conformation of the distillation column to construct a deviation evolution characteristic body, including: Based on the aforementioned thermodynamic coupling relationship, a parameter table for characterizing operating condition deviation is generated; Based on the operating condition deviation characterization parameter table, the endogenous state parameters are mapped to the multidimensional thermodynamic steady-state topological boundary space to obtain the operating condition drift trajectory that characterizes the current operating condition deviation from the steady-state equilibrium range, and the deviation evolution feature body is constructed in conjunction with the external excitation disturbance parameters.

[0031] Preferably, in establishing the thermodynamic coupling relationship between endogenous state parameters and external excitation disturbance parameters based on the full-cycle process operation characteristic conformation of the distillation column, the thermodynamic coupling relationship describing the response characteristics of the internal state of the distillation column to external excitation is first constructed based on the temporal correspondence between the endogenous state vector and the external excitation vector contained in the full-cycle process operation characteristic conformation of the distillation column. Since the distillation column is a typical multivariable nonlinear thermodynamic system in the light hydrocarbon separation process, the endogenous state parameters such as the real-time values ​​of the top pressure, bottom pressure, top temperature, and bottom temperature, and the external excitation disturbance parameters such as the values ​​of the feed flow rate disturbance, reflux flow rate disturbance, and bottom liquid level disturbance are not simply linearly superimposed, but follow the coupling constraints of the gas-liquid two-phase fluid dynamics equilibrium equation and the heat and mass transfer equation. Therefore, establishing the thermodynamic coupling relationship essentially involves systematically identifying and modeling the endogenous state vector and the external excitation vector at each moment in the full-cycle process operation characteristic configuration of the distillation column, in order to quantify the intensity and direction of the influence of external disturbances on the internal state drift. Unlike traditional approaches that treat each parameter as an independent variable for threshold comparison, this application, by establishing the thermodynamic coupling relationship, treats the distillation column as a complete kinetic system, thereby enabling the capture of early signs of operating condition deviations at the system level.

[0032] Preferably, based on the thermodynamic coupling relationship established by the full-cycle process operation characteristic conformation of the distillation column, a deviation parameter table for quantifying the degree of deviation of the current operating state of the distillation column is further generated. The generation process of the deviation parameter table is as follows: First, the endogenous state vector sequence corresponding to the steady-state operation period of the distillation column (e.g., a period when the feed composition is stable, the processing load is stable, and there is no external interference) is extracted from the full-cycle process operation characteristic conformation of the distillation column. The statistical expectation values ​​of each dimension component of the endogenous state vector (i.e., the real-time characterization values ​​of the top pressure, bottom pressure, top temperature, and bottom temperature) within the steady-state operation period are calculated, and a reference vector composed of these statistical expectation values ​​is defined as the steady-state equilibrium reference point. Subsequently, for each synchronous sampling moment in the full-cycle process operation characteristic conformation of the distillation column, the degree of deviation between the endogenous state vector at that moment and the steady-state equilibrium reference point is calculated. The deviation is calculated using a weighted Euclidean distance metric, where the weighting coefficients for each dimension are pre-configured based on the sensitivity of the distillation column to various endogenous state parameters during the evolution of the flooding critical state (for example, the weighting coefficient for the real-time characterization value of the bottom pressure is higher than that of the real-time characterization value of the top pressure, because the bottom pressure is more sensitive to changes in the liquid holdup within the column). Finally, the weighted Euclidean distance values ​​calculated at each sampling time are correlated and arranged with the external excitation vector at the same time, forming a two-dimensional mapping table indexed by the time series. This two-dimensional mapping table is the operating condition deviation characterization parameter table. Each row of the operating condition deviation characterization parameter table records a timestamp of a sampling time, the corresponding quantified value of the deviation of the endogenous state vector, and the instantaneous values ​​of each dimension of the external excitation vector, thus providing a structured quantitative basis for subsequent operating condition drift trajectory analysis.

[0033] Preferably, after generating the operating condition deviation characterization parameter table, the endogenous state parameters are mapped to a multidimensional thermodynamic steady-state topological boundary space based on the operating condition deviation characterization parameter table to obtain the operating condition drift trajectory characterizing the current operating condition deviation from the steady-state equilibrium range. The multidimensional thermodynamic steady-state topological boundary space is a pre-constructed geometric model used to define the safe operating range of the distillation column. The construction method of the multidimensional thermodynamic steady-state topological boundary space is as follows: based on the design operating elasticity data of the distillation column, historical safe operating data, and flooding critical state data obtained based on gas-liquid two-phase fluid dynamics simulation, a closed hypersurface enveloping all safe operating state points is fitted in a four-dimensional state space with the real-time characterization values ​​of the column top pressure, column bottom pressure, column top temperature, and column bottom temperature as coordinate axes. The internal region of this closed hypersurface represents the safe operating envelope region of the distillation column, and the closed hypersurface itself is the steady-state topological boundary. Simultaneously, in this four-dimensional state space, with the steady-state equilibrium reference point as the origin, deviation tolerance thresholds in different directions are defined. The endogenous state vector at each moment in the operating condition deviation characterization parameter table is mapped as a state point to this multi-dimensional thermodynamic steady-state topological boundary space. As time progresses, these mapped state points form a continuous trajectory in the four-dimensional state space, which is defined as the operating condition drift trajectory. The operating condition drift trajectory intuitively reflects the dynamic process of the internal thermodynamic state of the distillation column gradually moving away from the steady-state equilibrium reference point and approaching the steady-state topological boundary under external excitation disturbances.

[0034] Preferably, after acquiring the operating condition drift trajectory, the operating condition drift trajectory is further associated with the external excitation disturbance parameters to construct the deviation evolution feature body. Specifically, the operating condition drift trajectory itself only describes the migration path of the endogenous state parameters in the state space, but does not explicitly contain the external disturbance information driving this migration. Therefore, it is necessary to bind each state point on the operating condition drift trajectory with the external excitation vector recorded in the operating condition deviation characterization parameter table at the same time. After the binding operation, each state point on the operating condition drift trajectory carries not only its position coordinates in the four-dimensional state space (i.e., the value of the endogenous state vector), but also the corresponding feed flow disturbance characterization value, reflux flow disturbance characterization value, and bottom liquid level disturbance characterization value at that time. Arranging this sequence of trajectory points with added external excitation information in chronological order constitutes a composite feature description with five or more dimensions (four endogenous state dimensions plus three external excitation dimensions, for a total of seven dimensions). This composite feature description is the aforementioned deviation evolution feature. The deviation evolution feature completely records the entire process of the distillation column's internal state drifting from the steady-state equilibrium point along the path and at what rate towards the steady-state topological boundary in the state space under external excitation disturbance.

[0035] Preferably, the construction process of the deviation evolution feature body further includes a quantitative description of the relative positional relationship between the state points on the drift trajectory of the operating condition and the steady-state topological boundary. For each time node in the deviation evolution feature body, the shortest spatial distance from the state point corresponding to that time node to the steady-state topological boundary is calculated, and this shortest spatial distance is defined as the instantaneous boundary distance. Simultaneously, the instantaneous drift velocity vector of that state point is calculated. This instantaneous drift velocity vector is obtained by the spatial difference of the state points at adjacent sampling times, and its direction points towards the direction of state migration inside the distillation column. Its magnitude represents the speed of state drift. Incorporating the instantaneous boundary distance and the instantaneous drift velocity vector as additional dimensions into the deviation evolution feature body ensures that the deviation evolution feature body not only includes the original observation data of the state and disturbances but also integrates dynamic features describing the drift trend. Unlike traditional solutions that can only provide a binary judgment of whether or not the limit has been exceeded, the deviation evolution feature provides rich information about "how far the current operating condition has deviated", "how fast it is approaching the danger boundary", and "what kind of external disturbance has dominated this drift process".

[0036] Preferably, the construction of the deviation evolution feature directly supports the advanced prediction of subsequent flooding risks. Since the deviation evolution feature fully couples the evolution trajectory of the internal state of the distillation column with the temporal excitation of external disturbances in its data form, by analyzing the changing trend of the instantaneous boundary distance and the directional distribution of the instantaneous drift velocity vector in the deviation evolution feature, it is possible to identify whether the current operating condition drift is dominated by a continuous increase in feed flow, improper adjustment of reflux flow, or abnormal fluctuations in the bottom liquid level. For example, when it is detected that the feed flow disturbance characterization value in the deviation evolution feature remains at a high level, and the direction of the instantaneous drift velocity vector points to the sensitive region related to flooding in the steady-state topological boundary (such as the quadrant formed by the direction of increase in the real-time characterization value of the bottom pressure and the direction of decrease in the real-time characterization value of the top pressure), even if the individual operating process parameters have not yet reached the preset fixed alarm threshold, it can be inferred that the distillation column is being driven by excessive feed load and is evolving towards a critical flooding state. Therefore, the construction of the deviation evolution feature body provides a dynamic feature carrier that can reveal the physical essence of the flooded risk evolution, so as to perform digital characterization processing on the deviation evolution feature body to generate risk situation evolution feature identifier.

[0037] Optionally, a deviation evolution feature body is constructed, including: Spatiotemporal dimension projection processing is performed on the drift trajectory of the operating condition, and path integration is performed based on the projection result to calculate the multidimensional composite operating condition deviation degree that characterizes the deviation intensity of physical quantities. Based on the deviation of the multidimensional composite working condition and the external excitation disturbance parameters, a deviation evolution feature body is constructed.

[0038] Preferably, in the process of performing spatiotemporal dimension projection processing on the operating condition drift trajectory, the operating condition drift trajectory is first taken as the processing object. This operating condition drift trajectory is a continuous state migration path pre-mapped in the multidimensional thermodynamic steady-state topological boundary space by the endogenous state vector sequence, and each trajectory point corresponds to an endogenous state vector at a synchronous sampling time. The purpose of spatiotemporal dimension projection processing is to map the operating condition drift state originally implicit in the state space to an observation subspace that can intuitively reflect the correlation between the drift trend and flooding risk, thereby facilitating subsequent quantitative calculation of the deviation intensity. Specifically, spatiotemporal dimension projection processing includes two associated steps: spatial dimension projection and temporal dimension projection. In spatial dimension projection, each trajectory point in the four-dimensional state space where the operating condition drift trajectory is located (the coordinate axes are the real-time characterization values ​​of the tower top pressure, the tower bottom pressure, the tower top temperature, and the tower bottom temperature, respectively) is orthogonally projected onto a predefined two-dimensional risk observation plane. The two-dimensional risk observation plane consists of the direction of the rate of change of the real-time characterization value of the pressure at the bottom of the tower and the direction of the rate of change of the real-time characterization value of the temperature at the top of the tower. These two directions are highly correlated with the accumulation trend of liquid holdup and the change trend of gas phase load in gas-liquid two-phase fluid dynamics, respectively, and are key projection axes for the early manifestation of flooding characteristics. After spatial dimension projection, the drift trajectory of the operating condition is transformed into a projection trajectory curve located within the two-dimensional risk observation plane.

[0039] Preferably, after completing the spatial dimension projection to obtain the projection trajectory curve in the two-dimensional risk observation plane, a time dimension projection processing is further performed on the projection trajectory curve. For each projection point on the projection trajectory curve, the time dimension projection processing calculates the time difference vector between the projection point and its preceding adjacent projection points. This time difference vector reflects the displacement change of the projection point on the two-dimensional risk observation plane within a unit sampling time interval. Simultaneously, for each projection point, its position vector relative to the steady-state equilibrium reference point on the two-dimensional risk observation plane is calculated. The magnitude of this position vector characterizes the spatial distance of the current projection point from the steady-state projection position. The time difference vector and the position vector are combined to form a local projection dynamic descriptor corresponding to each projection point. Arranging the local projection dynamic descriptors corresponding to all projection points on the projection trajectory curve according to a time series yields a time series feature set describing the instantaneous direction and instantaneous amplitude evolution of the operating condition drift on the two-dimensional risk observation plane. This time series feature set is defined as a spatiotemporal projection feature sequence. The spatiotemporal projection feature sequence not only preserves the evolution information of the drift trajectory of the operating condition in the time dimension, but also compresses the redundant dimensions in the original state space through spatial projection, focusing attention on the dynamic features closely related to the direction of flooding risk sensitivity.

[0040] Preferably, after obtaining the spatiotemporal projection feature sequence, a path integral operation is performed based on the spatiotemporal projection feature sequence to calculate the multidimensional composite working condition deviation degree characterizing the deviation intensity of physical quantities. The path integral operation processes each local projection dynamic descriptor in the spatiotemporal projection feature sequence and its arc length parameter along the projection trajectory curve. This path integral operation uses a designed multidimensional path integral operator, which accumulates and sums the deviation effects along the direction of the projection trajectory curve. Specifically, for each integral infinitesimal arc segment (i.e., the line segment between two adjacent projection points) on the projection trajectory curve, the local deviation contribution value on the infinitesimal arc segment is first calculated. The calculation of the local deviation contribution value integrates three aspects of information: first, the magnitude of the position vector at the starting point of the infinitesimal arc segment; a larger magnitude indicates that the current operating condition deviates further from the steady state in the risk-sensitive direction; second, the magnitude of the time difference vector within the infinitesimal arc segment, i.e., the instantaneous drift rate; a larger rate indicates a more severe drift trend; and third, the cosine value of the angle between the time difference vector and the position vector within the infinitesimal arc segment. This cosine value reflects whether the drift direction points away from the steady-state topological boundary or towards a return to the steady-state reference point. When the drift direction points towards the steady-state topological boundary (i.e., the angle is small, and the cosine value is close to one), a higher deviation weight is assigned. The local deviation contribution value of the infinitesimal arc segment is obtained by performing a weighted product operation on the above three factors. Subsequently, the local deviation contribution values ​​of all infinitesimal arc segments are summed along the entire projection trajectory curve to obtain a cumulative deviation metric. Furthermore, to reflect the differences in deviation effects across different time segments, the path integral operator introduces a time decay factor. This means that the closer a micro-element is to the current time, the greater its contribution to the final deviation; conversely, the farther away a micro-element is from the current time, the smaller its contribution. This time decay factor is configured according to an exponential decay law. The comprehensive quantitative result obtained after the above weighted cumulative summation is the multidimensional composite operating condition deviation.

[0041] Preferably, the calculation process of the multidimensional composite operating condition deviation not only generates a scalar value reflecting the overall deviation intensity at the current moment, but also simultaneously generates a local deviation contribution distribution curve along the projected trajectory curve. This local deviation contribution distribution curve describes the distribution of the contribution intensity of the deviation effect over time throughout the entire historical process from the start of steady-state operation to the current moment. The multidimensional composite operating condition deviation differs from the simple judgment method in traditional schemes that only compares a single parameter such as the top pressure or bottom pressure with a fixed threshold. It is a comprehensive quantification of the synergistic drift effect of the distillation column over a continuous period of time across multiple risk-sensitive dimensions. For example, when the feed flow rate disturbance characterization value exhibits intermittent fluctuations rather than a sustained high level, the traditional single-parameter threshold method may not trigger an alarm. However, the multidimensional composite operating condition deviation obtained by the path integral calculation in this application will show a gradually increasing trend due to the cumulative contribution of the instantaneous drift rate across multiple time micro-elements, thus revealing that the system is experiencing overall operating condition deterioration before a single parameter has significantly exceeded its limit. Therefore, the deviation of the multidimensional composite operating conditions can more sensitively and comprehensively reflect the overall progress of the distillation column's evolution from steady state to flooding critical state.

[0042] Preferably, after calculating the multidimensional composite operating condition deviation, the deviation evolution feature body is constructed based on the multidimensional composite operating condition deviation and the external excitation disturbance parameters. The construction process of the deviation evolution feature body is not a simple numerical concatenation of the multidimensional composite operating condition deviation and the external excitation disturbance parameters, but rather the establishment of a composite feature description structure that uses time as an index and couples the comprehensive intensity of internal state drift with the real-time state of external disturbances. Specifically, for each synchronous sampling moment, the multidimensional composite operating condition deviation calculated at that moment is combined with the corresponding external excitation vector (including the feed flow disturbance characterization value, the reflux flow disturbance characterization value, and the tower bottom liquid level disturbance characterization value) to form a seven-dimensional feature row vector describing the system deviation state and external driving conditions at that moment (where four dimensions come from the endogenous state vector, one dimension comes from the multidimensional composite operating condition deviation, and three dimensions come from the external excitation vector). Arranging the seven-dimensional feature row vectors corresponding to all synchronous sampling moments in timestamp order constitutes a time series feature matrix, which is the deviation evolution feature body. Each row of the deviation evolution feature body corresponds to a sampling moment, and each column corresponds to the feature value of a certain dimension at that moment, thus completely recording the synchronous evolution history of the deviation intensity and external excitation of the distillation column during its operating cycle.

[0043] Preferably, the construction of the deviation evolution feature body further includes explicit labeling of the causal relationship between external disturbances and internal deviations. In the deviation evolution feature body, by calculating the time-delay correlation coefficient between the multidimensional composite operating condition deviation at each sampling time and the components of each dimension of the external excitation vector at the previous time, the time delay and degree of influence of different external disturbance factors on subsequent deviation intensity changes can be quantified. This set of time-delay correlation coefficients is stored as additional metadata associated with the time series feature matrix, enabling the deviation evolution feature body to possess the ability for dynamic causal inference. For example, when the time-delay correlation coefficients reflected in the deviation evolution feature body indicate a high correlation between the feed flow disturbance characterization value and the multidimensional composite operating condition deviation delayed by several sampling periods, it can be inferred that the currently observed operating condition deviation is mainly caused by feed fluctuations. Unlike traditional solutions where alarm signals and disturbance sources are isolated and it is difficult to trace the root cause, the deviation evolution feature constructed in this application tightly couples the drift trajectory, deviation intensity, and disturbance source together. This provides a dynamic feature carrier with rich information, complete structure, and clear physical meaning for subsequent digital characterization processing of the deviation evolution feature to generate a risk situation evolution feature identifier.

[0044] In summary, the deviation evolution feature body described in this application achieves a multi-scale, multi-dimensional characterization of the flooding risk evolution process in distillation columns. First, the spatiotemporal projection processing focuses the complex four-dimensional state drift trajectory onto a two-dimensional risk observation plane highly correlated with the flooding physical mechanism, filtering out noise interference from secondary dimensions and making the drift trend clearer and more discernible. Second, the path integral operation, through the cumulative weighted summation of drift rate, drift direction, and deviation amplitude, condenses the dynamic drift effect within a historical time window into a comprehensive multi-dimensional composite operating condition deviation scalar, overcoming the potential for misjudgment or omission due to single instantaneous value assessment. Finally, the multi-dimensional composite operating condition deviation is synchronously coupled with the external excitation disturbance parameters in time series, constructing a composite feature body that includes internal response and external drive, allowing the risk evolution process to be fully recorded and described within a causal framework. Therefore, the construction of the deviation evolution feature body provides a reliable and physically interpretable data foundation for the subsequent steps of nonlinear spatiotemporal feature extraction by the pre-trained liquid flooding risk prediction model, thereby generating the risk situation evolution feature identifier that can characterize the spatial distribution of risk situation.

[0045] Optionally, the risk situation evolution feature identifiers that characterize the spatial distribution of risk situations include: The risk feature evolution carrier is input into the pre-trained flooded risk prediction model to perform nonlinear spatiotemporal feature extraction, so as to generate a flooded risk evolution thermodynamic feature map that characterizes the spatial distribution of risk. Based on the aforementioned flooding risk evolution thermodynamic feature map, a risk situation evolution feature identifier is generated to characterize the spatial distribution of risk situation.

[0046] Preferably, in the process of inputting the risk feature evolution carrier into the pre-trained flooding risk prediction model to generate a flooding risk evolution thermodynamic feature map representing the spatial distribution of risk, the risk feature evolution carrier is the deviation evolution feature body constructed in the previous step. It exists in the form of the time series feature matrix, where each row of the matrix corresponds to a synchronous sampling time, and each column corresponds to a seven-dimensional feature row vector composed of the endogenous state vector, the multi-dimensional composite operating condition deviation, and the external excitation vector at that time. Here, the “” in the time series feature matrix means that the number of dimensions of the state space carried by the matrix is ​​greater than or equal to three dimensions. In the application scenario of flooding risk prediction of the distillation tower of the light hydrocarbon separation unit in this application, the matrix specifically contains seven dimensions of feature information, which far exceeds the representation capability of a two-dimensional plane, and is therefore defined as a feature structure. The flooding risk prediction model is composed of an input feature regularization layer, a multi-scale temporal convolutional coding layer, a graph neural spatial association extraction layer, a gated attention fusion layer, a thermodynamic situation analysis layer, and a risk situation identification generation layer connected sequentially. The model receives the time series feature matrix on the input side and generates the flooded risk evolution thermodynamic feature map and the risk situation evolution feature identifier on the output side in sequence. The layers work together in a tightly coupled manner to complete the complete mapping from the original composite features to a quantifiable description of the risk situation.

[0047] Preferably, the input feature normalization layer serves as the entry layer of the flooding risk prediction model. It performs dimension alignment, missing value imputation, and standardization transformation on the input time-series feature matrix to adapt to the numerical stability requirements of subsequent deep network structures. Upon receiving the time-series feature matrix, this layer first checks whether there are missing data points at individual sampling moments due to momentary communication interruptions in the sensor monitoring array. If missing data exists, it uses linear interpolation based on adjacent moments to complete the missing feature row vectors, thereby ensuring the continuity of the time-series feature matrix in the time dimension. Subsequently, the input feature normalization layer performs standardization transformation on each column of the time-series feature matrix (i.e., the feature sequence of each dimension). Specifically, for each dimension of the feature sequence, it calculates its mean and standard deviation over the entire historical operating period. Each feature value is subtracted from the mean of that dimension and then divided by the standard deviation of that dimension, resulting in a statistical distribution where the transformed feature sequences of each dimension have a mean of zero and a standard deviation of one. After the input feature normalization layer is processed, a normalized feature matrix with the same dimension as the input but with a normalized numerical distribution is obtained. This normalized feature matrix serves as the direct input to the next layer structure.

[0048] Preferably, the multi-scale temporal convolutional coding layer is immediately following the input feature normalization layer, and it is used to extract temporal evolution features containing dependencies at different time scales from the normalized feature matrix. This layer consists of multiple parallel temporal convolutional kernel channels, and the length of the convolutional kernel along the time axis (i.e., the receptive field size) of each channel is different. For example, three convolutional kernel channels with lengths of three sampling periods, seven sampling periods, and fifteen sampling periods are configured respectively. Each channel independently performs a one-dimensional causal convolution operation on each dimension of the feature sequence of the normalized feature matrix. During the convolution process, only the feature values ​​of the current time and historical time are used, without involving information of future time, so as to ensure causality in real-time monitoring scenarios. The convolutional kernel channels with shorter receptive fields focus on capturing the rapid response characteristics of the multidimensional composite operating condition deviation to the instantaneous changes in the external excitation disturbance parameters, such as the immediate impact of short-term pulses of the feed flow rate disturbance value on the operating condition deviation. The convolutional kernel channels with longer receptive fields focus on extracting the cumulative trend characteristics of the gradual drift of the internal state of the distillation column under continuous external disturbances (such as the slow rise of the bottom liquid level disturbance value). The convolutional feature maps output from each channel are stitched together along the feature dimension to form a temporal encoded feature tensor that simultaneously contains short-term dynamic details and long-term trend information.

[0049] Preferably, the graph neural spatial correlation extraction layer is located after the multi-scale temporal convolutional coding layer. It is used to explicitly model and propagate the spatial topological dependencies between the state variables of each process node in the temporal coding feature tensor. In the distillation column system of a light hydrocarbon separation unit, the real-time characterization values ​​of the column top pressure and bottom pressure are correlated by a pressure gradient through the gas phase flow within the column. The real-time characterization values ​​of the column top temperature and reflux flow disturbance are thermodynamically coupled through the thermal balance of the column top condenser. These spatial correlations are not fully connected but rather sparse graph structures determined by the process flow. Therefore, this layer pre-constructs a process topology adjacency graph describing the physical connections and thermodynamic influences between the monitoring nodes of the distillation column. In this process topology adjacency graph, graph nodes correspond to the monitoring variables represented by the real-time values ​​of the top pressure, bottom pressure, top temperature, bottom temperature, feed flow disturbance, reflux flow disturbance, and bottom liquid level disturbance, respectively. Graph edges represent direct thermodynamic or hydrodynamic influence paths between two variables. The graph neural spatial association extraction layer uses each time slice of the temporally encoded feature tensor as the initial feature of each graph node and performs multi-layer graph convolution message passing along the graph edges. In each layer of graph convolution, the feature vector of each graph node is updated by aggregating the feature vectors of its neighboring nodes through a learnable weight matrix, so that the updated feature vector of each graph node not only contains its own historical temporal information but also incorporates the state information of its physically associated neighboring nodes. After multi-layer graph convolution, the feature vectors of each graph node are endowed with spatial association semantics, forming a spatial association enhanced feature tensor.

[0050] Preferably, the gated attention fusion layer receives the temporal coding feature tensor output from the multi-scale temporal convolutional coding layer and the spatial correlation enhancement feature tensor output from the graph neural spatial correlation extraction layer. It is used to adaptively weight and fuse the features from the two sources through a gating mechanism and a self-attention mechanism, capturing key time segments and key variable dimensions in the fused feature sequence that play a crucial role in flooding risk prediction. This layer first calculates a gated fusion coefficient matrix, where each element is between zero and one, generated by the temporal coding feature tensor and the spatial correlation enhancement feature tensor through a fully connected network and a non-linear activation function. The gated fusion coefficient matrix determines the weight ratio of the temporal coding feature and the spatial correlation enhancement feature in the final fused feature at the same time and in the same feature dimension. Subsequently, the gated attention fusion layer performs a multi-head self-attention transformation on the weighted fused feature sequence. In the multi-head self-attention transformation process, the feature vector at each time step generates a query vector, a key vector, and a value vector through linear projection. The dot product of the query vector and the key vector yields the attention weight distribution, representing the strength of the association between that time step and other time steps in the sequence. This attention weight distribution is then applied to the weighted sum of the value vectors to obtain the attention-enhanced feature vector for that time step. Since fluid flooding is usually accompanied by a continuous co-evolution of state variables towards a dangerous direction over a certain period, the self-attention mechanism can automatically focus on time segments where the deviation from the trend is significantly enhanced, assigning them higher feature response amplitudes. After processing by the gated attention fusion layer, a fused enhanced feature sequence is output. This fused enhanced feature sequence has the same length as the input in the time dimension but has undergone information filtering and recalibration in the feature dimension.

[0051] Preferably, the thermal situation analysis layer takes the fused enhanced feature sequence as input and maps the abstract fused enhanced feature sequence onto a visualized thermal feature map that intuitively reflects the spatial distribution of flooded risk. This layer contains a feature dimensionality reduction projection module and a thermal distribution rendering module. The feature dimensionality reduction projection module uses a designed nonlinear dimensionality reduction transformation network to map the fused enhanced feature vector corresponding to each time step to a coordinate point on a two-dimensional plane, which is defined as the risk situation projection plane. Here, the "" in the fused enhanced feature vector also refers to a state space with a dimension number greater than or equal to three, and the mapping to a two-dimensional plane is a dimensionality reduction operation. The two-dimensional plane constitutes a "low-dimensional" projection space with two dimensions, thus technically defining the dimensionality comparison relationship before and after dimensionality reduction. The training objective of the nonlinear dimensionality reduction transformation network is to make time points with similar risk evolution patterns in the original feature space closer to each other on the risk situation projection plane, while time points with significantly different risk evolution patterns are further apart. Specifically, the fused enhanced feature vectors corresponding to the steady-state safe operation of the distillation column are concentrated in a small safe cluster center area on the risk situation projection plane. As the operating condition gradually drifts towards the flooding critical state, the corresponding fused enhanced feature vectors extend outward along a dominant direction on the risk situation projection plane, forming a migration trajectory from the safe area to the danger boundary. After receiving the two-dimensional projection coordinate sequence output by the feature dimensionality reduction projection module from the fused enhanced feature sequence, the thermal distribution rendering module constructs a two-dimensional grid (grid resolution, for example, 100x100) on the risk situation projection plane. For each grid cell, it calculates the local kernel density estimate of the projection points falling within the neighborhood of that grid cell and maps this local kernel density estimate to the corresponding color value in the thermochromatogram (e.g., low density corresponds to cool color, high density corresponds to warm color). Finally, each grid cell on the risk situation projection plane is assigned a thermal color value representing the degree of risk aggregation at that location. The thermal color values ​​of all grid cells together constitute a two-dimensional color image, which is the flooding risk evolution thermal feature map. In the thermodynamic characteristic diagram of flooding risk evolution, the warm-colored bright areas represent the high-risk spatial areas where the distillation column is frequently in or approaching a high-risk state during its operating history, while the color gradient direction from the safe area to the warm-colored area intuitively presents the main evolution path of the operating condition drift.

[0052] Preferably, after obtaining the flooding risk evolution thermal feature map, the risk situation identification generation layer further generates the risk situation evolution feature identifier body representing the spatial distribution state of the risk situation based on the flooding risk evolution thermal feature map. This layer first performs image segmentation processing on the flooding risk evolution thermal feature map, extracting connected regions in the thermal feature map whose color values ​​exceed a preset thermal threshold, forming at least one high-risk cluster patch. For each high-risk cluster patch, its geometric center coordinates, coverage area, maximum thermal intensity, and main color gradient direction are calculated. The geometric center coordinates reflect the concentrated location of the risk on the risk situation projection plane, the coverage area reflects the breadth of the risk's impact range, the maximum thermal intensity reflects the peak severity of the risk, and the main color gradient direction reflects the main direction of risk evolution. Subsequently, the risk situation identification generation layer combines the morphological parameters of the high-risk cluster patches extracted from the flooding risk evolution thermal feature map into a structured descriptor, and defines this structured descriptor as the risk situation evolution feature identifier body. Furthermore, the risk situation evolution feature identifier also includes timestamp range labels associated with each high-risk cluster patch. These timestamp range labels are directly mapped from the original sampling time corresponding to each projection point in the flooding risk evolution thermodynamic feature map, thereby establishing a correspondence between the spatial distribution of risk and its temporal evolution process. Unlike traditional solutions that only output a single alarm signal, the risk situation evolution feature identifier provides a multi-dimensional quantitative description of the specific clustering location, scale, evolution direction, and corresponding time interval of flooding risk in the state space. This provides accurate and resolvable feature input for subsequent hierarchical clustering mapping processing to obtain risk efficiency level labels and generate a hierarchical distillation column flooding risk prediction and control decision-making scheme.

[0053] Preferably, the pre-training process of the flooding risk prediction model is first based on a training sample set consisting of a large amount of historical operating data. The data in this training sample set comes from historical operating process parameters collected by the sensor monitoring array over several past operating cycles of the light hydrocarbon separation unit's distillation column, and have undergone time-series synchronization calibration and variable analysis processing according to the steps described above in this application. For each historical operating cycle, a risk efficiency level label corresponding to each sampling moment within that cycle is assigned as a monitoring signal, based on whether a flooding failure event actually occurred, whether there were any instances of manual intervention triggered by significant deviations in operating conditions, and in conjunction with the subsequent process analysis report. The labeling process specifically includes: for periods of stable operation without any signs of flooding, all sampling times are labeled as low-risk energy efficiency levels; for periods showing early signs such as a continuous rise in the real-time value of the bottom pressure and abnormal fluctuations in the real-time value of the top temperature, they are labeled as medium-risk energy efficiency levels; for periods where signs intensify and are approaching but not yet reaching the critical state of flooding, they are labeled as high-risk energy efficiency levels; and for periods where flooding actually occurs or an interlock shutdown has been triggered, and several preceding times, they are labeled as extremely high-risk energy efficiency levels. The historical operating cycles labeled in this way constitute supervised training samples. Simultaneously, to ensure the model's generalization ability under different operating conditions, the training sample set also covers diverse operating scenarios under different feed compositions (e.g., variations in the proportions of ethane, propane, and butane in the feed), different processing loads (e.g., within 60% to 120% of the design load), and different seasonal environmental temperature influences.

[0054] Preferably, during the training sample construction phase, a fixed-length continuous sampling sequence is extracted from the supervised training samples using a sliding time window as a training instance. The input structure of a single training instance is a slice of the time-series feature matrix, containing several consecutive (e.g., 128) synchronous sampling moments. Each moment corresponds to a seven-dimensional feature row vector composed of the endogenous state vector, the multi-dimensional composite operating condition deviation, and the external stimulus vector. The output supervision signal of the training instance is the risk efficiency level label marked at the end of the time window, and optionally, the risk efficiency level label sequence for each moment within the time window. To enhance the model's perception of temporal causal relationships during the fluidization evolution process, data augmentation strategies are also introduced when constructing training instances. For example, random amplitude scaling (scaling factor between 0.8 and 1.2) and translational perturbations on the time series of the original feed flow rate disturbance characterization values ​​and reflux flow rate disturbance characterization values ​​are applied to simulate feed fluctuations and reflux adjustment scenarios of different amplitudes and timings. Gaussian white noise with a signal-to-noise ratio conforming to the statistical characteristics of industrial field measurement noise is superimposed on the time series of the real-time bottom pressure characterization values ​​and the real-time top temperature characterization values ​​to improve the model's robustness to measurement noise. After data augmentation, the total number of training samples is expanded, and the sample diversity more closely reflects the uncertainties of a real industrial environment.

[0055] Preferably, the pre-training of the flooded risk prediction model is performed in an end-to-end manner, that is, the input part of the training instance (slices of the time series feature matrix) is directly fed into the entry layer of the model, and each layer of the model performs forward propagation calculations in sequence. Finally, the prediction result of the risk energy efficiency level label is output by the risk situation label generation layer. In each forward propagation process, the input feature normalization layer performs missing value imputation and normalization transformation on the input slice; the multi-scale temporal convolutional coding layer extracts multi-scale temporal evolution features from the normalized feature matrix; the graph neural spatial association extraction layer enhances the spatial association of the temporal coding feature tensor based on the process topology adjacency graph; the gated attention fusion layer adaptively fuses the temporal coding features and spatial association enhancement features and captures key time segments; the thermal situation analysis layer maps the fused enhanced feature sequence to the risk situation projection plane and generates the flooded risk evolution thermal feature map; and the risk situation label generation layer generates a risk situation evolution feature label based on the flooded risk evolution thermal feature map. The prediction results output by the risk situation identification generation layer include the category probability distribution of the risk energy efficiency level label at the end of the input time window. This category probability distribution is a four-dimensional vector, with four components corresponding to the predicted probabilities of low-risk, medium-risk, high-risk, and extremely high-risk energy efficiency levels, respectively, and the sum of each component is one.

[0056] Preferably, to drive the parameters of the flooded risk prediction model to update in the direction of minimizing prediction error during pre-training, a loss function matching the model output structure needs to be defined. Since the risk efficiency level label is a multi-class discrete label, this application uses a multi-class cross-entropy loss function as the main optimization objective during the pre-training stage. For each training instance, the true risk efficiency level label of that instance is converted into a one-hot encoded vector, and the cross-entropy loss value between this one-hot encoded vector and the class probability distribution output by the risk situation identification generation layer is calculated. Furthermore, considering that the evolution of flooded risk is a continuous process, and that adjacent risk levels have an order relationship (low risk < medium risk < high risk < extremely high risk), the loss function also includes an order regression constraint term. This order regression constraint term ensures that for training instances with higher true labels, the model outputs a correspondingly higher cumulative probability (e.g., the sum of the probability of predicting a high-risk efficiency level and the probability of predicting an extremely high-risk efficiency level), thereby avoiding prediction biases that violate the order of levels, such as outputting a low risk level probability for obvious flooded precursor states. The total loss function is formed by weighted summation of the cross-entropy loss value and the sequential regression constraint term. The cross-entropy loss term has a higher weight than the sequential regression constraint term in the total loss function, ensuring that the primary goal of classification accuracy dominates during training.

[0057] Preferably, after calculating the total loss function value in each forward propagation, the gradient of the total loss function value with respect to each learnable parameter in the flooded risk prediction model is backpropagated layer by layer using the backpropagation algorithm. Each learnable parameter includes: the kernel weight parameters of each temporal convolutional kernel channel in the multi-scale temporal convolutional coding layer; the learnable weight matrix of each layer of graph convolution in the graph neural spatial association extraction layer; the fully connected network parameters that generate the gated fusion coefficient matrix in the gated attention fusion layer and the linear projection matrix parameters in each head self-attention transformation; the nonlinear dimensionality reduction transformation network parameters of the feature dimensionality reduction projection module in the thermal situation analysis layer; and the logical parameters used for image segmentation and morphological parameter extraction in the risk situation identification generation layer. With minimizing the total loss function value as the optimization objective, the designed adaptive moment estimation optimization algorithm updates the above learnable parameters. In each round of iterative training, the optimization algorithm adaptively adjusts the learning rate of each parameter based on the first-order and second-order moment estimates of the historical gradients, thereby finding the optimal parameter in the parameter space at a relatively fast convergence speed. To stabilize the training process and prevent overfitting, weight decay regularization is applied during parameter updates, whereby all learnable parameters are multiplied by a decay factor slightly less than one with each update. Furthermore, a random deactivation mechanism is introduced after the multi-scale temporal convolutional coding layer and the gated attention fusion layer. During each forward propagation, a subset of neurons' outputs are randomly masked with a preset probability (e.g., 0.1). These masked neurons do not participate in subsequent calculations in that iteration, thus forcing the model to learn more robust feature representations.

[0058] Preferably, the pre-training of the liquid flooding risk prediction model employs a mini-batch stochastic gradient descent approach for iterative training. Within each training cycle, all training instances constructed from the training sample set are randomly shuffled and divided into several mini-batches, each containing a fixed number (e.g., thirty-six) of training instances. For each mini-batch, the aforementioned forward propagation, loss calculation, backpropagation, and parameter update steps are executed sequentially. After completing a training cycle, the generalization performance of the model under the current parameter state is evaluated using a validation sample set independent of the training sample set. The validation sample set is constructed in the same way as the training sample set, but it originates from historical running cycles different from the training sample set to ensure the independence of the validation. During validation, the model's classification accuracy on the validation sample set (i.e., the proportion of predicted risk efficiency level labels that match the actual labeled labels) and the category-weighted macro-average score are calculated. Simultaneously, the trend of the flooding risk evolution probability score index output by the model on the validation sample set as the operating conditions change is recorded and compared with the risk evolution trend marked by process experts based on the actual operation records during the validation period, in order to evaluate the consistency and rationality of the model's characterization of the risk situation. When the classification accuracy on the validation sample set no longer shows a significant improvement within several consecutive training cycles (e.g., five), and the total loss function value has converged to a small fluctuation range, it can be determined that the pre-training process has reached the convergence state, and the iteration is terminated. At this time, the parameters of each layer of the obtained flooding risk prediction model are fixed and saved, forming the pre-trained flooding risk prediction model. This model can then be directly used to perform nonlinear spatiotemporal feature extraction on the time series feature matrix collected online in real time from the distillation column of the light hydrocarbon separation unit, in order to generate the flooding risk evolution thermodynamic feature map and the risk situation evolution feature identifier.

[0059] Optionally, based on the flooding risk evolution thermodynamic feature map, a risk situation evolution feature identifier representing the spatial distribution of risk situation is generated, including: Based on the aforementioned flooding risk evolution thermodynamic feature map, the collaborative evolution law of the endogenous state parameters under the influence of the external excitation disturbance parameters is deduced; By performing probability density function mapping on the aforementioned co-evolutionary laws, a risk situation evolution feature identifier representing the spatial distribution of risk situations is generated.

[0060] Preferably, in the process of generating the risk situation evolution feature identifier based on the flooding risk evolution thermodynamic feature map, the flooding risk evolution thermodynamic feature map is first used as the direct analysis object. This flooding risk evolution thermodynamic feature map is a two-dimensional color image output by the aforementioned thermodynamic situation analysis layer. It visually presents the trajectory of the distillation column's operating condition drift during its operational history and its current position on the risk situation projection plane through warm-colored bright areas and color gradient directions. Based on this, the collaborative evolution law of the endogenous state parameters under the influence of external excitation disturbance parameters is deduced. The core processing action is to reconstruct the dynamic process leading to the formation of the thermal distribution from the static thermal distribution of the flooding risk evolution thermodynamic feature map. Specifically, based on the original timestamp label attached to each grid cell in the flooding risk evolution thermodynamic feature map, projection points belonging to different time periods are traced back in chronological order, thereby reconstructing a migration trajectory sequence extending from the safety cluster center region to the warm-colored bright area on the risk situation projection plane. Each location point in the migration trajectory sequence corresponds to a specific sampling time, and the two-dimensional coordinates of this location point are obtained by mapping the fused enhanced feature vector at that time through the feature dimensionality reduction projection module. By analyzing the geometric shape, directional changes, and color value evolution along the trajectory of the migration trajectory sequence on the two-dimensional plane, it is possible to deduce how the endogenous state parameters, such as the real-time values ​​of the column top pressure, column bottom pressure, column top temperature, and column bottom temperature, collaboratively deviate from their steady-state equilibrium reference point under the temporal changes of the external excitation disturbance parameters, such as the feed flow rate disturbance value, the reflux flow rate disturbance value, and the column bottom liquid level disturbance value. This collaborative evolution law is specifically manifested in the temporal changes of the curvature, extension rate, and local thermal accumulation pattern of the migration trajectory on the flooding risk evolution thermodynamic feature map.

[0061] Preferably, after the co-evolution law is deduced, the risk situation evolution feature identifier is generated by further performing probability density function mapping on the co-evolution law. This probability density function mapping does not directly statistically analyze the original endogenous state parameter sequence, but rather performs probabilistic modeling on the spatial distribution characteristics of the migration trajectory sequence on the risk situation projection plane. Specifically, using the two-dimensional projection coordinates corresponding to each sampling time on the migration trajectory sequence as the sample point set, a continuous probability density distribution function is established on the risk situation projection plane. This probability density distribution function can be constructed using a kernel density estimation method, where for any position on the two-dimensional risk situation projection plane, its probability density value is obtained by the weighted sum of the kernel function contributions of all sample points in the neighborhood of that position, and the bandwidth parameter of the kernel function is adaptively selected according to the dispersion of the sample point set. After the probability density function mapping process, the originally discrete migration trajectory sequence is transformed into a continuous probability density surface covering the entire risk situation projection plane. The value of each point on the probability density surface represents the probability density of the distillation column operating condition at the corresponding state space location. This value is defined as the probability density distribution value of flooding risk evolution.

[0062] Preferably, after obtaining the continuous probability density surface formed by the probability density distribution values ​​of the flooding risk evolution, feature parameters are further extracted from the probability density surface to form structured information that can quantify the risk situation. First, at least one probability density maxima is identified on the probability density surface. Each probability density maxima represents the most frequently resident or traversed cluster center of the distillation column operating condition in the state space. For each probability density maxima, its two-dimensional coordinates, peak probability density, and the area enclosed by the probability density contour lines surrounding the maxima are calculated. The two-dimensional coordinates reflect the specific location of the cluster center on the risk situation projection plane, the peak probability density reflects the frequency of the operating condition occurring at the cluster center, and the area enclosed by the contour lines reflects the distribution breadth of the cluster center in the state space. Second, the direction vector from the probability density maxima corresponding to the safe cluster center region to the probability density maxima corresponding to the warm-colored bright region is calculated. The direction of this direction vector is the dominant migration direction of risk evolution. In addition, the mean probability density gradient along the dominant migration direction is calculated, which reflects the severity of the change from a safe state to a dangerous state.

[0063] Preferably, the feature parameters extracted from the probability density distribution value of the flooded risk evolution are combined and encapsulated to generate the risk situation evolution feature identifier. The risk situation evolution feature identifier is a structured descriptor containing at least three components. The first part is a set of cluster center descriptors, which includes the two-dimensional coordinates, peak probability density, and distribution area of ​​each probability density maxima. The second part is a migration trend descriptor, which includes the direction vector of the dominant migration direction of the risk evolution and the mean of the probability density gradient along that direction vector. The third part is a temporal association label, which records the time interval range corresponding to each cluster center. This time interval range is obtained by statistically analyzing the sampling times associated with the original projection points falling within the neighborhood of the cluster center. For example, when the set of clustering center descriptors in the risk situation evolution feature identifier shows that there are two main clustering centers, located in the safe area and the dangerous boundary area respectively, and the migration trend descriptor indicates that there is a high probability density gradient mean from the safe area to the dangerous boundary area, even if the individual operating process parameters have not yet exceeded the preset threshold, it can be objectively inferred that the operating condition of the distillation column is being continuously driven by the external excitation disturbance parameter and is evolving towards the flooding critical state along a certain dominant direction.

[0064] Preferably, unlike the discrete binary judgment method in traditional schemes that only outputs over-limit alarm signals, this application provides a continuous, multi-dimensional quantitative description of the distribution pattern, aggregation characteristics, and migration trend of flooding risk in the state space by generating a risk situation evolution feature identifier through probability density function mapping processing of the aforementioned co-evolution law. The risk situation evolution feature identifier not only reflects the severity of the risk at the current moment but also reveals the historical trajectory and future direction of risk evolution through the distribution and migration direction of aggregation centers. This structured identifier, as input for subsequent hierarchical clustering mapping processing, makes the determination of risk efficiency level labels no longer dependent on simple comparisons of instantaneous thresholds, but rather on a comprehensive assessment of the global distribution characteristics and dynamic migration trends of the operating conditions in the state space. This provides accurate, reliable, and physically meaningful feature basis for generating a hierarchical distillation column flooding risk prediction and control decision-making scheme.

[0065] Optionally, by performing probability density function mapping on the co-evolutionary laws, a risk situation evolution feature identifier representing the spatial distribution of risk situations is generated, including: By performing probability density function mapping on the aforementioned co-evolution law, the probability density distribution value of flooding risk evolution, which characterizes the change in risk intensity, is obtained, and the flooding risk evolution probability score index is calculated accordingly. The probability intensity of the distillation column evolving into the flooding critical state is calibrated based on the flooding risk evolution probability score index, thereby generating a risk situation evolution feature identifier that characterizes the spatial distribution of the risk situation.

[0066] Preferably, in the process of performing probability density function mapping on the co-evolution law to obtain the flooding risk evolution probability density distribution value characterizing the change in risk intensity, the co-evolution law obtained from the previous steps is first used as the processing object. Mathematically, the co-evolution law is represented by the distribution of two-dimensional coordinate points of the migration trajectory sequence on the risk situation projection plane over time. This distribution implies the spatial aggregation and diffusion pattern of the endogenous state parameters, such as the real-time values ​​of the column top pressure, column bottom pressure, column top temperature, and column bottom temperature, which co-deviate from the steady-state equilibrium reference point under the temporal drive of external excitation disturbance parameters such as the feed flow disturbance value, the reflux flow disturbance value, and the column bottom liquid level disturbance value. The purpose of the probability density function mapping is to transform the discrete migration trajectory sequence into a continuous probability density surface defined on the risk situation projection plane, thereby enabling a smooth and normalized characterization of the frequency of occurrence of the distillation column operating conditions at various positions in the state space. This processing employs a kernel density estimation method. A pre-defined two-dimensional kernel function (e.g., a Gaussian kernel function) is placed centered on each two-dimensional projected coordinate point in the migration trajectory sequence. This kernel function has its maximum value at the center point and smoothly decays with increasing distance from the center point. For any position to be estimated on the risk situation projection plane, its flooding risk evolution probability density distribution value is equal to the sum of the contributions of the kernel functions corresponding to all two-dimensional projected coordinate points at that position. The bandwidth parameter of the kernel function controls the smoothness of the probability density surface. This bandwidth parameter is adaptively calculated based on the overall dispersion of the sample points in the migration trajectory sequence, for example, using the Scott criterion or Silverman criterion based on the sample point covariance matrix. After the above mapping processing, the sparsely distributed projection points at each sampling time on the risk situation projection plane are transformed into a continuously defined scalar field. The value of each point in this continuous scalar field is the flooding risk evolution probability density distribution value; a larger value indicates a higher probability of the distillation column operating condition occurring in the corresponding state space region.

[0067] Preferably, after obtaining the continuous probability density surface formed by the probability density distribution values ​​of the flooding risk evolution, the global distribution characteristics of the probability density distribution values ​​of the flooding risk evolution on the risk situation projection plane are further analyzed to calculate the flooding risk evolution probability score index. This calculation process first performs watershed segmentation or gradient-based local maximum search on the continuous probability density surface to locate at least one probability density maximum point. Each probability density maximum point corresponds to a coordinate position on the risk situation projection plane, which is the clustering center of the distillation column operating condition in the state space, reflecting a local equilibrium region where the endogenous state tends to revert or remain under external disturbances. For each clustering center, its peak probability density value and the closed contour area enclosed by contour lines drawn from the maximum point and where the probability density drops to a preset proportion (e.g., 60%) of the peak value are extracted. This closed area is defined as a risk cluster patch, and the coverage area of ​​the risk cluster patch is calculated. Subsequently, based on the pre-marked safe zone and dangerous zone boundaries on the risk situation projection plane, at least one dangerous cluster center located within the dangerous zone and at least one safe cluster center located within the safe zone are identified. The calculation of the flooding risk evolution probability score index integrates the peak probability density, coverage area, and probability density gradient comparison between the safe and dangerous cluster centers. Specifically, the product of the peak probability density and coverage area of ​​the dangerous cluster center is used as the basic measure of the dangerous situation intensity; the ratio of the peak probability density of the dangerous cluster center to that of the safe cluster center is used as the risk salience factor; and the mean of the probability density change gradient along the line connecting the dangerous and safe cluster centers is used as the evolution urgency factor. The basic measure, risk salience factor, and evolution urgency factor are weighted and summed to obtain a dimensionless scalar value between zero and one hundred, which is the flooding risk evolution probability score index. The higher the score index, the more significant the high-risk cluster has formed in the state space under the current operating conditions of the distillation column, and the more it is evolving towards a more dangerous area along a larger probability density gradient.

[0068] Preferably, after calculating the flooding risk evolution probability score index, the probability intensity of the distillation column evolving towards the flooding critical state is calibrated based on the flooding risk evolution probability score index, thereby generating the risk situation evolution characteristic identifier. This calibration action first establishes a probability intensity grading scale, mapping the continuous value range of the flooding risk evolution probability score index to discrete probability intensity levels. For example, a score index between 0 and 30 is designated as low probability intensity, indicating that the distillation column's operating conditions are mainly concentrated within the safe zone, without a significant tendency to drift towards the danger boundary; a score index between 30 and 60 is designated as medium probability intensity, indicating that secondary clustering centers have emerged and are spreading towards the danger zone, but a dominant danger cluster has not yet formed; a score index between 60 and 85 is designated as high probability intensity, indicating that the probability density of danger clustering centers is significant, and there is a clear migration channel between them and safe clustering centers; a score index between 85 and 100 is designated as extremely high probability intensity, indicating that the operating conditions are highly concentrated near the danger boundary and may cross the steady-state topological boundary at any time, entering a flooding state. For the current sampling time, the corresponding flooding risk evolution probability score index is converted into a probability intensity level label according to this grading scale. Subsequently, this probability intensity level label is combined and encapsulated with the various clustering center feature parameters extracted from the continuous probability density surface. Specifically, the risk situation evolution feature identifier, as a structured descriptor, comprises the following three components: the first part is a probability intensity level label, used to concisely characterize the risk severity level at the current moment; the second part is a set of cluster center descriptors, which includes the two-dimensional coordinates, peak probability density, coverage area, and region type (safe area or dangerous area) of each cluster center extracted from the continuous probability density surface; the third part is a migration trend descriptor, which includes the dominant migration direction vector from the safe cluster center to the dangerous cluster center and the mean of the probability density gradient along the direction vector.

[0069] Preferably, the generation process of the risk situation evolution feature identifier further includes embedding temporal evolution information. Since each two-dimensional projected coordinate point in the migration trajectory sequence carries its corresponding original sampling time timestamp, when extracting cluster centers from the continuous probability density surface, the timestamp distribution of the two-dimensional projected coordinate points covered by each cluster center can be traced. For each cluster center, the minimum and maximum values ​​of the timestamps corresponding to all two-dimensional projected coordinate points in its neighborhood are calculated to form a time interval range label for that cluster center. This time interval range label is appended to the corresponding cluster center descriptor in the risk situation evolution feature identifier, so that each cluster center not only has spatial location attributes but also a clear temporal persistence attribute. For example, when the risk situation evolution feature identifier shows the existence of a danger cluster center located in a dangerous area, and its time interval range label indicates that the danger cluster center was formed within the most recent sampling periods (e.g., the most recent ten minutes), and the flooding risk evolution probability score index is in the high probability intensity range, even if the single endogenous state parameter such as the real-time characterization value of the bottom pressure or the real-time characterization value of the top temperature has not yet reached its respective safety alarm threshold, this application can still objectively determine that the distillation column is in a high-risk stage of accelerated evolution towards the flooding critical state based on the evolution trend of the global probability distribution in the state space.

[0070] Preferably, unlike the traditional binary judgment mechanism that triggers discrete alarm signals solely based on comparing the instantaneous value of a single parameter with a fixed threshold, this application provides a continuous, multi-dimensional, and hierarchically quantifiable descriptive framework for flooded risks by generating the risk situation evolution feature identifier through probability density function mapping and probability scoring index calculation. The risk situation evolution feature identifier not only reflects the overall severity of the current risk using the flooded risk evolution probability scoring index as a comprehensive scalar, but also reveals the specific clustering patterns, evolution directions, and temporal processes of the risk in the state space through the distribution of cluster centers, migration trends, and time-domain labels. This structured identifier serves as input for subsequent hierarchical clustering mapping processing, enabling the determination of the risk efficiency level label to be based on a comprehensive assessment of the global probability distribution and dynamic migration trends in the state space, rather than an isolated response to an instantaneous threshold. Therefore, the risk situation evolution feature identifier generated in this application provides a highly physically interpretable and risk-foresight-based basis for the subsequent steps of generating the flooding risk prediction and control decision scheme for the graded distillation tower, which can better adapt to the actual needs of light hydrocarbon separation unit distillation tower for advanced prediction and graded control of flooding risk.

[0071] Preferably, in the process of generating the risk situation evolution characteristic identifier, the selection of the kernel function and the setting of the bandwidth parameter in the kernel density estimation method are closely related to the operating characteristics of the distillation column of the light hydrocarbon separation unit. Since the distribution of the endogenous state vector of the distillation column in the state space during steady-state operation is usually an approximately ellipsoidal Gaussian distribution around the steady-state equilibrium reference point, the selection of a two-dimensional Gaussian kernel function can more closely approximate the natural aggregation form of the operating conditions. The determination of the bandwidth parameter comprehensively considers the sampling frequency of the sensor monitoring array and the typical time constant of the distillation column operating conditions. For example, when the sampling frequency is high (one hundred times per second), the state difference between adjacent sampling points is small. If the bandwidth parameter is selected too large, the probability density surface will be too smooth and will mask the local risk aggregation details; if the bandwidth parameter is selected too small, the probability density surface will have too many false local extrema. Therefore, the bandwidth parameter is adaptively adjusted according to the local density of sample points in the migration trajectory sequence. A smaller bandwidth is used in the safe area where sample points are dense to retain details, and a larger bandwidth is used in the dangerous boundary area where sample points are sparse to ensure the stability of the estimation. By employing this probability density estimation strategy adapted to the dynamic characteristics of the distillation column, the probability density distribution value of flooding risk evolution can accurately reflect the true probability distribution of the operating condition in the state space, thereby enabling the flooding risk evolution probability scoring index and the risk situation evolution characteristic identifier to have high reliability and sensitivity.

[0072] Optionally, a staged flooding risk prediction and control decision-making scheme for flooding failures in light hydrocarbon separation units can be generated accordingly, including: Logical matching processing is performed on the risk efficiency level labels using a pre-built preset control strategy mapping model to generate a candidate control strategy sequence; The candidate control strategy sequence is correlated and verified with the control effect evaluation model built on the historical risk control case library, and a graded distillation tower flood risk prediction and control decision scheme is generated based on the verification results for flooding failure of the distillation tower in the light hydrocarbon separation unit.

[0073] Preferably, in the process of generating a candidate control strategy sequence by performing logical matching processing on the risk efficiency level labels through a pre-constructed preset control strategy mapping model, the internal module composition of the preset control strategy mapping model and its relationship with the risk efficiency level labels are first clarified. The risk efficiency level labels are discrete level identifiers obtained from the hierarchical clustering mapping process in the preceding steps, used to characterize the severity of the current flooding risk in the distillation column. For example, they can be divided into low-risk efficiency level, medium-risk efficiency level, high-risk efficiency level, and extremely high-risk efficiency level. The preset control strategy mapping model, as a two-layer decision support structure based on rule reasoning and case matching, is sequentially composed of a risk level triggering module, a strategy rule parsing module, a case similarity retrieval module, and a strategy ranking and synthesis module. This model receives the risk efficiency level labels on the input side and generates the candidate control strategy sequence on the output side. The modules collaborate in a tightly coupled manner to complete the complete mapping from risk level identifiers to a preliminary set of control recommendations. The risk level triggering module, acting as an entry point, receives the risk efficiency level tag and first converts it into a response level instruction based on a preset correspondence table between risk levels and control response levels (e.g., low-risk efficiency level corresponds to monitoring and observation level response, medium-risk efficiency level corresponds to early warning and prompting level response, high-risk efficiency level corresponds to proactive intervention level response, and extremely high-risk efficiency level corresponds to emergency response level response). This response level instruction determines the scope and depth of subsequent strategy retrieval and matching, avoiding mismatches such as invoking overly aggressive control strategies in low-risk states or only providing observation suggestions in extremely high-risk states.

[0074] Preferably, the strategy rule parsing module follows the risk level triggering module. It is used to extract basic control rule entries matching the response level from a pre-built process control rule library based on the response level instruction. The process control rule library is a structured knowledge base pre-built based on the design and operation manual of the distillation column of the light hydrocarbon separation unit, historical fault handling experience, and process safety analysis reports. In this process control rule library, each control rule entry is stored in the form of "if condition then action." The condition part involves the numerical range or trend characteristics of the endogenous state parameter and the external excitation disturbance parameter, while the action part contains specific process adjustment suggestions. For example, a rule entry corresponding to a high-risk energy efficiency level is: if the rate of increase of the real-time characterization value of the column bottom pressure exceeds a preset rate threshold (e.g., 5% increase every ten minutes) and the characterization value of the feed flow disturbance remains above 110% of the design load, then the suggested actions include reducing the feed flow setpoint and increasing the column top reflux setpoint. After receiving the response level instruction, the strategy rule parsing module uses a forward reasoning mechanism to filter out all control rule entries from the process control rule base whose conditions match the current operating state of the distillation column. It then extracts the action components from these control rule entries to form a set of rule-driven control actions. Simultaneously, the case similarity retrieval module also retrieves information from the historical risk control case base associated with the process control rule base. This historical risk control case base records the operating state characteristics of the distillation column, the control measures taken, and the final evaluation results of the treatment effects in past flooding risk events. The case similarity retrieval module uses the set of cluster center descriptors and migration trend descriptors in the current risk situation evolution feature identifier as retrieval feature vectors. It performs a nearest neighbor search based on cosine similarity or Euclidean distance in the historical risk control case base, returning several candidate historical cases most similar to the current risk situation, and extracts the control measure sequence recorded in each candidate historical case.

[0075] Preferably, the strategy ranking and synthesis module receives the rule-driven control actions from the strategy rule parsing module and the candidate historical case control measure sequences from the case similarity retrieval module. It is used to deduplicate, combine, and prioritize the control actions from both sources to generate the candidate control strategy sequence. This module first performs semantic normalization on the rule-driven control actions and the actions in each candidate historical case control measure sequence, mapping process adjustment actions with different expressions but identical substance to a unified standard action coding system. Subsequently, the module calculates the comprehensive recommendation strength score for each appearing standard action code. The comprehensive recommendation strength score is calculated by integrating information from three dimensions: first, whether the action appears in the rule-driven control actions; if it does, a higher rule confidence weight is assigned; second, the frequency and position of the action in the control measure sequence of each candidate historical case (actions closer to the beginning of the measure sequence in a case indicate higher intervention priority), based on which case support weight is assigned; and third, the corresponding treatment effect evaluation result of the action in the historical risk control case database. If the risk situation was effectively suppressed after the implementation of the action in a historical case (e.g., a decrease in risk efficiency level), a positive effect feedback weight is assigned. The rule confidence weight, case support weight, and positive effect feedback weight are weighted and summed to obtain the comprehensive recommendation strength score for each standard action code. The standard action codes are sorted in descending order according to the comprehensive recommendation strength score, and actions with scores below a preset adoption threshold are removed, ultimately forming an ordered list of control actions. This ordered list of control actions is the candidate control strategy sequence. Each element in the candidate control strategy sequence contains a specific process adjustment action name and its execution priority order, such as first priority to reduce the feed flow rate set value, second priority to increase the reflux flow rate set value, and third priority to reduce the tower reboiler heat load, etc.

[0076] Preferably, after generating the candidate control strategy sequence, the candidate control strategy sequence is further correlated and verified with the control effect evaluation model constructed based on the historical risk control case library, and a graded distillation column flooding risk prediction and control decision scheme is generated based on the verification results for the flooding failure of the distillation column in the light hydrocarbon separation unit. The control effect evaluation model is composed of a control action prediction module, a state evolution simulation module, and an effect quantification evaluation module connected sequentially. First, the control action prediction module receives the candidate control strategy sequence as input, and applies each control action in the sequence to a virtual process digital twin synchronized with the current real-time state of the distillation column according to its execution priority. This virtual process digital twin is a dynamic simulation environment jointly constructed based on the distillation column's mechanism model (such as material balance equation, energy balance equation, and gas-liquid phase balance equation) and the current state data in the time series feature matrix. For each control action, the control action prediction module modifies the corresponding boundary condition parameters in the virtual process digital twin according to the action type (such as adjusting the feed flow rate setpoint, adjusting the return flow rate setpoint, adjusting the reboiler heat load, etc.), and drives the virtual process digital twin to extrapolate a prediction time window (e.g., the next fifteen minutes).

[0077] Preferably, the state evolution simulation module receives the predicted state trajectory of the virtual process digital twin after each control action is executed by the control action prediction module. This trajectory is used to extract key feature parameters reflecting the flooding risk evolution trend. Specifically, at each virtual sampling moment within the prediction time window, the state evolution simulation module calculates the virtual multidimensional composite operating condition deviation, the virtual flooding risk evolution probability score index, and the virtual risk efficiency level label, following the same processing logic as described above. The change curves of these virtual indicators within the prediction time window are recorded to form a predicted risk evolution curve. This predicted risk evolution curve describes the expected change trend of the distillation column's flooding risk indicators over time after each control action in the candidate control strategy sequence is executed sequentially. The effect quantification and evaluation module then receives the predicted risk evolution curve and compares it with the baseline risk evolution curve under no intervention conditions (also obtained by free deduction of the virtual process digital twin without modifying boundary conditions). The comparative analysis includes: the difference in risk efficiency levels between the predicted risk evolution curve and the baseline risk evolution curve; the magnitude of the decrease in the risk score index; the time required for the risk indicator to fall back to the safe range; and the reduction ratio of the time integral area of ​​the risk indicator within the entire prediction time window. Based on these comparative indicators, the effectiveness quantification evaluation module calculates a pass confidence score between zero and one, which reflects the effectiveness of the candidate control strategy sequence in suppressing the current flooding risk.

[0078] Preferably, the effect quantification evaluation module is also responsible for retrieving past cases from the historical risk management case library that are similar to the core action combination in the current candidate management strategy sequence, and extracting the actual execution effect data of these past cases as auxiliary verification basis. The verification pass confidence score obtained based on the virtual process digital twin simulation is weighted and fused with the experience verification pass confidence score obtained based on historical case statistics to obtain the final verification pass confidence score. Subsequently, the verification pass confidence score is compared with a preset graded verification threshold. If the verification pass confidence score is higher than the preset adoption threshold (e.g., 0.7), the candidate management strategy sequence is marked as valid, and the candidate management strategy sequence is directly output as the core management action sequence in the graded distillation column liquid flooding risk prediction and management decision-making scheme. If the verification pass confidence score is lower than the preset adoption threshold but higher than the preset correction threshold (e.g., 0.4), the management effect evaluation model will trigger a strategy correction mechanism. The strategy correction mechanism is as follows: retrieve cases that are similar to the current risk situation and have been successfully handled from the historical risk management case library, extract supplementary management actions adopted in the successful cases but not included in the candidate management strategy sequence, insert the supplementary management actions into the appropriate positions in the candidate management strategy sequence according to their execution sequence in the successful cases, form a corrected candidate management strategy sequence, and perform a new round of simulation verification until the verification passes and the confidence score meets the adoption conditions.

[0079] Preferably, if the confidence score of the verification pass is still lower than the preset adoption threshold after a preset number of corrections, or if the confidence score of the verification pass is lower than the preset correction threshold from the beginning, then the control effect evaluation model determines that the candidate control strategy sequence is insufficient to effectively deal with the current risk, and instead retrieves several emergency response cases that are most similar to the current risk situation and have been successfully handled from the historical risk control case library, and outputs the control action sequence in the emergency response cases as an alternative strategy. Finally, the flooding risk prediction and control decision-making scheme for the staged distillation column is generated in the form of a structured data package. This data package includes: First, the current risk efficiency level label and its corresponding risk situation evolution characteristic summary information; Second, the final control strategy sequence determined after verification and optimization, where each control action is accompanied by execution priority, suggested adjustment range (e.g., a 5% to 10% reduction in feed flow rate), and expected effective time window; Third, the predicted risk evolution curve chart obtained from the virtual process digital twin simulation, which visually shows the expected decline trajectory of risk indicators after executing different control actions at different time points; Fourth, a summary of similar successful cases extracted from the historical risk control case library, serving as background information for operator decision-making reference. Unlike traditional solutions that can only output a single alarm prompt without providing an actionable intervention strategy, this application generates a flooding risk prediction and control decision-making scheme for the staged distillation column through the collaborative work of the preset control strategy mapping model and the control effect evaluation model. This scheme can provide specific process control command sequences that match the risk level and have been verified by both simulation and historical experience at different stages of flooding risk evolution. This provides clear and reliable action guidelines for operators or automatic control systems, helping to take effective intervention measures in a timely manner before flooding failures occur to suppress the further development of risks.

[0080] Optionally, based on the verification results, a staged flooding risk prediction and control decision-making scheme for flooding failures in the distillation column of a light hydrocarbon separation unit is generated, including: Based on the verification results, a preliminary simulation assessment is performed to generate a hierarchical risk management decision description. The hierarchical risk management decision description is converted into an instruction, thereby generating a hierarchical distillation column flooding risk prediction and management decision scheme for flooding failures in light hydrocarbon separation units.

[0081] Preferably, in the process of generating a hierarchical risk management decision description by performing a predictive simulation evaluation based on the verification results, the verification results output by the management effectiveness evaluation model in the preceding steps are first used as the processing object. The verification results are a data structure containing multi-dimensional evaluation information, which at least includes the verification pass confidence score, the expected effective time window of each management action in the candidate management strategy sequence, the comparison data between the predicted risk evolution curve and the baseline risk evolution curve, and a summary of similar successful cases retrieved from the historical risk management case library. The predictive simulation evaluation is not a simple repetition of the forward extrapolation completed by the virtual process digital twin in the correlation verification processing stage, but rather uses the candidate management strategy sequence (or its modified version) finally determined in the verification results as the core input, and performs robustness and adaptability verification within an extended time scale and operating condition disturbance range. Specifically, each management action in the finally determined candidate management strategy sequence is applied sequentially to an enhanced virtual process digital twin synchronized with the current real-time status of the distillation column, according to its execution priority and the expected effective time window. This enhanced virtual process digital twin differs from the conventional virtual process digital twin used in the correlation verification stage. It introduces an additional simulation module to address the future uncertainty of the external excitation disturbance parameters, building upon the mechanistic model (material balance equation, energy balance equation, and gas-liquid phase balance equation). This uncertainty simulation module generates several possible external disturbance evolution trajectories based on the historical fluctuation statistical characteristics (such as variance and autocorrelation function) of the feed flow rate disturbance characterization value, the reflux flow rate disturbance characterization value, and the bottom liquid level disturbance characterization value, thus forming a set of disturbance scenarios encompassing various potential operating conditions.

[0082] Preferably, during the application of the candidate control strategy sequence to the enhanced virtual process digital twin, for each external disturbance evolution trajectory in the disturbance scenario set, the enhanced virtual process digital twin is driven to perform a longer-term prediction and extrapolation, such as extrapolating to the next 30 to 60 minutes, to fully observe the medium- and long-term effects of the control measures. During the extrapolation of each external disturbance evolution trajectory, following the same processing logic as the aforementioned steps, the virtual multidimensional composite operating condition deviation, the virtual flooding risk evolution probability score index, and the virtual risk efficiency level label are calculated, and the changes of these virtual risk indicators over time are recorded. Thus, for each external disturbance evolution trajectory, a corresponding post-control risk evolution prediction curve is generated. By summarizing the post-control risk evolution prediction curves corresponding to all external disturbance evolution trajectories in the disturbance scenario set, a statistical distribution description of the future risk state of the distillation column under the action of the candidate control strategy sequence can be obtained. Specifically, at each future time point within the estimated projection time window, the frequency percentage of each level of the virtual risk efficiency level label under different perturbation trajectories is calculated, along with the mean and confidence interval of the virtual flooded risk evolution probability score index. Simultaneously, the proportion of external perturbation evolution trajectories under which the virtual risk efficiency level label can fall below the low-risk efficiency level within a preset time threshold (e.g., fifteen minutes) is also recorded. These statistical distribution information collectively constitute a quantitative assessment of the robustness of the candidate control strategy sequence in uncertain environments.

[0083] Preferably, based on the risk status statistical distribution information obtained from the above-mentioned pre-simulation assessment, the hierarchical risk management decision description is further generated. The hierarchical risk management decision description is a structured natural language or formatted information statement for operators or upper-level automatic control systems, aiming to convey the content and expected effects of the management plan in a clear and hierarchical manner. The process of generating this description first determines the main theme of the decision description based on the verification pass confidence score and the risk reduction ratio in the risk status statistical distribution information. For example, if the verification pass confidence score is higher than a preset high confidence threshold (e.g., 0.85) and the risk reduction ratio is high (e.g., the risk can reduce within 15 minutes in more than 90% of disturbance scenarios), then the main theme of the decision description is determined to be recommended execution; if the verification pass confidence score is at a medium level or the risk reduction ratio is medium, then the main theme is determined to be suggested execution with enhanced observation; if the verification pass confidence score is passed but close to the lower threshold, then the main theme is determined to be cautious execution with backup plans. Subsequently, following a pre-defined hierarchical description template, each control action in the candidate control strategy sequence, its suggested adjustment range, and execution priority order are filled into the action description paragraph of the template. The action description paragraphs are generated in descending order of priority. Each control action description includes the action name, adjustment target, adjustment range, and expected effective time window. For example, the generated description statement might be: "First priority: Reduce the feed flow rate setpoint by 5% to 10%, expected to begin suppressing the upward trend of bottom pressure within 5 to 10 minutes; Second priority: Increase the reflux flow rate setpoint by 3% to 8%, expected to synergistically improve the gas-liquid phase load distribution within the column within 10 to 15 minutes." Preferably, the tiered risk management decision description also includes an embedded paragraph stating the expected effects based on the estimated simulation evaluation results. This paragraph does not use subjective inference but rather objectively reflects quantitative data from the statistical distribution information of the risk status. For example, when statistical results show that the virtual risk efficiency level can drop to a low-risk level within 20 minutes under 95% of the disturbance scenarios, the expected effect paragraph is described as: "Based on the statistical results of simulating possible future external disturbance scenarios, after implementing the above management strategy, in 95% of the simulated disturbance scenarios, the distillation column flooding risk index is expected to fall back to the low-risk range within 20 minutes." Furthermore, if a strategy correction mechanism is triggered during the verification process of the management effect evaluation model and supplementary management actions are introduced from the historical risk management case library, the tiered risk management decision description will also include a note paragraph clearly indicating the source of the supplementary action and its historical success rate. For example: "Note: The action of reducing the heat load of the tower bottom in the current strategy is supplemented by historical successful cases with similar risk situations (several case numbers). In these cases, the risk situation was effectively suppressed after this measure was implemented." In this way, the hierarchical risk management decision description not only conveys specific operational instructions, but also provides the simulation basis and historical experience reference behind the instructions, thereby enhancing the credibility and interpretability of the decision scheme.

[0084] Preferably, after generating the hierarchical risk management decision description, the hierarchical risk management decision description is further processed into an instruction-based conversion process to generate a hierarchical distillation column flooding risk prediction and control decision scheme for flooding failures in the light hydrocarbon separation unit. The purpose of the instruction-based conversion process is to transform the hierarchical risk management decision description in natural language form into a machine-readable instruction sequence that can be directly recognized and executed by the distributed control system or advanced process control system of the distillation column. This process first performs semantic parsing and structured extraction on the action description paragraphs in the hierarchical risk management decision description. Through a preset process control instruction mapping table, the action names in natural language (such as "reduce feed flow rate setpoint") are mapped to the corresponding control tag number and operation type code. For example, "feed flow rate setpoint" is mapped to the setpoint register address of the flow control loop on the feed pipeline of the light hydrocarbon separation unit, and "reduce" is mapped to a write operation and negative increment mode. The adjustment range in the description statement (such as "five percent to ten percent") is parsed into a target adjustment range and decomposed into a step-by-step adjustment instruction sequence according to a preset adjustment step size strategy. For example, the first step of adjustment is carried out with a small amplitude (such as 3%). After observing the expected effective time window, if the trend is in line with expectations, the current adjustment amount is maintained. If the deviation is still large, the second step of adjustment is applied until the upper limit of the amplitude range is reached.

[0085] Preferably, the instruction conversion process also converts the execution priority order in the hierarchical risk management decision description into execution time sequence markers for the instruction sequence. Control instructions corresponding to high-priority management actions are assigned earlier execution timestamps or immediate execution markers, while control instructions corresponding to low-priority management actions are set to conditionally triggered execution. That is, after the preceding high-priority instruction has been executed and the expected effective time window has passed, if the real-time operating process parameters fed back by the sensor monitoring array indicate that the decline in risk indicators has not reached the expected level, execution is automatically triggered. Through this conversion, a complete industrial control instruction sequence is formed, including instruction opcodes, operation object register addresses, operation values, execution time sequence markers, and triggering conditions. Finally, the current risk efficiency level label, the summary of the cluster center descriptor subset in the risk situation evolution feature identifier, the verification pass confidence score, the predicted risk evolution curve chart, and the aforementioned industrial control instruction sequence are encapsulated into a unified data package, which is the hierarchical distillation column flooding risk prediction and management decision scheme. The flooding risk prediction and control decision-making scheme of the staged distillation column can be sent to the distillation column control system of the light hydrocarbon separation unit via standard industrial communication protocols (such as process control standards on network communication protocols or Ethernet factory automation protocols), or presented to the operator in a visual manner through a human-machine interface, and then manually executed by the operator after confirmation.

[0086] Preferably, unlike traditional solutions that rely solely on operator experience for tentative adjustments after an alarm is triggered and lack prediction of the adjustment effect, this application transforms optimized control strategies obtained from digital twin simulation and historical case verification into precise control commands that can directly guide the process system through two closely linked stages: prediction simulation evaluation and command conversion processing. The staged distillation column flooding risk prediction and control decision-making scheme not only includes robustly verified specific adjustment actions but also step-by-step implementation strategies and expected effect information based on simulation statistics. This shifts intervention measures for flooding faults in light hydrocarbon separation unit distillation columns from passive response and experience-driven to a proactive prediction, data-driven, and model-driven approach. This technical path provides a highly operable, reliable, and forward-looking decision support method for reducing the risk of unplanned downtime caused by flooding, minimizing capacity losses, and ensuring the safe operation of the unit.

[0087] like Figure 2 As shown in the figure, a flooding risk prediction and control decision-making device for a distillation column in a light hydrocarbon separation unit is provided in this application embodiment, comprising: The feature configuration module is used to collect real-time operating process parameters corresponding to each process node during the operating cycle of the distillation column by a sensor monitoring array deployed at each process node of the distillation column of the light hydrocarbon separation unit, so as to generate a full-cycle process operation feature configuration of the distillation column that characterizes the full-dimensional operating characteristics of the unit based on the real-time operating process parameters. The deviation evolution module is used to establish the thermodynamic coupling relationship between endogenous state parameters and external excitation disturbance parameters based on the full-cycle process operation characteristic conformation of the distillation column, so as to construct the deviation evolution feature body. The deviation evolution feature body is used to characterize the degree of drift of the current operating condition of the distillation column relative to the steady-state equilibrium point. The identifier module is used to perform digital characterization processing on the deviation evolution feature body to obtain the risk feature evolution carrier and generate a risk situation evolution feature identifier body that represents the spatial distribution state of the risk situation. The clustering module is used to perform hierarchical clustering mapping processing on the risk situation evolution feature identifier to obtain a risk energy efficiency level label that represents the current risk severity and generate a hierarchical distillation column flooding risk prediction and control decision scheme for flooding failure of light hydrocarbon separation unit.

[0088] like Figure 3 As shown, this application discloses an electronic device including a processor and a memory; The memory is used to store computer programs; When the processor executes the program stored in the memory, it implements the flooding risk prediction and control decision-making method for distillation towers in light hydrocarbon separation units as described in any of the preceding claims.

[0089] Figures 2-3 For an exemplary description, please refer to the above. Figure 1 This will not be elaborated upon here.

Claims

1. A method for predicting and controlling the decision of flooding risk of a rectifying column of a light hydrocarbon separation device, characterized in that, include: Step 1: Collect real-time operating process parameters corresponding to each process node during the operating cycle of the distillation column by a sensor monitoring array deployed at each process node of the light hydrocarbon separation unit, so as to generate a full-cycle process operation characteristic profile of the distillation column that characterizes the full-dimensional operating characteristics of the unit based on the real-time operating process parameters. Step 2: Based on the full-cycle process operation characteristic conformation of the distillation column, establish the thermodynamic coupling relationship between endogenous state parameters and external excitation disturbance parameters to construct a deviation evolution characteristic body. The deviation evolution characteristic body is used to characterize the degree of drift of the current operating condition of the distillation column relative to the steady-state equilibrium point. Step 3: Perform digital characterization processing on the deviation evolution feature body to obtain the risk feature evolution carrier and generate a risk situation evolution feature identifier body that represents the spatial distribution of risk situation. Step 4: Perform hierarchical clustering mapping on the risk situation evolution feature identifier to obtain a risk energy efficiency level label that represents the current risk severity and generate a hierarchical distillation column flooding risk prediction and control decision scheme for flooding failure of light hydrocarbon separation unit.

2. The method for predicting and controlling the flooding risk of a rectifying column of a light hydrocarbon separation device according to claim 1, characterized in that, The process of generating a full-cycle process operation characteristic profile of the distillation column based on the real-time operating process parameters, representing the full-dimensional operating characteristics of the device, includes: All real-time operating process parameters collected are subjected to time-series synchronization calibration to generate digital characterization data covering the full range of operating conditions of the distillation column; The digital characterization data is analyzed for variables to generate a full-cycle process operation characteristic profile of the distillation column that characterizes the full-dimensional operation features of the device.

3. The method for predicting and controlling the flooding risk of a rectifying column of a light hydrocarbon separation device according to claim 2, characterized in that, The step of performing variable analysis on the digital characterization data to generate a full-cycle process operation characteristic profile of the distillation column representing the full-dimensional operational characteristics of the device includes: The digital characterization data is analyzed to classify and obtain the endogenous state parameters characterizing the top pressure, bottom pressure, top temperature, and bottom temperature of the column, as well as the external excitation disturbance parameters characterizing the feed flow rate, reflux flow rate, and bottom liquid level. Based on the endogenous state parameters and the external excitation disturbance parameters, a full-cycle process operation characteristic profile of the distillation column is generated, representing the full-dimensional operation characteristics of the device.

4. The method for predicting and controlling the flooding risk of a rectifying column of a light hydrocarbon separation device according to claim 3, characterized in that, Based on the full-cycle process operation characteristic conformation of the distillation column, a thermodynamic coupling relationship between endogenous state parameters and external excitation disturbance parameters is established to construct a deviation evolution characteristic body, including: Based on the aforementioned thermodynamic coupling relationship, a parameter table for characterizing operating condition deviation is generated; Based on the operating condition deviation characterization parameter table, the endogenous state parameters are mapped to the multidimensional thermodynamic steady-state topological boundary space to obtain the operating condition drift trajectory that characterizes the current operating condition deviation from the steady-state equilibrium range, and the deviation evolution feature body is constructed in conjunction with the external excitation disturbance parameters.

5. The method for predicting and controlling the flooding risk of a rectifying column of a light hydrocarbon separation device according to claim 4, characterized in that, Constructing the deviation evolution feature body, including: Spatiotemporal dimension projection processing is performed on the drift trajectory of the operating condition, and path integration is performed based on the projection result to calculate the multidimensional composite operating condition deviation degree that characterizes the deviation intensity of physical quantities. Based on the deviation of the multidimensional composite working condition and the external excitation disturbance parameters, a deviation evolution feature body is constructed.

6. The method for predicting and controlling flooding risks in a distillation column of a light hydrocarbon separation unit according to claim 5, characterized in that, Based on this, risk situation evolution feature identifiers representing the spatial distribution of risk situation are generated, including: The risk feature evolution carrier is input into the pre-trained flooded risk prediction model to perform nonlinear spatiotemporal feature extraction, so as to generate a flooded risk evolution thermodynamic feature map that characterizes the spatial distribution of risk. Based on the aforementioned flooding risk evolution thermodynamic feature map, a risk situation evolution feature identifier is generated to characterize the spatial distribution of risk situation.

7. The method for predicting and controlling flooding risks in a distillation column of a light hydrocarbon separation unit according to claim 6, characterized in that, Based on the aforementioned flooding risk evolution thermodynamic feature map, a risk situation evolution feature identifier representing the spatial distribution of risk situation is generated, including: Based on the aforementioned flooding risk evolution thermodynamic feature map, the collaborative evolution law of the endogenous state parameters under the influence of the external excitation disturbance parameters is deduced; By performing probability density function mapping on the aforementioned co-evolutionary laws, a risk situation evolution feature identifier representing the spatial distribution of risk situations is generated.

8. The method for predicting and controlling flooding risks in a distillation column of a light hydrocarbon separation unit according to claim 7, characterized in that, By performing probability density function mapping on the aforementioned co-evolutionary laws, a risk situation evolution feature identifier representing the spatial distribution of risk situations is generated, including: By performing probability density function mapping on the aforementioned co-evolution law, the probability density distribution value of flooding risk evolution, which characterizes the change in risk intensity, is obtained, and the flooding risk evolution probability score index is calculated accordingly. The probability intensity of the distillation column evolving into the flooding critical state is calibrated based on the flooding risk evolution probability score index, thereby generating a risk situation evolution feature identifier that characterizes the spatial distribution of the risk situation.

9. The method for predicting and controlling flooding risks in a distillation column of a light hydrocarbon separation unit according to claim 8, characterized in that, Based on this, a staged flooding risk prediction and control decision-making scheme for flooding failures in light hydrocarbon separation units is generated, including: Logical matching processing is performed on the risk efficiency level labels using a pre-built preset control strategy mapping model to generate a candidate control strategy sequence; The candidate control strategy sequence is correlated and verified with the control effect evaluation model built on the historical risk control case library, and a graded distillation tower flood risk prediction and control decision scheme is generated based on the verification results for flooding failure of the distillation tower in the light hydrocarbon separation unit.

10. The method for predicting and controlling flooding risks in a distillation column of a light hydrocarbon separation unit according to claim 9, characterized in that, Based on the validation results, a staged flooding risk prediction and control decision-making scheme is generated for flooding failures in the distillation column of a light hydrocarbon separation unit, including: Based on the verification results, a preliminary simulation assessment is performed to generate a hierarchical risk management decision description. The hierarchical risk management decision description is converted into an instruction, thereby generating a hierarchical distillation column flooding risk prediction and management decision scheme for flooding failures in light hydrocarbon separation units.

Images