A Machine Learning-Based Dual-Reflux Control Method and System for a Two-Stage Refrigeration Centrifugal Compressor

By introducing a fast reflux branch and a main reflux branch into the refrigeration centrifugal compressor and combining them with a machine learning model, a rapid response and steady-state control of the refrigeration centrifugal compressor at low flow rates is achieved, solving the lag problem of traditional reflux control methods and improving the safety and efficiency of the unit.

CN123110128APending Publication Date: 2026-11-17SHANGHAI JIAOTONG UNIV
View PDF 1 Cites 0 Cited by

Patent Information

Application Number
CN202611618958.6
Authority / Receiving Office
CN · China
Patent Type
Applications(China)
Current Assignee / Owner
Filing Date
2026-10-20
Publication Date
2026-11-17

AI Technical Summary

Technical Problem

In existing technologies, refrigeration centrifugal compressors are prone to stall and surge at low flow rates. Traditional reflux control methods have a lag response and are difficult to generate effective flow at the compressor inlet, increasing the risk of unit operation, especially in the case of unsteady and random stall conditions.

Method used

A two-stage refrigeration centrifugal compressor dual reflux control system based on machine learning is adopted. By connecting a fast reflux branch and a main reflux branch in parallel, the fast reflux branch is used to achieve rapid flow increase and vibration suppression, while the main reflux branch is used for steady-state reflux. Combined with multi-point data acquisition and machine learning models, the flow distribution strategy is optimized in real time to achieve adaptive control.

Benefits of technology

It effectively broadens the safe operating range of the refrigeration centrifugal compressor at low flow rates, reduces the risk of surge, improves the safety and efficiency of unit operation, and reduces the compressor inlet temperature rise and efficiency loss caused by backflow.

✦ Generated by Eureka AI based on patent content.

Smart Images

  • Figure CN123110128A_ABST
    Figure CN123110128A_ABST
Patent Text Reader

Abstract

This invention provides a machine learning-based dual-reflux control method and system for a two-stage refrigeration centrifugal compressor, comprising: a rapid reflux branch and a main reflux branch connected in parallel between the compressor's discharge pipe and intake pipe; the rapid reflux branch has a small volume and does not have a dedicated cooling heat exchanger, enabling transient rapid flow increase under low-flow conditions; the main reflux branch is equipped with a heat exchanger to handle steady-state reflux cooling, and both reflux gases are led to the upstream intake pipe of the compressor. The system is equipped with a steady-state sensor and a high-frequency transient pressure sensor to collect unit operating data and construct a machine learning proxy model. After reflux is triggered, the rapid reflux branch is activated first, and the reflux flow is smoothly switched to the main reflux branch under flow conservation constraints. The proxy model is iteratively optimized online, and pressure fluctuations and efficiency thresholds are monitored to achieve fault alarm protection. This invention balances rapid response with reduced reflux efficiency loss, expanding the safe operating range of the unit under low-flow conditions.
Need to check novelty before this filing date? Find Prior Art

Description

Technical Field

[0001] This invention relates to the field of optimized design and safety control technology for refrigeration centrifugal compressors, specifically to a dual reflux control method and system for a two-stage refrigeration centrifugal compressor based on machine learning. Background Technology

[0002] Centrifugal compressors may experience stalling and surge at low flow rates. Stalling and surge cause significant periodic fluctuations in inlet and outlet pressures and flow rates, and can lead to shaft vibration, seal damage, impeller fatigue, and process interruptions. This is especially true for refrigeration centrifugal compressors, which operate at high speeds and often employ magnetic or air-bearing bearings to ensure smooth operation. Vibration in the unit can cause devastating damage to these high-speed bearings. Therefore, for refrigeration centrifugal compressors, it is crucial to continuously improve the compressor's unloading capacity while ensuring stable and safe operation of the unit at low flow rates.

[0003] Typically, installing a return pipe and regulating valve between the exhaust and intake pipes is the most common practice. However, in practical applications, the regulating valve's action is delayed, failing to immediately create effective flow at the compressor inlet, thus increasing the operational risk of the unit. Furthermore, the length and diameter of the return pipe, gas compressibility, valve actuator speed, and heat exchanger all contribute to pure hysteresis in the return flow. Additionally, for refrigeration centrifugal compressors, there is a design that directly introduces the return gas into the diffuser. However, on the one hand, two-stage centrifugal refrigeration compressors often use bladeless diffusers; on the other hand, stall occurrences are often unsteady and random. With changes in operating conditions and system conditions, the initial stall could occur at the inlet guide vanes or even at the first-stage moving vanes. Therefore, this design does not significantly expand the operating range for centrifugal compressors with bladeless diffusers.

[0004] Patent document CN215370159U discloses a surge prevention system for a centrifugal compressor and a refrigeration system. The surge prevention system includes: a refrigerant receiver, a multi-stage compressor, an evaporator, and a flash evaporator connected in sequence to form a closed loop; it also includes a gas supply line and a gas release line. The first end of the gas supply line is connected between adjacent stages of the multi-stage compressor, and the second end is connected to the flash evaporator. The first end of the gas release line is connected between adjacent stages of the multi-stage compressor, and the second end is connected to the evaporator. A gas supply valve and a gas release valve are respectively provided on the gas supply line and the gas release line.

[0005] Although patent document CN215370159U also has two pipelines, one is a gas supply pipeline, intended to improve efficiency (not to control surge), while the other is a gas venting pipeline, which is used for surge control. This patent still uses traditional surge control methods. In practical applications, the regulating valve often lags in its action, and the placement of the heat exchanger (evaporator) in the pipeline also introduces a pure lag in the return flow. Therefore, this method cannot immediately create an effective flow increase at the compressor inlet, thus increasing the operational risk of the unit. Furthermore, the gas venting circuit design in this patent involves interstage gas venting; however, compressor stall has strong unsteady and random characteristics. If stall occurs first in a later stage of the compressor, the gas venting return device cannot effectively alleviate the gas runoff, thus failing to achieve the purpose of controlling surge.

[0006] Therefore, there is a market need for a machine learning-based dual reflux system for two-stage centrifugal refrigeration compressors that can quickly respond to low-flow conditions and promptly resolve risks based on the operating characteristics of centrifugal compressors used in refrigeration, thereby ensuring the safe operation of the unit. Summary of the Invention

[0007] To address the shortcomings of existing technologies, the present invention aims to provide a dual reflux control method and system for a two-stage refrigeration centrifugal compressor based on machine learning, in order to solve the problem of balancing rapid response and continuous reflux of reflux gas, and to continuously improve the adaptive allocation strategy of dual-channel flow through machine learning.

[0008] According to the present invention, a machine learning-based dual-reflux control system for a two-stage refrigeration centrifugal compressor includes: the two-stage refrigeration centrifugal compressor comprising a first-stage impeller, a second-stage impeller, a first-stage bladeless diffuser, a second-stage bladeless diffuser, a volute, an exhaust pipe, and an intake pipe; A rapid return branch and a main return branch are connected in parallel between the exhaust pipe and the intake pipe; The rapid return branch is equipped with a rapid return branch regulating valve; the main return branch is equipped with a main return branch regulating valve and a main return branch heat exchanger; the return gas from both the rapid return branch and the main return branch is delivered to the intake pipe upstream of the compressor or an independent intake mixing chamber, and the return gas is not micro-jet injected towards the impeller outlet, diffuser, or blade tip region; the rapid return branch and the main return branch are respectively led out from different gas inlets of the exhaust pipe; The system also includes multiple sets of measurement sensors and controllers; The controller is configured with a machine learning agent model, which includes agent relationships f1, f2, and f3; agent relationship f1 is used to output the compressor's variable efficiency η. polTransient pressure fluctuation coefficient ΔP; proxy relationship f2 is used to output the valve position W1 of the fast return branch control valve; proxy relationship f3 is used to output the valve position W2 of the main return branch control valve; The controller is configured with the following control logic: after the return current is triggered, the fast return current branch is first activated to achieve rapid current increase and vibration suppression, and then gradually switched to the main return current branch. During the switching process, the decrease in return current of the fast return current branch is kept equal to the increase in return current of the main return current branch.

[0009] Preferably, the measuring sensor includes a steady parameter sensor and a high-frequency transient pressure sensor; The constant parameter sensor is used to collect the compressor inlet pressure P. in Inlet temperature T in Exhaust pressure P out Outlet temperature T out Import flow Q in Speed ​​N, pressure P1, temperature T1, flow rate Q1, and valve position W1 of the rapid return branch, and pressure P2, temperature T2, flow rate Q2, and valve position W2 of the main return branch; The high-frequency transient pressure sensors are respectively arranged at the compressor inlet, the first-stage impeller inlet, the first-stage diffuser inlet, the second-stage impeller inlet, and the second-stage diffuser inlet; the high-frequency transient pressure sensors have a frequency response greater than 100kHz and are used to collect multiple transient pressure P signals. t1 To P t5 And calculate the transient pressure fluctuation coefficient ΔP.

[0010] Preferably, the maximum flow rate of the fast return branch is 10% of the compressor's design flow rate. 30%; the maximum flow rate of the main return branch can reach 100% of the compressor's rated inlet mass flow rate.

[0011] Preferably, the machine learning agent model supports offline pre-training and online incremental updates; In the offline phase, the original samples are obtained through design calculations, one-dimensional system simulation, three-dimensional unsteady numerical simulation, and control valve flow characteristics to complete model training and solidification. During the online operation phase, the agent relationships f1, f2, and f3 are fine-tuned and updated using the incremental effective sample data from the actual unit measurements.

[0012] Preferably, the controller has a built-in alarm protection unit; when the transient pressure fluctuation coefficient ΔP exceeds the maximum vibration threshold ΔP, the controller will detect an alarm. max Or the variable efficiency η of the compressor pol When the efficiency falls below the preset threshold, the controller will issue an alarm, stop the branch flow switching operation, and increase the opening of the compressor inlet regulating valve to reduce the unit unloading rate.

[0013] According to the present invention, a machine learning-based dual-reflux control method for a two-stage refrigeration centrifugal compressor, employing the dual-reflux control system, includes the following steps: Step S1: Construct machine learning agent relations f1, f2, and f3 offline and complete model training and solidification; Step S2: Real-time acquisition of steady-state operating parameters of the compressor and its two return branches, as well as multiple transient pressure data, to calculate the compressor's multivariable efficiency η. pol And the transient pressure fluctuation coefficient ΔP; Step S3: When the real-time intake manifold flow rate drops to 1.05 times the minimum surge flow rate, or the transient pressure fluctuation coefficient ΔP exceeds the maximum vibration threshold ΔP max The backflow control is triggered in time; the fast backflow branch regulating valve is rapidly opened, while the main backflow branch regulating valve is gradually opened at the same time. Step S4: After the system traffic stabilizes at the target offload traffic level, perform branch traffic switching: reduce the traffic of the fast return branch and simultaneously increase the traffic of the main return branch, maintaining this adjustment process. ΔQ1=ΔQ2 until the fast return branch is completely closed, and all the return gas is cooled by the heat exchanger of the main return branch and then sent to the upstream of the compressor. Step S5: Collect valid running samples online and continuously update the proxy relationships f1, f2, and f3 incrementally; continuously monitor ΔP and η during the operation. pol If the threshold is triggered, a protection alarm will be executed.

[0014] Preferably, in the offline training phase of step S1, a set of working condition samples is generated by using Latin hypercube sampling or orthogonal sampling. The samples cover working conditions such as fast return branch fully open, two branches working simultaneously, fast branch switching to main branch, and near the anti-surge boundary. The samples are divided into training set, validation set, and test set, and the time window samples generated in the same continuous adjustment process belong to only a single dataset.

[0015] Preferably, the flow constraint is satisfied during the return flow trigger operation:

[0016] When the system flow rate drops to the required unloading flow rate level and gradually stabilizes, the main return branch control valve also reaches a certain opening degree. At this point... And satisfy:

[0017] The flow constraint is satisfied after the fast return branch is completely closed:

[0018] in, These represent the real-time traffic within the fast return branch and the main return branch, respectively.

[0019] Preferably, during incremental updates of the online model, the current online model version is replaced only if the updated surrogate model's prediction error on the independent validation dataset is not inferior to that of the current online model, and the prediction error for the anti-surge boundary condition meets the preset index; otherwise, the original model is used.

[0020] Preferably, when a fault is detected in the fast return branch regulating valve, the controller immediately outputs an alarm signal and actively increases the opening of the compressor inlet regulating valve to reduce the unit unloading rate.

[0021] Compared with the prior art, the present invention has the following beneficial effects: 1. This invention adopts a dual-recirculation architecture with a fast recirculation branch and a main recirculation branch connected in parallel. The recirculated gas from both branches is led to the upstream intake pipe of the compressor or an independent intake mixing chamber. Compared with the traditional single recirculation pipeline scheme, the fast recirculation branch is used to achieve rapid flow increase and vibration suppression under transient conditions, while the main recirculation branch undertakes steady-state continuous recirculation. This solves the contradiction between rapid response and continuous recirculation that is difficult to balance in a single recirculation system, and broadens the safe operating range of the two-stage centrifugal refrigeration compressor under low flow conditions.

[0022] 2. This invention configures a heat exchanger on the main return branch. After the operating conditions stabilize, the return load is gradually switched to the main return branch. Compared with the return scheme without cooling, the return gas is cooled down, reducing the compressor inlet temperature rise caused by the return and effectively compensating for the reduction in unit efficiency caused by the return.

[0023] 3. This invention uses multi-point acquisition of unit operation data, constructs a machine learning agent model, and iterates and optimizes the dual-branch flow distribution strategy in real time. Compared with the fixed valve position logic control method, it can dynamically match the opening degree of the two branches according to the actual operating conditions, and at the same time monitor pressure fluctuations and efficiency thresholds to realize abnormal alarms, thereby improving the control adaptability of the dual-return system and the safety of unit operation. Attached Figure Description

[0024] Other features, objects, and advantages of the present invention will become more apparent from the following detailed description of non-limiting embodiments with reference to the accompanying drawings: Figure 1 This is a structural diagram of a two-stage centrifugal compressor with a dual reflux device; Figure 2 This is a schematic diagram of a refrigeration system; Figure 3 The relationship between flow rate and efficiency of a two-stage centrifugal compressor unit.

[0025] in, Figure 1 The labels in the attached figures are as follows: 1. Main reflux branch regulating valve; 2. Fast reflux branch regulating valve; 3. Main reflux branch heat exchanger; 4. Fast reflux branch; 5. Main reflux branch; 6. Exhaust pipe; 7. Volute; 8. First-stage impeller; 9. First-stage bladeless diffuser; 10. Second-stage impeller; 11. Second-stage bladeless diffuser; 12. Inlet pipe. Detailed Implementation The present invention will now be described in detail with reference to specific embodiments. These embodiments will help those skilled in the art to further understand the present invention, but do not limit the invention in any way. It should be noted that those skilled in the art can make several changes and improvements without departing from the concept of the present invention. These all fall within the scope of protection of the present invention.

[0026] This invention directly delivers the parallel fast return branch and main return branch to the intake pipe upstream of the compressor impeller or an independent intake mixing chamber, without micro-jet injection to the impeller outlet, diffuser, or blade tip region. Through machine learning, it achieves adaptive adjustment of the dual-channel flow rate. This method effectively avoids the risks of surge and stall while increasing unloading capacity without altering the compressor impeller, diffuser, or volute, thus expanding the operating range of the centrifugal compressor.

[0027] According to the present invention, a machine learning-based dual-reflux system for a two-stage centrifugal refrigeration compressor is provided. The centrifugal compressor has two moving blades, two bladeless diffusers, a volute, an exhaust pipe, and an intake pipe. Specifically, as... Figure 1 As shown, the system includes a first-stage impeller 8, a second-stage impeller 10, a first-stage bladeless diffuser 9, a second-stage bladeless diffuser 11, a volute 7, an exhaust pipe 6, and an intake pipe 12. The first-stage impeller, second-stage impeller, first-stage bladeless diffuser, and second-stage bladeless diffuser are all installed inside the volute. One end of the volute is connected to the intake pipe, and the other end is connected to the exhaust pipe.

[0028] A rapid return branch 4 and a main return branch 5 are connected in parallel between the exhaust pipe and the intake pipe, which will be referred to as the rapid branch and the main branch respectively below.

[0029] The rapid recirculation branch has a small gas volume and a short response time, including the rapid recirculation branch regulating valve 2, and does not pass through a heat exchanger dedicated to cooling the recirculation gas. The gas volume of the rapid recirculation branch can be reduced by shortening the pipe length, reducing unnecessary cavities, or placing the rapid valve close to the inlet mixing point, but the pipe diameter of the rapid recirculation branch is not limited to a micro-nozzle size. The rapid recirculation branch valve can be a pneumatic, hydraulic, or electric rapid regulating valve. The rapid recirculation branch does not have a heat exchanger dedicated to reducing the temperature of the recirculation gas, but low-volume components for safety isolation and noise reduction are permitted. The maximum design flow rate of the rapid recirculation branch is 10%-30% of the compressor design flow rate, determined based on the compressor characteristic diagram, allowable inlet temperature, pipeline pressure rating, and maximum process disturbance.

[0030] The main return branch has a large flow capacity and includes a main return branch regulating valve 1 and a main return branch heat exchanger 3 located upstream or downstream of the main regulating valve. The maximum design flow rate of the main return branch is 100% of the rated inlet mass flow rate.

[0031] Both branches send the return gas to the intake pipe upstream of the compressor impeller or an independent intake mixing chamber, without micro-jet injection into the impeller outlet, diffuser, or blade tip region. The rapid return branch and the main return branch are led out from different gas inlets on the exhaust pipe and return to the mixing section on the intake pipe, respectively. By opening two gas inlets on the exhaust pipe, different gas inlets are achieved, facilitating the control of the two circuits. This effectively expands the operating range of the centrifugal refrigeration compressor without changing the internal flow channels of the compressor impeller and diffuser.

[0032] like Figure 2 As shown, the inlet of the two-stage centrifugal refrigeration compressor is connected to the inlet pipe 12, and the outlet is connected to the outlet pipe 6. The rapid return branch 4 and the main return branch 5 are led out from different air intakes on the outlet pipe 6 and return to the mixing section on the inlet pipe 12, respectively. A main return branch regulating valve 1 and a rapid return branch regulating valve 2 are respectively installed before the mixing section on each branch. A straight pipe section or rectifying component is retained between the mixing section and the impeller inlet to meet the requirements of flow field homogenization. An inlet total pressure probe, an inlet total temperature probe, and a flow meter are installed at the inlet; an outlet total pressure probe and an outlet total temperature probe are installed at the outlet. Pressure and temperature probes and a flow meter are installed on the rapid return branch; pressure and temperature probes and a flow meter are installed on the main return branch. Simultaneously, high-frequency response transient pressure sensors are installed at the compressor inlet, the first-stage impeller inlet, the first-stage diffuser inlet, the second-stage impeller inlet, and the second-stage diffuser inlet, respectively. The frequency response of these pressure sensors is greater than 100kHz.

[0033] The collected compressor steady-state data includes the compressor inlet pressure P. in Inlet temperature T in Exhaust pressure P out Outlet temperature Tout Import flow Q in The compressor's rotational speed N, inlet pressures P1 and P2 of the two return flow branches, inlet temperatures T1 and T2, flow rates Q1 and Q2, and valve position feedback W1 and W2 of the two channels are also collected. Here, subscript 1 indicates the rapid return flow branch, and subscript 2 indicates the main return flow branch. Simultaneously, the transient pressure P at the compressor inlet channel is also collected. t1 The transient pressure P at the inlet of the first-stage impeller t2 The transient pressure P at the inlet of the first-stage diffuser t3 The transient pressure P at the inlet of the second-stage impeller t4 The transient pressure P at the inlet of the second-stage diffuser t5 .

[0034] compressor variable efficiency η pol Calculate according to the following formula: Where κ represents the coolant variability index.

[0035] The transient pressure fluctuation coefficient ΔP of the compressor is calculated according to the following formula:

[0036] In the above formula It is the average transient pressure at the compressor inlet channel during the data acquisition period. It is the average transient pressure at the inlet of the first-stage impeller of the compressor during the data acquisition period. It is the average transient pressure at the inlet of the first-stage diffuser of the compressor during the data acquisition period. It is the average transient pressure at the inlet of the second-stage impeller of the compressor during the data acquisition period. This is the average transient pressure during the data acquisition time at the second-stage diffuser of the compressor. The transient data acquisition time can be determined based on the frequency response of the transient pressure sensor. (Function) This represents the maximum absolute value of all elements in the sequence.

[0037] Based on design calculations, one-dimensional system simulation, three-dimensional unsteady numerical simulation, control valve flow characteristics, and compressor operation data, the following surrogate model is established using machine learning methods:

[0038]

[0039]

[0040] In the above formula, N represents the compressor speed. For centrifugal compressors used in refrigeration, the flow rate is usually reduced by decreasing the speed. Here, f1 represents the functional relationship established through machine learning between compressor efficiency, transient pressure fluctuation coefficient, speed, inlet flow rate, pressure, temperature, and parameters of both loops; f2 represents the functional relationship established through machine learning between the position of the fast regulating valve in the fast return branch and the return flow rate, pressure, and temperature; f3 represents the functional relationship established through machine learning between the position of the regulating valve in the main loop branch and the return flow rate, pressure, and temperature.

[0041] The proxy relationship f1 can be provided by design calculations, one-dimensional system simulation, and three-dimensional unsteady numerical simulation calculations before the compressor starts running. After the compressor starts running, the relevant valid operating data is stored and the proxy relationship is updated. Proxy relationships f2 and f3 can be provided by the control valve flow characteristics and design calculations before the compressor starts running. After the compressor starts running, the relevant valid operating data is stored and the proxy relationship is updated.

[0042] For refrigeration compressors, the minimum flow rate Q that will not cause surge can be determined by using the condensing and evaporating temperatures of the refrigerant and the compressor's anti-surge line. 喘 Assume the system requires a certain offloading traffic volume of Q. 卸载 When the system's required offload traffic meets At that time, the reflux system will be triggered.

[0043] As the rotational speed decreases, when or ( This refers to the real-time flow rate of the intake manifold or the real-time flow rate of the system. Q is the vibration threshold of the unit. 喘 (This refers to the minimum flow rate to prevent surge). The rapid return branch and main return branch open simultaneously. The rapid return branch quickly reaches 100% control valve opening to stabilize the emergency situation and prevent unit vibration. Simultaneously, the main return branch gradually opens. As the system's real-time flow rate further decreases, the rapid opening of the rapid return branch control valve prevents increased pressure pulsation due to the reduced flow. When the system flow rate drops to the required unloading flow level and gradually stabilizes, the main return branch control valve also reaches a certain opening. It should be noted that if the rapid regulating valve in the rapid return branch malfunctions, an alarm will immediately sound and the compressor inlet regulating valve will be actively controlled to increase its opening, thereby reducing the unloading rate. Furthermore, when controlling the opening of the rapid branch regulating valve and the main branch regulating valve, the decrease in return flow rate in the rapid branch must equal the increase in return flow rate in the main circuit.

[0044] at this time, And satisfy:

[0045] in, These represent the real-time traffic within the fast return branch and the main return branch, respectively.

[0046] To compensate for the efficiency loss caused by backflow, all backflow gas should pass through the heat exchanger to reduce the impeller inlet temperature. Therefore, gradually decrease the opening of the backflow branch control valve and gradually increase the opening of the main flow branch control valve. Calculate the relationship between the opening and flow rate based on the proxy equations f2 and f3. It is essential to ensure that the following conditions are always met during the adjustment process:

[0047] Until the rapid return branch is completely closed. Here, -ΔQ1 represents the decrease in return flow rate in the rapid branch, and ΔQ2 represents the increase in return flow rate in the main loop. At this point, all return gas enters the intake pipe upstream of the impeller through the main return branch and the heat exchanger, thereby maximally compensating for the efficiency loss caused by the return flow. At this point, the following conditions are met:

[0048] Based on monitoring at time f1 When the adjustment process occurs When the time is reached, the adjustment operation is stopped and an alarm is issued. Simultaneously, the opening of the compressor inlet regulating valve is increased to reduce the unloading rate. During the adjustment process, monitoring is performed based on time f1. In principle, this reflux device can achieve 100% gas reflux. However, as the reflux flow rate (or unloading rate) increases, the compressor efficiency will decrease sharply. Therefore, an efficiency threshold needs to be set. When the compressor efficiency is lower than this threshold, the adjustment operation will stop and an alarm will be issued.

[0049] During the adjustment process, real-time measurement data is continuously recorded, and agent relations f1, f2, and f3 are trained based on the newly measured real-time data to achieve a reasonable allocation of traffic between the two branches.

[0050] Furthermore, as a preferred example, when the condensing temperature of the refrigerant in the refrigeration compressor is 36°C and the evaporating temperature is 4°C, the compressor pressure ratio is determined. The compressor's design flow rate is 4.5 kg / s. Based on the compressor's pressure ratio and anti-surge line, the minimum flow rate Qsurge that prevents surge at these condensing and evaporating temperatures is determined to be 2.5 kg / s.

[0051] Assuming the refrigeration system requires an unloading flow rate of Q_unloading = 1.35 kg / s, then... The reflux system will be triggered.

[0052] For centrifugal compressors used in refrigeration, the flow rate is typically reduced by decreasing the rotational speed. As the rotational speed decreases, when... or ( This refers to the real-time flow rate of the intake manifold or the real-time flow rate of the system. ), triggered by the reflux system. In this embodiment, when... When the flow rate reaches 2.62 kg / s, the reflux system is triggered. Both the rapid reflux branch and the main reflux branch open simultaneously. The rapid reflux branch quickly reaches 100% opening of its control valve to alleviate the emergency and prevent unit vibration. Simultaneously, the main reflux branch gradually opens. As the system's real-time flow rate further decreases, the rapid opening of the rapid reflux branch's control valve prevents increased pressure pulsation due to the reduced flow. When the refrigeration system's flow rate further decreases to the required unloading flow rate of 1.35 kg / s and gradually stabilizes, the main reflux branch control valve also reaches a certain opening. At this point: And satisfy:

[0053] In the above formula These are the real-time traffic flows within the fast return branch and the main return branch, respectively. The compressor's unloading flow rate has been greatly increased, meeting the requirements of the refrigeration system.

[0054] To compensate for the efficiency loss caused by backflow, all backflow gas should pass through the heat exchanger to reduce the impeller inlet temperature. Therefore, gradually decrease the opening of the backflow branch control valve and gradually increase the opening of the main flow branch control valve. Calculate the relationship between opening and flow rate based on the agent relationships f2 and f3, ensuring that the following is always satisfied during the adjustment process:

[0055] Until the rapid recirculation branch is completely closed, all recirculated gas enters the intake pipe upstream of the impeller through the main recirculation branch and the heat exchanger, thus maximally compensating for the efficiency loss caused by recirculation. At this point, the following conditions are met:

[0056] During the adjustment process, monitoring is performed based on time f1. In this embodiment, The pressure remains below the pressure alarm threshold. This occurs during the adjustment process. When the condition is reached, the adjustment operation is stopped and an alarm is issued. At the same time, the opening of the compressor inlet regulating valve is increased to reduce the unloading rate.

[0057] During the adjustment process, monitoring is performed based on time f1. In principle, this reflux device can achieve 100% gas reflux, but as the reflux flow rate increases (or the unloading flow rate increases), the compressor efficiency decreases sharply. Therefore, an efficiency threshold needs to be set. In this embodiment, the compressor's variable efficiency threshold is set to 60%. Throughout the adjustment process, such as... Figure 3 As shown, the efficiency is 62.2% when the unloading flow rate reaches 1.35 kg / s. When the time comes, stop the adjustment operation and issue an alarm.

[0058] Furthermore, the specific operation flow of the machine learning agent model in this invention is described as follows: Step S1: Determine the sample space and acquire raw samples in the offline phase. Based on the compressor's design speed range, inlet flow range, inlet and outlet boundary conditions corresponding to condensing and evaporating temperatures, and the allowable opening range of the regulating valves for the fast return branch and main return branch, multiple operating condition combinations are generated using orthogonal design, Latin hypercube sampling, or hierarchical grid sampling. Each operating condition combination covers the states of two return branches being open separately, simultaneously open, the fast branch being fully open, and switching from the fast branch to the main branch, and also covers the stable operating region near the anti-surge line. Sample data before compressor operation is provided by one or more of the following: design calculations, one-dimensional system simulation, three-dimensional unsteady numerical simulation, and regulating valve flow characteristics. Each raw sample includes at least the compressor inlet pressure P. in Inlet temperature T in Import flow Q in Speed ​​N; inlet pressure P1, temperature T1, flow rate Q1, and valve position W1 of the rapid return branch; inlet pressure P2, temperature T2, flow rate Q2, and valve position W2 of the main return branch; and transient pressure P t1 To P t5 The transient pressure is segmented according to a preset time window, and the P value within each time window is calculated accordingly. t1 To P t5 Calculate the transient pressure fluctuation coefficient ΔP, based on P in T in Calculate P out T out and the compressor's variable efficiency η pol .

[0059] Step S2: Perform time synchronization, validity checks, and feature construction on the original samples. Align constant measurements according to the same timestamp, and mark or remove samples that are missing, exceed the measurement range, have sensor malfunctions, or do not meet mass conservation. As a preferred example, the sampling frequency of the transient pressure sensor is greater than 100 kHz, the time window length is 1 s, and the overlap rate of adjacent time windows is 50%; the time window can also be adjusted to cover at least a predetermined number of pressure fluctuation cycles. Use [P1, T1, Q1] as the input of surrogate model f2 and W1 as the output of f2, and use [P2, T2, Q2] as the input of surrogate model f3 and W2 as the output of f3; use Q... in P in T in The two loop parameters output by f2 and f3 are used as inputs to the surrogate model f1, and [η] pol , ΔP, N] are the outputs of f1.

[0060] Step S3: Construct training, validation, and test sets. Using complete operating condition combinations as the grouping unit, divide the effective samples into training, validation, and test sets at ratios of 80%, 10%, and 10%, respectively. This ensures that time windows generated by the same continuous adjustment process belong to only one subset of data, avoiding data leakage caused by adjacent time windows. Standardize each input and output quantity using the mean and standard deviation of the training set, and save the mean and standard deviation. The validation and test sets use the same standardization parameters as the training set. Samples near the surge line, in the fully open fast branch state, and in the two-branch switching state should be covered in all three subsets.

[0061] Step S4: Train initial surrogate models f1, f2, and f3. The surrogate models can employ neural networks, support vector regression, random forests, gradient boosting trees, or combinations thereof. Taking a fully connected neural network as an example, f1 has two hidden layers with 64 and 32 neurons respectively; f2 and f3 each have two hidden layers with 32 and 16 neurons respectively; the hidden layers use the ReLU activation function, and the output layers use a linear activation function. Weighted mean squared error is used as the loss function, where the three outputs of f1 are standardized with equal weights, and f2 and f3 each use the single-output mean squared error; the optimizer is Adam, with an initial learning rate of 0.001, a batch size of 64, an L2 regularization coefficient of 0.00001, and a maximum training epoch of 500. Training is stopped early when the validation set loss does not decrease for 30 consecutive epochs, and the model parameters with the minimum validation set loss are saved. The above network structure and training parameters are a preferred implementation and can be adjusted according to the number of samples, control cycle, and unit characteristics.

[0062] Step S5: Test and solidify the initial proxy model. Calculate f1 versus η respectively. polThe prediction errors of ΔP and N, as well as the prediction errors of f2 and f3 for W1 and W2, are considered. As a preferred criterion, the coefficient of determination R² of each output on the test set is not less than 0.95 and the normalized root mean square error is not greater than 5%. If the criterion is not met, samples are added in the region near the speed, flow rate, valve position, or anti-surge line where the prediction error is large, and the process returns to step S3 for retraining. After the criterion is met, the model structure, weights or tree parameters, normalized parameters, applicable input range, training time, and version number are saved to form an initial proxy model for the online controller to call.

[0063] Step S6: During the online operation phase, the controller reads real-time measurement data of the compressor and the two return branches according to a preset control cycle. The real-time measurement data includes P in T in P out T out Q in N, P1, T1, Q1, W1, P2, T2, Q2, W2 and P t1 To P t5 Simultaneously read the target offloading traffic Q 卸载 The minimum flow rate Q that does not cause surge 喘 Pressure fluctuation alarm threshold ΔP max Variable efficiency threshold η min And the status of the regulating valve. The controller performs the same time synchronization, range checking, time window calculation, and standardization processing on the real-time measurement data as in the offline phase, and calculates the current η. pol And ΔP. When there is a critical input missing, the model is outside the applicable range of the fixed model, or the control valve feedback is abnormal, the output of the current model is not used as the control command, but the previous safety command is retained and an alarm is issued.

[0064] Step S7: Store the inputs, model outputs, control commands, and measured results that have passed the validity check during online operation into the online sample buffer at the same timestamp for updating f1, f2, and f3. As a preferred example, when the number of newly added valid samples reaches 256 time windows or the preset update cycle is reached, an incremental training set is formed using the newly added samples and some offline samples to fine-tune the current model. Fine-tuning uses the Adam optimizer with an initial learning rate of 0.0001, a batch size of 64, and a maximum number of training epochs of 50. The process stops early when the validation set loss does not decrease for 10 consecutive epochs. The updated model only replaces the current online model if its error on independent validation data is no higher than that of the current online model, and the ΔP prediction error and valve position prediction error of samples near the surge line meet preset requirements; otherwise, the previous version is retained. Thus, while maintaining safety constraints and model traceability, the surrogate relationships f1, f2, and f3 are gradually corrected using actual operating data.

[0065] The sample division ratio, number of network layers, number of neurons, learning rate, batch size, number of training rounds, number of early stopping rounds, error threshold, and online update triggering conditions mentioned above are all optional implementation parameters and do not constitute a limitation on the type of machine learning algorithm or the range of parameters.

[0066] Specific embodiments of the present invention have been described above. It should be understood that the present invention is not limited to the specific embodiments described above, and those skilled in the art can make various changes or modifications within the scope of the claims, which do not affect the essence of the present invention. Unless otherwise specified, the embodiments and features described in this application can be arbitrarily combined with each other.

Claims

1. A machine learning-based dual-reflux control system for a two-stage refrigeration centrifugal compressor, characterized in that, include: The two-stage refrigeration centrifugal compressor includes a first-stage impeller, a second-stage impeller, a first-stage bladeless diffuser, a second-stage bladeless diffuser, a volute, an exhaust pipe, and an intake pipe. The first-stage impeller, the second-stage impeller, the first-stage bladeless diffuser, and the second-stage bladeless diffuser are all installed inside the volute. One end of the volute is connected to the intake pipe, and the other end is connected to the exhaust pipe. A rapid return branch and a main return branch are connected in parallel between the exhaust pipe and the intake pipe; The rapid return branch is equipped with a rapid return branch regulating valve; The main return branch is equipped with a main return branch regulating valve and a main return branch heat exchanger; The return gas from the rapid return branch and the main return branch is delivered to the intake pipe upstream of the compressor or to the independent intake mixing chamber, and the return gas is not micro-jet injected into the impeller outlet, diffuser or blade tip area; the rapid return branch and the main return branch are respectively led out from different gas inlets of the exhaust pipe; The system also includes multiple sets of measurement sensors and controllers; The controller is configured with a machine learning agent model to output the valve position of the return branch regulating valve; The controller is configured with the following control logic: after the return current is triggered, the fast return current branch is first activated to achieve rapid current increase and vibration suppression, and then gradually switched to the main return current branch. During the switching process, the decrease in return current of the fast return current branch is kept equal to the increase in return current of the main return current branch.

2. The machine learning-based dual-reflux control system for a two-stage refrigeration centrifugal compressor according to claim 1, characterized in that, The measurement sensors include a constant parameter sensor and a high-frequency transient pressure sensor; The steady parameter sensors are used to collect the compressor inlet pressure P in , inlet temperature T in , exhaust pressure P out , outlet temperature T out , inlet flow Q in , rotating speed N, the pressure P1, temperature T1, flow Q1 of the quick reflux branch, the valve position W1, the pressure P2, temperature T2, flow Q2 of the main reflux branch, the valve position W2; The high-frequency transient pressure sensors are respectively arranged at the compressor inlet, the first-stage impeller inlet, the first-stage diffuser inlet, the second-stage impeller inlet, and the second-stage diffuser inlet; the high-frequency transient pressure sensors have a frequency response greater than 100kHz and are used to collect multiple transient pressure P signals. t1 To P t5 And calculate the transient pressure fluctuation coefficient ΔP.

3. The machine learning-based dual-reflux control system for a two-stage refrigeration centrifugal compressor according to claim 1, characterized in that, The maximum flow rate of the fast return branch is designed to be 10% of the compressor's design flow rate. 30%; the maximum flow rate of the main return branch can reach 100% of the compressor's rated inlet mass flow rate.

4. The machine learning-based dual-reflux control system for a two-stage refrigeration centrifugal compressor according to claim 1, characterized in that, The machine learning agent model supports offline pre-training and online incremental updates; In the offline phase, original samples are obtained based on design calculations, one-dimensional system simulation, three-dimensional unsteady numerical simulation, and the flow characteristics of regulating valves to complete model training and solidification. The machine learning surrogate model includes surrogate relations f1, f2, and f3. The surrogate relation f1 is used to output the compressor's variable efficiency η. pol Transient pressure fluctuation coefficient ΔP; proxy relationship f2 is used to output the valve position W1 of the fast return branch control valve; proxy relationship f3 is used to output the valve position W2 of the main return branch control valve; During the online operation phase, the agent relationships f1, f2, and f3 are fine-tuned and updated using the incremental effective sample data from the actual unit measurements.

5. The machine learning-based dual-reflux control system for a two-stage refrigeration centrifugal compressor according to claim 1, characterized in that, The controller has a built-in alarm protection unit; when the transient pressure fluctuation coefficient ΔP exceeds the maximum vibration threshold ΔP, it will detect that the transient pressure fluctuation coefficient ΔP exceeds the maximum vibration threshold ΔP. max Or the variable efficiency η of the compressor pol When the efficiency falls below the preset threshold, the controller will issue an alarm, stop the branch flow switching operation, and increase the opening of the compressor inlet regulating valve to reduce the unit unloading rate.

6. A dual-reflux control method for a two-stage refrigeration centrifugal compressor based on machine learning, characterized in that, The dual reflux control system according to any one of claims 1 to 5 includes the following steps: Step S1: Construct machine learning agent relations f1, f2, and f3 offline and complete model training and solidification; Step S2: Real-time acquisition of steady-state operating parameters of the compressor and its two return branches, as well as multiple transient pressure data, to calculate the compressor's multivariable efficiency η. pol And the transient pressure fluctuation coefficient ΔP; Step S3: When the real-time intake manifold flow rate drops to 1.05 times the minimum surge flow rate, or the transient pressure fluctuation coefficient ΔP exceeds the maximum vibration threshold ΔP max The backflow control is triggered in time; the fast backflow branch regulating valve is rapidly opened, while the main backflow branch regulating valve is gradually opened at the same time. Step S4: After the system traffic stabilizes at the target offload traffic level, perform branch traffic switching: reduce the traffic of the fast return branch and simultaneously increase the traffic of the main return branch, maintaining this adjustment process. ΔQ1=ΔQ2 until the fast return branch is completely closed, and all the return gas is cooled by the heat exchanger of the main return branch and then sent to the upstream of the compressor. Step S5: Collect valid running samples online and continuously update the proxy relationships f1, f2, and f3 incrementally; continuously monitor ΔP and η during the operation. pol If the threshold is triggered, a protection alarm will be executed.

7. The machine learning-based dual-reflux control method for a two-stage refrigeration centrifugal compressor according to claim 6, characterized in that, In step S1, during the offline training phase, Latin hypercube sampling or orthogonal sampling is used to generate a set of working condition samples. The samples cover working conditions such as fast return branch fully open, two branches working simultaneously, fast branch switching to main branch, and working conditions near the anti-surge boundary. The samples are divided into training set, validation set, and test set, and the time window samples generated in the same continuous adjustment process belong to only a single dataset.

8. The machine learning-based dual-reflux control method for a two-stage refrigeration centrifugal compressor according to claim 6, characterized in that, Backflow-triggered runtime satisfies flow constraints: When the system flow rate drops to the required unloading flow rate level and gradually stabilizes, the main return branch control valve also reaches a certain opening degree. At this point... And satisfy: The flow constraint is satisfied after the fast return branch is completely closed: in, These represent the real-time traffic within the fast return branch and the main return branch, respectively.

9. The machine learning-based dual-reflux control method for a two-stage refrigeration centrifugal compressor according to claim 6, characterized in that, When the online model is incrementally updated, the current online model version will only be replaced if the updated surrogate model has a prediction error that is not inferior to the current online model on the independent validation dataset and the prediction error of the anti-surge boundary condition meets the preset index; otherwise, the original model will be used.

10. The machine learning-based dual-reflux control method for a two-stage refrigeration centrifugal compressor according to claim 6, characterized in that, When a fault is detected in the fast return branch regulating valve, the controller immediately outputs an alarm signal and actively increases the opening of the compressor inlet regulating valve to reduce the unit's unloading rate.

Citation Information

Patent Citations

  • Anti-surge system of refrigeration centrifugal compressor and refrigeration system

    CN215370159U