Multi-stage distillation control method and system for fatty acid ester refining

Abstract

The invention discloses a multi-stage distillation control method and system for fatty acid ester refining, and the method comprises the steps: collecting the temperature, pressure, component concentration and flow data of each tower of a multi-stage distillation system in real time, and obtaining a real-time operation state vector representing the overall operation situation of the current system; according to the real-time operation state vector, obtaining a global optimization instruction set containing optimal set temperature, set pressure, reflux ratio and inter-tower energy flow instructions of each tower; according to the inter-tower energy flow instruction, obtaining a real-time energy flow map for realizing cross-tower energy cascade matching; according to the global optimization instruction set and the real-time energy flow atlas, obtaining a stable rectification operation state matched with the optimization target and the energy flow atlas; and obtaining a system self-healing state for maintaining continuous and stable output of the high-purity fatty acid ester according to abnormal disturbance characteristics continuously monitored in a stable rectification operation state. By utilizing the embodiment of the invention, global cooperative control of multi-stage distillation can be realized, and the separation purity and the energy efficiency ratio are improved.

Description

Technical Field

[0001] The invention belongs to the technical field of distillation control, and in particular to a multi-stage distillation control method and system for refining fatty acid esters. Background Art

[0002] The refining process of fatty acid esters usually adopts a multi-stage distillation system, and its process complexity and energy consumption have always been difficulties in the industry. Traditional control methods rely on the independent adjustment of a single tower, which is difficult to cope with dynamic conditions such as component concentration fluctuations and energy coupling between towers, resulting in low separation efficiency and high energy consumption. In the existing technology, the control strategy based on fixed parameters lacks the dynamic response capability for multi-objective collaborative optimization, and the heat recovery network is mostly static in design, which cannot achieve energy cascade utilization. In addition, the system's anti-disturbance ability is insufficient, and abnormal operating conditions can easily cause a decrease in product purity. Summary of the Invention

[0003] The purpose of the present invention is to provide a multi-stage distillation control method and system for refining fatty acid esters, so as to solve the deficiencies in the prior art, realize global coordinated control of multi-stage distillation, and improve separation purity and energy efficiency.

[0004] One embodiment of the present application provides a multi-stage distillation control method for refining fatty acid esters, the method comprising: Real-time data collection of temperature, pressure, component concentration, and flow rate of each tower in a multi-stage distillation system is used to perform dynamic operation state modeling based on artificial intelligence, thereby obtaining a real-time operation state vector that represents the overall operation status of the current system. Based on the real-time operation state vector, a multi-objective collaborative optimization process is performed based on a reinforcement learning algorithm to obtain a global optimization instruction set including the optimal set temperature, set pressure, reflux ratio, and inter-tower energy flow instructions for each tower, wherein the optimization objectives include maximizing separation purity, minimizing energy consumption, and maximizing system stability margin; According to the inter-tower energy flow instructions, the dynamic topology reconstruction of the heat recovery network is performed, and the high-grade tower top steam waste heat is preferentially directed to the low-grade heat demand point through intelligent valve switching, thereby obtaining a real-time energy flow map that realizes cross-tower energy ladder matching; According to the global optimization instruction set and the real-time energy flow map, each tower actuator including a heater, a reflux controller, and a pressure regulating valve is coordinated and controlled to obtain a stable distillation operation state that matches the optimization target and the energy flow map; According to the abnormal disturbance characteristics continuously monitored under the stable distillation operation state, adaptive parameter fine-tuning and emergency strategy triggering processing are performed based on the pre-trained fault mode library to obtain a system self-healing state that maintains continuous and stable output of high-purity fatty acid esters.

[0005] Optionally, the real-time collection of temperature, pressure, component concentration, and flow rate data of each tower of the multi-stage distillation system, and the implementation of dynamic operation state modeling based on artificial intelligence to obtain a real-time operation state vector representing the overall operation status of the current system include: The temperature, pressure, component concentration and flow rate data of each tower are collected, and multi-source data fusion is performed based on Kalman filtering to obtain a high-confidence operating condition data set after noise suppression; Inputting the high-confidence operating condition data set into a dynamic component concentration soft-sensing model based on a long short-term memory network to predict real-time component concentration values to obtain a supplementary data set of component concentrations; Based on the high-confidence operating condition data set and the supplementary data set, a material-energy coupling relationship between towers is constructed to obtain a dynamic system topology diagram representing the state association of each tower; According to the dynamic system topology diagram, multi-tower collaborative state embedding driven by graph neural network is performed to obtain a 128-dimensional real-time operation state vector containing system-level spatiotemporal features.

[0006] Optionally, the multi-objective collaborative optimization process is performed based on the reinforcement learning algorithm according to the real-time operation state vector to obtain a global optimization instruction set including the optimal set temperature, set pressure, reflux ratio and inter-tower energy flow instructions for each tower, wherein the optimization objectives include maximizing separation purity, minimizing energy consumption and maximizing system stability margin, including: According to the real-time operation state vector, a multi-objective reward function is constructed based on a deep deterministic policy gradient algorithm to obtain a purity-energy consumption-stability ternary reward function; Based on the reward function, offline pre-training of the policy network and online real-time policy fine-tuning are performed to obtain a dynamic policy parameter set for multi-objective optimization; According to the dynamic strategy parameter set, a distributed parallel strategy is executed to obtain a local optimization instruction set of independent operating parameters of each tower; Performing cross-tower energy flow game theory optimization based on the local optimization instruction set and the real-time operation state vector to obtain an inter-tower energy flow instruction matrix that minimizes global entropy increase; According to the local optimization instruction set and the inter-tower energy flow instruction matrix, instruction conflicts are resolved and collaboratively packaged to obtain a global optimization instruction set including temperature, pressure, reflux ratio setting values and energy flow instructions.

[0007] Optionally, the dynamic topology reconstruction of the heat recovery network is performed according to the inter-tower energy flow instruction, and high-grade tower top steam waste heat is preferentially directed to low-grade heat demand points through intelligent valve switching, thereby obtaining a real-time energy flow map that realizes cross-tower energy cascade matching, including: According to the inter-tower energy flow instructions, waste heat grade classification and heat demand priority sorting are performed to obtain a matching priority list of heat sources and heat sinks; According to the matching priority list, dynamic path planning of the thermal network based on directed graph theory is performed to obtain an optimal heat transfer path sequence with a minimum heat transfer temperature difference; generating a valve switch instruction matrix for controlling steam flow according to the optimal heat transfer path sequence; According to the actual heat transfer data after executing the valve switching instruction matrix, real-time evaluation of thermodynamic efficiency and graph visualization processing are performed to obtain a real-time energy flow graph with efficiency labels.

[0008] Optionally, the step of collaboratively regulating the tower actuators including the heater, the reflux controller, and the pressure regulating valve according to the global optimization instruction set and the real-time energy flow map to obtain a stable distillation operation state that matches the optimization target and the energy flow map includes: According to the tower parameters in the global optimization instruction set, feedforward-feedback composite adjustment based on model predictive control is performed to obtain a reference control signal for each tower actuator; Performing dynamic compensation calculation of reboiler heating power based on cross-tower heat transfer efficiency data in the real-time energy flow map to obtain a thermal coupling compensation coefficient; According to the reference control signal and the thermal coupling compensation coefficient, the control amount of each tower actuator is collaboratively corrected to obtain an optimized control amount set under the anti-overshoot constraint; Perform multivariable decoupling control and actuator response lag compensation according to the optimized control quantity set to obtain synchronized execution instructions; After executing the synchronized execution instruction to drive each actuator, the operation state stability index processing is verified in real time to obtain a stable distillation operation state that meets the Lyapunov stability condition.

[0009] Optionally, according to the abnormal disturbance characteristics continuously monitored under the stable distillation operation state, adaptive parameter fine-tuning and emergency strategy triggering processing based on a pre-trained fault mode library are performed to obtain a system self-healing state that maintains continuous and stable output of high-purity fatty acid esters, including: Extracting disturbance features based on wavelet packet transform according to the real-time monitoring data under the stable distillation operation state to obtain an abnormal disturbance fingerprint vector; Based on the abnormal disturbance fingerprint vector, similarity matching is performed with the pre-trained fault mode library to obtain the fault type identification and confidence level; According to the fault type identification and confidence level, an emergency strategy retrieval based on case reasoning is performed to obtain an emergency operation instruction set including a parameter adjustment range and a valve safety interlock; According to the emergency operation instruction set, parameter gradual adjustment and system status monitoring are performed under safety constraints, and a system self-healing state is obtained in which the system automatically recovers to an optimized state after the disturbance is eliminated.

[0010] Another embodiment of the present application provides a multi-stage distillation control system for refining fatty acid esters, the system comprising: The acquisition module is used to collect temperature, pressure, component concentration and flow rate data of each tower in the multi-stage distillation system in real time, perform dynamic operation state modeling based on artificial intelligence, and obtain a real-time operation state vector that represents the overall operation status of the current system; an optimization module for performing multi-objective collaborative optimization processing based on the real-time operation state vector and a reinforcement learning algorithm to obtain a global optimization instruction set including the optimal set temperature, set pressure, reflux ratio, and inter-tower energy flow instructions for each tower, wherein the optimization objectives include maximizing separation purity, minimizing energy consumption, and maximizing system stability margin; A reconstruction module is used to dynamically reconstruct the topology of the heat recovery network according to the inter-tower energy flow instructions, and to preferentially direct high-grade tower overhead steam waste heat to low-grade heat demand points through intelligent valve switching, thereby obtaining a real-time energy flow map that achieves cross-tower energy cascade matching; a control module for collaboratively controlling the actuators of each column, including the heater, the reflux controller, and the pressure regulating valve, according to the global optimization instruction set and the real-time energy flow map, to obtain a stable distillation operation state that matches the optimization target and the energy flow map; The processing module is used to perform adaptive parameter fine-tuning and emergency strategy triggering based on a pre-trained fault mode library according to the abnormal disturbance characteristics continuously monitored under the stable distillation operation state, so as to obtain a system self-healing state that maintains continuous and stable output of high-purity fatty acid esters.

[0011] Yet another embodiment of the present application provides a storage medium, wherein the storage medium stores a computer program, wherein the computer program is configured to execute any of the above methods when run.

[0012] Yet another embodiment of the present application provides an electronic device, comprising a memory and a processor, wherein the memory stores a computer program, and the processor is configured to run the computer program to execute any of the above methods.

[0013] Compared with the prior art, the present invention provides a multi-stage distillation control method for refining fatty acid esters, which collects temperature, pressure, component concentration and flow data of each tower of the multi-stage distillation system in real time to obtain a real-time operation state vector that characterizes the overall operation status of the current system; based on the real-time operation state vector, a global optimization instruction set including the optimal set temperature, set pressure, reflux ratio and inter-tower energy flow instructions of each tower is obtained; based on the inter-tower energy flow instructions, a real-time energy flow spectrum that realizes cross-tower energy ladder matching is obtained; based on the global optimization instruction set and the real-time energy flow spectrum, a stable distillation operating state that matches the optimization target and the energy flow spectrum is obtained; based on the abnormal disturbance characteristics continuously monitored under the stable distillation operating state, a system self-healing state that maintains continuous and stable output of high-purity fatty acid esters is obtained, thereby realizing global coordinated control of multi-stage distillation and improving separation purity and energy efficiency. BRIEF DESCRIPTION OF THE DRAWINGS

[0014] Figure 1 A hardware structure block diagram of a computer terminal for a multi-stage distillation control method for refining fatty acid esters provided in an embodiment of the present invention; Figure 2 A schematic flow chart of a multi-stage distillation control method for refining fatty acid esters provided in an embodiment of the present invention; Figure 3 A schematic structural diagram of a multi-stage distillation control system for refining fatty acid esters provided in an embodiment of the present invention. DETAILED DESCRIPTION

[0015] The embodiments described below with reference to the accompanying drawings are exemplary and are only used to explain the present invention, and are not to be construed as limiting the present invention.

[0016] The embodiment of the present invention first provides a multi-stage distillation control method for refining fatty acid esters. The method can be applied to electronic devices such as computer terminals, specifically ordinary computers.

[0017] The following describes it in detail by taking running on a computer terminal as an example. Figure 1 The hardware structure block diagram of a computer terminal for a multi-stage distillation control method for refining fatty acid esters provided by an embodiment of the present invention. Figure 1 As shown, the computer device includes a processor, a memory, and a network interface connected via a system bus, wherein the memory may include a non-volatile storage medium and an internal memory.

[0018] The non-volatile storage medium can store an operating system and a computer program. The computer program includes program instructions, and when the program instructions are executed, the processor can execute any multi-stage distillation control method for refining fatty acid esters.

[0019] The processor is used to provide computing and control capabilities and support the operation of the entire computer equipment.

[0020] The internal memory provides an environment for the operation of the computer program in the non-volatile storage medium. When the computer program is executed by the processor, the processor can execute any multi-stage distillation control method for refining fatty acid esters.

[0021] The network interface is used for network communication, such as sending assigned tasks, etc. Those skilled in the art will understand that Figure 1 The structure shown in the figure is only a block diagram of a part of the structure related to the solution of the present application, and does not constitute a limitation on the computer device to which the solution of the present application is applied. The specific computer device may include more or fewer components than shown in the figure, or combine certain components, or have a different component arrangement.

[0022] It should be understood that the processor may be a central processing unit (CPU), other general-purpose processors, digital signal processors (DSP), application-specific integrated circuits (ASIC), field-programmable gate arrays (FPGA), other programmable logic devices, discrete gate or transistor logic devices, discrete hardware components, etc. The general-purpose processor may be a microprocessor or any conventional processor, etc.

[0023] See also Figure 2 The embodiment of the present invention provides a multi-stage distillation control method for refining fatty acid esters, which may include the following steps: S201, collecting temperature, pressure, component concentration, and flow rate data of each tower of the multi-stage distillation system in real time, performing dynamic operation state modeling based on artificial intelligence, and obtaining a real-time operation state vector representing the overall operation status of the current system; Specifically, the temperature, pressure, component concentration and flow rate data of each tower can be collected, and multi-source data fusion can be performed based on Kalman filtering to obtain a high-confidence operating condition data set after noise suppression; Data Collection Specifications Temperature collection: K-type thermocouples (temperature sensors) are deployed at the top, middle, and bottom of each distillation tower, with a sampling frequency of 1 Hz (once per second), a measurement range of 0-300°C (Celsius), and an accuracy of ±0.5°C. Pressure collection: Install a piezoresistive transmitter (pressure sensor) at the key section of the tower body, with a measuring range of 0-1.0 MPa (megapascal) and an output of 4-20 mA current signal; Component concentration collection: The concentration of C16-C22 fatty acid esters was obtained every 5 minutes by online gas chromatograph (GC-Analyzer), with a measurement error of ≤±1.5%; Flow collection: A Coriolis mass flowmeter was used to monitor the feed / discharge flow rate with an accuracy of 0.2 (allowable error 0.2%).

[0024] Kalman filter execution process State variable definition: Taking tower 1 as an example, the state vector contains [tower top temperature, tower bottom pressure, C18 ester concentration, feed flow rate]; Prediction-correction dual loop: In the prediction phase, the current value is estimated based on the previous state and system dynamics equations (such as the heat conservation model). In the correction phase, the predicted value is corrected using actual sensor data, with the weights dynamically adjusted by the noise covariance matrix (a statistical indicator describing measurement error). Noise suppression example: When a sudden disturbance occurs to the tower top temperature sensor (such as a transient ±3°C fluctuation caused by steam condensation), the filter controls the output fluctuation to within ±0.8°C by analyzing the correlation between pressure and flow data.

[0025] ‌High Confidence Dataset Generation‌ Data fields: timestamp (accurate to milliseconds), tower number, temperature (unit: °C), pressure (unit: kPa), concentration (unit: wt%), flow rate (unit: kg / h); Confidence label: A confidence score (Confidence_Score, range 0-1) is marked for each data point. For example, the confidence score of the concentration value after three repeated measurements on the chromatograph is 0.98; Output format: Time series database storage (such as InfluxDB time series database). A single tower generates 86,400 records per hour (1 Hz × 3,600 seconds).

[0026] Inputting the high-confidence operating condition data set into a dynamic component concentration soft-sensing model based on a long short-term memory network to predict real-time component concentration values to obtain a supplementary data set of component concentrations; Soft sensor model architecture Input layer: Receives time series operating condition data, including 8-dimensional features (Feature_Dimension, i.e. 8 parameter types) such as temperature, pressure, and flow rate; LSTM layer (Long Short-Term Memory Network layer): contains 128 memory units and the time step is set to 60 (i.e., using the past 60 seconds of data for prediction); Output layer: The fully connected network is mapped to the concentration value and outputs the concentrations of 7 fatty acid esters from C16 to C22.

[0027] Real-time prediction mechanism Online prediction trigger: When the gas chromatograph is in the sampling interval (e.g., no measured data in the 3rd minute), the soft measurement model is automatically started; Dynamic compensation logic: If the feed flow rate suddenly changes by more than 10%, the emergency prediction mode is activated (the time step is compressed to 20 seconds); when the tower pressure fluctuation is more than 5kPa, the pressure characteristic weight coefficient is increased to 1.3 times; Forecast example: When the feed flow rate of Tower 2 increases from 5000 kg / h to 5500 kg / h, the predicted C20 ester concentration decreases from 85.2% to 83.7%, which deviates by 0.2% from the actual chromatographic detection value of 83.9%.

[0028] ‌Supplementary dataset construction‌ Data fusion rules: When gas chromatography data are available, the measured values (confidence score = 1.0) are used directly; when there are no measured data, the soft measurement predicted values (confidence score = 0.92, verified by historical data) are used. Abnormal handling: When the predicted value and the previous data suddenly deviate by more than 8%, the manual review flag is triggered; Output increment: Increase the concentration sampling frequency from 5 minutes / time to 1 second / time.

[0029] Based on the high-confidence operating condition data set and the supplementary data set, a material-energy coupling relationship between towers is constructed to obtain a dynamic system topology diagram representing the state association of each tower; ‌Coupling Relationship Modeling‌ Material flow coupling: the bottom discharge of the front tower = the feed of the rear tower (for example, the C18 enriched liquid at the bottom of Tower 1 is used as the feed of Tower 2); establish flow balance: the feed flow fluctuation threshold of Tower 2 is set to ±3% (exceeding it will trigger an alarm) Energy flow coupling: The steam from the top of Tower 1 (temperature 182°C) preheats the feed to Tower 3 through a heat exchanger; real-time calculation of energy transfer efficiency: actual heat transfer / theoretical maximum heat transfer.

[0030] Dynamic topology construction Node definition: Each distillation tower is abstracted as a topological node (Node), which contains a set of state attributes [temperature, pressure, concentration, flow rate]; Edge definition: Material edge: The arrow direction points from the upstream tower to the downstream tower, and the weight = the material flow rate (unit: kg / h); Energy edge: The arrow direction points from the energy supply tower to the energy receiving tower, and the weight = the heat transfer power (unit: kW); Real-time update mechanism: Whenever a key parameter of a tower changes by more than a set threshold (such as temperature ±2°C), topology reconstruction is triggered.

[0031] ‌Visual Graph Generation‌ Graphic elements: The diameter of the circular nodes increases with the tower throughput (diameter range is 20-50 pixels); the material edge is displayed as a solid blue arrow (line width is proportional to the flow rate); the energy edge is displayed as a dashed red arrow (line width is proportional to the heat transfer power); Auxiliary analysis: Clicking a node displays real-time operating parameters (e.g., current pressure of tower 3 = 85.3 kPa); hovering over an edge displays coupling efficiency (e.g., energy coupling efficiency of towers 1→3 = 78.5%).

[0032] According to the dynamic system topology diagram, multi-tower collaborative state embedding driven by graph neural network is performed to obtain a 128-dimensional real-time operation state vector containing system-level spatiotemporal features.

[0033] ‌Graph Neural Network Configuration‌ Model selection: GraphSAGE algorithm (graph sampling and aggregation algorithm) is used, and the number of neighbor samples = 8; Aggregation function (Aggregation_Function): Mean aggregation (Mean_Aggregation), which averages the features of neighboring nodes; Hidden layer design: 3-layer neural network, with the number of neurons in each layer being [256, 192, 128] (gradual dimensionality reduction).

[0034] Feature Embedding Process Node initialization: Normalize the 10-dimensional original features of a single tower (temperature × 3 points, pressure × 2 points, concentration × 4 components, flow rate) to the range [0, 1]; Spatial feature extraction: First-level aggregation: integrating the states of directly adjacent towers (e.g., Tower 2 aggregates the features of Towers 1 and 3); Second-level aggregation: extending to indirectly adjacent towers (e.g., Tower 2 further aggregates the features of adjacent towers of Tower 1); Temporal feature fusion: The current topology map and the historical map of the past 60 seconds are input into the Conv-LSTM (Convolutional Long Short-Term Memory Module) to capture dynamic evolution.

[0035] ‌128-dimensional vector generation‌ Feature compression: compress the 512-dimensional intermediate features to 128 dimensions through the fully connected layer (Full_Connected_Layer); Key feature explanation: Vector dimensions 1-32: characterize the material balance level (e.g., a value > 0.7 indicates good flow matching); vector dimensions 33-64: characterize energy utilization efficiency (e.g., a value < 0.3 triggers waste heat recovery); vector dimensions 65-96: characterize separation purity trends (e.g., a continuous decrease in the value indicates product quality warning); vector dimensions 97-128: characterize system stability (e.g., a value fluctuation > 0.1 / minute triggers stabilization control); Output application: The vector is updated once per second for real-time call by the optimization module.

[0036] This step utilizes a sensor network deployed at key nodes of the distillation tower to acquire core process parameters such as temperature and pressure in real time. Artificial intelligence (AI) technology is then used to fuse and analyze multi-source heterogeneous data. Using time series models such as long-short-term memory networks, discrete sensor data is converted into high-dimensional state vectors with temporal and spatial correlations, enabling digital mapping of the operating status of the entire distillation system. This establishes a precise system state representation model, providing a data foundation for subsequent optimization decisions. Using AI to process complex operating data overcomes the limitations of traditional approaches to modeling nonlinear dynamic processes, significantly improving the real-time and accuracy of condition monitoring.

[0037] S202, performing multi-objective collaborative optimization processing based on the real-time operation state vector and a reinforcement learning algorithm to obtain a global optimization instruction set including the optimal set temperature, set pressure, reflux ratio, and inter-tower energy flow instructions for each tower, wherein the optimization objectives include maximizing separation purity, minimizing energy consumption, and maximizing system stability margin; Specifically, a multi-objective reward function can be constructed based on the real-time operation state vector and a deep deterministic policy gradient algorithm to obtain a purity-energy consumption-stability ternary reward function; The Deep Deterministic Policy Gradient algorithm consists of four core components: Actor Network (decision generator): takes as input a 128-dimensional real-time operational state vector and outputs preliminary optimization suggestions for each tower's operational parameters. Critic_Network (Effect Evaluator): Evaluates the expected return of the actor network output solution; ‌Target_Network(Stable Trainer)‌: Copies the actor and critic network parameters for computing stable target values; Experience Replay Buffer (data storage): stores 100,000 historical operation records for model learning.

[0038] ‌Ternary reward function construction process:‌ Purity Reward Term: Calculation basis: Deviation between the measured value of the online chromatograph and the target purity (e.g. C18 ester ≥ 99.2%); Reward rules: When the deviation is ≤0.5%, the reward is +2 points; when the deviation is >1%, the penalty is -1 point; Dynamic weighting: When the system load is greater than 90%, the purity weight increases by 50%.

[0039] Energy Reward Term: Calculation basis: ratio of steam consumption (unit: tons / hour) to the benchmark value (e.g. 5.8 tons / hour); Reward rules: 5% energy saving reward +1 point; 10% over-consumption penalty -2 points; Itemized measurement: reboiler heating energy consumption accounts for 60%, and condenser cooling energy consumption accounts for 40%.

[0040] Stability Reward Term: Calculation basis: fluctuation amplitude of key parameters (pressure, liquid level) (unit: fluctuation standard deviation); Reward rules: If the fluctuation is less than 0.5% for 30 consecutive minutes, +1 point will be awarded; if the fluctuation is greater than 3%, a penalty of -3 points will be imposed; Margin protection: When the pressure approaches the safety threshold (such as 85% of the design value), the stability weight is automatically increased to 200%.

[0041] Function fusion mechanism: Normalization: Purity (0-100%), energy consumption (tons / hour), and stability (0-1 fluctuation coefficient) are uniformly mapped to the [0,10] score range; Weight distribution: basic weight ratio = purity: energy consumption: stability = 4:3:3, automatically adjusted to 5:2:3 when a sudden change in raw material composition is detected; Final output example: Ternary reward value = 0.4 × purity score + 0.3 × energy consumption score + 0.3 × stability score. For example, under certain working conditions, the calculated total score is: 7.2 (purity) + 6.5 (energy consumption) + 8.1 (stability) = 7.26.

[0042] Based on the reward function, offline pre-training of the policy network and online real-time policy fine-tuning are performed to obtain a dynamic policy parameter set for multi-objective optimization; Offline pre-training phase: ‌Training data preparation‌: Historical database: retrieves operation records from the past three years, containing 2 million state-operation-result triplets; data enhancement: performs 10x oversampling of key operating conditions (such as sudden changes in feed composition); label generation: an expert system annotates each record with a ternary reward value.

[0043] ‌Network training configuration‌: Actor network structure: 8-layer fully connected neural network (number of neurons: 128 → 256 → 256 → 128 → 64 → 32 → 16 → 8); training parameters: learning rate = 0.001, batch size (Batch_Size) = 512; stopping condition: reward value increase < 0.1% for 10 consecutive training cycles (epochs).

[0044] Pre-training results: Basic Strategy Library: Generates 10 optimal parameter sets for typical operating conditions (such as 85% load standard operation). Performance Verification: In simulator tests, it achieved 12.7% energy savings and a 0.3% purity improvement over traditional PID control.

[0045] ‌Online real-time fine-tuning:‌ ‌Fine-tuning trigger conditions‌: Environmental mutation: feed flow rate change rate > 5% / minute; Performance degradation: the actual reward value is 15% lower than the expected value for 5 consecutive minutes.

[0046] ‌Incremental learning mechanism‌: Real-time data streaming: A new set of state vectors is collected every 10 seconds and stored in the replay pool. Lightweight training: Only the weights of the last two layers of the network are updated, and the training time is controlled within 300 milliseconds. Safety constraints: Parameter adjustment is limited to ±5% to prevent violent oscillations.

[0047] Dynamic parameter set output: Data structure: JSON format stores key parameters of each tower, for example: { "Tower 1": {"Temperature setting value": 185.2, "Pressure setting value": 83.4, "Reflux ratio": 4.2}, "Tower 2": {"Temperature Setpoint": 168.7, "Pressure Setpoint": 77.1, "Reflux Ratio": 5.8}}.

[0048] Update frequency: A new parameter set is generated every minute, and the version number is recorded incrementally.

[0049] According to the dynamic strategy parameter set, a distributed parallel strategy is executed to obtain a local optimization instruction set of independent operating parameters of each tower; ‌Distributed parallel architecture:‌ Computing node allocation: Each distillation tower is equipped with an independent edge computing unit. Communication protocol: OPC UA (industrial communication standard) is used to achieve millisecond-level data interaction. Load balancing: The main controller dynamically allocates computing tasks.

[0050] ‌Local optimization process‌: Parameter distribution: The main controller broadcasts the dynamic strategy parameter set to each tower edge unit; Local optimization: Each edge unit performs secondary optimization based on the real-time status of the tower, including: temperature optimization: fine-tuning within the range of ±1°C using fuzzy control; reflux ratio optimization: dynamic adjustment based on the concentration gradient change rate (adjustment step size 0.1 / time); Constraint processing: When the optimization result of the edge unit exceeds the safety range (such as pressure > 90% of the design value), it automatically falls back to the main control parameters.

[0051] ‌Instruction Set Encapsulation‌: Data fields: timestamp (accurate to milliseconds); tower number (1-N); operating parameters (temperature, pressure, reflux ratio); confidence level (the degree of match between local optimization results and the master control strategy).

[0052] Example output: Tower 3 local instructions: Temperature = 172.4°C (confidence 0.93); Pressure = 80.2 kPa (confidence 0.88); Reflux Ratio = 6.3 (confidence 0.95).

[0053] Performing cross-tower energy flow game theory optimization based on the local optimization instruction set and the real-time operation state vector to obtain an inter-tower energy flow instruction matrix that minimizes global entropy increase; Game Theory Optimization Model: ‌Participant Definition‌: Energy supply tower: a distillation tower with a top steam temperature greater than 150°C (such as Tower 1); energy demand tower: a distillation tower with a reboiler heat load greater than 200kW (such as Tower 3).

[0054] Revenue function design: Revenue for energy suppliers = revenue from heat sales - losses in the pipeline network. For example, the unit price of heat sales is graded by steam quality: 180°C steam = 200 yuan / ton, 150°C steam = 150 yuan / ton. Revenue for energy consumers = energy savings × electricity price - depreciation of heat exchange equipment.

[0055] ‌Nash equilibrium solution: Iterative process: Each tower edge unit exchanges quotes and demands; Convergence condition: The change in quotes for three consecutive rounds is less than 1%; Entropy increase constraint: The total entropy increase of the network is required to be less than 5kW / K (a measure of energy quality loss).

[0056] ‌Instruction Matrix Generation:‌ Matrix structure design: Row: Energy supply tower number (such as T1, T4); Column: Energy demand tower number (such as T2, T3); Element value: Steam flow rate (unit: kg / h).

[0057] Dynamic Adjustment Mechanism: When the feed to Tower 3 increases by 20%, the T1 supply rate is automatically increased from 850kg / h to 1050kg / h. When the pipeline pressure fluctuation is greater than 10kPa, the flow smoothing strategy is activated (maximum adjustment rate 50kg / h / min).

[0058] According to the local optimization instruction set and the inter-tower energy flow instruction matrix, instruction conflicts are resolved and collaboratively packaged to obtain a global optimization instruction set including temperature, pressure, reflux ratio setting values and energy flow instructions.

[0059] Conflict detection and resolution: ‌Conflict type identification‌: Resource conflicts: For example, Tower 1 needs to reduce the reflux ratio to save energy, but the energy flow instruction requires it to increase steam production; Timing conflicts: Pressure regulation requires 5 minutes to stabilize, but the energy flow switching must be completed within 2 minutes.

[0060] ‌Resolution strategy library‌: Priority rule: purity target > energy consumption target > stability target; Dynamic compromise algorithm: When purity and energy consumption conflict, a 1% increase in energy consumption is allowed in exchange for a 0.2% increase in purity; a virtual buffer tower is introduced to temporarily store excess energy.

[0061] Typical digestion cases: Conflict scenario: Tower 3 needs to be heated by 5°C (purity requirement), but the energy flow matrix shows insufficient waste heat; Solution: 1) Temporarily activate the backup gas boiler to add 200kW of heat (cost: 3% increase in energy consumption); 2) Reduce the reflux ratio of Tower 2 from 5.0 to 4.7 to save steam for Tower 3; 3) After 15 minutes, wait for the steam in Tower T1 to recover and shut down the backup boiler.

[0062] ‌Co-packaging:‌ ‌Command Hierarchy‌: Level 1: Tower parameters (temperature / pressure / reflux ratio); Level 2: Energy flow instructions (source tower → target tower, flow rate); Level 3: Timing constraints (such as "T1→T3 steam delivery must be completed before 15:00").

[0063] Communication protocol encapsulation: OPC UA XML format (industry standard data format); Data packet segmentation check: Each frame contains a CRC-32 cyclic redundancy check code (data transmission check code); Transmission cycle: Broadcast updates every 500 milliseconds.

[0064] ‌Global instruction set example‌: { "timestamp": "2023-08-15 14:25:30.450", "instructions": [ { "tower": "T1", "params": {"temp": 185.2, "pressure": 83.4, "reflux_ratio": 4.2}, "energy_out": [{"target": "T3", "flow": 850}] }, { "tower": "T3", "params": {"temp": 172.4, "pressure": 80.2, "reflux_ratio": 6.3}, "energy_in": [{"source": "T1", "flow": 850}] } ] }.

[0065] This step employs a deep reinforcement learning framework to transform the multi-objective optimization problem of the distillation process into a Markov decision process. By designing a composite reward function encompassing purity, energy consumption, and stability, the intelligent agent is trained to automatically generate optimal operating parameters under complex constraints, achieving coordinated optimization of process conditions across the towers. This resolves the conflicting process parameters associated with traditional single-objective optimization, significantly reducing energy consumption while ensuring product purity. The self-learning nature of reinforcement learning enables the system to continuously adapt to uncertainties such as raw material changes and equipment aging.

[0066] S203, based on the inter-tower energy flow instructions, dynamically reconfigure the topology of the heat recovery network, and preferentially direct high-grade tower overhead steam waste heat to low-grade heat demand points through intelligent valve switching, thereby obtaining a real-time energy flow map that achieves cross-tower energy cascade matching; Specifically, the waste heat grade classification and heat demand priority ranking can be performed according to the inter-tower energy flow instruction to obtain a heat source-heat sink matching priority list; ‌Waste heat grade classification rules:‌ Temperature range division: High-grade heat source: tower top steam temperature ≥180℃ (degrees Celsius); medium-grade heat source: 150℃ ≤ temperature <180℃; low-grade heat source: temperature <150℃.

[0067] Example: Tower 1 top steam temperature 182°C → high-grade heat source Heat source quality coefficient calculation: Quality factor = (actual temperature - 150°C) / 30°C × 100%. Additional correction: If the proportion of non-condensable gases is greater than 5%, the coefficient is reduced by 20%. Example: Steam temperature in Tower 2 is 165°C → Quality factor = (165-150) / 30 × 100% = 50%.

[0068] ‌Hot demand prioritization rules:‌ Urgency Level: Level 1 Demand: The reboiler heat medium temperature is less than the lower limit of the process requirement (for example, Tower 3 requires 160°C but is currently only 155°C). Level 2 Demand: The heat medium temperature is within the process requirement range but energy conservation is required. Level 3 Demand: Non-core heating points (such as pipeline heating).

[0069] Priority Weight Calculation: Priority score = Basic weight (Level 1 = 100, Level 2 = 60, Level 3 = 20) + Energy Saving Gain Factor × 10. Energy Saving Gain Factor: The factor increases by 0.1 for every 1 MW (megawatt) of waste heat recovered. Example: The reboiler in Tower 4 has Level 1 demand and can save 1.5 MW → Priority score = 100 + 1.5 × 10 = 115 points.

[0070] Matching priority list generation: Matrix matching table, an example is shown in Table 1.

[0071] Table 1

[0072] Dynamic update mechanism: rescan the status of each tower every 30 seconds; immediately update the ranking when the feed rate of a tower suddenly changes by more than 10%; prioritize towers with a difference of less than 5 and rank them at the same level.

[0073] According to the matching priority list, dynamic path planning of the thermal network based on directed graph theory is performed to obtain an optimal heat transfer path sequence with a minimum heat transfer temperature difference; ‌Directed graph model construction (Directed_Graph_Model):‌ Node Definition: Heat source node (Source_Node): a red circle with a diameter proportional to the steam flow rate (e.g. 1 kg / s corresponds to 10 mm); heat sink node (Sink_Node): a blue square with a side length proportional to the heat demand (e.g. 1 MW corresponds to 15 mm); heat exchange station node (Exchange_Node): a yellow diamond, used for multi-stage heat exchange.

[0074] Edge Definition: A directed edge arrow points from a heat source to a heat sink. Edge Weight = Pipeline Pressure Drop (kPa) + Heat Transfer Temperature Loss (°C) × Conversion Factor. Note: The conversion factor is 0.5 kPa / °C, which converts the temperature loss into an equivalent pressure drop.

[0075] ‌Path planning algorithm execution:‌ ‌Improved Dijkstra algorithm (improved shortest path algorithm)‌: Core goal: Find the path with the minimum weight; Constraints: Heat transfer temperature difference ≥ 20°C (to avoid condensation); Pipeline flow rate ≤ 30 m / s (to prevent erosion). Example: Calculate the path weights from tower T1 to tower T3. Path 1: T1 → Heat Exchange Station A → T3. Weight = 12 kPa + (25°C × 0.5) = 24.5. Path 2: Direct connection from T1 to T3. Weight = 18 kPa + (18°C × 0.5) = 27. Select Path 1.

[0076] Sequence generation rules: Step 1: Select the heat source-heat sink pair with the highest priority. Step 2: Generate its optimal path and lock the path resources. Step 3: Continue to select the next highest priority pair from the remaining resources.

[0077] Example of the final output path sequence: ①T1→T3 (flow rate 850kg / h)→②T2→T4 (flow rate 430kg / h).

[0078] generating a valve switch instruction matrix for controlling steam flow according to the optimal heat transfer path sequence; ‌Valve command generation logic:‌ ‌Valve Status Definition‌: Opening degree 100%: steam is fully open (red indicator light is always on); opening degree 0%: completely closed (green indicator light is always on); opening degree 30%-70%: adjustment state (yellow indicator light flashes).

[0079] The instruction matrix structure includes information examples as shown in Table 2.

[0080] Table 2

[0081] Security Control Strategy: Progressive adjustment: For large-diameter valves (DN300 and above), the opening change rate is ≤5% / second; for small-diameter valves (DN150 and below), the opening change rate is ≤10% / second. Example: A V-101 valve requires at least 100% / 5% = 20 seconds to fully open from closed.

[0082] Multiple interlock protection: Pressure interlock: When the target tower pressure is greater than 90% of the design value, the opening is stopped; Temperature interlock: When the pipe wall temperature is less than the dew point temperature + 5℃, the water is automatically drained to prevent condensation; Flow balance interlock: When the inlet and outlet flow difference is continuously greater than 10%, the alarm is triggered.

[0083] ‌Command issuance mechanism:‌ Timestamp synchronization, including: control cycle: refresh instructions every 500 milliseconds; timing requirements: the interval between adjacent valve actions is ≥ 2 seconds to prevent hydraulic shock; abnormal handling plan, including: valve response timeout: if the target opening is not reached within 5 seconds, start the backup valve; communication interruption: automatically maintain the current opening and switch to local manual mode.

[0084] According to the actual heat transfer data after executing the valve switching instruction matrix, real-time evaluation of thermodynamic efficiency and graph visualization processing are performed to obtain a real-time energy flow graph with efficiency labels.

[0085] Thermodynamic Efficiency Evaluation ‌Core indicator calculation‌: Waste heat recovery rate: actual recovered heat / theoretical maximum recoverable heat × 100%. Example: 1.8MW of steam can be recovered from the top of tower T1, but 1.5MW is actually utilized → recovery rate 83.3%; Heat transfer end difference: The temperature difference between the heat source outlet and the heat sink inlet (target value 20-50℃). Example: T1 steam outlet 162℃ vs T3 reboiler inlet 140℃ → end difference 22℃.

[0086] ‌Efficiency Label Generation‌: Green label: waste heat recovery rate ≥80% and end difference ≤30℃; yellow label: recovery rate 60%-80% or end difference 30-45℃; red label: recovery rate <60% or end difference >45℃.

[0087] ‌Graph visualization processing (Visualization_Processing):‌ Dynamic Topology Rendering: Pipe coloring rules: Red: steam temperature > 160°C; Orange: 120-160°C; Blue: < 120°C; Flow animation: The thickness of the arrow is proportional to the flow rate (1m / s corresponds to 1px thickness).

[0088] ‌Efficiency Label Implantation‌: Location: A circular label is displayed in the middle of the heat exchange pipeline; Content: Recovery rate: 83%, End difference: 22°C, Status: Green.

[0089] ‌Real-time refresh mechanism‌: Data source: Temperature / pressure transmitters upload data every second; Rendering engine: WebGL (Web Graphics Library) drives 3D topology maps; Refresh rate: 25 frames / second (screen updates every 40 milliseconds).

[0090] ‌Graph interaction function:‌ ‌Details Drilldown‌: Click on a label to pop up the historical efficiency curve (including a 30-day trend chart); ‌Warning Tip‌: For example, when a path displays red labels three times in a row, an optimization suggestion will automatically pop up: "It is recommended to clean the T1-T3 heat exchanger tube bundle (fouling coefficient > 0.0002)."

[0091] This step uses graph theory algorithms to analyze the heat source quality and hotspot demand of each tower in real time, dynamically planning the optimal heat integration path. Through intelligent valve configuration adjustments, an energy recovery network adapts to changing operating conditions, achieving precise matching and cascaded utilization of waste heat resources. This overcomes the energy efficiency bottleneck of fixed heat exchange networks and improves the average system heat recovery rate. Dynamic topology reconfiguration enables the energy network to self-organize, effectively responding to production load fluctuations.

[0092] S204, based on the global optimization instruction set and the real-time energy flow map, coordinately control the actuators of each column, including the heater, the reflux controller, and the pressure regulating valve, to obtain a stable distillation operation state that matches the optimization target and the energy flow map; Specifically, a feedforward-feedback composite adjustment based on model predictive control can be performed according to the tower parameters in the global optimization instruction set to obtain a reference control signal for each tower actuator; Model Predictive Control (MPC controller) execution process: Dynamic Modeling: Establish a three-input and three-output model for each tower: Input variables: reboiler steam valve opening (unit %), reflux valve opening (unit %), tower pressure regulating valve opening (unit %); Output variables: tower bottom temperature (unit: °C), tower top product concentration (unit: wt%), tower pressure (unit: kPa); Model parameter identification: Obtain the response curve through step test (Step_Test). For example, a 10% increase in the steam valve opening causes the tower temperature to rise by 8.5°C (delay of 30 seconds).

[0093] Feedforward Control For measurable disturbances: When the feed flow rate suddenly increases by 15%, the reboiler steam valve opening is increased by 8% 5 seconds in advance; Feedforward coefficient table: generated based on historical data, for example, for every 1 ton / hour increase in feed flow, the steam valve needs to be opened 0.6%.

[0094] Feedback Control Rolling optimization: Calculate the optimal operation sequence for the next 3 minutes every 10 seconds; Deviation correction: When the measured value of the tower top concentration is 0.3% lower than the set value, the reflux ratio is automatically increased by 0.2.

[0095] Output reference signal example: Tower 1 benchmark control signal: reboiler steam valve opening = 62.3%; reflux valve opening = 55.7%; pressure regulating valve opening = 38.4%.

[0096] Performing dynamic compensation calculation of reboiler heating power based on cross-tower heat transfer efficiency data in the real-time energy flow map to obtain a thermal coupling compensation coefficient; Thermal coupling compensation mechanism: Efficiency data analysis: Extract label data from the energy flow graph: Waste Heat Recovery Ratio = 83.5% (green label); Approach Temperature Difference = 22°C (temperature difference).

[0097] Energy flow topology: Confirm that the reboiler of Tower 3 receives 850 kg / h of steam from Tower 1 ‌Compensation calculation logic‌: Compensation coefficient (Compensation_Coefficient) = Base heating power × (1 - waste heat contribution rate); Waste heat contribution ratio (Waste_Heat_Contribution_Ratio) = actual recovered heat / reboiler theoretical heat requirement.

[0098] Dynamic Correction: When the waste heat recovery rate exceeds 80%, the compensation coefficient decreases by 0.15. When the heat transfer temperature difference exceeds 30°C, the compensation coefficient increases by 0.1 (to compensate for heat loss). Example: The theoretical heat demand of the reboiler in Tower 3 is 1.8MW, and Tower 1 supplies 1.5MW → Waste heat contribution rate = 83.3% → Compensation coefficient = 1 - 0.833 = 0.167.

[0099] ‌Real-time update rules‌: Scanning cycle: the compensation coefficient is updated every 20 seconds; abnormal handling: when the steam flow fluctuation is greater than 10%, the coefficient value is frozen and recalculated after stabilization.

[0100] According to the reference control signal and the thermal coupling compensation coefficient, the control amount of each tower actuator is collaboratively corrected to obtain an optimized control amount set under the anti-overshoot constraint; Collaborative Revision Strategy: Correction algorithm execution: Final reboiler steam valve opening = reference opening × (1 + thermal coupling compensation coefficient). Example: Reference opening 62.3% × (1 + 0.167) = 72.6%. Reflux valve correction: When waste heat utilization is greater than 70%, the reflux ratio is reduced by 0.3 (energy saving effect).

[0101] ‌Anti-overshoot_Constraint‌: Single adjustment limit: valve opening change ≤5% / time; temperature set point change ≤2℃ / minute; cumulative change constraint: the total adjustment within 10 consecutive minutes shall not exceed 15% of the design range; case: the steam valve of Tower 3 needs to be increased from 50% to 72.6%, which is adjusted in 5 times (+4.52% each time).

[0102] ‌Optimize control volume set generation‌: Data structure example: { "Tower Number": "T3", "Reboiler steam valve": 72.6, / / Unit: % "Return valve": 52.1, / / Unit: % "Pressure valve": 41.3, / / Unit: % "ConstraintState": "Active"}.

[0103] Validity verification: Check whether the valve position exceeds the limit (if > 95%, an error will be reported).

[0104] Perform multivariable decoupling control and actuator response lag compensation according to the optimized control quantity set to obtain synchronized execution instructions; ‌Multivariable Decoupling Control (Multivariable_Decoupling_Control):‌ Elimination of coupling relationship: Temperature-pressure decoupling: When the pressure regulating valve is actuated, the reboiler steam volume is automatically compensated (compensation coefficient 0.7); Example: the pressure valve opening increases by 10% → the steam valve opening is temporarily reduced by 7%; Reflux ratio-concentration decoupling: When the top concentration drops by 1%, the reflux ratio increase is limited to within 0.8 (to prevent flooding).

[0105] ‌Response_Lag_Compensation‌: Actuator Response Time Library: Electric control valve (Electric_Control_Valve): 500 milliseconds to reach 90% opening; pneumatic diaphragm valve (Pneumatic_Diaphragm_Valve): 2 seconds to reach 90% opening.

[0106] Calculation of command lead time: Commands for slow valves are sent in advance. Lead time = target adjustment / valve rate. Example: A pneumatic valve needs to open 10% (rate 5% / second) → send the command 2 seconds in advance.

[0107] Synchronize command generation: Time alignment: All instructions are bound to the same timestamp (error < 50 milliseconds) Timing arrangement example: 14:25:30.000: Open the T1→T3 steam main shut-off valve; 14:25:30.500: Adjust the T3 reboiler steam valve to 72.6%; 14:25:31.000: Adjust the T3 reflux valve to 52.1%.

[0108] Hardware synchronization: Use the PTP precision time protocol (Precision_Time_Protocol) to calibrate the controller clock.

[0109] After executing the synchronized execution instruction to drive each actuator, the operation state stability index processing is verified in real time to obtain a stable distillation operation state that meets the Lyapunov stability condition.

[0110] Stability Verification System: ‌Lyapunov function construction (Lyapunov function)‌: State variables selected: tower bottom temperature deviation ΔT, tower pressure deviation ΔP, concentration deviation ΔC; Energy function definition: V = (ΔT)² + 0.8×(ΔP)² + 1.2×(ΔC)²; Stability condition: dV / dt<0 (the function value decreases with time).

[0111] ‌Real-time monitoring mechanism‌: Data acquisition: Calculate 10 sets of state variables per second; Trend analysis: dV / dt>0 for 5 consecutive sampling points → yellow warning; dV / dt>0 for 10 consecutive sampling points → red alarm; Example: ΔT = 1.2°C, ΔP = 0.8 kPa, ΔC = 0.3% → V = 1.44 + 0.51 + 0.43 = 2.38 → At the next moment, V = 2.15 (satisfying dV / dt = -0.23 < 0).

[0112] Stable state determination: Level 1 stability: All parameters remain within the range of ±1% of the set value for 10 minutes; Level 2 stability: The energy function V value decreases continuously and the fluctuation is less than 0.5 within 30 minutes.

[0113] Output flags: Stable distillation operation state: stability index = 0.92 (range 0-1); continuous stability time = 42 minutes; Lyapunov condition: satisfied.

[0114] Abnormal recovery logic: When a yellow warning is triggered, the over-adjustment prevention constraint is automatically tightened (the adjustment range is limited to 2%); when a red alarm is triggered, the preset safety parameters are activated (such as forcibly increasing the reflux ratio by 0.5).

[0115] This step employs a strategy combining model predictive control with decoupling control to translate optimization instructions into coordinated actions of the actuators. A feedforward-feedback composite control algorithm compensates for system inertia and coupling effects, ensuring that process parameters quickly and stably track setpoints. This enables precise control of multivariable, tightly coupled systems and minimizes fluctuations in key parameters. This coordinated control mechanism effectively avoids overall performance degradation caused by local optimization.

[0116] S205 , based on the abnormal disturbance characteristics continuously monitored under the stable distillation operation state, adaptive parameter fine-tuning and emergency strategy triggering processing based on the pre-trained fault mode library are performed to obtain a system self-healing state that maintains continuous and stable output of high-purity fatty acid esters.

[0117] Specifically, the abnormal disturbance fingerprint vector can be obtained by extracting disturbance features based on wavelet packet transform according to the real-time monitoring data under the stable distillation operation state; Wavelet Packet Transform (Wavelet_Packet_Transform, multi-resolution signal analysis technology) execution process: Signal acquisition and preprocessing: Monitoring parameters: tower bottom temperature (sampling rate 10Hz), tower pressure (sampling rate 5Hz), reflux liquid flow rate (sampling rate 1Hz); data purification: use sliding window filtering (window width = 5 seconds) to eliminate measurement spikes; example: detect a momentary temperature jump of 3°C (duration <0.5 seconds) → determine it as instrument interference → automatically eliminate it.

[0118] Multi-scale eigendecomposition: Decomposition level: 4-layer wavelet packet decomposition; frequency band division: divide the 0-5 Hz signal into 16 frequency bands (each bandwidth is 0.3125 Hz).

[0119] Feature extraction: Energy entropy: Calculates the energy proportion of each frequency band signal; Singular value: Extracts the intensity of the main fluctuation mode; Case: Feed pump vibration fault characteristics are concentrated in the 1.25-1.56Hz frequency band (energy proportion >35%).

[0120] Fingerprint vector generation: Vector dimension: contains 24 eigenvalues (8 features for each parameter); normalization processing: maps temperature fluctuation amplitude, pressure change rate, etc. to the [0,1] interval.

[0121] Typical fingerprint examples: Abnormal perturbation fingerprint vector = [ Temperature band 3 energy: 0.83, temperature singular value: 0.62, / / temperature-related features Pressure Band 5 Energy: 0.91, Pressure Gradient: 0.78, / / Pressure-related characteristics Traffic mutation count: 0.47, traffic fluctuation entropy: 0.56 / / Traffic-related features ].

[0122] Based on the abnormal disturbance fingerprint vector, similarity matching is performed with the pre-trained fault pattern library to obtain the fault type identification and confidence level; ‌Pre-trained failure pattern library construction:‌ ‌Fault knowledge base architecture‌: Fault type: 12 typical faults (such as reboiler fouling, reflux pump surge, and feed composition mutation).

[0123] Data source: Historical case: 300 fault data accumulated from 5 years of maintenance records; simulation experiment: 20 controllable faults were manually triggered and data collected; example: reboiler fouling fault fingerprint characteristics: temperature band 2 energy > 0.8, pressure gradient > 0.7.

[0124] ‌Similarity matching algorithm‌: Core algorithm: Improved cosine similarity Calculation process: Multiply the current fingerprint vector by each fault mode vector in the library; introduce weight coefficients: for example, the temperature feature weight is 0.4, the pressure is 0.3, and the flow rate is 0.3; the similarity score = weighted dot product value / (current vector modulus length × fault vector modulus length); example: the similarity between the current vector and the "reboiler fouling" mode is 0.86 (the full score is 1.0).

[0125] Fault indicator output: Identification rules: Similarity ≥ 0.9 → high confidence match (red alert); 0.7 ≤ similarity < 0.9 → medium confidence (yellow warning); similarity < 0.7 → unknown fault (blue prompt).

[0126] Output format example: { "Fault Type": "Reboiler Fouling", "Confidence": 86%, "Feature Matching Point": ["Temperature Band 2 Energy", "Pressure Gradient"]}.

[0127] According to the fault type identification and confidence level, an emergency strategy retrieval based on case reasoning is performed to obtain an emergency operation instruction set including a parameter adjustment range and a valve safety interlock; Case-Based Reasoning (experience-based decision-making technology) execution mechanism: ‌Case Library Search Rules‌: First-level screening: matching the same fault type (e.g., all "reboiler fouling" cases); second-level sorting: sorting by descending operating condition similarity (current load rate ±5%); example: selecting 5 cases that meet the current 85% load from 38 fouling cases.

[0128] Strategy extraction and fusion: Parameter adjustment range generation: Take the median adjustment value of historical successful cases ± 20% as the safety margin; example: In historical cases, increasing the steam valve opening by 8-12% is effective → the current recommended adjustment is 9-11%.

[0129] Valve safety interlock setting: Mandatory constraint: It is prohibited to open the steam valve when the pressure is greater than 85% of the design value; Action sequence: Step 1: Increase the reboiler steam valve opening by 10%; Step 2: Check the temperature rise rate after 30 seconds (expected to be >0.5℃ / minute); Step 3: If ineffective, start the backup heater.

[0130] ‌Instruction set package structure‌: { "fault handling strategy": { "Parameter adjustment": [ {"Equipment": "T3 Reboiler Steam Valve", "Operation": "Opening +10%", "Range": "8-12%"}, {"Equipment": "T3 Reflux Valve", "Operation": "Opening -3%", "Range": "-2~-4%"} ], "Safety Interlock": [ {"Trigger Condition": "T3 Pressure>830kPa", "Action": "Steam Valve Locks Current Opening"}, {"Trigger condition": "Bottom liquid level < 15%", "Action": "Close the discharge valve"} ] } }.

[0131] According to the emergency operation instruction set, parameter gradual adjustment and system status monitoring are performed under safety constraints, and a system self-healing state is obtained in which the system automatically recovers to an optimized state after the disturbance is eliminated.

[0132] Safety-constrained_Adjustment mechanism: ‌Gradual Adjustment Rules‌: Adjust the rate control: Temperature setting value: Maximum change rate ±2°C / minute; Valve opening: Maximum change rate ±5% / time (interval ≥30 seconds); Example: Steam valve needs to increase by 10% → adjust in 3 steps (+3.5% → +3.5% → +3.0%).

[0133] Monitoring indicator system: Key recovery indicators: Concentration deviation recovery degree = |current value - set value| / allowable deviation; Energy function convergence: dV / dt < 0 for five consecutive calculations; Example: Concentration deviation drops from 1.2% to 0.3% (allowable deviation 0.5%) → Recovery degree = (1.2-0.3) / 0.5 = 180%.

[0134] ‌Self-healing state transition logic:‌ Disturbance elimination judgment: Core conditions: Key parameters remain within ±1% of the set value for three consecutive minutes; wavelet packet characteristic energy entropy returns to ±10% of the baseline level; Example: Temperature fluctuation band energy drops from 0.83 to 0.12 (baseline = 0.15) → determined to be eliminated.

[0135] ‌Optimized state recovery‌: Recovery strategy: Step-by-step withdrawal of emergency adjustments (50% each time); monitor stability for 30 seconds after each withdrawal; completely switch back to the global optimization instruction set; Example: Steam valve opening gradually withdraws from the emergency value of 72%: 72% → 66% → 60% (original optimized value 58%).

[0136] Self-healing status output: System self-healing status report, examples include: Current status: Normal optimization mode; Troubleshooting time: 18 minutes; Concentration stability: 99.2±0.15% (lasting 25 minutes); Energy consumption level: 5.7 tons of steam / hour (better than the benchmark value of 6.0 tons).

[0137] This step uses wavelet analysis to capture abnormal characteristics in real time and, combined with case-based reasoning techniques, quickly matches historical failure patterns. A progressive parameter adjustment strategy, under safety constraints, achieves fault suppression and system recovery without stopping the system. This significantly improves the system's anti-interference capabilities and reduces unplanned downtime. A self-healing mechanism ensures production continuity, avoiding millions of yuan in annual losses due to abnormal operating conditions.

[0138] It can be seen that the temperature, pressure, component concentration and flow rate data of each tower of the multi-stage distillation system are collected in real time to obtain a real-time operation state vector that characterizes the overall operation status of the current system; based on the real-time operation state vector, a global optimization instruction set including the optimal set temperature, set pressure, reflux ratio and inter-tower energy flow instructions of each tower is obtained; based on the inter-tower energy flow instructions, a real-time energy flow spectrum that realizes cross-tower energy ladder matching is obtained; based on the global optimization instruction set and the real-time energy flow spectrum, a stable distillation operation state that matches the optimization target and the energy flow spectrum is obtained; based on the abnormal disturbance characteristics continuously monitored under the stable distillation operation state, a system self-healing state that maintains continuous and stable output of high-purity fatty acid esters is obtained, thereby realizing global coordinated control of multi-stage distillation and improving separation purity and energy efficiency.

[0139] Another embodiment of the present invention provides a multi-stage distillation control system for refining fatty acid esters, see Figure 3 , the system may include: The acquisition module 301 is used to collect temperature, pressure, component concentration, and flow rate data of each tower in the multi-stage distillation system in real time, perform dynamic operation state modeling based on artificial intelligence, and obtain a real-time operation state vector representing the overall operation status of the current system; An optimization module 302 is configured to perform multi-objective collaborative optimization processing based on the real-time operation state vector and a reinforcement learning algorithm to obtain a global optimization instruction set including the optimal set temperature, set pressure, reflux ratio, and inter-tower energy flow instructions for each tower, wherein the optimization objectives include maximizing separation purity, minimizing energy consumption, and maximizing system stability margin; Reconstruction module 303 is used to dynamically reconstruct the topology of the heat recovery network according to the inter-tower energy flow instruction, and preferentially direct the high-grade tower overhead steam waste heat to the low-grade heat demand point through intelligent valve switching, thereby obtaining a real-time energy flow map that achieves cross-tower energy ladder matching; a control module 304 for collaboratively controlling the actuators of each column, including the heater, reflux controller, and pressure regulating valve, based on the global optimization instruction set and the real-time energy flow map, to obtain a stable distillation operation state that matches the optimization target and the energy flow map; The processing module 305 is used to perform adaptive parameter fine-tuning and emergency strategy triggering based on the pre-trained fault mode library according to the abnormal disturbance characteristics continuously monitored under the stable distillation operation state, so as to obtain a system self-healing state that maintains continuous and stable output of high-purity fatty acid esters.

[0140] An embodiment of the present invention further provides a storage medium storing a computer program, wherein the computer program is configured to execute the steps of any one of the above method embodiments when running.

[0141] Specifically, in this embodiment, the above-mentioned storage medium may be configured to store a computer program for performing the following steps: S201, collecting temperature, pressure, component concentration, and flow rate data of each tower of the multi-stage distillation system in real time, performing dynamic operation state modeling based on artificial intelligence, and obtaining a real-time operation state vector representing the overall operation status of the current system; S202, performing multi-objective collaborative optimization processing based on the real-time operation state vector and a reinforcement learning algorithm to obtain a global optimization instruction set including the optimal set temperature, set pressure, reflux ratio, and inter-tower energy flow instructions for each tower, wherein the optimization objectives include maximizing separation purity, minimizing energy consumption, and maximizing system stability margin; S203, based on the inter-tower energy flow instructions, dynamically reconfigure the topology of the heat recovery network, and preferentially direct high-grade tower overhead steam waste heat to low-grade heat demand points through intelligent valve switching, thereby obtaining a real-time energy flow map that achieves cross-tower energy cascade matching; S204, based on the global optimization instruction set and the real-time energy flow map, coordinately control the actuators of each column, including the heater, the reflux controller, and the pressure regulating valve, to obtain a stable distillation operation state that matches the optimization target and the energy flow map; S205 , based on the abnormal disturbance characteristics continuously monitored under the stable distillation operation state, adaptive parameter fine-tuning and emergency strategy triggering processing based on the pre-trained fault mode library are performed to obtain a system self-healing state that maintains continuous and stable output of high-purity fatty acid esters.

[0142] An embodiment of the present invention further provides an electronic device, comprising a memory and a processor, wherein the memory stores a computer program, and the processor is configured to run the computer program to perform the steps in any one of the above method embodiments.

[0143] Specifically, the electronic device may further include a transmission device and an input / output device, wherein the transmission device is connected to the processor, and the input / output device is connected to the processor.

[0144] Specifically, in this embodiment, the processor may be configured to execute the following steps through a computer program: S201, collecting temperature, pressure, component concentration, and flow rate data of each tower of the multi-stage distillation system in real time, performing dynamic operation state modeling based on artificial intelligence, and obtaining a real-time operation state vector representing the overall operation status of the current system; S202, performing multi-objective collaborative optimization processing based on the real-time operation state vector and a reinforcement learning algorithm to obtain a global optimization instruction set including the optimal set temperature, set pressure, reflux ratio, and inter-tower energy flow instructions for each tower, wherein the optimization objectives include maximizing separation purity, minimizing energy consumption, and maximizing system stability margin; S203, based on the inter-tower energy flow instructions, dynamically reconfigure the topology of the heat recovery network, and preferentially direct high-grade tower overhead steam waste heat to low-grade heat demand points through intelligent valve switching, thereby obtaining a real-time energy flow map that achieves cross-tower energy cascade matching; S204, based on the global optimization instruction set and the real-time energy flow map, coordinately control the actuators of each column, including the heater, the reflux controller, and the pressure regulating valve, to obtain a stable distillation operation state that matches the optimization target and the energy flow map; S205 , based on the abnormal disturbance characteristics continuously monitored under the stable distillation operation state, adaptive parameter fine-tuning and emergency strategy triggering processing based on the pre-trained fault mode library are performed to obtain a system self-healing state that maintains continuous and stable output of high-purity fatty acid esters.

[0145] The above describes in detail the structure, features and effects of the present invention based on the embodiments shown in the drawings. The above is only a preferred embodiment of the present invention, but the scope of implementation of the present invention is not limited to what is shown in the drawings. Any changes made in accordance with the concept of the present invention, or modifications to equivalent embodiments with equivalent changes, which do not exceed the spirit covered by the description and drawings, should be within the scope of protection of the present invention.

Claims

1. A multi-stage distillation control method for refining fatty acid esters, characterized in that: The method comprises: Real-time data collection of temperature, pressure, component concentration, and flow rate of each tower in a multi-stage distillation system is used to perform dynamic operation state modeling based on artificial intelligence, thereby obtaining a real-time operation state vector that represents the overall operation status of the current system. Based on the real-time operation state vector, a multi-objective collaborative optimization process is performed based on a reinforcement learning algorithm to obtain a global optimization instruction set including the optimal set temperature, set pressure, reflux ratio, and inter-tower energy flow instructions for each tower, wherein the optimization objectives include maximizing separation purity, minimizing energy consumption, and maximizing system stability margin; According to the inter-tower energy flow instructions, the dynamic topology reconstruction of the heat recovery network is performed, and the high-grade tower top steam waste heat is preferentially directed to the low-grade heat demand point through intelligent valve switching, thereby obtaining a real-time energy flow map that realizes cross-tower energy ladder matching; According to the global optimization instruction set and the real-time energy flow map, each tower actuator including a heater, a reflux controller, and a pressure regulating valve is coordinated and controlled to obtain a stable distillation operation state that matches the optimization target and the energy flow map; According to the abnormal disturbance characteristics continuously monitored under the stable distillation operation state, adaptive parameter fine-tuning and emergency strategy triggering processing are performed based on the pre-trained fault mode library to obtain a system self-healing state that maintains continuous and stable output of high-purity fatty acid esters.

2. The method according to claim 1, characterized in that The real-time collection of temperature, pressure, component concentration and flow rate data of each tower of the multi-stage distillation system is performed to perform dynamic operation state modeling based on artificial intelligence to obtain a real-time operation state vector representing the overall operation status of the current system, including: The temperature, pressure, component concentration and flow rate data of each tower are collected, and multi-source data fusion is performed based on Kalman filtering to obtain a high-confidence operating condition data set after noise suppression; Inputting the high-confidence operating condition data set into a dynamic component concentration soft-sensing model based on a long short-term memory network to predict real-time component concentration values to obtain a supplementary data set of component concentrations; Based on the high-confidence operating condition data set and the supplementary data set, a material-energy coupling relationship between towers is constructed to obtain a dynamic system topology diagram representing the state association of each tower; According to the dynamic system topology diagram, multi-tower collaborative state embedding driven by graph neural network is performed to obtain a 128-dimensional real-time operation state vector containing system-level spatiotemporal features.

3. The method according to claim 2, characterized in that According to the real-time operation state vector, a multi-objective collaborative optimization process is performed based on a reinforcement learning algorithm to obtain a global optimization instruction set including the optimal set temperature, set pressure, reflux ratio and inter-tower energy flow instructions for each tower, wherein the optimization objectives include maximizing separation purity, minimizing energy consumption and maximizing system stability margin, including: According to the real-time operation state vector, a multi-objective reward function is constructed based on a deep deterministic policy gradient algorithm to obtain a purity-energy consumption-stability ternary reward function; Based on the reward function, offline pre-training of the policy network and online real-time policy fine-tuning are performed to obtain a dynamic policy parameter set for multi-objective optimization; According to the dynamic strategy parameter set, a distributed parallel strategy is executed to obtain a local optimization instruction set of independent operating parameters of each tower; Performing cross-tower energy flow game theory optimization based on the local optimization instruction set and the real-time operation state vector to obtain an inter-tower energy flow instruction matrix that minimizes global entropy increase; According to the local optimization instruction set and the inter-tower energy flow instruction matrix, instruction conflicts are resolved and collaboratively packaged to obtain a global optimization instruction set including temperature, pressure, reflux ratio setting values and energy flow instructions.

4. The method according to claim 3, characterized in that According to the inter-tower energy flow instruction, the dynamic topology reconstruction of the heat recovery network is performed, and the high-grade tower top steam waste heat is preferentially directed to the low-grade heat demand point through intelligent valve switching, thereby obtaining a real-time energy flow map that realizes cross-tower energy ladder matching, including: According to the inter-tower energy flow instructions, waste heat grade classification and heat demand priority sorting are performed to obtain a matching priority list of heat sources and heat sinks; According to the matching priority list, dynamic path planning of the thermal network based on directed graph theory is performed to obtain an optimal heat transfer path sequence with a minimum heat transfer temperature difference; generating a valve switch instruction matrix for controlling steam flow according to the optimal heat transfer path sequence; According to the actual heat transfer data after executing the valve switching instruction matrix, real-time evaluation of thermodynamic efficiency and graph visualization processing are performed to obtain a real-time energy flow graph with efficiency labels.

5. The method according to claim 4, characterized in that The method includes: coordinating and controlling the actuators of each column including the heater, the reflux controller, and the pressure regulating valve according to the global optimization instruction set and the real-time energy flow map to obtain a stable distillation operation state that matches the optimization target and the energy flow map, including: According to the tower parameters in the global optimization instruction set, feedforward-feedback composite adjustment based on model predictive control is performed to obtain a reference control signal for each tower actuator; Performing dynamic compensation calculation of reboiler heating power based on cross-tower heat transfer efficiency data in the real-time energy flow map to obtain a thermal coupling compensation coefficient; According to the reference control signal and the thermal coupling compensation coefficient, the control amount of each tower actuator is collaboratively corrected to obtain an optimized control amount set under the anti-overshoot constraint; Perform multivariable decoupling control and actuator response lag compensation according to the optimized control quantity set to obtain synchronized execution instructions; After executing the synchronized execution instruction to drive each actuator, the operation state stability index processing is verified in real time to obtain a stable distillation operation state that meets the Lyapunov stability condition.

6. The method according to claim 5, characterized in that According to the abnormal disturbance characteristics continuously monitored under the stable distillation operation state, adaptive parameter fine-tuning and emergency strategy triggering processing based on the pre-trained fault mode library are performed to obtain a system self-healing state that maintains continuous and stable output of high-purity fatty acid esters, including: Extracting disturbance features based on wavelet packet transform according to the real-time monitoring data under the stable distillation operation state to obtain an abnormal disturbance fingerprint vector; Based on the abnormal disturbance fingerprint vector, similarity matching is performed with the pre-trained fault pattern library to obtain the fault type identification and confidence level; According to the fault type identification and confidence level, an emergency strategy retrieval based on case reasoning is performed to obtain an emergency operation instruction set including a parameter adjustment range and a valve safety interlock; According to the emergency operation instruction set, parameter gradual adjustment and system status monitoring are performed under safety constraints, and a system self-healing state is obtained in which the system automatically recovers to an optimized state after the disturbance is eliminated.

7. A multi-stage distillation control system for refining fatty acid esters, characterized in that: The system comprises: The acquisition module is used to collect temperature, pressure, component concentration and flow rate data of each tower in the multi-stage distillation system in real time, perform dynamic operation state modeling based on artificial intelligence, and obtain a real-time operation state vector that represents the overall operation status of the current system; an optimization module for performing multi-objective collaborative optimization processing based on the real-time operation state vector and a reinforcement learning algorithm to obtain a global optimization instruction set including the optimal set temperature, set pressure, reflux ratio, and inter-tower energy flow instructions for each tower, wherein the optimization objectives include maximizing separation purity, minimizing energy consumption, and maximizing system stability margin; A reconstruction module is used to dynamically reconstruct the topology of the heat recovery network according to the inter-tower energy flow instructions, and to preferentially direct high-grade tower overhead steam waste heat to low-grade heat demand points through intelligent valve switching, thereby obtaining a real-time energy flow map that achieves cross-tower energy cascade matching; a control module for collaboratively controlling the actuators of each column, including the heater, the reflux controller, and the pressure regulating valve, according to the global optimization instruction set and the real-time energy flow map, to obtain a stable distillation operation state that matches the optimization target and the energy flow map; The processing module is used to perform adaptive parameter fine-tuning and emergency strategy triggering based on a pre-trained fault mode library according to the abnormal disturbance characteristics continuously monitored under the stable distillation operation state, so as to obtain a system self-healing state that maintains continuous and stable output of high-purity fatty acid esters.

8. The system according to claim 7, characterized in that The acquisition module is specifically used to: The temperature, pressure, component concentration and flow rate data of each tower are collected, and multi-source data fusion is performed based on Kalman filtering to obtain a high-confidence operating condition data set after noise suppression; Inputting the high-confidence operating condition data set into a dynamic component concentration soft-sensing model based on a long short-term memory network to predict real-time component concentration values to obtain a supplementary data set of component concentrations; Based on the high-confidence operating condition data set and the supplementary data set, a material-energy coupling relationship between towers is constructed to obtain a dynamic system topology diagram representing the state association of each tower; According to the dynamic system topology diagram, multi-tower collaborative state embedding driven by graph neural network is performed to obtain a 128-dimensional real-time operation state vector containing system-level spatiotemporal features.

9. A storage medium, characterized in that: The storage medium stores a computer program, wherein the computer program is configured to execute the method according to any one of claims 1 to 6 when executed.

10. An electronic device comprising a memory and a processor, characterized in that: A computer program is stored in the memory, and the processor is configured to run the computer program to perform the method according to any one of claims 1 to 6.

Images