Heat exchanger outlet temperature control method based on world model and digital twinning

Abstract

The invention discloses a heat exchanger outlet temperature control method based on world model and digital twinning, and relates to the field of industrial process control. According to the method, firstly, a digital twinborn body based on a reproduction state space model is constructed, a heat balance constraint is introduced to train a submerged space dynamic model, and nonlinear thermal hysteresis and random disturbance are accurately represented; secondly, performing zero-risk strategy exploration and optimization in the virtual submerged space by utilizing reinforcement learning; in a real-time operation stage, the strategy network directly outputs an optimal control instruction according to a real-time potential state, and continuous iteration is realized in combination with the cloud edge collaborative architecture; and finally, PID safety takeover is implemented through a distributed external detection mechanism based on variational free energy. According to the method, high-precision self-adaptive control of the heat exchanger under the complex working condition can be achieved, and the physical consistency, the real-time performance and the industrial operation safety of the algorithm are considered.

Description

Technical Field

[0001] This invention belongs to the field of industrial process automatic control technology, and in particular relates to an intelligent control method for heat exchanger outlet temperature based on a world model and digital twin fusion architecture. Background Technology

[0002] Heat exchangers are crucial heat transfer devices in process industries such as petroleum, chemical, and pharmaceutical. The stability of their outlet temperature directly affects the quality of the final product, production safety, and overall energy efficiency. However, as a complex physical system with large inertia, pure time delay, strong nonlinearity, and extreme susceptibility to external multi-source disturbances, the precise control of heat exchangers faces significant challenges.

[0003] Currently, the mainstream industrial field still uses proportional-integral-derivative (PID) control and its improved algorithms. However, when faced with variable operating conditions and large time delay characteristics, PID control often has a slow response and large overshoot, which cannot meet the stringent high-precision temperature control requirements.

[0004] While Model Predictive Control (MPC) can handle multivariable constraints and system lag problems well, its control performance is highly dependent on an extremely accurate prior mathematical mechanism model. In the long-term operation of actual equipment, heat exchangers inevitably experience phenomena such as fouling on the tube walls and metal fatigue, which leads to a gradual change in the thermal response characteristics of the system. This evolution of physical characteristics causes the performance of traditional MPC based on fixed parameters to degrade significantly, and the cost and technical threshold for re-identifying the mechanism model in the industrial field are extremely high.

[0005] In recent years, Deep Reinforcement Learning (DRL) has shown great potential in the field of adaptive control of complex systems. However, there are insurmountable obstacles to directly applying traditional DRL to high-value industrial heat exchanger applications: First, the trial-and-error cost is extremely high, and real physical equipment cannot withstand the serious safety risks brought about by the algorithm's initial random exploration. Second, the purely data-driven "black box" model lacks physical boundary constraints and is prone to outputting dangerous commands that violate the common sense of heat transfer when encountering extreme operating conditions with few samples. Finally, in the face of long-term gradual changes such as equipment fouling, conventional DRL models are prone to "catastrophic forgetting" in incremental learning, that is, in order to adapt to the current aging and derailed operating conditions, they completely lose the ability to control the initial clean operating conditions.

[0006] Digital twins, as a core technology for realizing the integration of cyber-physical systems, provide a new path to solve the aforementioned risks of trial and error by constructing a real-time synchronous mapping between virtual and physical entities. However, existing digital twin technologies are mostly limited to passive monitoring of equipment status, fault diagnosis, or offline simulation.

[0007] How to deeply integrate digital twins with world models that have long-term probabilistic extrapolation capabilities, and construct a "trial and error sandbox" in virtual space that both follows the physical mechanism of energy conservation and can overcome operating condition drift and catastrophic forgetting, and use this to safely drive reinforcement learning agents to achieve millisecond-level online real-time closed-loop control, is a core technical problem that urgently needs to be solved in the field of complex industrial thermal control. Summary of the Invention

[0008] To overcome the shortcomings of existing technologies, this invention provides a method for precise control of heat exchanger outlet temperature based on a world model and digital twin. This method constructs a digital twin based on a Recurrent State Space Model (RSSM), combines it with physical law constraints, performs risk-free training of reinforcement learning strategies in the latent space, and introduces a distributed out-of-systems detection mechanism to achieve high-precision, adaptive, and safe control of the heat exchanger.

[0009] A heat exchanger outlet temperature control method based on a world model and digital twin, characterized by comprising the following steps: S1. Establish a digital twin of the heat exchanger based on a world model architecture; The world model architecture includes a perceptual coding network, a latent space dynamics model, and an observation decoding network. The sensing coding network will collect the high-dimensional real-time operating state vector of the heat exchanger. Mapping to a low-dimensional latent space, we can extract the latent states that imply the system's thermal inertia and heat transfer attenuation characteristics. ; The latent space dynamics model learns by minimizing the training loss function given the operating variables. Under incentives, the current potential state Next time potential state The evolution probability distribution is used to characterize the nonlinear dynamic response process of the heat exchanger outlet temperature under complex operating conditions. The observation decoding network maps the latent state back to the original observation space and outputs a reconstructed system state value including the heat exchanger outlet temperature, so as to verify the latent state characterization capability and physical constraint verification. S2. Deploy reinforcement learning control agents in the latent space of the world model construction: Use the latent space dynamics model to perform multi-step recursive reasoning to generate virtual evolution trajectories composed of predicted latent state sequences. The control agent explores and interacts in the virtual evolution trajectory with the core objective of minimizing the cumulative tracking error between the predicted exit temperature and the set target value. The action execution policy network is optimized by gradient descent. S3. Execute closed-loop control during real-time operation: Acquire current sensor data of the heat exchanger during real-time operation and input it into the sensing coding network to obtain the real-time latent state. The optimal action execution policy network, trained to its full potential, is based on the real-time latent state. The direct mapping output provides compensation control commands for the operated variables, which are then sent to the actuator to achieve precise closed-loop regulation of the heat exchanger outlet temperature.

[0010] Due to the adoption of the above technical solution, the present invention has the following beneficial effects: 1. This invention uses RSSM to construct the latent space dynamics model of the digital twin. Through its internal dual-path parallel architecture of determinism and stochasticity, it can simultaneously and accurately capture the long-term thermal hysteresis characteristics of the heat exchanger caused by heat capacity, as well as the system thermal balance uncertainty caused by external flow peaks, ambient temperature changes, etc., which significantly improves the system's ability to predict the evolution of complex and strongly nonlinear operating conditions.

[0011] 2. This invention innovatively introduces a thermal balance consistency constraint term based on the law of energy conservation into the training loss function of the world model; this design forces the state deduction of the digital twin to strictly conform to the basic physical laws of heat transfer, fundamentally eliminating the unreliability of pure data-driven "black box" models when faced with few samples or unseen operating conditions, which are prone to outputting dangerous instructions that violate physical common sense.

[0012] 3. In the practical stage, this invention introduces an out-of-distribution (OOD) extreme condition detection mechanism based on variational free energy. When the system detects that the real-time characteristics deviate from the known distribution, it can instantly become alert and automatically trigger a safety takeover, seamlessly switching control to a preset traditional PID robust controller. While pursuing the ultimate artificial intelligence optimized control, it safeguards the safety bottom line of industrial production.

[0013] 4. In the reinforcement learning strategy training stage, this invention combines a full life cycle experience playback mechanism. By proportionally mixing and extracting historical samples from the entire life cycle, such as cleaning benchmarks, gradual fouling, and sudden disturbances, for joint training, it effectively prevents the deep learning model from "catastrophic forgetting" when adapting to gradual characteristics such as equipment aging and parameter drift, and gives the system seamless adaptive capability throughout the entire life cycle of the heat exchanger.

[0014] 5. This invention adopts an asynchronous deployment architecture of cloud-edge collaboration. It utilizes the massive computing power of the cloud to carry out high-intensity offline heavy training and periodic iteration of the world model and control strategy, while only a lightweight network is deployed at the edge close to the field to carry out millisecond-level latent state feature extraction and control command inference. This perfectly balances the advanced nature of the continuous evolution of complex AI algorithms with the ultimate real-time performance of the closed-loop control of the industrial foundation.

[0015] 6. When deploying reinforcement learning agents in the latent space, this invention sets the time domain length of the virtual evolution trajectory to be greater than the thermal inertia time constant of the heat exchanger. This mechanism forces the action execution strategy network to learn "advanced compensation" control for large time delay systems through millions of trials and errors, completely overcoming the slow response of traditional control strategies, significantly reducing temperature overshoot and shortening the adjustment and stabilization time. Attached Figure Description

[0016] Figure 1 This is a flowchart illustrating the overall control method of the present invention. Figure 2 This is a block diagram of the digital twin structure based on the world model architecture of this invention; Figure 3 This is a schematic diagram illustrating the principle of reinforcement learning agent deduction and training within the latent space of this invention. Figure 4 This is a diagram of the cloud-edge collaborative deployment and security takeover architecture of the system of this invention. Detailed Implementation

[0017] The exemplary embodiments of the present invention will be described in detail below with reference to the accompanying drawings. It should be emphasized that the exemplary embodiments described are merely illustrative and do not represent all possible implementations covered by this disclosure. Those skilled in the art can make equivalent substitutions or improvements based on the principles of the present invention without departing from the spirit and scope of the present invention.

[0018] Example: This example uses the global control of natural gas outlet temperature in a heat exchanger network (HEN) within a natural gas differential energy utilization project as a specific application scenario. In natural gas differential energy power generation or pressure regulation processes, to ensure system safety and gas supply quality, the heat exchanger network typically includes two core nodes: 1. Preheater before expansion: It is used to heat the natural gas before it enters the expander to do work, so as to prevent the equipment from freezing or forming hydrate blockage due to drastic temperature drop. This part faces extremely strong random disturbances in the upstream gas source pressure and flow, and the requirements for the safety bottom line and disturbance resistance of the control are extremely high.

[0019] 2. Post-expansion temperature controller: Used to perform final heating on the low-temperature natural gas after it has been used to ensure that the outlet temperature of the natural gas delivered to the city pipeline network or user end is stable within the national standard range; the requirements for temperature control stability and high-precision tracking are extremely high here.

[0020] Faced with the complex large inertia system composed of a preheater and a thermostat connected in series, traditional PID controllers are prone to coupled oscillations and slow response. To address this, this embodiment adopts a control method based on a world model and digital twin to construct a "virtual sandbox" that can provide a comprehensive overview and accurately predict the future, thereby achieving advanced collaborative control.

[0021] The flowchart for constructing the digital twin of the heat exchanger in this embodiment is as follows: Figure 1 As shown, it includes: S1: Establish a digital twin of the heat exchanger based on a world model architecture; This step aims to construct a dynamic digital twin of the HEN in latent space, including the preheater and thermostat, using a deep learning network. Its digital twin structure, based on a world model architecture, is as follows: Figure 2 As shown; the world model architecture mainly consists of three parts: a perceptual coding network, a latent space dynamics model, and an observation decoding network. Its specific implementation details are as follows: During the real-time operation of the heat exchanger network, the data acquisition system obtains the global high-dimensional real-time operating state vector at a fixed sampling period. In this embodiment, Specifically, this includes: upstream natural gas inlet temperature and instantaneous flow rate, preheater heat source inlet temperature, expander inlet and outlet temperatures and pressure difference, thermostat heat source inlet temperature, final natural gas outlet temperature, and ambient temperature. The sensing-coding network receives the above-mentioned multi-source heterogeneous sensor data. It is then fused with the memory state from the previous moment, mapped to a low-dimensional latent space, and the posterior latent state is output. This global latent state It is a highly compressed feature vector that filters out field instrument noise and implicitly extracts and stores the overall heat storage state (system thermal inertia) of the preheater and thermostat pipe walls, as well as the degree of scaling on the pipe walls of each device due to operation (heat transfer attenuation characteristics).

[0022] To simultaneously address the long-term deterministic thermal response of the global system and the short-term stochastic changes in the upstream natural gas source, the latent space dynamics model in this embodiment adopts the RSSM architecture, which contains two parallel paths: 1. Deterministic state path: Using gated cyclic units, the deterministic hidden states are calculated. This path is responsible for storing the large hysteresis physical inertia of the superimposed series heat exchanger networks; among them, the operational variables are... It is a control vector that contains the preheater heat source valve opening command and the thermostat heat source valve opening command applied at the previous moment.

[0023] 2. Stochastic State Path: Based on Deterministic Hidden States Output the prior probability distribution of the current state evolution. This approach endows the twin with the ability to represent uncertainties when facing unknown disturbances such as drastic fluctuations in pipeline flow; through the above mechanism, the dynamic model can learn the characteristics of a given joint operating variable. Under incentives, the transition probability of state evolution This allows for the nonlinear deduction of the future temperature evolution trajectory of the entire heat exchange network within the virtual latent space.

[0024] To verify the latent state To determine whether it truly reflects physical laws, the system is equipped with an observation-decoding network to handle hidden states. Remap back to physical space and output the reconstructed observations. (The core includes the reconstructed preheated natural gas temperature forecast and the final natural gas outlet temperature forecast) ).

[0025] To ensure that the model strictly adheres to energy conservation when extrapolating this complex series system and avoids prediction distortions that violate the first law of thermodynamics, this invention employs a joint optimization objective incorporating physical mechanism constraints during the offline training phase of the model. Its total loss function... The calculation formula is as follows: Among them, the basic reconstruction loss in the above formula Global observations used to compute decoded reconstruction Errors between the data and those acquired by real sensors are considered to ensure the accuracy of the underlying feature mapping; distribution difference loss. The KL divergence between the prior distribution predicted by the dynamic model and the posterior distribution extracted by the sensing network is used to calculate the digital twin model's long-term predictive capability; the mechanistic constraint term for thermal equilibrium consistency. The theoretical outlet temperature of the natural gas after preheating and temperature regulation is calculated using basic formulas of heat transfer and thermodynamics, based on the collected natural gas mass flow rate, specific heat capacity at constant pressure, flow rate of each heat source fluid, and inlet and outlet temperature difference of each heat exchanger. This constraint is specifically defined as the final outlet temperature after reconstruction. The residual penalty between the above theoretical temperature and the temperature is... Furthermore, in the formula , , These are all corresponding weight coefficients; during model training, the system adjusts... The weighting of values ​​is used to ensure the rigidity of physical mechanism constraints, ensuring that the model achieves the best balance between fitting accuracy to complex data and strictness of physical laws.

[0026] S2: Deploy reinforcement learning control agents in the latent space of world model construction; The principle behind the reinforcement learning agent's multi-step inference and policy optimization within the latent space in this step is as follows: Figure 3 As shown; at any training moment in the virtual sandbox The control agent (i.e., the action execution policy network) determines the global latent state based on the currently extracted data. Output a tentative joint control action (Including the opening degree of the heat source valves of the preheater and thermostat); the latent space dynamics model receives this action and, combining its own deterministic and stochastic paths, predicts the latent state at the next moment. By iterating this process repeatedly, the system can rapidly generate a virtual evolution trajectory consisting of a sequence of predicted latent states within the algorithm's virtual environment in a very short time. To completely overcome the enormous thermal inertia and pure time lag characteristics generated by the superposition of series heat exchange networks, the time domain length of the virtual evolution trajectory in this embodiment is... It is strictly set to be greater than the thermal inertia time constant of the entire heat exchange network.

[0027] The optimal objective for controlling the agent's exploration within the virtual trajectory is achieved by continuously optimizing the action execution policy network using gradient descent to maximize the cumulative reward function. To balance the high-precision tracking of natural gas post-heating with the anti-icing safety of preheating, the cumulative reward function in this embodiment... The specific calculation formula is as follows: ; Specifically, the formula covers three core engineering objectives: the first is a temperature control tracking reward, which aims to calculate the final predicted natural gas outlet temperature reconstructed by the observation decoding network. Set temperature target value with urban pipe network The squared error between the two terms enables the thermostat to achieve extremely high control precision; the second term is a smoothness penalty term, in which... The term used to characterize the rate of change of action of each actuator (control valve) is used to severely punish severe valve vibration by calculating its corresponding norm square term, thereby protecting the mechanical life of the real physical valve and preventing water hammer effect in the pipeline network; the third term is a risk penalty term for extreme latent states. This is specifically designed to prevent icing in preheaters, when in submerged state When the mapped internal physical state approaches the freezing point or the temperature boundary formed by hydrates, the system will impose a significant negative penalty to prevent dangerous actions from occurring at the source of the strategy; where As a time discount factor, , , These are all weighting coefficients, used to balance the above multidimensional optimization objectives.

[0028] During long-term high-pressure operation, fouling inevitably accumulates on the tube bundle surface of natural gas heat exchangers, leading to a decrease in the overall system heat transfer coefficient and a slower thermal response. Conventional deep learning models, when incrementally learning and adapting to older equipment with fouling, are prone to losing their ability to control the rapid response characteristics of clean equipment, i.e., "catastrophic forgetting." To address this, this embodiment incorporates a full lifecycle experience playback mechanism in strategy training. The system sets up a sample buffer in the cloud containing typical operating conditions throughout the heat exchanger's entire lifecycle. When constructing training batches, historical data is extracted and jointly trained according to a preset ratio (in this embodiment, 30% clean baseline operating condition samples, 50% gradually fouling operating condition samples, and 20% natural gas flow change disturbance operating condition samples are used). Through continuous strategy updates on the mixed samples, the forced action execution strategy network retains dual memory of the rapid thermal response characteristics under clean conditions and the large hysteresis characteristics under fouling conditions, ensuring that the control agent always has robust global adaptive capabilities throughout the heat exchanger's several-year service life.

[0029] S3: Perform closed-loop control during the real-time operation phase; After offline strategy training is completed, the control system enters the real-time closed-loop operation phase to implement online control of the natural gas heat exchange network. To meet the stringent requirements of industrial sites for real-time response and operational safety of the control algorithm, this embodiment constructs a cloud-edge collaborative asynchronous deployment architecture and embeds a distributed external detection and security takeover mechanism. Its system cloud-edge collaborative deployment and security takeover architecture is as follows: Figure 4 As shown, the specific implementation method is as follows: Millisecond-level closed-loop control based on cloud-edge collaborative architecture: In view of the high-frequency time-varying characteristics of natural gas pipeline network operation and the high computing power requirements for training digital twin models, the world model and control strategy described in this invention adopt an asynchronous deployment architecture of cloud-edge collaboration; wherein, the cloud computing layer utilizes massive historical operating data to perform high-intensity offline training of the world model and periodic iteration of the action execution strategy network; the edge computing layer only carries the lightweight processing of the perception encoding network, observation decoding network, and optimal action execution strategy network.

[0030] During the real-time operation phase, the edge computing layer acquires the current heterogeneous sensor data of the heat exchange network at a millisecond sampling period and inputs it into the sensing coding network to extract the global latent state in real time. Subsequently, the optimal action execution strategy network, issued from the cloud, is based on this real-time latent state. It directly maps and outputs compensation control commands for the operating variables (i.e., the opening degree of the heat source valves of the preheater and thermostat), and immediately sends them to the underlying actuators, thereby achieving rapid and precise closed-loop regulation of the natural gas outlet temperature. This architecture avoids the communication delay caused by complex online optimization calculations and takes into account both the algorithm evolution capability and the real-time control requirements of the industrial underlying layer.

[0031] Out-of-distribution detection and safe takeover mechanism based on variational free energy: Although the action execution strategy network has completed optimization in the latent space through mixed samples, actual industrial sites may still face rare out-of-distribution conditions not represented by the world model, such as mechanical failure of actuators or extreme weather conditions. In order to avoid the risk of system instability and pipeline ice blockage caused by the output of unbounded control commands by the pure data-driven model under such out-of-distribution conditions, this embodiment adds an out-of-distribution detection link in the edge control cycle: During real-time operation, the system synchronously calculates the variational free energy of the perception coding network for the current input data, or calculates the reconstruction error output by the observation decoding network. This index quantitatively characterizes the degree to which the current physical condition deviates from the distribution of the model's historical training data. When the variational free energy or reconstruction error exceeds a preset safety threshold, the system determines that the current system has entered an out-of-distribution extreme condition that has not been learned by the world model. At this time, the system automatically triggers a safety takeover mechanism, immediately blocking the control output of the action execution strategy network and smoothly switching the control of the actuator to a preset traditional PID robust controller. At the same time, the edge computing layer synchronously uploads the high-value abnormal condition data that triggered the out-of-distribution alarm to the cloud computing layer, triggering online incremental learning of the world model. After the cloud completes the parameter update for the extreme condition, the latest network weights are asynchronously sent to the edge, thereby realizing continuous closed-loop iteration of the control system under the premise of ensuring absolute safety.

[0032] The above description is merely a specific embodiment of this application, but the scope of protection of this application is not limited thereto; any variations or substitutions that can be easily conceived by those skilled in the art within the scope of the technology disclosed in this application should be included within the scope of protection of this application; therefore, the scope of protection of this application should be determined by the scope of the claims.

Claims

1. A method for controlling the outlet temperature of a heat exchanger based on a world model and digital twin, characterized in that, Includes the following steps: S1. Establish a digital twin of the heat exchanger based on a world model architecture; The world model architecture includes a perceptual coding network, a latent space dynamics model, and an observation decoding network. The sensing coding network will collect the high-dimensional real-time operating state vector of the heat exchanger. Mapping to a low-dimensional latent space, we can extract the latent states that imply the system's thermal inertia and heat transfer attenuation characteristics. ; The latent space dynamics model learns by minimizing the training loss function given the operating variables. Under incentives, the current potential state Next time potential state The evolution probability distribution is used to characterize the nonlinear dynamic response process of the heat exchanger outlet temperature under complex operating conditions. The observation decoding network maps the latent state back to the original observation space and outputs a reconstructed system state value including the heat exchanger outlet temperature, so as to verify the latent state characterization capability and physical constraint verification. S2. Deploy reinforcement learning control agents in the latent space of the world model construction: Use the latent space dynamics model to perform multi-step recursive reasoning to generate virtual evolution trajectories composed of predicted latent state sequences. The control agent explores and interacts in the virtual evolution trajectory with the core objective of minimizing the cumulative tracking error between the predicted exit temperature and the set target value. The action execution policy network is optimized by gradient descent. S3. Execute closed-loop control during real-time operation: Acquire current sensor data of the heat exchanger during real-time operation and input it into the sensing coding network to obtain the real-time latent state. The optimal action execution policy network, trained to its full potential, is based on the real-time latent state. The direct mapping output provides compensation control commands for the operated variables, which are then sent to the actuator to achieve precise closed-loop regulation of the heat exchanger outlet temperature.

2. The heat exchanger outlet temperature control method based on world model and digital twin according to claim 1, characterized in that, The latent space dynamics model described in step S1 adopts the Recurrent State Space Model (RSSM) architecture. The RSSM architecture includes deterministic state paths and stochastic state paths. The deterministic state paths are used to capture the long-term thermal hysteresis characteristics of the heat exchanger caused by the heat capacity of the tube wall and the fluid, ensuring the time consistency of the model in long-term time domain predictions. The stochastic state paths are used to characterize the system thermal balance uncertainty caused by inlet flow fluctuations and ambient temperature disturbances, enhancing the robustness of the model to random disturbances.

3. The heat exchanger outlet temperature control method based on world model and digital twin as described in claim 1, characterized in that, The training loss function of the world model described in step S1 is the total loss function. The calculation formula is as follows: in, The reconstruction loss based on the basic representation is used to calculate the error between the reconstructed observations output by the observation decoding network and the real sensor data. The distribution difference loss of the latent space dynamics model is used to calculate the KL divergence between the prior distribution of the predicted state and the posterior distribution of the output of the perceptual coding network. This is a thermal balance consistency constraint term; , , All are weighting coefficients.

4. The heat exchanger outlet temperature control method based on world model and digital twin according to claim 1, characterized in that, The optimization objective of the control strategy described in step S2 is to maximize the cumulative reward function. The calculation formula is as follows: in, The time domain length of the virtual evolution trajectory; Discount factor; For the predicted time; To set the target temperature value; The predicted exit temperature is reconstructed from the world model through an observation decoding network. The norm square term is used to characterize the rate of change of the actuator's action, and its corresponding standard square term serves as a smoothing penalty term. This is a risk penalty item targeting extreme potential states; , , All are weighting coefficients; the time domain length The time constant is set to be greater than the thermal inertia time constant of the heat exchanger to ensure that the strategy network can learn a lead compensation strategy for the large time lag system.

5. The heat exchanger outlet temperature control method based on world model and digital twin according to claim 1, characterized in that, The strategy training described in step S2 incorporates a full lifecycle experience replay mechanism. A sample buffer containing typical operating conditions throughout the heat exchanger's lifecycle is set up. When constructing the training batch, clean baseline operating condition samples, gradual fouling operating condition samples, and sudden disturbance operating condition samples are mixed and extracted according to a preset ratio. By updating the strategy on the mixed samples, the action execution strategy network is forced to retain the memory of the rapid thermal response characteristics under clean conditions, preventing the model from catastrophically forgetting when adapting to the hysteresis characteristics of fouling.

6. The heat exchanger outlet temperature control method based on world model and digital twin according to claim 1, characterized in that, In step S3, an out-of-distribution detection step based on variational free energy is added: the variational free energy or reconstruction error of the perceptual coding network for the current input data is calculated in real time; when the variational free energy exceeds a preset safety threshold, it is determined that the current system has entered an out-of-distribution extreme condition that has not been learned by the world model. The system automatically triggers a safety takeover mechanism, switching control to a preset PID robust controller and simultaneously recording current operating data to trigger online incremental learning of the world model.

7. The heat exchanger outlet temperature control method based on world model and digital twin according to claim 1, characterized in that, The world model and control strategy adopt an asynchronous deployment architecture that combines cloud and edge: The cloud computing layer utilizes massive amounts of historical operational data for intensive offline training of the world model and periodic updates of the action execution policy network; The edge computing layer deploys a lightweight perception encoding network, observation decoding network, and action execution strategy network, which calculates out-of-distribution detection indicators while executing millisecond-level control command inference. The cloud and edge terminals collaborate through asynchronous parameter distribution and high-value sample uploading mechanisms to achieve continuous iteration of the control strategy.

Images