Oil-free magnetic suspension low-temperature refrigerating unit control method and storage medium

Abstract

The invention discloses an oil-free magnetic suspension low-temperature refrigerating unit control method and a storage medium, and relates to the technical field of intelligent refrigerating unit control, cooperative control between units is achieved through a multi-agent reinforcement learning framework, and a system can autonomously recognize and restrain the pressure fluctuation coupling phenomenon between parallel units; coherence and phase difference of pressure signals between units are monitored in real time by adopting a frequency domain analysis method, so that the system can predict surge risk in advance and adopt a phase desynchronization strategy, and coupling oscillation is avoided fundamentally; the design of the collaborative reward function embeds the anti-surge priority principle into the agent decision process, it is ensured that the anti-surge control weight is automatically enhanced when the system approaches the surge boundary, and meanwhile energy efficiency optimization and operation stability are considered; a near-end strategy optimization algorithm based on an attention mechanism enables each agent to effectively learn an interaction relationship between units, thereby generating a coordinated control instruction.

Description

Technical Field

[0001] This invention relates to the field of intelligent refrigeration unit control technology, and in particular to a control method and storage medium for an oil-free magnetic levitation cryogenic refrigeration unit. Background Technology

[0002] In the petrochemical and metallurgical industries, refrigeration units are needed to cool wastewater or exhaust gas emissions to separate target components from the total emissions and meet emission standards. Due to the large processing volume, multiple low-temperature refrigeration units need to be operated in parallel. As the processing volume is dynamic, multiple units need to work together to meet dynamic load requirements and improve system redundancy and reliability. Magnetic levitation compressors, with their oil-free operation and high-speed characteristics, provide a new technical approach for low-temperature conditions.

[0003] When multiple units are operating in parallel, the surge frequencies of each compressor may have a coupling effect. When the system load changes abruptly or the operating conditions approach the surge boundary, the pressure fluctuations of a single unit will be transmitted to other units through the shared pipeline, triggering a chain reaction of surge phenomena. This coupled vibration will cause periodic pressure oscillations in the system, affecting the stable operation of the entire refrigeration system. Most existing control schemes adopt independent surge protection strategies and lack a collaborative prevention and control mechanism between units.

[0004] In existing technologies, some cases use model predictive control algorithms to predict surge boundaries and use phase difference control strategies to stagger the operating frequencies of each unit; some systems also introduce real-time data sharing mechanisms so that each unit can obtain the overall operating status of the parallel system; such methods alleviate surge coupling problems to a certain extent, but there is still a response lag when dealing with rapid load changes. Summary of the Invention

[0005] In view of the aforementioned existing problems, the present invention is proposed.

[0006] The technical problem to be solved by this invention is to overcome the shortcomings of the prior art and provide a control method and storage medium for an oil-free magnetic levitation cryogenic refrigeration unit, which solves the problem of system oscillation caused by surge frequency coupling when multiple units are running in parallel, and the lack of an effective collaborative prevention and control mechanism in the existing control strategy.

[0007] To solve the above-mentioned technical problems, the present invention provides the following technical solution: This invention provides a control method for an oil-free magnetic levitation cryogenic refrigeration unit, comprising: Construct a multi-agent reinforcement learning framework with one agent corresponding to one generator unit; The operating status of each unit and shared pipeline is acquired in real time by the status observer; The joint reward is calculated based on a collaborative reward function that includes surge coupling suppression, system energy efficiency, and operational stability. A proximal strategy with centralized training and distributed execution is adopted to optimize and update the policies of each agent. The unit's execution unit issues and executes commands under a time synchronization mechanism to coordinate and adjust the compressor speed, guide vane opening, recirculation valve opening, and magnetic bearing control interface parameters.

[0008] As a preferred embodiment of the oil-free magnetic levitation cryogenic refrigeration unit control method described in this invention, the state observer acquires at least: the intake and exhaust pressures of each unit and their ratio, the fluctuation amplitude and phase difference of the pressure signals of the main pipe and the branch pipes, the dominant fluctuation frequency of the refrigerant flow in the common pipes, the rotational speed of each unit and its deviation from the target, the spectral characteristic value of the rotor vibration, and the inlet and outlet temperatures and flow rates of cooling water / chilled water.

[0009] As a preferred embodiment of the oil-free magnetic levitation cryogenic refrigeration unit control method described in this invention, the cooperative reward function is composed of a weighted sum of surge suppression reward term, energy efficiency reward term, and stability reward term; Each item is measured based on the correlation between the predicted surge margin and the inter-machine pressure, the unit cooling capacity power or energy efficiency ratio, and the main pipe pressure fluctuation and the actuator saturation degree. In the aforementioned collaborative reward function, at each control step, the total reward is written as a weighted sum of three terms: , Among them, R t For the joint reward at time t, These are the normalized scores for the three terms—surging resistance, energy efficiency, and stability—at time t. For the corresponding weights, and ; Based on the principle of prioritizing asthma prevention over energy efficiency, and energy efficiency over stability, a risk-driven linear weighting priority is adopted as follows: , , , , in, Let M be the predicted surge margin of the i-th unit. ref For safety reference margin, q t As an indicator of the risk of relapse, These are the upper and lower limits of the weights for the three items. This means truncating X to the interval ; The minimum surge margin penalty and the inter-machine pressure correlation penalty are combined into an anti-surge term, as follows: , in, For the anti-breathing item, For the composite weights, , , Let be the coherence coefficient of the exhaust pressure of unit i and j in the dominant frequency band. To correspond to the intra-band phase difference, the dominant frequency band is determined by a narrow band near the peak frequency of the main pipe pressure; Define energy efficiency terms with COP improvement as a positive benefit: , Where, r e (t) represents the energy efficiency term, COP. t For current comprehensive energy efficiency, COP t ref For rolling reference under similar loads and environments, COP max This serves as an upper limit for the design or an upper bound for historical quantiles. The stability term r is obtained by penalizing the variance of the main pressure, the dominant frequency band energy, the actuator saturation, and the total motion variation. s (t), written as: , ,in, , , and λ represents the penalty components of the main pressure variance, dominant frequency band energy, actuator saturation, and total motion variation, respectively. v , λ b , λ u , λ tv The weights of each penalty component within the stabilization term; In the formula: , , , , in, For the main pressure, V represents the variance of a sliding window of length W, where W can be either a fixed engineering window or an exponential equivalent window. ref For reference variance, f is the power spectral density. s B is the sampling frequency. t Dominant narrowband Where K is the power spectral density and K is the number of actuators. and The opening degree and upper limit of the travel of the Kth actuator Strategy action volume Normalized scale of movement stride.

[0010] In a preferred embodiment of the oil-free magnetic levitation cryogenic refrigeration unit control method described in this invention, the calculation of the surge suppression bonus includes: Monitor the margin from the operating point of each unit to the surge boundary; assess the correlation and coherence of pressure fluctuations in the parallel system; increase the anti-surge suppression weight when a coupled oscillation trend is identified, and trigger phase desynchronization control when the phase difference of pressure fluctuations is lower than a preset critical value; The method for identifying the coupling oscillation trend is as follows: Using the main pipe pressure and the exhaust pressure of each unit as subjective measurements, Welch spectrum estimation is used to obtain the power spectrum and cross power spectrum. The dominant frequency band is a narrow band around the peak frequency, and an in-band weighting function is set. First, the amplitude squared coherence of each generator pair is weighted and aggregated within the dominant frequency band to obtain the coupling strength index at time t: , in, B is an index of coupling strength. t The dominant frequency band at time t. The amplitude square coherence of units i and j at frequency f is calculated from the cross spectrum and the auto spectrum, and w(f) is the in-band weighting function. The phase difference representative value is defined as follows: the argument of the weighted integral within the cross-spectral band is used as the representative value of the paired phase differences, and the most in-phase pair is taken as the system phase representative: , Among them, Ф t G represents the phase difference at time t. ij (f) represents the cross-power spectrum of units i and j; A dual-threshold and phase hysteresis approach is used to suppress jitter, and an adaptive phase threshold value is given: , , , , , in, The upper threshold of coupling strength is represented by the set of uncoupled baselines B. t upper half , Indexing historical events Press and The coupling strength index calculated for the same aperture Describe set B t Element B in t The set of uncoupled baseline times up to time t. For the upper quantile operator, , The lower threshold , is the proportional coefficient for the desynchronization off threshold. The phase threshold for entering desynchronization. , Defined lower / upper bounds To exit the phase threshold of synchronization, , To indicate the state of desynchronization switch, α is the diameter of the diameter. When D t When =1, choose to achieve The units with the highest values ​​are prioritized, and phase desynchronization is achieved by applying a small phase delay or frequency fine-tuning to one of the units to increase Ф. t .

[0011] Furthermore, the near-end policy optimization includes a policy network and a value network, adopts a centralized training and distributed execution architecture, introduces an attention mechanism to model the interaction between groups, and uses pruning, importance sampling and experience replay for policy updates.

[0012] Furthermore, the actions output by the policy network include at least: The compressor speed setpoint increment, guide vane opening fine adjustment, recirculation valve or hot gas bypass valve opening, excitation current bias or target gap correction of the magnetic bearing control interface, and cooling water or chilled water flow ratio are all considered. Constraints are determined in the order of anti-surge priority over energy efficiency, and energy efficiency priority over comfort.

[0013] Furthermore, this includes surge warning: Based on the online updated surge boundary dynamic model, the surge margin of each unit is estimated in real time, and multi-level early warning thresholds are set. When multiple units are detected approaching the boundary simultaneously, the system enters a coordinated anti-surge mode, prioritizing the implementation of phase desynchronization and recirculation valve pre-opening strategies.

[0014] Furthermore, the unit controller shares status information through a high-speed network, and control commands are issued via timestamp-synchronized broadcast or multicast methods, with command transmission achieved through clock synchronization and redundant communication.

[0015] Furthermore, a safety shielding layer is set between the strategy output and execution to impose constraints on the actions and automatically fall back to the independent anti-surge control of each unit in the event of communication abnormalities or model mismatch.

[0016] Furthermore, the agent employs a hybrid training approach combining digital twin and real-world training. First, the digital twin model is pre-trained offline, then supervised fine-tuning is performed based on historical running data, and after deployment, the model is updated online in small steps and the playback data is recorded.

[0017] The present invention also provides a computer-readable storage medium, characterized in that the computer-readable storage medium includes a stored program, wherein the program, when running, controls the device where the computer-readable storage medium is located to execute the above-described control method.

[0018] The present invention also provides an oil-free magnetic levitation cryogenic refrigeration unit, which is composed of multiple magnetic levitation refrigeration units connected in parallel, and the magnetic levitation refrigeration units are controlled by the control method described above.

[0019] The beneficial effects of this invention are as follows: This invention achieves coordinated control between generating units through a multi-agent reinforcement learning framework. The system can autonomously identify and suppress pressure fluctuation coupling phenomena between parallel generating units. Frequency domain analysis is used to monitor the coherence and phase difference of pressure signals between units in real time, enabling the system to predict surge risks in advance and adopt phase desynchronization strategies, thus preventing coupled oscillations at their source. The design of the cooperative reward function embeds the anti-surge priority principle into the agent decision-making process, ensuring that the system automatically strengthens anti-surge control weights when approaching the surge boundary, while also considering energy efficiency optimization and operational stability. The proximal policy optimization algorithm based on the attention mechanism enables each agent to effectively learn the interaction relationships between generating units, thereby generating coordinated control commands. The distributed execution architecture combined with a time synchronization mechanism ensures the synchronization and consistency of control commands for multiple generating units, avoiding control conflicts caused by command asynchrony. The security shield provides multiple protections, automatically reverting to an independent anti-surge control mode in case of communication anomalies or model mismatch, ensuring that the system always operates within a safe range. The hybrid training strategy of digital twin and real machine further enhances the system's adaptability to different operating conditions, enabling the control strategy to be continuously optimized and updated.

[0020] This invention not only solves the surge coupling problem in parallel operation of multiple generating units, but also significantly improves the overall energy efficiency and operational stability of the system. By intelligently coordinating the operating states of each unit, it avoids the conservative operating strategies commonly found in traditional control, enabling the units to operate near their optimal operating points. Simultaneously, the system's adaptive capabilities allow it to cope with different load conditions and environmental changes, maintaining a highly efficient and stable operating state. Attached Figure Description

[0021] To more clearly illustrate the technical solutions of the embodiments of this application, the accompanying drawings used in the embodiments will be briefly described below. It should be understood that the following drawings only show some embodiments of this application and should not be regarded as a limitation on the scope of this application.

[0022] Figure 1 This is a flowchart illustrating the control method for the oil-free magnetic levitation cryogenic refrigeration unit in the embodiment. Detailed Implementation

[0023] To make the objectives, technical solutions, and advantages of this application clearer, the following detailed description is provided in conjunction with the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are merely illustrative and not intended to limit the scope of this application.

[0024] All terms used in this application (including technical and scientific terms) have the meanings commonly understood by those skilled in the art, unless otherwise defined. It should be noted that the terms used herein should be interpreted in a manner consistent with the context of this specification, and not in an idealized or overly rigid way.

[0025] For example, the terms “first” and “second” used in this application are only used to distinguish and describe similar objects, to differentiate the first object from another object, and are not used to describe a specific order or sequence, nor should they be interpreted as indicating or implying relative importance.

[0026] This application proposes a control method for an oil-free magnetic levitation cryogenic refrigeration unit, combined with... Figure 1 As shown, the method includes: Step S1: Construct a multi-agent reinforcement learning framework with one unit corresponding to one agent; Step S2: The status observer acquires the operating status of each unit and the shared pipeline in real time.

[0027] The oil-free magnetic levitation cryogenic refrigeration unit in this application is composed of multiple magnetic levitation refrigeration units connected in parallel, and the magnetic levitation refrigeration units are controlled by the control method described above. The magnetic levitation refrigeration unit can be a magnetic levitation centrifugal refrigeration unit.

[0028] In this embodiment, the state observer refers to a hardware and software set that synchronously collects and filters suction / discharge pressure, main and branch pressure, refrigerant flow rate, speed, and vibration. It can be constructed using existing sensors and data acquisition units, with denoising and time alignment completed within the edge controller. Specifically, the sampling period is 100ms by default and can be dynamically adjusted within the range of 50-200ms according to the field network and actuators. Synchronization alignment uses a combination of timestamp alignment and linear interpolation, allowing a maximum jitter of ±1 sampling period as the criterion for valid samples. Furthermore, pressure and flow signals undergo first- or second-order IIR denoising, with the cutoff frequency defaulting to 1.5-2.0 times the upper limit of the dominant oscillation frequency to avoid weakening the recognition bandwidth. Optionally, if some units do not have high-frequency vibration acquisition capabilities, this embodiment allows only speed, pressure, and flow rate to form a minimum observation set, maintaining the computable coupling criterion. Necessary anomaly handling includes: when any key quantity is continuously missing for more than 3 sampling periods, freezing the spectral estimation of that channel and maintaining the previous valid statistic, while marking the unit as a deweighted participant to reduce the risk of misjudgment.

[0029] Step S3: Calculate the joint reward based on the synergistic reward function that includes surge coupling suppression, system energy efficiency, and operational stability; Step S4: Optimize and update the policies of each agent using a proximal strategy of centralized training and distributed execution. In step S5, the unit execution unit issues and executes commands under the time synchronization mechanism to coordinate and adjust the compressor speed, guide vane opening, recirculation valve opening, and magnetic bearing control interface parameters to suppress coupled oscillations and improve energy efficiency. The recirculation valve is installed on the recirculation branch between the exhaust pipe and suction pipe of each unit compressor to regulate the refrigerant flow from the exhaust side to the suction side. The magnetic bearing control interface is used to send parameters such as excitation current bias and target gap correction output by the unit controller to the magnetic bearing control system to adjust the suspension state and vibration level of the compressor rotor.

[0030] Furthermore, in this embodiment, the recirculation valve is installed on the recirculation branch between the compressor discharge pipe and the suction pipe of each unit. It is used to return a portion of the high-pressure refrigerant from the compressor discharge side to the unit's suction side under low load or near surge boundary conditions, thereby expanding the stable operating range of single-unit and multi-unit parallel systems. Preferably, the recirculation valve is arranged on the anti-surge bypass pipe near the compressor outlet, and can be shared with or connected in parallel with the hot gas bypass valve. The multi-agent control strategy adjusts the recirculation valve opening in conjunction with compressor speed and guide vane opening adjustments to achieve fine-grained control of unit flow rate and pressure ratio, thereby suppressing surge-coupled oscillations caused by load disturbances.

[0031] In one embodiment, the state observer acquires at least: the intake and exhaust pressures of each unit and their ratio, the fluctuation amplitude and phase difference of the pressure signals of the main pipe and the branches, the dominant fluctuation frequency of the refrigerant flow in the common pipeline, the rotational speed of each unit and its deviation from the target, the spectral characteristic value of the rotor vibration, and the inlet and outlet temperatures and flow rates of cooling water / chilled water. Specifically, the dominant fluctuation frequency refers to the frequency component with the highest energy proportion and local peak power spectral density within the current sliding time window; the phase difference is the representative phase value obtained by cross-spectral argument on the selected dominant frequency band; spectral characteristic values ​​include indicators such as the main peak frequency, peak amplitude, and in-band energy proportion. In terms of numerical scope, the default sliding window length for spectral estimation is 32-64 s, the window step size is 2-4 s, the number of Welch segments is 8-16, and the overlap rate is 50-75%. If the sampling frequency is 10 Hz, the dominant frequency band search range is generally set to 0.05-2 Hz to cover common main pipe pressure self-excited oscillations. Optionally, spectral estimation can also use a multi-window or multi-cone method to maintain the same output scope as Welch. In abnormal situations, when the flow meter enters the lower limit of its range or the valve is fully closed, causing flow distortion, a pressure-derived virtual flow is used as a substitute input for frequency estimation, but it is not used for energy efficiency calculation.

[0032] In one embodiment, the cooperative reward function is composed of a weighted sum of a surge suppression reward term, an energy efficiency reward term, and a stability reward term; Each of the measures is based on the correlation between the predicted surge margin and the inter-machine pressure, the unit cooling capacity power or energy efficiency ratio, and the main pipe pressure fluctuation and the actuator saturation degree.

[0033] In the collaborative reward function, at each control step, the total reward is written as a weighted sum of three terms: , Among them, R t For the joint reward at time t, These are the normalized scores for the three terms—surging resistance, energy efficiency, and stability—at time t, with larger scores indicating better performance. , where is the corresponding weight, and ; In this embodiment, the surge risk index is generated based on the surge margin of each unit, where the margin is estimated by an online boundary model or interpolation table, and the reference margin is the unit-level safety lower limit. Numerically, the upper and lower limits of the weights can be set to surge prevention 0.4-0.9, energy efficiency 0.05-0.5, and stability 0.05-0.4, with the surge prevention weight increasing monotonically within the high-risk range. Furthermore, the weights are allowed to drift slowly every 1-5 seconds to avoid frequent switching. Optionally, when there is an external energy-saving priority command, the upper limit of the energy efficiency weight can be temporarily increased, but the lower limit of the surge prevention weight will not be exceeded. If necessary, if the confidence level of the margin model is lower than the threshold, the system will revert to a fixed weight configuration.

[0034] Based on the principle of prioritizing asthma prevention over energy efficiency, and energy efficiency over stability, a risk-driven linear weighting priority is adopted as follows: , , , , in, Let M be the predicted surge margin of the i-th unit. ref For safety reference margin, q t As an indicator of the risk of relapse, These are the upper and lower limits of the weights for the three items. This means truncating x to the interval .

[0035] The minimum surge margin penalty and the inter-machine pressure correlation penalty are combined into an anti-surge term, as follows: , in, For the anti-breathing item, For the composite weights, , , Let be the coherence coefficient of the exhaust pressure of unit i and j in the dominant frequency band. To correspond to the phase difference within the band, the dominant frequency band is determined by a narrow band near the peak frequency of the main pipe pressure.

[0036] In this embodiment, the narrowband width is defaulted to ±10-20% of the relative bandwidth of the main peak frequency, and the in-band weights can adopt a triangular or Gaussian shape to emphasize the center frequency. Numerically, the engineering default coherence threshold is 0.6-0.8, and the minimum effective energy percentage within the band is not less than 10%; the phase difference uses the principal value [-π, π] and its absolute value is used in the representative value calculation. In the formula parameters, the inter-unit coherence coefficient and phase difference are derived from the standard estimation of the cross spectrum and autospectrum. Optionally, when there is multi-peak competition, the frequency band with a higher energy percentage and continuous occurrence in adjacent windows is preferentially selected as the dominant band. In abnormal situations, such as when the peak frequency drift exceeds the bandwidth limit, the center of the dominant band is automatically reset and the in-band quantities are recalculated.

[0037] Define energy efficiency items with COP improvement as a positive benefit (if the on-site metering is in kW / RT, it can be converted to equivalent COP before substitution): , Where, r e (t) represents the energy efficiency term, COP. t For current comprehensive energy efficiency, COP tref For rolling reference under similar loads and environments, COP max This serves as an upper limit for the design or an upper bound for historical quantiles. In this embodiment, the comprehensive COP reference value adopts a two-dimensional rolling baseline of similar cooling capacity and outdoor wet-bulb temperature. The default window is data from the same load range over the past 7-14 days, with the 80-90th percentile as the upper bound. When only kW / RT metering is used, the equivalent COP is first obtained through standard conversion before normalization. In terms of numerical caliber, the reference baseline is updated every 30-60 minutes, and rapid updates are allowed when external conditions change drastically. Optionally, if the accuracy of power meters for multiple units is inconsistent, a light correction can be performed according to the type test curve to eliminate systematic bias. In case of anomalies, if the baseline sample is insufficient, it reverts to a dated version of the manufacturer's rated curve and reduces the weight of the energy efficiency item.

[0038] The stability term r is obtained by penalizing the variance of the main pressure, the dominant frequency band energy, the actuator saturation, and the total motion variation. s (t), written as: , ,in, , , and λ represents the penalty components of the main pressure variance, dominant frequency band energy, actuator saturation, and total motion variation, respectively. v , λ b , λ u , λ tv To stabilize the weights of each penalty component within a given item, they can be calibrated offline and fine-tuned online.

[0039] In the formula: , , , , in, For the main pressure, V represents the variance of a sliding window of length W, where W can be either a fixed engineering window or an exponential equivalent window. ref For reference variance, f is the power spectral density. s B is the sampling frequency. t Dominant narrowband Where K is the power spectral density and K is the number of actuators. and The opening degree and upper limit of the travel of the Kth actuator For strategy action quantity, The normalized scale for the action step length; in this embodiment, the actuator saturation is defined based on the physical stroke, speed, and protection limit of each actuator, and the total action variation is measured by the cumulative incremental change of the strategy output. Numerically, the speed ramp rate is assumed to be ≤5% / s, guide vane opening change ≤10% / s, recirculation valve change ≤15% / s, and magnetic bearing interface offset change ≤5% / s; the sliding window variance reference value can be taken as the 50-70th percentile of the historical steady-state low-vibration condition. In the formula parameters, K is the number of actuators participating in the decision-making. This represents the upper limit of available actuators for the k-th actuator. Optionally, the smoothness of the action can be relaxed at night when the load is low and tightened during the day when the load is high, achieving intraday adaptation. If necessary, once any actuator remains saturated for more than a preset duration (e.g., 2-5 seconds), the policy output is pruned and fallback logic is triggered.

[0040] In this embodiment, the surge suppression term is jointly measured based on the margin of the weakest unit and the correlation between inter-unit pressure; the energy efficiency term is evaluated using the online calculated comprehensive COP or equivalent EER; the stability term characterizes operational stability using main pipe pressure fluctuation, dominant frequency band energy, actuator saturation, and action smoothness. Numerically, the reward calculation and weight update default to the same control step size (100ms-500ms), while a first-order exponential smoothing of 0.8-0.95 is used for single-step rewards to suppress jitter. In the formula parameters, the comprehensive COP can be obtained from online metering of unit power and cooling capacity, and the reference value is generated using the rolling quantile or regression baseline of approximate operating conditions. Optionally, when only kW / RT is provided on-site, an energy efficiency score with the same caliber as COP can be generated through equivalent conversion. In case of anomalies, failure of energy efficiency measurement does not affect the calculation of surge prevention and stability components; the system automatically reduces the weight of the energy efficiency term to a preset lower limit.

[0041] Specifically, the reward quantification scheme is designed for multi-unit coupled scenarios. Structurally, it adopts a three-weighted approach, incorporating risk-priority control objectives alongside operational economy and stability, facilitating the switching of priorities between different operating conditions. The anti-surge term measures the margin of the weakest unit and inter-unit coherence, covering both individual unit safety and suppressing resonance tendencies. The energy efficiency term uses rolling references of similar operating conditions for comparison, avoiding evaluation bias caused by external environmental and load fluctuations. The stability term constructs penalties from four perspectives: statistics, frequency domain energy, hardware availability, and operational smoothness, reducing the probability of the system entering a high-oscillation and high-wear state. The weight allocation is driven by risk indicators and provides upper and lower limits to meet the engineering rules of prioritizing anti-surge. The overall expression remains differentiable and normalizable, which is conducive to the unified convergence of near-end strategy optimization in online and offline stages.

[0042] In one embodiment, the calculation of the surge suppression reward includes: Monitor the margin from the operating point of each unit to the surge boundary; assess the correlation and coherence of pressure fluctuations in the parallel system; increase the anti-surge suppression weight when a coupled oscillation trend is identified; and trigger phase desynchronization control to stagger the dominant oscillations between units when the phase difference of pressure fluctuations is lower than a preset critical value.

[0043] The method for identifying coupled oscillation trends is as follows: Using the main pipe pressure and the exhaust pressure of each unit as subjective measurements, Welch spectrum estimation is used to obtain the power spectrum and cross power spectrum. The dominant frequency band is a narrow band around the peak frequency (e.g., ±10% relative bandwidth), and an in-band weighting function is set for subsequent weighting.

[0044] In this embodiment, the window type for Welch spectral estimation can be either a Hanning or Heman window to reduce spectral leakage; the center of the in-band weighting function is aligned with the main peak frequency, and the half-maximum width is consistent with the narrowband width. Numerically, a spectral resolution of ≥0.02-0.05 Hz is recommended to ensure the stability of phase estimation. Optionally, when power frequency and its harmonics interference exist at the measurement point, notch filtering is allowed at fixed frequencies outside the band to improve the signal-to-noise ratio of the dominant band. Anomaly handling includes: if coherence cannot be calculated within two consecutive windows (e.g., due to insufficient samples), the coupling strength determination is paused and the previous valid state is maintained.

[0045] First, the amplitude squared coherence of each generator pair is weighted and aggregated within the dominant frequency band to obtain the coupling strength index at time t: , in, B is an index of coupling strength. t The dominant frequency band at time t. The amplitude square coherence of units i and j at frequency f is calculated from the cross spectrum and the autospectrum, and w(f) is the in-band weighting function.

[0046] The representative value of the phase difference is defined as follows: the argument of the weighted integral within the cross-spectral band is used as the representative value of the paired phase difference, and the most in-phase pair is taken as the representative of the system phase: , Among them, Ф t The phase difference at time t is represented by the value in radians; the smaller the value, the more in-phase the phase. G ij (f) represents the cross-power spectrum of units i and j.

[0047] A dual-threshold and phase hysteresis approach is used to suppress jitter, and an adaptive phase threshold value is given: , , , , in, The upper threshold of coupling strength is represented by the set of uncoupled baselines B. t upper half , Indexing historical events Press and The coupling strength index calculated for the same aperture Describe set B t The element in B t The set of uncoupled baseline times up to time t. For the upper quantile operator, , lower threshold , is the proportional coefficient for the desynchronization off threshold. The phase threshold for entering desynchronization. , Defined lower / upper bounds To exit the phase threshold of synchronization, , In this embodiment, the uncoupled baseline set is derived from steady-state periods far from the surge boundary and weighted by time decay to reflect seasonal and operational condition drift. Numerically, the upper threshold quantile can be 90-95%, and the lower threshold is given as 0.7-0.9 times the upper threshold. The lower limit of the phase entry threshold is 10-20° by default, the upper limit is 45-60°, and the exit threshold is 1.2-1.5 times the entry threshold to form a hysteresis band. Optionally, the threshold refresh cycle can be set to 5-10 minutes, triggering an immediate refresh when a significant operational condition transition occurs. If necessary, when the baseline length is insufficient or the quality is substandard, the system uses a fixed engineering threshold and limits the desynchronization action amplitude to a smaller range.

[0048] When D t When =1, choose to achieve The generator set with the highest value is prioritized, and phase desynchronization is achieved by applying a small phase delay or frequency adjustment (constrained within the upper limit of action) to one of the generator sets to increase Ф. t To avoid overlapping of the dominant frequency bands.

[0049] In this embodiment, phase desynchronization is achieved by applying a small perturbation to the selected unit's speed setpoint or a fine adjustment to the guide vane opening, and the duration and amplitude of the perturbation are constrained by the safety shielding layer. Numerically, it is recommended that a single phase perturbation not exceed ±0.5-1.0% of the target speed, last for 0.5-2 seconds, and slowly return within 1-3 seconds; when the desynchronization action causes an efficiency drop exceeding a preset threshold, the surge prevention weight is increased as compensation. Optionally, units with more sufficient margin are preferentially selected as the objects to be regulated. If necessary, if the desynchronization action triggers any hard interlocks or excessive vibration, it is immediately revoked and enters backoff control.

[0050] Specifically, the core observation is based on frequency band-weighted coherence and cross-spectral phase. The coupling strength is first aggregated, and then the cross-spectral argument is used to form a representative phase value, taking into account the coupling criteria on both sides of the amplitude and phase. The triggering logic adopts dual thresholds and phase hysteresis. The entry condition emphasizes the simultaneous satisfaction of strong coherence and near-in-phase, while the exit condition is relaxed to weak coherence or significant phase decoherence, thereby suppressing frequent switching. The phase threshold adaptively shrinks with the coupling strength. Strong coupling requires a larger phase misalignment, while weak coupling reduces the intervention intensity. The baseline threshold comes from the rolling quantile, which can adapt to seasonal and operating condition drift and facilitates connection with subsequent phase desynchronization execution strategies.

[0051] In one embodiment, proximal policy optimization includes a policy network and a value network, employs a centralized training and distributed execution architecture, introduces an attention mechanism to model the interaction between groups, and uses pruning, importance sampling, and experience replay for policy updates to enhance convergence stability.

[0052] In this embodiment, the policy and value networks can adopt a two- to three-layer feedforward structure with observation normalization performed before the attention layer; the training batch and time span are set according to the control step size and actuator lag. Numerically, the policy shearing coefficient is set to 0.1-0.3 by default, and the advantage function uses generalized advantage estimation with a discount factor of 0.95-0.99; the experience replay window is recommended to cover the past 10-30 minutes of running data, with an update cycle of 1-5 seconds. Optionally, the value network can share some feature extraction layers to reduce the number of parameters, but the output heads remain independent. If necessary, when the estimated advantage variance increases abnormally, the learning rate is automatically reduced and the target network update interval is extended.

[0053] In one embodiment, the actions output by the policy network include at least: The compressor speed setpoint increment, guide vane opening fine adjustment, recirculation valve or hot gas bypass valve opening, excitation current bias or target gap correction of the magnetic bearing control interface, and cooling water or chilled water flow ratio are all considered. Constraints are determined in the order of anti-surge priority over energy efficiency, and energy efficiency priority over comfort.

[0054] In a preferred embodiment, the recirculation valve of each unit is installed in the recirculation branch between the exhaust pipe and the suction pipe of the corresponding unit compressor. The opening degree of the recirculation valve or hot gas bypass valve refers to the control of the valve opening degree in the anti-surge bypass branch. The object of control is the refrigerant flow rate returning from the exhaust side to the suction side.

[0055] In this invention, the magnetic bearing control interface refers to the control signal interface between the unit's upper-level controller and the magnetic bearing control system. This interface can be a fieldbus communication interface and / or an analog or digital interface, used to send setpoints or biases related to rotor suspension to the magnetic bearing control system. Preferably, the control parameters sent through this magnetic bearing control interface include at least the magnetic bearing excitation current bias, rotor target clearance (target air gap) correction amount and / or equivalent stiffness, damping adjustment amount, etc., used to fine-tune the target suspension position of the compressor rotor in the radial and / or axial directions, limit the rotor vibration amplitude, ensure a safe clearance between the rotor and stationary parts, and work in conjunction with anti-surge, vibration suppression, and energy efficiency optimization strategies.

[0056] Specifically, in the actions output by the strategy network, the action components related to the magnetic bearing control interface are used to make small-scale online corrections to the suspension target of the magnetic bearing control system: when a significant increase in energy is detected in the dominant frequency band of the rotor vibration spectrum of a certain unit, and it approaches the surge coupling tendency, the multi-agent control strategy can adjust the excitation current bias and / or target gap of the corresponding unit through the magnetic bearing control interface. This allows the unit to change its rotor dynamic response characteristics while meeting mechanical safety and magnetic bearing capacity constraints, thereby working with phase desynchronization and recirculation valve regulation to jointly suppress coupled oscillations. The above adjustments are constrained by the safety shielding layer, and their amplitude and rate of change are limited to the range allowed by the magnetic bearing manufacturer, without changing the basic structure and operating mode of the magnetic bearing.

[0057] In this embodiment, the upper and lower limits of action and the rate limit are set according to the equipment nameplate and protection strategy, and are uniformly decided within the safety shielding layer. Numerically, priority decisions are made to meet the following sequence: minimum margin → vibration limitation → energy efficiency maintenance → smooth action. If multiple constraints are triggered simultaneously, the most stringent one is used, and the degree of violation is recorded for subsequent weight fine-tuning. Optionally, the cooling water / chilled water flow ratio action can be enabled within a gradually changing load range and frozen within a rapidly changing load range to avoid coupling amplification. If necessary, when communication lag exceeds the upper limit, only conservative adjustments to the local speed and recirculation valve are allowed.

[0058] In one embodiment, surge warning is also included: Based on an online-updated surge boundary dynamic model, the surge margin of each unit is estimated in real time, and multi-level early warning thresholds are set. In this embodiment, the early warning levels include at least three levels: attention, alarm, and emergency, and are linked to the collaborative reward weights. Numerically, the attention threshold can be set to 1.5-2.0 times the reference margin, the alarm threshold to 1.0-1.5 times, and the emergency threshold to ≤1.0 times; the thresholds are updated on a rolling basis every 5-10 minutes. Optionally, if the confidence level of the boundary model is lower than the set value, the early warning threshold is increased proportionally to increase conservatism. If necessary, the emergency level triggers forced load reduction and an increase in the pre-opening degree of the recirculation valve, and temporarily prohibits energy efficiency optimization actions.

[0059] When multiple units are detected approaching the boundary at the same time, the coordinated anti-surge mode is entered, and the phase desynchronization and recirculation valve pre-opening strategy is prioritized, and the load of some units is reduced when necessary. In one embodiment, the unit controller shares status information through a high-speed network, and control commands are issued in a timestamp-synchronized broadcast or multicast manner. Commands are transmitted through clock synchronization and redundant communication. Clock synchronization and redundant communication are required to comply with a high-precision time protocol to ensure that commands can be synchronized at the microsecond level.

[0060] In this embodiment, time synchronization is based on a high-precision time synchronization protocol. The master clock and slave clock automatically switch according to priority and periodically perform delay calibration. Numerically, the clock deviation is controlled within ±5-20µs, and the control command issuance frequency is 2-10Hz. Redundant channels can adopt dual-port or ring network topology and set the link heartbeat timeout to 100-300ms. Optionally, when the synchronization quality degrades but is still within a controllable range, the control system reduces the desynchronization action amplitude to reduce the impact of phase error. If necessary, once synchronization is lost or the link degrades to a single channel, it switches to local timing and enters conservative control.

[0061] In one embodiment, a safety shielding layer is set between the strategy output and execution, and constraints are imposed on the action to meet the limits such as minimum surge margin, maximum vibration, speed rise rate, actuator stroke and temperature change rate, and automatically fall back to the independent anti-surge control of each unit in the event of communication abnormality or model mismatch. In this embodiment, the safety shielding layer performs dual clipping on the amplitude and rate of the strategy output and verifies its consistency with the device protection logic. The rollback strategy retains the most recent stable setting and uses linear ramp regression for independent surge prevention. Numerically, the minimum surge margin and maximum vibration threshold are determined according to the device type profile and acceptance testing, and are by default not lower than the manufacturer's recommended safety lower limit. The minimum duration of rollback triggering is 1-3 seconds to filter transient spikes. Optionally, the shielding layer can set graded limiting for different actuators to balance response and protection. If necessary, if the rollback continues beyond the set duration, the system issues a maintenance prompt and maintains conservative control until manual confirmation.

[0062] In one embodiment, the agent employs a hybrid digital twin-machine training approach: First, the digital twin model is pre-trained offline, then supervised fine-tuning is performed based on historical operating data. After deployment, the model is updated online in small steps and the playback data is recorded to continuously adapt to different operating conditions and refrigerant charging status.

[0063] In this embodiment, the digital twin should reproduce the main thermodynamic and aerodynamic characteristics and be consistent with the noise levels of key sensors. The offline phase aims to cover common and boundary conditions; the supervised fine-tuning phase uses high-quality historical data from the most recent 1-3 months. Numerically, the online update learning rate is 1 / 5 to 1 / 10 of the offline phase, the experience replay window is 10-30 minutes by default, and a first-in-first-out update is used; the update frequency is every 5-15 minutes to avoid interfering with real-time control. Optionally, online updates can be enabled during off-peak hours and frozen during peak hours. If necessary, when the online evaluation index deteriorates beyond a threshold, the most recent parameter update is revoked, and the system reverts to the previous stable checkpoint.

[0064] This invention provides a computer-readable storage medium including a stored program, wherein the program controls the device where the computer-readable storage medium is located to execute the above-described control method when it is running.

[0065] This invention provides a processor for running a program, wherein the program executes a control method for an oil-free magnetic levitation cryogenic refrigeration unit.

[0066] This invention provides a device including a processor, a memory, and a program stored in the memory and executable on the processor. When the processor executes the program, it implements the steps of a control method for an oil-free magnetic levitation cryogenic chiller. The device described herein may be a server, PC, etc.

[0067] It will be apparent to those skilled in the art that the steps of the present invention described above can be implemented using general-purpose computing devices. They can be centralized on a single computing device or distributed across a network of multiple computing devices. They can be implemented using computer-executable program code, and thus can be stored in a storage device for execution by a computing device. In some cases, the steps shown or described can be performed in a different order than those presented herein, or they can be fabricated as separate integrated circuit modules, or multiple modules or steps can be fabricated as a single integrated circuit module. Thus, the present invention is not limited to any particular combination of hardware and software.

[0068] Those skilled in the art will understand that embodiments of this application can be provided as methods, systems, or computer program products. Therefore, this application can take the form of a completely hardware embodiment, a completely software embodiment, or an embodiment combining software and hardware aspects. Furthermore, this application can take the form of a computer program product embodied on one or more computer-usable storage media (including but not limited to disk storage, CD-ROM, optical storage, etc.) containing computer-usable program code.

[0069] This application is described with reference to flowchart illustrations and / or block diagrams of methods, apparatus (systems), and computer program products according to embodiments of this application. It will be understood that each block of the flowchart illustrations and / or block diagrams, and combinations of blocks in the flowchart illustrations and / or block diagrams, can be implemented by computer program instructions. These computer program instructions can be provided to a processor of a general-purpose computer, special-purpose computer, embedded processor, or other programmable data processing apparatus to produce a machine, such that the instructions, which execute via the processor of the computer or other programmable data processing apparatus, generate instructions for implementing the flowchart... Figure 1 One or more processes and / or boxes Figure 1 A device that provides the functions specified in one or more boxes.

[0070] These computer program instructions may also be stored in a computer-readable storage medium that can direct a computer or other programmable data processing device to function in a particular manner, such that the instructions stored in the computer-readable storage medium produce an article of manufacture including instruction means, which are implemented in a process Figure 1 One or more processes and / or boxes Figure 1 The function specified in one or more boxes.

[0071] These computer program instructions may also be loaded onto a computer or other programmable data processing equipment to cause a series of operational steps to be performed on the computer or other programmable equipment to produce a computer-implemented process, thereby providing instructions that execute on the computer or other programmable equipment for implementing the process. Figure 1 One or more processes and / or boxes Figure 1 The steps of the function specified in one or more boxes.

[0072] In a typical configuration, a computing device includes one or more processors (CPU), input / output interfaces, network interfaces, and memory.

[0073] Memory may include non-persistent memory in computer-readable media, such as random access memory (RAM) and / or non-volatile memory, such as read-only memory (ROM) or flash RAM. Memory is an example of computer-readable media.

[0074] Computer-readable media includes both permanent and non-permanent, removable and non-removable media that can store information using any method or technology. Information can be computer-readable instructions, data structures, modules of programs, or other data. Examples of computer storage media include, but are not limited to, phase-change memory (PRAM), static random access memory (SRAM), dynamic random access memory (DRAM), other types of random access memory (RAM), read-only memory (ROM), electrically erasable programmable read-only memory (EEPROM), flash memory or other memory technologies, CD-ROM, digital versatile optical disc (DVD) or other optical storage, magnetic tape, magnetic disk storage or other magnetic storage devices, or any other non-transferable medium that can be used to store information accessible by a computing device. As defined herein, computer-readable media does not include transient computer-readable media, such as modulated data signals and carrier waves.

[0075] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of this application, and are not intended to limit them. Although this application has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some or all of the technical features therein. Such modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the scope of the technical solutions of the embodiments of this application.

[0076] Furthermore, those skilled in the art will understand that although some embodiments herein include certain features included in other embodiments but not others, combinations of features from different embodiments are meant to be within the scope of this application and form different embodiments. For example, all the embodiments above can be used in any combination. The information disclosed in this background section is intended only to enhance the understanding of the general background of this application and should not be construed as an admission or in any way implying that such information constitutes prior art known to those skilled in the art.

Claims

1. A control method for an oil-free magnetic levitation cryogenic refrigeration unit, characterized in that, include, Construct a multi-agent reinforcement learning framework with one agent corresponding to one generator unit; The operating status of each unit and shared pipeline is acquired in real time by the status observer; The joint reward is calculated based on a collaborative reward function that includes surge coupling suppression, system energy efficiency, and operational stability. A proximal strategy with centralized training and distributed execution is adopted to optimize and update the policies of each agent. The unit's execution unit issues and executes commands under a time synchronization mechanism to coordinate and adjust the compressor speed, guide vane opening, recirculation valve opening, and magnetic bearing control interface parameters.

2. The control method for an oil-free magnetic levitation cryogenic refrigeration unit as described in claim 1, characterized in that, The condition observer must acquire at least the following: the intake and exhaust pressures of each unit and their ratio, the fluctuation amplitude and phase difference of the pressure signals of the main pipe and the branches, the dominant fluctuation frequency of the refrigerant flow in the common pipeline, the speed of each unit and its deviation from the target, the spectral characteristic value of the rotor vibration, and the inlet and outlet temperatures and flow rates of cooling water / chilled water.

3. The control method for an oil-free magnetic levitation cryogenic refrigeration unit as described in claim 1, characterized in that, The collaborative reward function is composed of a weighted average of a surge suppression reward term, an energy efficiency reward term, and a stability reward term; Each item is measured based on the correlation between the predicted surge margin and the inter-machine pressure, the unit cooling capacity power or energy efficiency ratio, and the main pipe pressure fluctuation and the actuator saturation degree. In the aforementioned collaborative reward function, at each control step, the total reward is written as a weighted sum of three terms: ， Among them, R t For the joint reward at time t, These are the normalized scores for the three terms—surging resistance, energy efficiency, and stability—at time t. For the corresponding weights, and ; Based on the principle of prioritizing asthma prevention over energy efficiency, and energy efficiency over stability, a risk-driven linear weighting priority is adopted as follows: ， ， ， ， in, Let M be the predicted surge margin of the i-th unit. ref For safety reference margin, q t As an indicator of the risk of relapse, These are the upper and lower limits of the weights for the three items. This means truncating x to the interval ; The minimum surge margin penalty and the inter-machine pressure correlation penalty are combined into an anti-surge term, as follows: ， in, For the anti-breathing item, For the composite weights, , , Let be the coherence coefficient of the exhaust pressure of unit i and j in the dominant frequency band. To correspond to the intra-band phase difference, the dominant frequency band is determined by a narrow band near the peak frequency of the main pipe pressure; Define energy efficiency terms with COP improvement as a positive benefit: ， Where, r e (t) represents the energy efficiency term, COP. t For current comprehensive energy efficiency, COP t ref For rolling reference under similar loads and environments, COP max This serves as an upper limit for the design or an upper bound for historical quantiles. The stability term r is obtained by penalizing the variance of the main pressure, the dominant frequency band energy, the actuator saturation, and the total motion variation. s (t), written as: ， ,in, , , and λ represents the penalty components of the main pressure variance, dominant frequency band energy, actuator saturation, and total motion variation, respectively. V , λ b , λ u , λ tv The weights of each penalty component within the stability term; In the formula: ， ， ， ， in, For the main pressure, V represents the variance of a sliding window of length W, where W can be either a fixed engineering window or an exponential equivalent window. ref For reference variance, f is the power spectral density. s B is the sampling frequency. t As the dominant narrowband, Where is the power spectral density, and k is the number of actuators. and The opening degree and upper limit of the travel of the k-th actuator For strategy action quantity, Normalized scale of movement stride.

4. The control method for an oil-free magnetic levitation cryogenic refrigeration unit as described in claim 3, characterized in that, The calculation of the surge suppression reward includes: Monitor the margin from the operating point of each unit to the surge boundary; assess the correlation and coherence of pressure fluctuations in the parallel system; increase the anti-surge suppression weight when a coupled oscillation trend is identified, and trigger phase desynchronization control when the phase difference of pressure fluctuations is lower than a preset critical value; The method for identifying the coupling oscillation trend is as follows: Using the main pipe pressure and the exhaust pressure of each unit as subjective measurements, Welch spectrum estimation is used to obtain the power spectrum and cross power spectrum. The dominant frequency band is a narrow band around the peak frequency, and an in-band weighting function is set. First, the amplitude squared coherence of each generator pair is weighted and aggregated within the dominant frequency band to obtain the coupling strength index at time t: ， in, B is an index of coupling strength. t The dominant frequency band at time t. The amplitude square coherence of units i and j at frequency f is calculated from the cross spectrum and the auto spectrum, and w(f) is the in-band weighting function. The phase difference representative value is defined as follows: the argument of the weighted integral within the cross-spectral band is used as the representative value of the paired phase differences, and the most in-phase pair is taken as the system phase representative: ， Among them, Ф t Let G be the representative value of the phase difference at time t. ij (f) represents the cross-power spectrum of units i and j; A dual-threshold and phase hysteresis approach is used to suppress jitter, and an adaptive phase threshold value is given: ， ， ， ， ， in, The upper threshold of coupling strength is represented by the set of uncoupled baselines B. t upper half , Indexing historical events Press and The coupling strength index calculated for the same aperture Describe set B t The element in B t The set of uncoupled baseline times up to time t. For the upper quantile operator, , The lower threshold , is the proportional coefficient for the desynchronization off threshold. The phase threshold for entering desynchronization. ， Defined lower / upper bounds To exit the phase threshold of synchronization, , To indicate the state of desynchronization switch, α is the diameter of the diameter. When D t When =1, choose to reach The units with the highest values ​​are prioritized, and phase desynchronization is achieved by applying a small phase delay or frequency fine-tuning to one of the units to increase Ф. t .

5. The control method for an oil-free magnetic levitation cryogenic refrigeration unit as described in claim 1, characterized in that, Proximal policy optimization includes a policy network and a value network. It adopts a centralized training and distributed execution architecture, introduces an attention mechanism to model the interaction between groups, and uses pruning, importance sampling and experience replay for policy updates.

6. The control method for an oil-free magnetic levitation cryogenic refrigeration unit as described in claim 1, characterized in that, The actions output by the policy network should include at least the following: The compressor speed setpoint increment, guide vane opening fine adjustment, recirculation valve or hot gas bypass valve opening, excitation current bias or target gap correction of the magnetic bearing control interface, and cooling water or chilled water flow ratio are all considered. Constraints are determined in the order of anti-surge priority over energy efficiency, and energy efficiency priority over comfort.

7. The control method for an oil-free magnetic levitation cryogenic refrigeration unit as described in claim 1, characterized in that, It also includes surge warning: Based on the online updated surge boundary dynamic model, the surge margin of each unit is estimated in real time, and multi-level early warning thresholds are set. When multiple units are detected approaching the boundary simultaneously, the system enters a coordinated anti-surge mode, prioritizing the implementation of phase desynchronization and recirculation valve pre-opening strategies.

8. The control method for an oil-free magnetic levitation cryogenic refrigeration unit as described in claim 1, characterized in that, A safety shield is set up between strategy output and execution to impose constraints on actions and automatically fall back to the independent anti-surge control of each unit in the event of communication abnormalities or model mismatch.

9. The control method for an oil-free magnetic levitation cryogenic refrigeration unit as described in claim 1, characterized in that, The intelligent agent employs a hybrid training approach combining digital twin and real-machine training. First, the digital twin model is pre-trained offline, then supervised fine-tuning is performed based on historical running data, and after deployment, the model is updated online in small steps and the playback data is recorded.

10. A computer-readable storage medium, characterized in that, The computer-readable storage medium includes a stored program, wherein, when the program is executed, it controls the device on which the computer-readable storage medium is located to perform the control method according to any one of claims 1 to 9.

Images