A Reinforcement Learning-Based Real-Time Optimization and Control Method and System for Multi-Stage Ammonia Refrigeration Coupling

Abstract

This invention discloses a real-time optimization and control method and system for multi-stage ammonia refrigeration coupling based on reinforcement learning, relating to the field of multi-stage refrigeration technology. The method includes: acquiring the current operating context parameters of each stage of the ammonia liquid separator corresponding to the carbon dioxide distillation requirements; establishing a set of key feature points for low-dimensional spatiotemporal trajectories; outputting a timing sequence of feedforward adjustment actions; determining the endpoint phase change observation parameters; obtaining a set of key feature points for corrected low-dimensional spatiotemporal trajectories; determining the corresponding corrected feedforward adjustment action timing sequence, and converting it into an electrical signal to be sent to the hardware actuator to perform cooling capacity adjustment. This invention solves the technical problem in the prior art where the multi-stage ammonia control strategy cannot be adaptively adjusted according to the dynamic changes in distillation load, making it difficult to maintain stable operation of the ammonia refrigeration system under conditions of drastic load fluctuations. It achieves the technical effect of improving the adaptability of multi-stage ammonia refrigeration coupling to operating condition fluctuations.

Description

Technical Field

[0001] This invention relates to the field of multi-stage refrigeration technology, and in particular to a method and system for real-time optimization and control of multi-stage ammonia refrigeration coupling based on reinforcement learning. Background Technology

[0002] Currently, carbon dioxide distillation processes in large-scale chemical plants typically employ multi-stage ammonia refrigeration systems to provide cooling capacity. Each stage of the ammonia separator provides different temperature levels of refrigeration load to the distillation column through throttling expansion and evaporative heat absorption processes. The mainstream control scheme adopts an independent cascaded regulation method based on PID loops. This involves independently controlling the variable frequency compressor unit and the pressure cascaded regulating valves based on the liquid level and pressure signals of each ammonia separator to maintain the parameters of each cooling capacity output node near the setpoint.

[0003] However, this technology has significant shortcomings, specifically: there is a strong thermodynamic coupling relationship within the multi-stage ammonia refrigeration system. Fluctuations in the cooling demand of one stage will be transmitted to adjacent stages through pipeline pressure and temperature. Existing independent PID loops cannot effectively decouple such inter-stage coupling disturbances, leading to pressure oscillations and cooling supply-demand imbalances when the distillation load changes rapidly. Traditional control strategies rely solely on the current liquid level and pressure feedback for adjustment, lacking the ability to predict the future state of the system. When the dynamic cooling load envelope of the distillation column changes rapidly, the control response lags behind the evolution of the thermodynamic state, resulting in a significant cooling supply-demand gap and energy waste. Summary of the Invention

[0004] This application provides a method and system for real-time optimization and control of multi-stage ammonia refrigeration coupling based on reinforcement learning, which solves the technical problem in the prior art that it is impossible to adaptively adjust the multi-stage ammonia control strategy according to the dynamic changes of the distillation load, and it is difficult to maintain the stable operation of the ammonia refrigeration system under the condition of drastic load fluctuation.

[0005] The first aspect of this application provides a real-time optimization and control method for multi-stage ammonia refrigeration coupling based on reinforcement learning, the method comprising: Obtain the current operating context parameters of each stage of ammonia separator corresponding to the carbon dioxide distillation requirements; based on the current operating context parameters, establish a low-dimensional spatiotemporal trajectory key feature point set in the multidimensional thermodynamic state plane to characterize the imbalance between cooling supply and demand of each stage of ammonia separator; input the current operating context parameters and the low-dimensional spatiotemporal trajectory key feature point set into a dimensionality reduction mapping control network, and output the feedforward adjustment action timing sequence corresponding to the hardware actuators on the pipelines of each stage of ammonia separator; input the feedforward adjustment action timing sequence into the state-frozen thermodynamic time-delay evolution prediction model to extrapolate the endpoint phase change observation parameters of the internal state of the refrigeration station within a set future time period; Reconstruction planning calculations are performed in the feature search space formed by the key feature points of the low-dimensional spatiotemporal trajectory. Based on the perceived deviation between the endpoint phase change observation parameters and the target distillation load, the key feature points of the low-dimensional spatiotemporal trajectory are iteratively corrected to obtain the corrected key feature points of the low-dimensional spatiotemporal trajectory. The dimensionality reduction mapping control network is used to determine the corresponding correction feedforward adjustment action timing sequence based on the corrected key feature points of the low-dimensional spatiotemporal trajectory and the current operating context parameters, and converts it into an electrical signal and sends it to the hardware actuator to perform cooling capacity adjustment.

[0006] A second aspect of this application provides a multi-stage ammonia refrigeration coupling real-time optimization and control system based on reinforcement learning, the system comprising: The system includes a context parameter acquisition module for acquiring the current operating context parameters of each stage of ammonia separator corresponding to the carbon dioxide distillation requirements; a feature point group construction module for establishing a low-dimensional spatiotemporal trajectory key feature point group characterizing the supply-demand imbalance of cooling capacity in each stage of ammonia separator within a multidimensional thermodynamic state plane based on the current operating context parameters; an action timing sequence output module for inputting the current operating context parameters and the low-dimensional spatiotemporal trajectory key feature point group into a dimensionality reduction mapping control network, which outputs a feedforward adjustment action timing sequence corresponding to the hardware actuators on the pipelines of each stage of ammonia separator; and an observation parameter acquisition module for inputting the feedforward adjustment action timing sequence into a state-freezing thermodynamic time-delay evolution prediction model to extrapolate the endpoint phase change observation parameters of the internal state of the refrigeration station within a set future time period. The key feature point group acquisition module is used to perform reconstruction planning calculations in the feature search space formed by the key feature point group of the low-dimensional spatiotemporal trajectory. Based on the perceived deviation between the endpoint phase change observation parameters and the target distillation load, iteratively corrects the key feature point group of the low-dimensional spatiotemporal trajectory to obtain the corrected key feature point group of the low-dimensional spatiotemporal trajectory. The cooling capacity adjustment module is used to use a dimensionality reduction mapping control network to determine the corresponding corrected feedforward adjustment action timing sequence based on the corrected key feature point group of the low-dimensional spatiotemporal trajectory and the current operating context parameters, and convert it into an electrical signal to be sent to the hardware actuator to perform cooling capacity adjustment.

[0007] One or more technical solutions provided in this application have at least the following technical effects or advantages: This application obtains the current operating context parameters of each stage of ammonia liquid separator corresponding to the carbon dioxide distillation demand; based on the current operating context parameters, it establishes a set of key feature points of low-dimensional spatiotemporal trajectory in a multidimensional thermodynamic state plane to characterize the imbalance between cooling supply and demand of each stage of ammonia liquid separator; it inputs the current operating context parameters and the set of key feature points of low-dimensional spatiotemporal trajectory into a dimension-reduced mapping control network, which outputs a timing sequence of feedforward adjustment actions corresponding to the hardware actuators on the pipelines of each stage of ammonia liquid separator; and it inputs the timing sequence of feedforward adjustment actions into a state-frozen thermodynamic time-delay evolution prediction model to extrapolate the final phase change observation parameters of the internal state of the refrigeration station within a set future time period. Reconstruction planning calculations are performed within a feature search space comprised of key feature points of a low-dimensional spatiotemporal trajectory. Based on the perceived deviation between the endpoint phase change observation parameters and the target distillation load, the key feature point set of the low-dimensional spatiotemporal trajectory is iteratively corrected to obtain a corrected set. A dimensionality-reduced mapping control network is then used to determine the corresponding timing sequence of corrected feedforward adjustment actions based on the corrected set of key feature points and the current operating context parameters. This sequence is converted into electrical signals and sent to the hardware actuator to perform cooling capacity adjustment. This achieves the technical effect of improving the adaptability of multi-stage ammonia refrigeration coupling to operating condition fluctuations. Attached Figure Description

[0008] To more clearly illustrate the technical solutions in the embodiments of the present invention, the accompanying drawings used in the description of the embodiments will be briefly introduced below. Obviously, the accompanying drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0009] Figure 1 This is a flowchart illustrating the multi-stage ammonia refrigeration coupling real-time optimization and control method based on reinforcement learning provided in the embodiments of this application.

[0010] Figure 2 This is a diagram illustrating the reconstruction planning iterative convergence process of the multi-stage ammonia refrigeration coupling real-time optimization and control method based on reinforcement learning provided in the embodiments of this application.

[0011] Figure 3 This is a schematic diagram of the structure of a multi-stage ammonia refrigeration coupling real-time optimization and control system based on reinforcement learning provided in an embodiment of this application.

[0012] Figure labeling: Context parameter acquisition module 11, feature point group construction module 12, action time sequence output module 13, observation parameter acquisition module 14, key feature point group acquisition module 15, cooling capacity adjustment module 16. Detailed Implementation

[0013] This application provides a method and system for real-time optimization and control of multi-stage ammonia refrigeration coupling based on reinforcement learning, which solves the technical problem in the prior art that it is impossible to adaptively adjust the multi-stage ammonia control strategy according to the dynamic changes of the distillation load, and it is difficult to maintain the stable operation of the ammonia refrigeration system under the condition of drastic load fluctuation.

[0014] The technical solutions of the embodiments of this application will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only a part of the embodiments of this application, and not all of them. All other embodiments obtained by those skilled in the art based on the embodiments of this application without creative effort are within the scope of protection of this application.

[0015] It should be noted that the terms "first," "second," etc., in the specification and accompanying drawings of this application are used to distinguish similar objects and are not necessarily used to describe a specific order or sequence. It should be understood that such data can be interchanged where appropriate so that the embodiments of this application described herein can be implemented in orders other than those illustrated or described herein. Furthermore, the terms "comprising" and "having," and any variations thereof, are intended to cover non-exclusive inclusion; for example, a process, method, system, product, or server that includes a series of steps or units is not necessarily limited to those steps or units explicitly listed, but may include other steps or modules not explicitly listed or inherent to such processes, methods, products, or devices.

[0016] Example 1, as Figure 1 As shown, a multi-stage ammonia refrigeration coupling real-time optimization and control method based on reinforcement learning is described, wherein the method includes: Obtain the current operating context parameters of each stage of the ammonia separator corresponding to the carbon dioxide distillation requirements; Furthermore, the operating context parameters include real-time pressure gradients, saturation temperature differences, and time-series data characterizing load fluctuations of the main unit for each stage of the ammonia separator.

[0017] In one possible embodiment, this application is used for real-time balance control of cooling supply and demand in a multi-stage ammonia refrigeration system during carbon dioxide distillation. Taking a 600,000-ton-per-year carbon dioxide distillation unit in a petrochemical base as an example, its supporting ammonia refrigeration system adopts a three-stage ammonia liquid separator architecture, corresponding to a high-pressure stage with an evaporation temperature of -10℃, a medium-pressure stage with an evaporation temperature of -30℃, and a low-pressure stage with an evaporation temperature of -50℃. The three stages are connected by a pressure cascade regulating valve, and the total installed capacity of the variable frequency compressor unit is 8000kW. The load of the distillation column can change rapidly within the range of 60%-110% of the rated value as the load of the upstream unit fluctuates, with a maximum load change rate of up to 5% per minute. Under this condition, the liquid level fluctuation of each stage of the ammonia liquid separator exceeds ±30% of the set value, and the pressure fluctuation exceeds ±15% of the design pressure, frequently triggering the compressor loading and unloading protection, and the system cannot operate stably.

[0018] This application improves the adaptability of multi-stage ammonia refrigeration coupling to operating condition fluctuations by performing multi-dimensional sensing of the actual operation of ammonia liquid separators at each stage, and then executing a cycle of prediction, optimization and control.

[0019] Preferably, the ammonia separators at each stage refer to multiple gas-liquid separation containers in a multi-stage ammonia refrigeration system, divided according to different evaporation temperature levels. In this embodiment, there are three stages: a high-pressure stage ammonia separator, a medium-pressure stage ammonia separator, and a low-pressure stage ammonia separator. The gas phase outlet of each stage ammonia separator is connected to the corresponding compressor inlet via a return gas pipeline, and the liquid phase outlet is connected to the next stage ammonia separator or the distillation column heat exchanger via a throttling valve.

[0020] The current operating context parameters refer to a set of multi-dimensional real-time data reflecting the operating status of each stage of the ammonia separator and the cooling demand of the distillation column within the current control cycle. This includes real-time pressure gradients, saturation temperature differences, and time-series data characterizing the load fluctuations of the main unit for each stage of the ammonia separator. The real-time pressure gradient refers to the gas phase pressure difference between two adjacent ammonia separator stages. For a three-stage system, the pressure difference between the high-pressure stage and the medium-pressure stage is defined. and the pressure difference between the medium-pressure stage and the low-pressure stage ,in , , These are the real-time saturated pressure values ​​at the gas phase outlet of the three-stage ammonia separator, in kPa. When the cooling load of a certain stage suddenly increases, the evaporation pressure of that stage rises, and the inter-stage pressure difference deviates from the normal range. This is a sensitive indicator for judging whether the cooling supply and demand are balanced.

[0021] The saturation temperature difference refers to the difference between the current evaporation temperature of each stage of the ammonia separator and the design temperature required for the corresponding refrigerant outlet. The time-series data characterizing the load fluctuation of the main unit refers to continuous time-series signals reflecting the overall heat demand of the distillation process, such as the feed flow rate of the carbon dioxide distillation column, the reflux flow rate at the top of the column, and the heat load of the reboiler at the bottom of the column.

[0022] The aforementioned operating context parameters are collected in real time by sensors distributed throughout the ammonia separators at each stage. For example, pressure transmitters with a range of 0-2 MPa and an accuracy of 0.075 are installed at the gas phase outlet of each ammonia separator; platinum resistance temperature transmitters (Pt100, accuracy class A) are installed in the liquid phase zone; mass flow meters are installed on the feed and reflux lines of the distillation column; and vortex flow meters are installed on the reboiler steam line. Temperature and pressure compensation is used to calculate the heat load. All collected signals are transmitted to the controller via a fieldbus. The controller filters, denoises, and normalizes the dimensions of the raw signals to form the operating context parameter vector for the current cycle.

[0023] By collecting operating context parameters in real time at intervals preset by those skilled in the art during each control cycle, the controller is able to accurately capture minute changes in the distillation load, thus achieving the technical effect of providing reliable data support for subsequent regulation.

[0024] Based on the current operating context parameters, a set of key feature points of low-dimensional spatiotemporal trajectory is established in the multidimensional thermodynamic state plane to characterize the imbalance between cooling supply and demand of each stage of ammonia liquid separator. Preferably, the multidimensional thermodynamic state plane refers to a state space constructed with pressure, specific enthalpy, and temperature as coordinate axes. For ammonia refrigerant, each point on this plane uniquely determines a thermodynamic state, such as saturated liquid, saturated gas, superheated, or subcooled. In this embodiment, the core thermodynamic nodes of the multi-stage ammonia refrigeration system include: 3 gas phase outlet nodes of the three-stage ammonia liquid separator, 3 liquid phase outlet nodes, 6 inlet and outlet nodes of the inter-stage throttling valves, 6 compressor suction and discharge nodes (2 for each stage of the compressor), and 1 condenser outlet node, totaling 19 core thermodynamic nodes.

[0025] The low-dimensional spatiotemporal trajectory key feature point group refers to a set of feature points with a dimension significantly lower than the original state space extracted from the thermodynamic states of the aforementioned 19 nodes. This group is used to characterize the degree of imbalance between cooling supply and demand in each stage of the ammonia liquid separator under current operating conditions. In this embodiment, the dimension of the feature point group is set to 6, with one gas phase feature point and one liquid phase feature point selected for each stage.

[0026] Furthermore, based on the current operating context parameters, a set of key feature points of low-dimensional spatiotemporal trajectory characterizing the imbalance between cooling supply and demand in each stage of the ammonia liquid separator is established in the multidimensional thermodynamic state plane. This embodiment of the application also includes: Extract the dynamic cooling load envelope of the target carbon dioxide distillation column; Calculate the theoretical pressure-enthalpy-temperature distribution trajectory caused by the dynamic cooling load envelope at each core thermodynamic node of the multi-stage ammonia refrigeration system; Projecting the theoretical distribution trajectory of pressure-enthalpy-temperature onto the space of the operating context parameters of the current cycle, the key feature point group of the low-dimensional spatiotemporal trajectory falling in the current operating coordinate system is obtained.

[0027] Furthermore, the embodiments of this application also include calculating the theoretical pressure-enthalpy-temperature distribution trajectory caused by the dynamic cooling load envelope at each core thermodynamic node of the multi-stage ammonia refrigeration system. The dynamic cooling load envelope is input as a heat load boundary condition into the preset gas-liquid two-phase flow phase change kinetic equation; Based on the latent heat of vaporization and gas-liquid equilibrium constant in each stage of the ammonia liquid separator, the throttling expansion enthalpy change, evaporation endothermic phase change flow rate, and compressor multivariable compression temperature rise of the multi-stage ammonia refrigeration system under the heat load boundary conditions are dynamically solved. By simultaneously solving the mass and energy conservation equations for each core thermodynamic node, the real-time saturated pressure, specific enthalpy, and thermodynamic temperature values ​​of each core thermodynamic node as they evolve over time are obtained. After mapping the connecting lines, the theoretical distribution trajectory of pressure-enthalpy-temperature is obtained.

[0028] Furthermore, the set of key feature points of the low-dimensional spatiotemporal trajectory includes gas phase state feature points characterizing the gas phase return pressure of each stage of the ammonia liquid separator and liquid phase state feature points characterizing dynamic liquid level fluctuations.

[0029] In one possible embodiment, the target carbon dioxide distillation column is divided into three sections: a high-pressure section (top of the column), operating at approximately 2.5 MPa and a temperature of approximately -10°C; a medium-pressure section (middle of the column), operating at approximately 1.2 MPa and a temperature of approximately -30°C; and a low-pressure section (near the bottom of the column), operating at approximately 0.5 MPa and a temperature of approximately -50°C. Each section exchanges heat with the corresponding ammonia separator through its own heat exchanger. The dynamic cooling load envelope refers to the curve showing the change in required cooling load over time for each section of the distillation column under current feed conditions and operating parameters.

[0030] For the distillation column part, This corresponds to the high-pressure / medium-pressure / low-pressure segment, and this segment is currently... Required cooling load The load is determined by the gas phase load and liquid phase load of this section: .

[0031] in, For the first Gas flow rate of the segment, This refers to the latent heat of vapor-phase condensation under the operating pressure of this section. This is the liquid phase reflux flow rate for this section. The sensible heat (kJ / kmol) that needs to be removed during liquid phase subcooling. and The value is obtained based on the gas-liquid balance calculation inside the distillation column, and the input is the current feed flow rate. Feed composition and return flow rate .

[0032] Arranging the three cooling load segments along the time axis yields a complete expression of the dynamic cooling load envelope. For example, under a certain steady-state condition, the high-pressure section requires cooling... The medium-pressure section requires cooling capacity. Low-pressure section requires cooling capacity The total cooling load was 13.5 MW. When the feed flow rate suddenly increased from 130 t / h to 150 t / h, the cooling load of each segment rose to approximately 3.9 MW, 5.8 MW, and 5.3 MW respectively within 5-10 minutes, with the total load increasing to 15.0 MW, an increase of approximately 11%. By calculating the segmented cooling demand of the distillation column and then connecting the segments or fitting the envelope surface on the time axis, a dynamic cooling load envelope that dynamically reflects the overall fluctuation trend of the cooling demand of the distillation column can be drawn in the time dimension. This achieves the technical effect of transforming the dynamic changes of the distillation load into a specific and quantifiable sequence of heat load boundary conditions.

[0033] Furthermore, the dynamic cooling load envelope is used as the transient heat load boundary condition and input into the preset gas-liquid two-phase flow phase change kinetic equation. Combining the inherent latent heat of vaporization and gas-liquid equilibrium constants of each ammonia separator at the current operating pressure, the phase change parameters of each node are dynamically solved, including the throttling expansion enthalpy change of liquid ammonia passing through the cascade valve, the evaporation endothermic phase change mass flow rate per second, and the variable compression temperature rise caused by the sudden change in the suction volume of the variable frequency compressor. Next, the system establishes a set of mass and energy conservation differential equations for each core separator cavity and compressor port within the multi-stage refrigeration system. Using the fourth-order Runge-Kutta method with forward integration at 50ms steps, the real-time saturated pressure, specific enthalpy, and thermodynamic temperature values ​​of each node over time are calculated. These three-dimensional coordinate points are then fitted using cubic spline interpolation to obtain a continuous theoretical pressure-enthalpy-temperature distribution trajectory.

[0034] Specifically, based on the inherent latent heat of vaporization and gas-liquid equilibrium constant within the corresponding ammonia separator, the flow rate of the resulting endothermic evaporation phase change is determined using the heat balance equation. The heat balance equation is: ; in, The flow rate is for the evaporative endothermic phase change. For the current moment Required cooling load, Regarding the current operating pressure The nonlinear function is determined by interpolation from the saturated property table of ammonia.

[0035] Furthermore, when liquid ammonia passes through the cascaded regulating valve between adjacent ammonia separators, it undergoes an adiabatic throttling expansion process. Due to the sudden pressure drop, the total specific enthalpy remains unchanged after throttling, but due to the flash evaporation effect generating two-phase flow, its throttling expansion enthalpy changes to... ;in, The specific enthalpy of the liquid ammonia flowing in from the upstream, and The operating pressure of the lower level The enthalpy of saturated liquid and the enthalpy of saturated gas.

[0036] Simultaneously, the abrupt change in the phase change mass flow rate leads to a change in the gas volumetric flow rate entering the compressor. The variable compression temperature rise ΔT caused by the abrupt change in the intake volume of the variable frequency compressor satisfies the polytropic compression state equation for a nonideal gas: ;in, This refers to the compressor suction temperature. and These are the intake pressure and the exhaust pressure, respectively. It is a variable compression index, dynamically corrected based on the current operating frequency of the variable frequency compressor and the ammonia adiabatic index.

[0037] The combined mass conservation differential equation is: ; The energy conservation differential equation is: .

[0038] in, The total control volume mass within the current separator. The total internal energy is the specific internal energy. For the inflow of supplies from higher authorities, The mass flow rate of the gas phase pumped away by the compressor; This refers to the heat loss due to heat dissipation into the environment.

[0039] By using dynamic load to phase transition equations, then to conservation integrals, and a three-dimensional trajectory data loop, the energy evolution trend of each internal node is calculated both inside and outside the digital space, achieving the technical effect of providing low-dimensional trajectory feature guidance for subsequent control networks.

[0040] The current running context parameters and the low-dimensional spatiotemporal trajectory key feature point group are input to the dimension reduction mapping control network, and the dimension reduction mapping control network outputs the feedforward adjustment action timing sequence corresponding to the hardware actuator on the pipeline of each ammonia liquid separator. Furthermore, embodiments of this application also include: The key feature point group of the low-dimensional spatiotemporal trajectory is superimposed as a constraint label onto the current running context parameters to generate a guidance control vector; The guiding control vector is fed into the parameter update channel of the control network for offline reinforcement learning training; A state feature occlusion mechanism is introduced to independently hide some dimensions of the features in the low-dimensional spatiotemporal trajectory key feature point group according to a set probability during training iterations until convergence, thereby obtaining the trained dimensionality reduction mapping control network.

[0041] In one embodiment, the dimensionality reduction mapping control network refers to a neural network model trained offline using deep reinforcement learning. Its core function is to efficiently transform low-dimensional key feature points abstracted from the high-dimensional thermodynamic space into physical actuator control actions. The constraint labels refer to the physical boundary conditions used as physical boundary conditions to guide the neural network's search direction, derived from the key gas-liquid two-phase feature points extracted in the preceding steps. The guiding control vector is a composite input vector generated by concatenating real-time acquired operating context parameters, such as pressure and temperature difference, with the constraint label feature points through a multi-dimensional matrix.

[0042] During the offline training phase, the extracted low-dimensional spatiotemporal trajectory key feature point groups are used as constraint labels. These are then concatenated into the current operating context parameters, which include real-time pressure gradients and saturation temperature differences of the ammonia separators at each stage. The concatenated vector generates a dimension-expanded guiding control vector, which serves as the current state and is fed into the parameter update channel of the control network. To enable the dimensionality-reduced mapping control network to withstand disturbances and errors in complex industrial environments, a state feature occlusion mechanism is introduced during training iterations. In each training loop, the system independently and randomly selects a portion of the guiding control vector to hide based on a set Bernoulli random probability, such as 15%. This is achieved by multiplying the corresponding elements of the input vector using a mask matrix, forcibly replacing the selected dimension components with zero. For example, at a certain iteration step, the low-pressure separator liquid level feature component in the network input is forcibly cleared to zero. At this point, the control network is forced to perform forward propagation and output feedforward actions even with incomplete input features. Simultaneously, it incorporates a multi-objective reward function constructed from the total power consumption of the refrigeration station, the variance of pressure deviations from the target in each separator stage, and the penalty term for liquid level exceeding the safe range. A proximal policy optimization algorithm is then used for backpropagation to update the network parameters. When some sensor data is missing, the network must rely on the remaining temperature difference and gas pressure features to automatically complete the internal parameter matrix and infer the correct hardware action. After training with hundreds of thousands of industrial fluctuation samples until the network's expected policy reward reaches its maximum and converges, the internal weight matrices of the fully connected layers and attention mechanisms within the network are locked, thus completing the offline construction of the dimensionality-reduced mapping control network.

[0043] In real-time control of online applications, the current operating context parameters and the latest calculated key feature point group of the low-dimensional spatiotemporal trajectory are reassembled into a guiding control vector input to the network. The dimensionality reduction mapping control network utilizes its multi-layer fully connected or temporal attention mechanism to complete nonlinear dimensionality reduction mapping within milliseconds, directly outputting the feedforward adjustment action timing sequence of the hardware actuators of each stage of the ammonia liquid separator pipeline in the next control cycle, such as the frequency adjustment amplitude of the variable frequency compressor and the opening step size of the pressure cascade regulating valve.

[0044] The feedforward adjustment action timing sequence is input into the thermodynamic time delay evolution prediction model of state freezing, and the end phase change observation parameters of the internal state of the freezing station in a set future time period are extrapolated. Furthermore, the state-freezing thermodynamic time-delay evolution prediction model is constructed based on the dynamic mathematical mechanism equations of a multi-stage refrigeration system, and is a nonlinear evolver in which the internal mapping weights are locked and not dynamically updated during the execution of the reconstruction planning calculation.

[0045] In one embodiment, the state-freezing thermodynamic time-delay evolution prediction model is a nonlinear digital evolver constructed based on the real physical architecture of the refrigeration station and multi-level cyclic mechanism equations. This model can simulate the complex energy conduction inside the refrigeration station, and within a specific trajectory reconstruction calculation cycle, the mechanism formula parameters or mapping weights representing physical properties such as physical loss, heat capacity, and heat transfer coefficient are forcibly locked and do not dynamically update with the current local optimization calculation.

[0046] The system acquires a feedforward control action sequence output by the control network at the current moment, containing twenty future control steps (e.g., thirty seconds per step, totaling ten minutes). This sequence specifically includes the target frequency increment of the variable frequency compressor unit at each future time point and the opening step of the pressure cascaded regulating valve on the connecting pipeline of the adjacent ammonia separator. Next, this action sequence is used as the control drive source and injected into a prediction model constructed based on the dynamic mathematical mechanism equations of a multi-stage refrigeration system and the differential equations for the conservation of mass, energy, and momentum. At this point, the prediction model enters a state-freeze mode, where all core physical parameters identified and fixed during system initialization, such as the dynamic heat transfer coefficients of each heat exchanger, valve flow characteristic coefficients, and constants such as the equivalent heat capacity of the medium, are completely locked and cannot be changed during the iterative process of this planning and refactoring calculation.

[0047] Within the parameter-locked nonlinear evolution engine, the computational engine uses the measured saturation pressure, temperature, and liquid level of each stage of the current refrigeration plant's separators as the initial state for integration. It then progressively incorporates the feedforward action sequence into the discretized state update equations for high-frequency time-domain recursion. After twenty consecutive steps of mechanistic extrapolation, the model overcomes the significant physical time delays caused by the long pipelines and slow phase changes in the multi-stage ammonia refrigeration system, ultimately directly calculating the endpoint phase change observation parameters of the refrigeration plant's internal state at the end of the prediction window. These endpoint parameters include the transient saturation pressure and saturation temperature at the gas phase outlet of each stage of the ammonia liquid separator at the end of the prediction window, as well as the actual liquid level in the liquid phase cavity after considering the elimination of spurious liquid levels.

[0048] Reconstruction planning calculations are performed in the feature search space formed by the set of key feature points of the low-dimensional spatiotemporal trajectory. Based on the perception deviation between the endpoint phase change observation parameters and the target distillation load, the set of key feature points of the low-dimensional spatiotemporal trajectory is iteratively corrected to obtain the corrected set of key feature points of the low-dimensional spatiotemporal trajectory. Furthermore, a reconstruction planning calculation is performed in the feature search space formed by the set of key feature points of the low-dimensional spatiotemporal trajectory. Based on the perceived deviation between the endpoint phase change observation parameters and the target distillation load, the set of key feature points of the low-dimensional spatiotemporal trajectory is iteratively corrected to obtain a corrected set of key feature points of the low-dimensional spatiotemporal trajectory. This embodiment of the application also includes: Calculate the residual heat gap between the endpoint phase change observation parameters and the target distillation load; Based on the heat gap residual value, the correction increment of the key feature point group of the low-dimensional spatiotemporal trajectory is calculated in reverse within the feature search space by exploring the gradient descent direction. The correction increment is superimposed on the key feature point group of the low-dimensional spatiotemporal trajectory in the current cycle for updating, and the dimensionality reduction mapping control network and the thermodynamic time delay evolution prediction model are repeatedly triggered until the heat gap residual value converges to within the set threshold, thereby obtaining the corrected low-dimensional spatiotemporal trajectory key feature point group.

[0049] In one embodiment, the heat gap residual value refers to the scalar quantity representing the energy deviation between the predicted future end-of-period cooling capacity and the actual target load of the downstream carbon dioxide distillation tower. The feature search space refers to a low-dimensional virtual control variable space comprised of all possible values ​​of key feature points in a low-dimensional spatiotemporal trajectory; its dimension is far lower than the total dimension of hundreds or thousands of hardware controllers on-site. Gradient descent direction exploration is a reverse mathematical feedback mechanism that seeks the optimal solution within the feature search space. By calculating the sensitivity of the residual to feature points, it guides the feature points to make incremental adjustments in the direction of reducing deviation.

[0050] Once the thermodynamic time-delay evolution prediction model for the frozen state outputs the observation parameters for the endpoint phase change at the end of the set future time period, the system compares these parameters with the target distillation load of the current carbon dioxide distillation column. The optimization module calculates the residual value of the heat gap between the two. For example, if the distillation column requires a target cooling capacity of 250 kW, but the predicted endpoint phase change state can only provide 220 kW of cooling capacity, the calculated residual value of the heat gap is a positive 30 kW. Instead of directly adjusting the valves or compressors on-site for the current 30 kW heat gap residual, the system explores the gradient descent direction in a low-dimensional feature search space composed of gas and liquid phase characteristic points.

[0051] The first-order partial derivative of the heat gap residual with respect to the current set of key feature points in the low-dimensional spatiotemporal trajectory, including the gas phase return pressure feature point and the dynamic liquid level fluctuation feature point, is calculated; this is the sensitivity gradient. Based on the direction of this gradient, multiplied by a preset step size coefficient, the required adjustment increment for the current feature point set is calculated in reverse. For example, the low-pressure stage gas phase return pressure feature point needs to be finely adjusted upwards by 0.01 MPa, and the liquid level fluctuation feature point needs corresponding amplitude correction.

[0052] The calculated correction increment is directly superimposed onto the key feature point set of the low-dimensional spatiotemporal trajectory in the current loop to obtain the updated feature point set. Then, the forward control logic is immediately triggered again: the updated feature point set is repackaged with the current running context parameters and re-fed into the dimensionality reduction mapping control network to output a new set of feedforward adjustment action sequences. These new action sequences are then injected back into the state-frozen thermodynamic time-delay evolution prediction model, and the new endpoint phase transition observation parameters ten minutes later are extrapolated and calculated again. The residual between the new endpoint parameters and the target load is recalculated, and the above gradient exploration and update process is repeated. Figure 2 As shown, this digital simulation cycle continues at an extremely high speed until the heat gap residual value shrinks to within a set safety threshold, such as a residual of less than 1 kW, or the maximum number of iterations (e.g., 15 times) is reached. At this point, convergence is achieved, and the final determined set of key feature points for the corrected low-dimensional spatiotemporal trajectory is obtained.

[0053] Since it is impossible to exhaustively capture all extreme process fluctuations in the field during offline training, the feedforward adjustment action of a single output may deviate from the actual requirements. This step constructs an embedded iterative closed loop in the low-dimensional feature space to calculate residuals, inversely derive gradients, update features, and re-control and predict. This forces the control signal to perform self-correction and approximation exercises in the digital space before it is actually converted into a hardware electrical signal and sent out.

[0054] The dimensionality reduction mapping control network determines the corresponding timing sequence of the correction feedforward adjustment action based on the key feature point group of the corrected low-dimensional spatiotemporal trajectory and the current operating context parameters, and converts it into an electrical signal and sends it to the hardware actuator to perform cooling adjustment.

[0055] Furthermore, the hardware actuator includes a variable frequency compressor unit installed in the multi-stage ammonia refrigeration architecture and a pressure cascade regulating valve installed on the connecting pipeline between adjacent ammonia liquid separators.

[0056] In one embodiment, the key feature point set of the corrected low-dimensional spatiotemporal trajectory, which includes feature vectors containing the optimal gas phase pressure inflection point and liquid phase fluctuation feature values, is reassembled with current operating context parameters such as the real-time pressure gradient and saturation temperature difference. This complete vector is used as the final state input to trigger the dimensionality reduction mapping control network a second time. The dimensionality reduction mapping control network utilizes the weight matrices of its self-attention layer and fully connected layer, trained offline by reinforcement learning, to perform a deterministic forward propagation inference without feature occlusion, outputting the final corrected feedforward adjustment action sequence within milliseconds. This sequence precisely plans the actions of the actuators at 30-second control steps within the next ten minutes, including, for example, a frequency correction sequence for the variable frequency compressor and a cascaded regulating valve opening correction sequence.

[0057] After acquiring this high-precision action timing sequence, the controller does not directly send discrete digital data. Instead, it uses a digital-to-analog (D / A) converter chip or an industrial control bus communication card to perform a low-level physical conversion. For variable frequency compressor units, the controller sends the analog voltage signal of the frequency digital command output by the main control algorithm to the signal input terminal of the compressor inverter.

[0058] For cascaded pressure control valves, the controller converts the opening command into a pulse width modulation (PWM) electrical signal or a continuous current drive signal, which is then sent to the electric or pneumatic valve positioner of the cascaded control valve. Upon receiving the continuous electrical signal, the field hardware actuator immediately generates a physical-mechanical response. In the multi-stage ammonia refrigeration architecture, the variable frequency compressor unit, driven by the inverter, smoothly increases its motor speed to the specified frequency, directly increasing the suction volume flow rate of the low-pressure stage separator and forcing the gas phase return pressure towards the set safety axis.

[0059] Meanwhile, the pressure cascade regulating valves installed on the connecting pipelines of adjacent ammonia liquid separators, upon receiving an electrical signal, control a servo motor or pneumatic diaphragm actuator to precisely push the valve core to the opening position, adjusting the effective cross-sectional area of ​​the valve orifice. This allows the subcooled liquid ammonia in the next stage separator to be flash-sprayed into the next cascade stage at the most stable mass flow rate. This achieves the technical effect of improving the adaptability and reliability of multi-stage ammonia refrigeration coupling control and operating conditions.

[0060] Example 2, as Figure 3 As shown, based on the same inventive concept as in Embodiment 1 above, this application provides a multi-stage ammonia refrigeration coupling real-time optimization and control system based on reinforcement learning, the system comprising: Context parameter acquisition module 11 is used to acquire the current operating context parameters of each level of ammonia liquid separator corresponding to the carbon dioxide distillation requirements; The feature point group construction module 12 is used to establish a low-dimensional spatiotemporal trajectory key feature point group that characterizes the imbalance between cooling supply and demand of each stage of ammonia liquid separator in the multidimensional thermodynamic state plane according to the current operating context parameters. The action timing sequence output module 13 is used to input the current running context parameters and the low-dimensional spatiotemporal trajectory key feature point group into the dimension reduction mapping control network, and the dimension reduction mapping control network outputs the feedforward adjustment action timing sequence corresponding to the hardware actuators on the pipelines of each ammonia liquid separator. The observation parameter acquisition module 14 is used to input the feedforward adjustment action timing sequence into the thermodynamic time delay evolution prediction model of state freezing, and extrapolate the final phase change observation parameters of the internal state of the freezing station within a set future time period. The key feature point group acquisition module 15 is used to perform reconstruction planning calculation in the feature search space formed by the key feature point group of the low-dimensional spatiotemporal trajectory, and iteratively correct the key feature point group of the low-dimensional spatiotemporal trajectory according to the perception deviation between the endpoint phase change observation parameters and the target distillation load, so as to obtain the corrected key feature point group of the low-dimensional spatiotemporal trajectory. The cooling capacity adjustment module 16 is used to determine the corresponding correction feedforward adjustment action timing sequence based on the key feature point group of the corrected low-dimensional spatiotemporal trajectory and the current operating context parameters using a dimensionality reduction mapping control network, and convert it into an electrical signal and send it to the hardware actuator to perform cooling capacity adjustment.

[0061] Furthermore, the operating context parameters include real-time pressure gradients, saturation temperature differences, and time-series data characterizing load fluctuations of the main unit for each stage of the ammonia separator.

[0062] Furthermore, the feature point group construction module 12 is used to perform the following steps: Extract the dynamic cooling load envelope of the target carbon dioxide distillation column; Calculate the theoretical pressure-enthalpy-temperature distribution trajectory caused by the dynamic cooling load envelope at each core thermodynamic node of the multi-stage ammonia refrigeration system; Projecting the theoretical distribution trajectory of pressure-enthalpy-temperature onto the space of the operating context parameters of the current cycle, the key feature point group of the low-dimensional spatiotemporal trajectory falling in the current operating coordinate system is obtained.

[0063] Furthermore, the feature point group construction module 12 is used to perform the following steps: The dynamic cooling load envelope is input as a heat load boundary condition into the preset gas-liquid two-phase flow phase change kinetic equation; Based on the latent heat of vaporization and gas-liquid equilibrium constant in each stage of the ammonia liquid separator, the throttling expansion enthalpy change, evaporation endothermic phase change flow rate, and compressor multivariable compression temperature rise of the multi-stage ammonia refrigeration system under the heat load boundary conditions are dynamically solved. By simultaneously solving the mass and energy conservation equations for each core thermodynamic node, the real-time saturated pressure, specific enthalpy, and thermodynamic temperature values ​​of each core thermodynamic node as they evolve over time are obtained. After mapping the connecting lines, the theoretical distribution trajectory of pressure-enthalpy-temperature is obtained.

[0064] Furthermore, the set of key feature points of the low-dimensional spatiotemporal trajectory includes gas phase state feature points characterizing the gas phase return pressure of each stage of the ammonia liquid separator and liquid phase state feature points characterizing dynamic liquid level fluctuations.

[0065] Furthermore, the action timing sequence output module 13 is used to perform the following steps: The key feature point group of the low-dimensional spatiotemporal trajectory is superimposed as a constraint label onto the current running context parameters to generate a guidance control vector; The guiding control vector is fed into the parameter update channel of the control network for offline reinforcement learning training; A state feature occlusion mechanism is introduced to independently hide some dimensions of the features in the low-dimensional spatiotemporal trajectory key feature point group according to a set probability during training iterations until convergence, thereby obtaining the trained dimensionality reduction mapping control network.

[0066] Furthermore, the key feature point group acquisition module 15 is used to perform the following steps: Calculate the residual heat gap between the endpoint phase change observation parameters and the target distillation load; Based on the heat gap residual value, the correction increment of the key feature point group of the low-dimensional spatiotemporal trajectory is calculated in reverse within the feature search space by exploring the gradient descent direction. The correction increment is superimposed on the key feature point group of the low-dimensional spatiotemporal trajectory in the current cycle for updating, and the dimensionality reduction mapping control network and the thermodynamic time delay evolution prediction model are repeatedly triggered until the heat gap residual value converges to within the set threshold, thereby obtaining the corrected low-dimensional spatiotemporal trajectory key feature point group.

[0067] Furthermore, the hardware actuator includes a variable frequency compressor unit installed in the multi-stage ammonia refrigeration architecture and a pressure cascade regulating valve installed on the connecting pipeline between adjacent ammonia liquid separators.

[0068] Furthermore, the state-freezing thermodynamic time-delay evolution prediction model is constructed based on the dynamic mathematical mechanism equations of a multi-stage refrigeration system, and is a nonlinear evolver in which the internal mapping weights are locked and not dynamically updated during the execution of the reconstruction planning calculation.

[0069] Although this application makes various references to certain modules in the system according to the embodiments of this application, any number of different modules can be used and run on user terminals and / or servers. The various units and modules included are only divided according to functional logic, but are not limited to the above division, as long as the corresponding functions can be achieved; in addition, the specific names of each functional unit are only for easy distinction between each other and are not used to limit the scope of protection of this invention.

[0070] The specific embodiments described above do not constitute a limitation on the scope of protection of this application. Those skilled in the art should understand that various modifications, combinations, and substitutions can be made according to design requirements and other factors. Any modifications, equivalent substitutions, and improvements made within the spirit and principles of this application should be included within the scope of protection of this application. In some cases, the actions or steps described in this application can be performed in a different order than that shown in the embodiments and still achieve the desired results. Furthermore, the processes depicted in the accompanying drawings do not necessarily require a specific or sequential order to achieve the desired results. In some embodiments, multitasking and parallel processing are also possible or may be advantageous.

Claims

1. A multi-stage ammonia refrigeration coupling real-time optimization and control method based on reinforcement learning, characterized in that, The method includes: Obtain the current operating context parameters of each stage of the ammonia separator corresponding to the carbon dioxide distillation requirements; Based on the current operating context parameters, a set of key feature points of low-dimensional spatiotemporal trajectory is established in the multidimensional thermodynamic state plane to characterize the imbalance between cooling supply and demand of each stage of ammonia liquid separator. The current running context parameters and the low-dimensional spatiotemporal trajectory key feature point group are input to the dimension reduction mapping control network, and the dimension reduction mapping control network outputs the feedforward adjustment action timing sequence corresponding to the hardware actuator on the pipeline of each ammonia liquid separator. The feedforward adjustment action timing sequence is input into the thermodynamic time delay evolution prediction model of state freezing, and the end phase change observation parameters of the internal state of the freezing station in a set future time period are extrapolated. Reconstruction planning calculations are performed in the feature search space formed by the set of key feature points of the low-dimensional spatiotemporal trajectory. Based on the perception deviation between the endpoint phase change observation parameters and the target distillation load, the set of key feature points of the low-dimensional spatiotemporal trajectory is iteratively corrected to obtain the corrected set of key feature points of the low-dimensional spatiotemporal trajectory. The dimensionality reduction mapping control network determines the corresponding timing sequence of the correction feedforward adjustment action based on the key feature point group of the corrected low-dimensional spatiotemporal trajectory and the current operating context parameters, and converts it into an electrical signal and sends it to the hardware actuator to perform cooling adjustment.

2. The multi-stage ammonia refrigeration coupling real-time optimization and control method based on reinforcement learning as described in claim 1, characterized in that, The operating context parameters include real-time pressure gradients, saturation temperature differences, and time-series data characterizing load fluctuations of the main unit for each stage of the ammonia separator.

3. The multi-stage ammonia refrigeration coupling real-time optimization and control method based on reinforcement learning as described in claim 1, characterized in that, Based on the current operating context parameters, a set of key feature points of low-dimensional spatiotemporal trajectory characterizing the imbalance between cooling supply and demand in each stage of the ammonia liquid separator is established in the multidimensional thermodynamic state plane, including: Extract the dynamic cooling load envelope of the target carbon dioxide distillation column; Calculate the theoretical pressure-enthalpy-temperature distribution trajectory caused by the dynamic cooling load envelope at each core thermodynamic node of the multi-stage ammonia refrigeration system; Projecting the theoretical distribution trajectory of pressure-enthalpy-temperature onto the space of the operating context parameters of the current cycle, the key feature point group of the low-dimensional spatiotemporal trajectory falling in the current operating coordinate system is obtained.

4. The multi-stage ammonia refrigeration coupling real-time optimization and control method based on reinforcement learning as described in claim 3, characterized in that, Calculate the theoretical pressure-enthalpy-temperature distribution trajectory caused by the dynamic cooling load envelope at each core thermodynamic node of the multi-stage ammonia refrigeration system, including: The dynamic cooling load envelope is input as a heat load boundary condition into the preset gas-liquid two-phase flow phase change kinetic equation; Based on the latent heat of vaporization and gas-liquid equilibrium constant in each stage of the ammonia liquid separator, the throttling expansion enthalpy change, evaporation endothermic phase change flow rate, and compressor multivariable compression temperature rise of the multi-stage ammonia refrigeration system under the heat load boundary conditions are dynamically solved. By simultaneously solving the mass and energy conservation equations for each core thermodynamic node, the real-time saturated pressure, specific enthalpy, and thermodynamic temperature values ​​of each core thermodynamic node as they evolve over time are obtained. After mapping the connecting lines, the theoretical distribution trajectory of pressure-enthalpy-temperature is obtained.

5. The multi-stage ammonia refrigeration coupling real-time optimization and control method based on reinforcement learning as described in claim 1, characterized in that, The set of key feature points of the low-dimensional spatiotemporal trajectory includes gas phase state feature points characterizing the gas phase return pressure of each stage of ammonia liquid separator and liquid phase state feature points characterizing dynamic liquid level fluctuations.

6. The multi-stage ammonia refrigeration coupling real-time optimization and control method based on reinforcement learning as described in claim 1, characterized in that, include: The key feature point group of the low-dimensional spatiotemporal trajectory is superimposed as a constraint label onto the current running context parameters to generate a guidance control vector; The guiding control vector is fed into the parameter update channel of the control network for offline reinforcement learning training; A state feature occlusion mechanism is introduced to independently hide some dimensions of the features in the low-dimensional spatiotemporal trajectory key feature point group according to a set probability during training iterations until convergence, thereby obtaining the trained dimensionality reduction mapping control network.

7. The multi-stage ammonia refrigeration coupling real-time optimization and control method based on reinforcement learning as described in claim 1, characterized in that, Reconstruction planning calculations are performed in the feature search space formed by the set of key feature points of the low-dimensional spatiotemporal trajectory. Based on the perceived deviation between the endpoint phase change observation parameters and the target distillation load, the set of key feature points of the low-dimensional spatiotemporal trajectory is iteratively corrected to obtain the corrected set of key feature points of the low-dimensional spatiotemporal trajectory, including: Calculate the residual heat gap between the endpoint phase change observation parameters and the target distillation load; Based on the heat gap residual value, the correction increment of the key feature point group of the low-dimensional spatiotemporal trajectory is calculated in reverse within the feature search space by exploring the gradient descent direction. The correction increment is superimposed on the key feature point group of the low-dimensional spatiotemporal trajectory in the current cycle for updating, and the dimensionality reduction mapping control network and the thermodynamic time delay evolution prediction model are repeatedly triggered until the heat gap residual value converges to within the set threshold, thereby obtaining the corrected low-dimensional spatiotemporal trajectory key feature point group.

8. The multi-stage ammonia refrigeration coupling real-time optimization and control method based on reinforcement learning as described in claim 1, characterized in that, The hardware actuators include a variable frequency compressor unit installed in the multi-stage ammonia refrigeration architecture and a pressure cascade regulating valve installed on the connecting pipeline between adjacent ammonia liquid separators.

9. The multi-stage ammonia refrigeration coupling real-time optimization and control method based on reinforcement learning as described in claim 1, characterized in that, The state-freezing thermodynamic time-delay evolution prediction model is constructed based on the dynamic mathematical mechanism equations of a multi-stage refrigeration system, and is a nonlinear evolver in which the internal mapping weights are locked and not dynamically updated during the execution of the reconstruction planning calculation.

10. A multi-stage ammonia refrigeration coupling real-time optimization and control system based on reinforcement learning, characterized in that, The system is used to implement the reinforcement learning-based multi-stage ammonia refrigeration coupling real-time optimization and control method according to any one of claims 1-9, the system comprising: The context parameter acquisition module is used to acquire the current operating context parameters of each stage of ammonia liquid separator corresponding to the carbon dioxide distillation requirements; The feature point group construction module is used to establish a low-dimensional spatiotemporal trajectory key feature point group that characterizes the imbalance between cooling supply and demand of each stage of ammonia liquid separator in the multidimensional thermodynamic state plane based on the current operating context parameters. The action timing sequence output module is used to input the current running context parameters and the low-dimensional spatiotemporal trajectory key feature point group into the dimension reduction mapping control network, and the dimension reduction mapping control network outputs the feedforward adjustment action timing sequence corresponding to the hardware actuators on the pipelines of each ammonia liquid separator. The observation parameter acquisition module is used to input the feedforward adjustment action time sequence into the thermodynamic time delay evolution prediction model of state freezing, and extrapolate the end-point phase change observation parameters of the internal state of the freezing station within a set future time period. The key feature point group acquisition module is used to perform reconstruction planning calculation in the feature search space formed by the key feature point group of the low-dimensional spatiotemporal trajectory, and iteratively correct the key feature point group of the low-dimensional spatiotemporal trajectory based on the perception deviation between the endpoint phase change observation parameters and the target distillation load to obtain the corrected key feature point group of the low-dimensional spatiotemporal trajectory. The cooling capacity adjustment module is used to determine the corresponding correction feedforward adjustment action timing sequence based on the key feature point group of the corrected low-dimensional spatiotemporal trajectory and the current operating context parameters using a dimensionality reduction mapping control network, and convert it into an electrical signal and send it to the hardware actuator to perform cooling capacity adjustment.

Images