A catalytic cracking mechanism integrated multi-agent working condition prediction and regulation method

Abstract

The application discloses a kind of catalytic cracking mechanism fusion multi-agent working condition prediction and regulation method, including, for the problem that existing multi-agent regulation scheme does not fully fuse mechanism constraint, state estimation uncertainty is not quantified, prone to resource preemption and constraint conflict, and safety fault tolerance capability is insufficient.The application completes the physical division of the device based on the catalytic cracking mechanism weighted network;Through the boundary state interaction and covariance cross fusion containing uncertainty parameters, consistent estimation of the hidden state of the whole device is realized;The physical conservation margin of the device physical capacity is quantified as the binding state snapshot and the effective period, and the action contract containing margin consumption and precondition is generated by each agent;After combination screening and control barrier function safety correction, the margin is atomized and the regulation instruction is issued.The application can realize efficient and safe distributed collaborative working condition regulation of catalytic cracking device.

Description

Technical Field

[0001] This invention relates to the field of intelligent control and operating condition optimization technology for petrochemical processes, and in particular to a multi-agent method for predicting and regulating operating conditions by integrating catalytic cracking mechanisms. Background Technology

[0002] Catalytic cracking is a core process in petroleum refining systems for converting heavy oil into lighter oil. Its units exhibit typical complex system characteristics of strong nonlinearity, multivariate strong coupling, and large inertial lag. Precise sensing and stable control of operating conditions directly determine product distribution, energy consumption levels, and long-term operational safety. With the intelligent upgrading of the refining industry, traditional control modes based on centralized PID control and steady-state mechanism simulation are no longer adequate for the refined management needs under dynamic conditions such as fluctuations in feedstock properties and equipment performance degradation. In recent years, multi-agent distributed decision-making technology has been gradually introduced into the process industry control field. By decomposing complex production systems into independent decision-making sub-nodes according to functional units, it effectively improves the control response rate and local fault tolerance. At the same time, data-driven state estimation and predictive methods are also widely used in the operation optimization of catalytic cracking units, providing a mature technical path for real-time operating condition sensing and pre-adjustment control.

[0003] However, existing multi-agent control schemes for catalytic cracking primarily focus on the distributed scheduling of decision-making tasks, failing to integrate the constraints of material and energy conservation mechanisms in the catalytic cracking process. Boundary information exchange between agents only transmits deterministic state values, without quantifying the uncertainty of state estimation, easily leading to the accumulation of global state consistency deviations and reducing the reliability of operating condition predictions. Furthermore, existing control methods lack quantitative control and dynamic quota management of the unit's physical capacity boundaries. When multiple agents make parallel decisions, resource contention and constraint conflicts easily arise, and it is difficult to ensure that control actions remain within a safe operating range during operating condition fluctuations. In addition, existing schemes lack adaptive reconfiguration capabilities in scenarios involving local equipment or communication failures, easily leading to global control disorder due to local failures, making it difficult to simultaneously meet the dual requirements of control efficiency and operational safety in refining and chemical units. Summary of the Invention

[0004] The purpose of this section is to outline some aspects of embodiments of the present invention and to briefly describe some preferred embodiments. Simplifications or omissions may be made in this section, as well as in the abstract and title of this application, to avoid obscuring the purpose of these documents; however, such simplifications or omissions should not be construed as limiting the scope of the invention.

[0005] In view of the aforementioned existing problems, this invention is proposed. Therefore, this invention provides a multi-agent operating condition prediction and control method integrating catalytic cracking mechanisms to solve the problems mentioned in the background art.

[0006] To address the aforementioned technical problems, this invention provides the following technical solution: a multi-agent operating condition prediction and control method integrating catalytic cracking mechanisms, comprising:

[0007] S1: Based on the preset mechanism model of the catalytic cracking device, the device is physically divided into multiple intelligent agents;

[0008] S2: The multiple intelligent agents perform distributed state estimation based on the real-time and historical data of their respective subsystems, and exchange boundary states containing uncertainty information to obtain a consistent estimate of the hidden states that cannot be directly measured in the entire device.

[0009] S3: Based on the consensus estimate, calculate the available margin of the catalytic cracking unit in one or more future time slices regarding coke combustion capacity, oxygen supply capacity, heat removal capacity, catalyst circulation capacity, compressor operating margin, or pollutant emission capacity, and quantify the available margin as a physical conservation allowance bound to a state snapshot and validity period.

[0010] S4: The multiple intelligent agents perform causal condition prediction based on the consensus estimate and candidate control actions, and generate an action contract that includes candidate actions, the physical conservation amount required to execute the action, and the preconditions required to execute the action.

[0011] S5: Select a set of action contract combinations that satisfy the physical conservation limit constraints and do not conflict with each other, and use a control obstacle function to perform safety correction on the joint action generated by the action contract combination to ensure that the device state can be kept within the preset safety set after the correction joint action is executed.

[0012] S6: Upon confirming the execution of the safety-modified joint action, the physical conservation quota consumed by the action contract combination is atomically deducted, and a control command is issued to the downstream control system.

[0013] Compared with the prior art, the beneficial effects of the present invention are as follows:

[0014] 1. This invention uses the conservation of materials, energy, and reaction kinetics in catalytic cracking as the core to delineate the boundaries of intelligent agents. It transmits the uncertain statistical parameters of the boundary states between intelligent agents and eliminates the interference of repeated measurement evidence through the covariance cross-fusion algorithm. This avoids the error accumulation caused by traditional deterministic numerical interaction and can achieve a globally consistent estimate of key hidden states that cannot be directly measured, such as catalyst activity and coke combustion rate, providing a reliable state basis for advanced operating condition control.

[0015] 2. This invention introduces a physical conservation limit mechanism, which transforms the physical capacity boundaries of devices such as coke combustion, oxygen supply, and heat removal into quantifiable, reserveable, and deductible dynamic management limits. Combined with a standardized action contract negotiation model, it transforms the strongly coupled multivariate control problem into a resource application and contract verification problem, thereby fundamentally avoiding resource preemption and constraint mutual exclusion in distributed parallel decision-making and improving the feasibility and execution efficiency of multi-agent collaborative control.

[0016] 3. This invention is based on dynamic matching of safety margins for state estimation uncertainty. It controls the joint action of obstacle function online correction to constrain the state to a safe range. In the event of local equipment or communication failure, it can automatically complete the intelligent agent reconstruction and switch to a safe and stable operation mode. While pursuing control benefits, it ensures the long-term safe operation of the device and has strong adaptability to working conditions and fault tolerance. Attached Figure Description

[0017] To more clearly illustrate the technical solutions of the embodiments of the present invention, the accompanying drawings used in the description of the embodiments will be briefly introduced below. Obviously, the 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. Wherein:

[0018] Figure 1 This is a flowchart illustrating the overall process of a multi-agent operating condition prediction and control method integrating catalytic cracking mechanisms, as described in one embodiment of the present invention. Detailed Implementation

[0019] To make the above-mentioned objects, features, and advantages of the present invention more apparent and understandable, specific embodiments of the present invention will be described in detail below with reference to the accompanying drawings. Obviously, the described embodiments are only a part of the embodiments of the present invention, and not all of them. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort should fall within the protection scope of the present invention.

[0020] Many specific details are set forth in the following description in order to provide a full understanding of the invention. However, the invention may also be practiced in other ways different from those described herein, and those skilled in the art can make similar extensions without departing from the spirit of the invention. Therefore, the invention is not limited to the specific embodiments disclosed below.

[0021] Secondly, the term "one embodiment" or "embodiment" as used herein refers to a specific feature, structure, or characteristic that may be included in at least one implementation of the present invention. The phrase "in one embodiment" appearing in different places in this specification does not necessarily refer to the same embodiment, nor is it a single or selective embodiment that is mutually exclusive with other embodiments.

[0022] Furthermore, in the description of this invention, it should be noted that the terms "upper," "lower," "inner," and "outer," etc., indicate the orientation or positional relationship based on the orientation or positional relationship shown in the accompanying drawings. These terms are used solely for the convenience of describing the invention and for simplifying the description, and do not indicate or imply that the device or element referred to must have a specific orientation, or be constructed and operated in a specific orientation. Therefore, they should not be construed as limitations on the invention. In addition, the terms "first," "second," or "third" are used for descriptive purposes only and should not be construed as indicating or implying relative importance.

[0023] Example 1

[0024] Reference Figure 1 This is the first embodiment of the present invention, which provides a multi-agent operating condition prediction and control method integrating catalytic cracking mechanisms, including:

[0025] S1: Based on the preset mechanism model of the catalytic cracking unit, the unit is physically divided into multiple intelligent agents.

[0026] Specifically, catalytic cracking units typically include core equipment such as reaction settling systems, regeneration systems, and main fractionation towers, exhibiting typical characteristics of strong multivariate coupling and nonlinearity. To avoid the dimensional curse and single point of failure caused by centralized control, this step, based on the material and energy conservation relationships within the unit and the control variable manipulation permissions, achieves automatic agent partitioning by constructing a mechanistic association network.

[0027] Furthermore, the nonlinear mechanism model of the catalytic cracking unit is abstracted into a network topology graph. This network contains two types of nodes:

[0028] One is a variable node: representing all state variables of the device (e.g., regenerator dense phase bed temperature, reactor stock, etc.) and control variables (e.g., regeneration slide valve opening, main air flow rate, etc.).

[0029] Another is the mechanism node: representing the material conservation equation, energy conservation equation, and reaction kinetic equation between units.

[0030] Furthermore, if a variable is involved in a specific mechanism equation (e.g., the main airflow is involved in the regenerator heat balance equation), then a connection edge is established between the corresponding variable node and the mechanism node.

[0031] Furthermore, to achieve a scientific physical partition, each connection edge in the aforementioned network needs to be assigned a weight, which characterizes the physical coupling depth between the variables and the mechanistic equations. The calculation of the weights mainly considers the following three dimensions:

[0032] Mechanism sensitivity: Calculates the sensitivity (i.e., partial derivatives or Jacobian matrix elements) of the mechanistic equation to small fluctuations in a specific variable under baseline conditions. Higher sensitivity indicates a more significant impact of that variable on the local material / energy conservation state.

[0033] Dimensional normalization coefficient: used to eliminate the difference in magnitude between different physical quantities (e.g., temperature in °C and flow rate in t / h), ensuring that the weights have a uniform comparison benchmark.

[0034] Operation permission penalty factor: Based on the actual hardware architecture of DCS (Distributed Control System), if a control variable is restricted by specific operation permissions or hard-wired logic (e.g., a certain slide valve can only be operated by the regenerator control station), then it is given an extremely high penalty factor.

[0035] It should be noted that the weight of the connecting edge is proportional to the product of the above-mentioned mechanism sensitivity, dimensional normalization coefficient and manipulation authority penalty factor.

[0036] It should be noted that by introducing a manipulation permission penalty factor, it is possible to force restricted control variables and their strongly correlated physical devices to be classified within the same intelligent entity, thereby avoiding high-frequency control command transmission delays across intelligent entities.

[0037] Furthermore, after constructing the weighted mechanism association network, a graph clustering algorithm (e.g., normalized cut algorithm) is used to divide the entire network, outputting multiple non-overlapping subnetworks, each corresponding to an agent. The partitioning principle (i.e., optimization objective) of this clustering algorithm is:

[0038] Minimize the cost of cross-agent coupling: minimize the total weight of the severed connection edges, that is, ensure that the physical coupling between the separated agents is relatively weak.

[0039] Maximize the cohesion within an agent: maximize the sum of the weights of the edges that remain within the same agent, thus ensuring that strongly coupled variables and mechanistic equations are grouped together.

[0040] It should be noted that, through the clustering algorithm described above, the system can automatically cluster strongly coupled physical units (e.g., reactors and settlers) into the same intelligent agent (e.g., a reaction agent), while dividing units with clear physical boundaries (e.g., the main fractionation tower) into independent intelligent agents (e.g., fractionation agents). This division method preserves the autonomy within the subsystem to the greatest extent possible.

[0041] Furthermore, based on the above-mentioned division principles, the variables corresponding to the severed connection edges are extracted and defined as boundary states. For example, the boundary states between the reactive agent and the regenerating agent include: the amount of catalyst to be generated, the amount of catalyst to be regenerated, and the sensible heat they carry. For these boundary states, a standardized information exchange interface is established between adjacent agents. This interface is configured not only to transmit the current measurement or estimate of the boundary state, but also to transmit the uncertainty statistical parameters of that state (e.g., mean and variance / covariance).

[0042] It should be noted that this interface design breaks the error accumulation problem caused by traditional deterministic numerical transmission, enabling upstream and downstream agents to perceive the confidence level of each other's operating condition fluctuations, thereby providing a reliable data channel for collaborative state estimation and control among multiple agents.

[0043] S2: Multiple agents perform distributed state estimation based on real-time and historical data from their respective subsystems, and exchange boundary states containing uncertainty information to obtain a consistent estimate of the hidden states that cannot be directly measured throughout the device.

[0044] Specifically, in catalytic cracking units, there are numerous critical latent states that cannot be directly measured by instruments (e.g., real-time catalyst activity in the reactor, coke combustion rate in the regenerator, etc.). To accurately acquire these latent states within a multi-agent architecture, this step employs a method of distributed state observation and covariance cross-fusion.

[0045] Furthermore, each agent (e.g., a reactive agent, a regenerative agent) first constructs a local state observer (preferably an extended Kalman filter or an unscented Kalman filter in this embodiment) based on the local mechanism model partitioned in S1. The input data includes: the control command at the current moment, real-time measurements from sensors within the subsystem (e.g., temperature, pressure, flow rate), and the state estimation results from the previous moment. The output data includes: the estimated mean of the local state at the current moment (including measurable states and unmeasurable hidden states), and the error covariance matrix characterizing the confidence level of the estimation result. This covariance matrix quantifies the current uncertainty (variance) of each state variable and characterizes the coupling correlation between different physical states through its off-diagonal elements.

[0046] Furthermore, to achieve global state consistency, adjacent agents need to exchange boundary states (i.e., cross-agent coupling variables, such as the amount of catalyst to be generated and the amount of heat carried). In this embodiment, the agents do not directly transmit the original covariance matrix, but instead convert the estimated mean of the boundary state and its corresponding local covariance matrix into parameterized data packets in an information filtering format to reduce communication bandwidth consumption. Moreover, to avoid the measurement data from the same sensor being repeatedly transmitted and counted in a multi-agent ring communication network, the agents attach a data source identifier when generating information packets. This identifier records the source node ID and timestamp of the core measurement data upon which the current state estimation depends.

[0047] Furthermore, when the local agent receives a boundary state information packet from a neighboring agent, it first compares the data source identifiers. If overlapping measurement evidence sources are found, the covariance cross-fusion algorithm is triggered.

[0048] Specifically, this algorithm does not require knowledge of the cross-correlation between the estimation errors of the two agents. Instead, it introduces a dynamic fusion weight coefficient to weight and combine the boundary state estimates of the local agent with those from neighboring agents. It is important to emphasize that this fusion weight coefficient is not a fixed value, but is determined by solving a real-time optimization problem. The optimization objective is to minimize the trace (i.e., the total uncertainty) of the fused boundary state covariance matrix. After optimization, the algorithm directly outputs the fused confidence boundary state mean and the updated covariance matrix with lower uncertainty. This approach ensures that the system's boundary state estimation is not overconfident, guaranteeing the stability of the filtering.

[0049] In an embodiment of the present invention, the calculation method for the covariance cross-fusion is as follows: Let the mean of the boundary state estimate of the local agent be... covariance is The mean estimate transmitted from neighboring agents is covariance is The state after fusion Covariance Calculated using the following formula:

[0050] Among them, the fusion weight coefficient By solving optimization problems Real-time determination, Represents the trace of a matrix.

[0051] Furthermore, after the boundary states are fused, each agent uses the fused result to perform a secondary correction on its internal hidden states.

[0052] Specifically, each agent outputs the fused confidence boundary state and its uncertainty as a virtual sensor measurement, feeding it back into its local state observer. Taking a reaction-regeneration system as an example, the heat introduced by the catalyst (boundary state) calculated by the regeneration agent is transferred to the reaction agent. The reaction agent uses this heat value as a virtual measurement and, combined with the reaction kinetic mechanism model, can deduce the true rate of catalyst activity decay (hidden state) inside the reactor. Through this data interaction and correction, a globally consistent estimate of the hidden state of the entire device is achieved without a centralized computing center.

[0053] S3: Based on consensus estimation, calculate the available margins of the catalytic cracking unit in one or more future time slices regarding coke combustion capacity, oxygen supply capacity, heat removal capacity, catalyst circulation capacity, compressor operating margin, or pollutant emission capacity, and quantify the available margins as physical conservation allowances bound to state snapshots and expiration dates.

[0054] It should be noted that in multi-agent distributed control, to prevent resource contention among agents when generating control actions (e.g., the reaction agent requests an increase in catalyst circulation, while the regeneration agent's main fan oxygen supply capacity has reached its limit), this step introduces the concept of a physical conservation limit. This limit is equivalent to a budget in economics, transforming the device's physical constraints into quantifiable and distributable digital assets.

[0055] Specifically, firstly, each agent receives globally consistent hidden state estimation results from S2 (e.g., current catalyst activity, coke content, etc.). Combining the equipment nameplate parameters with the built-in mechanistic model, the agent calculates the theoretical upper limit of a specific physical resource's capacity within the predicted future time slice. Simultaneously, based on the current basic control strategy for maintaining stable operation of the device, the expected utilization of this resource under baseline conditions is calculated. Taking oxygen supply capacity as an example, the theoretical upper limit is limited by the maximum speed of the main fan and the anti-surge boundary; while the expected utilization is the amount of oxygen required to burn off the current basic coke content in the regenerator.

[0056] Furthermore, due to fluctuations in future operating conditions, directly using deterministic predicted values ​​could lead to control actions exceeding safety limits when conditions worsen. Therefore, it is necessary to utilize uncertainty indicators (e.g., covariance matrix) included in the S2 state estimation results to assess the prediction error of future time-slice resource utilization. Based on a preset confidence level (e.g., 95%), the system transforms this prediction error into a dynamic safety margin. It should be noted that the greater the uncertainty of the state estimation (i.e., the more fuzzy the prediction of future operating conditions), the larger the safety margin reserved by the system must be. Therefore, this approach enables an adaptive safety mechanism where, in practice, control becomes more conservative as operating conditions become more ambiguous.

[0057] Furthermore, after obtaining the above parameters, the agent executes resource deduction logic to determine the final available reserve. The specific deduction rules are as follows:

[0058] First, subtract the expected occupancy of the baseline operating condition from the theoretical upper limit of the equipment. Then, subtract the dynamic safety margin obtained from the uncertainty assessment of the state estimation. Finally, subtract a preset emergency reserve (a buffer margin to deal with sudden hard failures, such as when a pump suddenly stops).

[0059] It should be noted that the above continuous deduction process ensures that the calculated available margin is a safe and available amount after comprehensively considering all known consumption, prediction errors and unknown risks, thereby eliminating the possibility of subsequent control actions causing the device to exceed its limits.

[0060] Furthermore, the calculated available margin value is encapsulated into a standardized data structure, namely the physical conservation margin. To ensure data consistency during asynchronous decision-making by multiple agents, the following two key attributes must be bound to this margin data structure:

[0061] State Snapshot: Records the globally consistent state vector and corresponding timestamp used to calculate the credit limit. If a subsequent agent, when invoking the credit limit generation action, finds that the actual state of the current device deviates from the state recorded in the snapshot by more than a set threshold, the credit limit will automatically become invalid and must be recalculated.

[0062] Validity period: Defines the future time span for which this credit limit applies (e.g., valid only within the next 5 minutes), and it will expire after that period.

[0063] S4: Multiple agents perform causal condition prediction based on consensus estimation and candidate control actions, and generate an action contract that includes candidate actions, the physical conservation amount required to execute the action, and the preconditions that need to be met to execute the action.

[0064] Specifically, in a multi-agent distributed control architecture, each agent independently generates a series of candidate regulatory actions to achieve local optimization goals (e.g., the reaction agent aims to maximize the yield of the target product, while the regeneration agent aims to minimize energy consumption). However, a catalytic cracking unit is a strongly coupled system, and local actions inevitably affect the global system. Therefore, before proposing any action, each agent must perform causal condition prediction and encapsulate the prediction results and resource requirements into a standardized action contract for subsequent global negotiation.

[0065] Furthermore, each agent, based on its own local optimization objective and combined with the current consensus state estimation results, uses a built-in local optimization solver (such as a model predictive control algorithm or a reinforcement learning policy network) to generate one or more candidate sequences of regulatory actions. For example, to improve the light oil yield, the reaction agent might generate a candidate action: within the next 10 minutes, increase the reaction temperature setpoint by 2°C and correspondingly increase the amount of catalyst to be generated.

[0066] In an implementable embodiment of the present invention, when a reinforcement learning policy network is used to generate candidate control actions, the input data (state space) of the policy network is set as follows: the current consensus state estimation vector, the current remaining physical conservation allowance, and the market price weight of the target product; the output data (action space) of the policy network is set as follows: the changes in each control variable (e.g., valve opening, setpoint) within future time slices; the training reward function of the policy network is set as follows: the difference between the economic benefit brought by the target product yield and the energy consumption cost brought by the control action, provided that the physical conservation allowance is not exceeded. Furthermore, in this embodiment, since the control actions (e.g., valve opening, setpoint changes) belong to a continuous action space, the reinforcement learning policy network is preferably constructed using the DDPG (Deep Deterministic Policy Gradient) algorithm. Specifically, the policy network includes an Actor network and a Critic network. The Actor network consists of an input layer, at least two fully connected hidden layers (using the ReLU activation function), and an output layer (using the Tanh activation function to map actions to the [-1, 1] interval, then linearly scaling them to the actual physical constraints). The Critic network takes the concatenation of the state vector and the action vector output by the Actor network as input, passes through at least two fully connected hidden layers, and outputs a single Q-value (i.e., the expected cumulative reward). The training steps for the reinforcement learning policy network are as follows:

[0067] Data collection: Offline simulation is performed using local mechanism models, or historical operating data of the device is collected, and (current state, action, reward, next state) is stored as experience tuples in the experience replay pool.

[0068] Network update: Randomly draw small batches of samples from the experience replay pool.

[0069] Critic network training: Calculate the target Q value, use MSE (mean squared error) as the loss function, and update the Critic network parameters using gradient descent.

[0070] Actor network training: With the goal of maximizing the Q-value of the Critic network output, the policy gradient is calculated using the chain rule, and the Actor network parameters are updated.

[0071] The parameters of the target Actor network and the target Critic network are updated periodically using a soft update method until the network converges.

[0072] Furthermore, for each candidate action, the agent needs to predict its causal impact on itself and adjacent subsystems. Therefore, this invention employs a causal inference method based on a mechanistic model. Specifically, the agent uses the current consensus state estimation result as initial conditions and the candidate control action as input perturbation, substituting them into the local mechanistic model constructed in S1 for forward integration simulation. Through this simulation, the agent can not only predict the future trajectory of its own state but also predict the change in the boundary state caused by the action. For example, the reaction agent predicts that increasing the catalyst circulation rate will lead to an increase in the amount of coke carried by the spent catalyst transferred to the regeneration agent. This mechanistic-based causal prediction allows the agent to anticipate the impact of its actions on the global material and energy balance, avoiding blind control.

[0073] Furthermore, based on the causal condition prediction results, the agent needs to quantify the consumption requirement of the candidate action on the physical conservation quota generated in S3.

[0074] Specifically, the agent extracts the peak occupancy of various physical resources in the predicted trajectory and subtracts the occupancy under the baseline condition to obtain the net consumption of the action. For example, if the reactive agent predicts that an increase in coke production requires the regenerative agent to provide more oxygen for burning the coke, the reactive agent must calculate how much oxygen supply capacity and heat removal capacity is needed for this action. If the required consumption exceeds the available margin published in S3, the candidate action is rejected locally and will not proceed to the subsequent negotiation stage. Then, for candidate actions that pass the local capacity verification, the agent encapsulates them into a standardized action contract. In this embodiment, the data structure of this action contract includes the following core elements:

[0075] Candidate actions: The specific sequence of control commands and their timing.

[0076] Credit Consumption List: Clearly lists the specific values ​​of various physical conservation credits required to perform this action.

[0077] Prerequisites: The device status requirements that must be met to perform this action. For example, the temperature of the dense phase bed in the regenerator must be below 700℃ or the anti-surge margin of the main fan must be greater than 15%.

[0078] Execution deadline: This action must be started before this timestamp, otherwise it will be invalid due to work status drift.

[0079] The predicted distribution of regulatory effects includes the mean and variance of the predicted state, showing other agents the expected benefits and risk range of the action.

[0080] Rollback action: A safety control command used to safely restore the device to the baseline operating condition when execution fails or the preconditions are suddenly not met.

[0081] It should be noted that by encapsulating the control intention into action contracts, this invention transforms the multivariable coupled control problem into a resource request and contract signing problem among multiple agents. The preconditions and fallback actions in the contracts enhance the safety of the distributed control system in the face of sudden fluctuations in operating conditions, and realize decoupling and cooperation among multiple agents.

[0082] S5: Select a set of action contract combinations that satisfy the physical conservation limit constraints and do not conflict with each other, and use the control obstacle function to perform safety correction on the joint action generated by the action contract combination to ensure that the corrected joint action can keep the device state within the preset safety set after execution.

[0083] Specifically, in multi-agent systems, the action contracts independently generated by each agent may lead to resource competition or physical constraint conflicts from a global perspective. To ensure the security and feasibility of global regulation, this invention introduces a negotiation mechanism based on the preparation phase and a low-level safety filtering mechanism based on a control barrier function.

[0084] Furthermore, to avoid resource over-exploitation caused by concurrent control, a preparation phase (similar to a two-phase commit protocol in a distributed system) is established before the formal execution of the action. A global coordinating node (or through a consensus algorithm among agents) collects all candidate action contracts and performs a combined screening. The screening principles include the following two points:

[0085] Resource verification: In the selected action contract combination, the total consumption of the same type of physical conservation quota (e.g., oxygen supply capacity) by all actions must be less than or equal to the available margin of that quota calculated in S3.

[0086] Logical verification: The preconditions and execution intentions of each action must not be logically mutually exclusive (e.g., one action cannot require increasing the reaction temperature while another action requires decreasing the reaction temperature).

[0087] Furthermore, once a preliminary set of feasible action contracts is selected, the system will reserve the physical conservation quota that these contracts will consume during the preparation phase (i.e., temporarily freeze this quota, and other unselected contracts cannot apply for it). If, during the preparation phase, negotiation times out due to communication delays, or if an agent detects a sudden change in local operating conditions that causes the preconditions to no longer be met, the system will immediately terminate the current process and release the reserved quota, thereby preventing the system from falling into a deadlock state.

[0088] It should be emphasized that when there are multiple action contract combinations that satisfy both resource verification and logic verification, the system will take maximizing global economic benefits (e.g., maximizing the total yield of the target product or minimizing global energy consumption) as the optimization objective and select the combination with the highest overall expected return as the final action contract combination.

[0089] Furthermore, for the initially selected action contract combinations, the control instructions they contain are merged into nominal joint actions (i.e., the actions the agent expects to execute). Due to the nonlinearity of the mechanistic model and the existence of prediction errors, directly executing the nominal joint actions may still pose a risk of exceeding limits. Therefore, a control barrier function needs to be introduced for safety verification.

[0090] Specifically, based on the process safety regulations of the catalytic cracking unit (e.g., the regenerator temperature must not exceed the set upper limit, and the main blower anti-surge margin must be greater than the set lower limit), the system defines a multi-dimensional set of safety states. It should be explained that the purpose of the control barrier function is to calculate the safety margin of the current state from the safety boundary in real time. In this invention, the control barrier function will enforce the following: when the unit state is inside the safety zone and far from the boundary, control actions are allowed to move the state towards the boundary; however, when the unit state approaches the safety boundary (i.e., the safety margin approaches zero), the control barrier function will limit the rate at which the state moves towards the boundary, and at the instant the boundary is reached, the control action must force the system state to turn back towards the safety zone.

[0091] In an implementable embodiment of the present invention, safety correction using a control barrier function transforms the correction process of the nominal joint action into a QP (quadratic programming) optimization problem. The objective function of this optimization problem is set as: minimizing the deviation (e.g., Euclidean distance) between the corrected joint action and the nominal joint action to preserve the original control intent to the greatest extent possible. The constraints of this optimization problem include: upper and lower bound constraints on the physical actuators; and a control barrier function constraint, requiring that the corrected action must result in the system state along the directional derivative of the control barrier function gradient, plus a term proportional to the current safety margin, being greater than or equal to zero. This constraint mathematically guarantees that when the state approaches the safety boundary, the correction action will force the generation of a control component moving towards the safety region.

[0092] In an embodiment of the present invention, the QP optimization is considered a problem. Let the local nonlinear control affine model of the catalytic cracking unit be... ,in For state vectors, To control actions, define the safety set as follows: ,in Let be a continuously differentiable control barrier function (e.g., representing the margin of the upper limit of temperature distance). Then the standard mathematical form of the QP optimization problem is:

[0093]

[0094] in, For the sake of joint action, The revised joint action; and They are respectively along and Li's derivative; For incremental extended K-class functions (e.g., ); and For the physical upper and lower limits of the actuator.

[0095] Furthermore, the system transforms the safety correction problem of the nominal joint action into a constrained real-time optimization problem. The input to this optimization problem is the current consistent state estimate and the nominal joint action, and the output is the corrected joint action ultimately issued to the actuator. Specifically, the solution logic for this optimization problem is to find the action that minimizes the deviation from the nominal joint action as the corrected joint action, while satisfying the physical limits of the actuator (e.g., maximum valve opening, maximum pump speed) and the safety boundary constraints of the control obstacle function.

[0096] It should be noted that by solving this optimization problem in real time, the system can fine-tune the actions negotiated by the multi-agents in a short period of time. If the nominal joint action itself is absolutely safe, the optimization solver will not make any modifications and will output directly; if the nominal joint action will lead to a violation of the limit in the future, the solver will automatically weaken or adjust the action under the mandatory intervention of safety constraints, and output a corrected action that can preserve the original agent's control intention to the greatest extent while ensuring the safety of the device.

[0097] Furthermore, under actual extreme operating conditions (such as sudden severe equipment failure or external interference), the above optimization problem may become unsolvable, that is, there is no action that can satisfy both the physical hardware constraints of the actuator and the safety constraints of the control obstacle function.

[0098] Specifically, when the optimization solver returns to a no-solution state, the system determines that the current action contract combination is unexecutable. At this time, the system will immediately activate the preset authentication rollback strategy. This strategy includes: the system actively abandoning all current multi-agent optimization and control intentions and refusing to execute nominal joint actions; directly invoking predefined rollback actions in the action contract (e.g., forcibly switching all critical control loops back to conservative baseline PID control, or keeping the current valve position unchanged); simultaneously, the system releases all reserved physical conservation allowances and issues a high-level operational risk alarm to the DCS (Distributed Control System) and the operator.

[0099] It should be noted that by introducing a control barrier function and an authentication fallback strategy, this invention, while pursuing economic benefits and maximizing regulatory performance among multiple agents, is equivalent to adding a security lock at the underlying level, thus solving the problem of system instability that is easily caused by distributed artificial intelligence control in refining scenarios.

[0100] S6: When confirming the execution of a safety-modified joint action, the physical conservation quota consumed by the action contract combination is atomically deducted, and a control instruction is issued to the downstream control system.

[0101] Specifically, in a multi-agent distributed control system, after the safety modification and feasibility verification of the action contract are completed, the system enters the final execution and state update phase.

[0102] Furthermore, when the system confirms the execution of the security-corrected joint action, it must formally deduct the reserved physical conservation quota. To prevent inconsistencies in quota data due to communication delays or node failures in a distributed network (e.g., multiple agents simultaneously vying for the same resource), this invention employs atomic deduction operations. Specifically, atomic operations mean that quota deduction is an indivisible and independent process at the system's underlying level. The system obtains the actual consumption of various physical conservation quotas by the corrected joint action (this consumption is usually less than the original request due to security correction) and directly subtracts it from the current available balance in the global resource ledger to obtain the updated remaining quota.

[0103] It should be noted that, through atomized deduction, the system ensures that the remaining physical conservation allowance within the current time slice accurately reflects the actual resource reserves of the device in real time. This provides reliable boundary constraints for decisions made by other agents that have not yet performed actions or in the next control cycle, eliminating resource conflicts during concurrent control by multiple agents.

[0104] Furthermore, after the quota deduction is completed, the system will convert the corrected joint action into specific low-level control commands (such as specific valve opening setpoints, pump speed setpoints, etc.), and send them to the downstream DCS or basic control loop for execution through standard industrial communication protocols.

[0105] Furthermore, in order to form a complete control system and address the prediction deviations of the mechanism model, this invention introduces a quota replenishment and contraction adjustment mechanism based on the actual operational response after issuing the control command.

[0106] Specifically, after the control command is issued and one control cycle has elapsed, each agent collects the actual operational response data of the device through sensors and obtains the actual operating state. The system compares this actual state with the control effect predicted in the action contract and calculates the state prediction deviation. Based on this state prediction deviation, the system dynamically adjusts the physical conservation allowance for the next time slice.

[0107] Credit replenishment: If the actual operation response shows that the actual consumption of a certain physical resource by the joint action is less than the predicted consumption (i.e. the device exhibits better performance or higher efficiency than expected), the system will replenish the over-deducted credit to the available margin in the next time slice, thereby freeing up more optimization space for the system and improving the economic efficiency of the device.

[0108] Limit Reduction: If the actual operation response shows that the actual consumption is greater than the predicted consumption, or the state prediction deviation shows a divergent trend (indicating that there is a large error in the mechanism model or that an unknown external disturbance has occurred), the system will forcibly reduce the physical conservation limit of the next time slice (i.e., artificially reduce the available margin).

[0109] It should be noted that this dynamic adjustment mechanism based on actual deviations is equivalent to introducing an adaptive risk preference into the multi-agent system. When the prediction is accurate, the system encourages agents to make full use of resources for economic optimization; when the prediction is biased or the environment deteriorates, the system will automatically tighten resource allocation, forcing agents to adopt more conservative control strategies, thereby achieving a dynamic balance between pursuing economic benefits and ensuring operational safety.

[0110] Furthermore, to avoid ineffective high-frequency computation and communication in the multi-agent system under steady-state conditions, this embodiment also employs an event-triggered mechanism for startup. Specifically, the event includes at least one of the following:

[0111] The predicted innovation of the state estimation exceeds the preset threshold: that is, the norm of the deviation (innovation vector) between the actual measurement value calculated by the local state observer in S2 and the predicted measurement value exceeds the set tolerance, indicating that an unmodeled dynamic change has occurred inside the device.

[0112] The availability of any physical conservation quota shrinks: When the quota is forcibly shrunk based on the actual response, it indicates that system resources are becoming strained and a new round of regulation and negotiation needs to be triggered immediately.

[0113] The predicted probability of operating conditions exceeds the preset limit: Based on historical data statistics or model prediction, it is found that the probability of the device in the chip touching the safety boundary in the future time is increasing.

[0114] Furthermore, to enhance the system's robustness under extreme conditions, the present invention also includes a fault reconfiguration and degradation operation mechanism. Specifically, when any intelligent agent (e.g., a computing node crash), communication network (e.g., a disconnection between nodes), or critical measuring instrument malfunctions, the system will perform the following operations:

[0115] First, based on the observability matrix of the remaining available information, assess whether the current network topology can still uniquely determine the hidden state of the entire device. If not, dynamically merge the affected agents (e.g., merge neighboring agents that have lost communication and have them uniformly taken over by the superior node) or repartition them. Then, before fault recovery, the system temporarily freezes the generation of control action contracts aimed at economic benefits (e.g., maximizing yield), forcing all agents to generate only conservative actions aimed at maintaining the safe and stable operation of the device (e.g., maintaining stable critical temperatures and pressures), thereby achieving safe degraded operation of the system.

[0116] It should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention and are not intended to limit it. Although the present invention has been described in detail with reference to preferred embodiments, those skilled in the art should understand that modifications or equivalent substitutions can be made to the technical solutions of the present invention without departing from the spirit and scope of the technical solutions of the present invention, and all such modifications or substitutions should be covered within the scope of the claims of the present invention.

Claims

1. A multi-agent method for predicting and controlling operating conditions by integrating catalytic cracking mechanisms, characterized in that, include: S1: Based on the preset mechanism model of the catalytic cracking device, the device is physically divided into multiple intelligent agents; S2: The multiple intelligent agents perform distributed state estimation based on the real-time and historical data of their respective subsystems, and exchange boundary states containing uncertainty information to obtain a consistent estimate of the hidden states that cannot be directly measured in the entire device. S3: Based on the consensus estimate, calculate the available margin of the catalytic cracking unit in one or more future time slices regarding coke combustion capacity, oxygen supply capacity, heat removal capacity, catalyst circulation capacity, compressor operating margin, or pollutant emission capacity, and quantify the available margin as a physical conservation allowance bound to a state snapshot and validity period. S4: The multiple intelligent agents perform causal condition prediction based on the consensus estimate and candidate control actions, and generate an action contract that includes candidate actions, the physical conservation amount required to execute the action, and the preconditions required to execute the action. S5: Select a set of action contract combinations that satisfy the physical conservation limit constraints and do not conflict with each other, and use a control obstacle function to perform safety correction on the joint action generated by the action contract combination to ensure that the device state can be kept within the preset safety set after the correction joint action is executed. S6: Upon confirming the execution of the safety-modified joint action, the physical conservation quota consumed by the action contract combination is atomically deducted, and a control command is issued to the downstream control system.

2. The multi-agent operating condition prediction and control method integrating catalytic cracking mechanism as described in claim 1, characterized in that, The division into multiple intelligent agents includes: Based on the mechanism factor diagram of the catalytic cracking unit, the material and energy conservation relationships between units, and the control permissions of the control variables, the division boundaries and information exchange interfaces of the multiple intelligent agents are determined.

3. The multi-agent operating condition prediction and control method integrating catalytic cracking mechanisms as described in claim 1, characterized in that, The exchange includes boundary states containing uncertain information, including: Exchange the parameterized representations of the mean, covariance, or information matrix of the boundary states, and the common factor identifiers used to avoid double-counting common measurement evidence during state fusion.

4. The multi-agent operating condition prediction and control method integrating catalytic cracking mechanisms as described in claim 1, characterized in that, The calculation method for the physical conservation limit includes: The confidence lower bound of the available margin is obtained by subtracting the resource consumption of the baseline operating condition, the safety margin related to the uncertainty of prediction, and the preset emergency reserve from the upper limit of the physical capacity of the equipment or environment.

5. The multi-agent operating condition prediction and control method integrating catalytic cracking mechanisms as described in claim 1, characterized in that, The generated action contract also includes at least one of the following: The execution deadline, the predicted distribution of control effects, and the backoff action to restore the device to a safe state in case of execution failure or failure to meet preconditions.

6. The multi-agent operating condition prediction and control method integrating catalytic cracking mechanism as described in claim 1, characterized in that, In S5, after selecting a set of action contracts that satisfy the physical conservation limit constraints and do not conflict with each other, and before using the control barrier function for safety correction, the following is also included: Establish a preparation phase in which the physical conservation allowance to be consumed by the selected action contract combination is reserved. If the preparation phase is terminated due to negotiation failure or execution timeout, the reserved quota will be released.

7. The multi-agent operating condition prediction and control method integrating catalytic cracking mechanisms as described in claim 1, characterized in that, The safety correction is performed using a control barrier function, including: If no feasible corrective action that satisfies all safety constraints can be found after solving the problem, the current action contract combination is determined to be unexecutable, and the preset authentication rollback strategy is activated.

8. The multi-agent operating condition prediction and control method integrating catalytic cracking mechanism as described in claim 1, characterized in that, The method is initiated by an event, which includes at least one of the following: The predicted innovation of the state estimation exceeds the preset threshold, the available amount of any physical conservation quota shrinks, or the predicted probability of the working condition risk exceeds the preset limit.

9. The multi-agent operating condition prediction and control method integrating catalytic cracking mechanism as described in claim 1, characterized in that, The method further includes: When a failure is detected in any agent, communication network, or key measuring instrument, the agents are dynamically merged or repartitioned based on the observability of the remaining available information, and the control aimed at economic efficiency is temporarily frozen, maintaining only the safe and stable operation of the device.

10. The multi-agent operating condition prediction and control method integrating catalytic cracking mechanisms as described in claim 1, characterized in that, The method further includes: After the control command is issued, the actual operating response of the device is compared with the predicted effect in the action contract, and the physical conservation allowance of subsequent time slices is adjusted by replenishing or shrinking based on the deviation between the two.

Images