A method and system for staged independent monitoring and surge prevention protection of multi-stage centrifugal compressors
By using a hierarchical independent monitoring system and an RL-CP collaborative protection model, the problems of delayed surge risk identification and unsuitable adjustment in multi-stage centrifugal compressors under varying operating conditions have been solved, achieving more accurate surge risk assessment and more efficient protection control.
Patent Information
- Application Number
- CN202611439920.2
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2026-09-16
- Publication Date
- 2026-11-17
AI Technical Summary
Existing multi-stage centrifugal compressors have difficulty accurately identifying the surge risk of each stage under varying operating conditions, resulting in delayed identification of local risks and incompatibility of uniform adjustment methods, which can easily lead to unnecessary backflow losses and excessive protection action range.
A tiered independent monitoring method is adopted. By collecting operating data from each stage and combining a one-dimensional thermodynamic model and a three-dimensional CFD simulation model, a surge boundary identification model is established. Fuzzy clustering and support vector machine classification are then used to identify surge states. Simultaneously, a tiered independent anti-surge controller and an RL-CP cooperative protection model are employed for local and cooperative control, utilizing feedforward gain and propagation delay for inter-stage compensation.
It improves the accuracy of surge risk assessment and the targeted nature of protection, reduces ineffective adjustments and energy loss, and enhances adaptability under varying operating conditions.
Smart Images

Figure CN123110122A_ABST
Abstract
Description
Technical Field
[0001] This invention belongs to the field of compressor control and fluid machinery safe operation technology. Specifically, it relates to a method and system for staged independent monitoring and anti-surge protection of multi-stage centrifugal compressors. Background Technology
[0002] Multistage tandem centrifugal compressors are widely used in petrochemical, natural gas transportation, air separation, refrigeration, and fuel cell air supply applications. They compress the working fluid sequentially through multiple compressor stages to achieve higher output pressures. Due to differences in impeller structure, inlet conditions, and aerodynamic characteristics among the stages, the operating points of each stage will shift to varying degrees under variable operating conditions such as start-up, load adjustment, and changes in inlet pressure or temperature. When a local operating point enters an unstable region, surge is prone to occur. Surge is typically accompanied by periodic fluctuations in flow and pressure and can cause decreased efficiency, increased vibration, and damage to blades, bearings, and seals. Therefore, appropriate surge monitoring and protection devices are required.
[0003] Existing anti-surge systems for multi-stage centrifugal compressors typically install pressure, temperature, and flow rate monitoring points at the compressor outlet or between stages. They assess surge risk based on fixed surge boundaries, empirical criteria, or calculated operating parameters. When the operating conditions reach preset levels, the compressor flow rate is adjusted by opening anti-surge valves, reflux valves, or venting devices. Some existing solutions can also collect inlet and outlet parameters for different stages and determine the operating status of the corresponding stages.
[0004] However, there is strong flow and pressure coupling between the stages of a multi-stage series compressor. When the operating state of one stage changes, it will affect the adjacent stages. Using only a few steady-state parameters, fixed thresholds, or uniform adjustment of the whole machine is not enough to fully reflect the actual stability differences between different stages under rapidly changing operating conditions. This can easily lead to problems such as delayed identification of local risks, excessive protection action range, and unnecessary backflow losses.
[0005] Therefore, there is still a need for further improvement in how to enhance the accuracy and adaptability of surge monitoring and protection for multi-stage centrifugal compressors under complex and variable operating conditions. Summary of the Invention
[0006] This invention addresses the problems in existing multi-stage centrifugal compressor surge monitoring systems, such as the difficulty in accurately reflecting the differences in operating states of each stage, the lag in identifying local surge risks, and the inability of a unified reflux or unified adjustment method to meet the actual needs of different stages. It proposes a multi-stage centrifugal compressor staged independent monitoring and surge protection method and system. This technical solution identifies and controls the operating state of each stage separately, coordinates the dynamic influence between adjacent stages, and implements multi-stage collaborative protection when necessary. This improves the targeting of surge risk assessment and protection actions under varying operating conditions, and reduces unnecessary overall adjustments.
[0007] To achieve the above objectives, on the one hand, the technical solution adopted by the present invention provides a method for staged independent monitoring and anti-surge protection of a multi-stage centrifugal compressor, the method comprising:
[0008] S1: Collect full-condition operating data of each stage of the multi-stage centrifugal compressor, construct the mass flow rate-pressure ratio characteristic curve of each stage by combining the one-dimensional thermodynamic model and the three-dimensional CFD simulation model, determine the surge boundary of each stage according to the surge state, extract pressure pulsation, low-frequency oscillation of mass flow rate and vibration characteristics, and classify the safe zone, warning zone and surge zone by combining fuzzy clustering and support vector machine to establish a hierarchical surge boundary identification model;
[0009] S2: Establish a hierarchical independent anti-surge controller for each stage, which has online surge identification, recirculation valve control, inlet guide vane adjustment and / or speed regulation, as well as inter-stage decoupling coordination rules;
[0010] S3: The local states of the machine level and adjacent machine levels are used as local observations of each anti-surge agent. The corresponding action space is determined according to the configuration of the actuators of each machine level. For cases where two or more machine levels share the same shaft system or the same recirculation valve, the corresponding speed regulation amount or recirculation valve opening regulation amount is used as a shared action constraint. Each machine level agent generates control requirements for the shared shaft system or the shared recirculation valve. The MADDPG algorithm is used to train each machine level agent, and the centralized evaluation network is used to evaluate based on the global state and joint actions to establish an RL-CP collaborative protection model.
[0011] S4: Real-time acquisition of operating data of each machine level, determination of surge margin and its changing trend based on the corresponding surge boundary, and determination of surge risk status and risk level of each machine level by combining pressure pulsation, low-frequency oscillation of mass flow, vibration characteristics, fuzzy logic reasoning and support vector machine classification results;
[0012] S5: The anti-surge agent corresponding to the risk level generates candidate control actions based on local observations, and combines the recirculation valve baseline control requirements formed by the corresponding hierarchical independent anti-surge controller in S2 with the inlet guide vane and / or speed coordination results to determine the actual local control actions of the risk level. After forming the actual local control actions, it forms feedforward compensation requirements to adjacent levels based on the inter-level feedforward gain and propagation delay pre-calibrated in S2. When a shared shaft system or a shared recirculation valve is involved, each relevant level forms control requirements for the corresponding shared actuator and sends the shared control requirements to S6 for coordination.
[0013] S6: Continuously monitor the surge risk after local control of the risk level. If the risk is effectively mitigated within the preset time window, maintain independent control of this level. If it is not effectively mitigated or further aggravated, switch to RL-CP collaborative protection mode. The upper-level coordinator coordinates the recirculation valve, inlet guide vane and / or speed of the relevant level according to the candidate actions of each agent and the global state evaluation results.
[0014] In the preferred implementation, further, in S5, after the i-th stage actually forms a local control action, the actually executed local action is used as the input for inter-stage disturbance calculation; based on the current corrected speed, corrected mass flow rate, pressure ratio, and inter-stage cooling state, the propagation delay between the i-th stage and the adjacent j-th stage, as well as the feedforward gain corresponding to the current action type q and the compensation actuator type r, are read or interpolated from the parameters calibrated and stored in S2. The feedforward compensation requirement of the j-th stage is determined according to the following formula:
[0015] ;
[0016] In the formula: This represents the feedforward compensation requirement generated by the i-th level local action and sent to the j-th level r-type actuator; This indicates the feedforward gain corresponding to the current operating condition; This indicates the type q normalized control action currently being executed at level i; the execution time of adjacent level j compensation actions is arranged according to the propagation delay, so that the compensation actions begin to play a regulating role when or before the pressure or mass flow disturbance caused by the local control action at level i reaches level j.
[0017] In the preferred implementation, further, in S6, when local control has effectively mitigated the risk but there is still a shared actuator control requirement, the local actions that the risk level can already execute independently are not changed. The control requirements of each level belonging to the same shared actuator are restricted to the allowable range of the actuator. A shared action candidate set is formed by the values of each requirement value and the values obtained by the actuator according to the minimum controllable increment of the actuator. Each shared action candidate value is combined with the currently determined independent actions to form a limited joint candidate action. The centralized Critic evaluation completed in S3 is used to select the shared action with the largest evaluation value as the actual control quantity of the shared actuator. When the risk level is not effectively mitigated or the risk is further aggravated within the preset time window, the RL-CP collaborative protection mode is switched.
[0018] In the preferred implementation, further, in S5, the normalized recirculation valve candidate opening increment given by the i-th stage Actor is denoted as... When the local stage is in a state of insufficient surge margin, the control quantity of the local independent recirculation valve is determined according to the following formula:
[0019] ;
[0020] In the formula: This represents the normalized opening increment of the actual output to the independent recirculation valve in this stage; This indicates the baseline control requirements for the recirculation valve obtained from S2; This represents the increment of the candidate opening of the recirculation valve obtained by the Actor; This indicates that the amplitude is limited according to the actual operating range of the recirculation valve; when the candidate recirculation opening of the Actor is lower than the baseline control requirement of the recirculation valve, or when it is required to reduce the necessary recirculation amount under insufficient surge margin, the baseline control requirement of the recirculation valve is adopted.
[0021] In the preferred implementation, further, in S2, the propagation delay and feedforward gain between the control actions of each machine level and the dynamic response of the directly adjacent machine level are pre-calibrated using the full-condition synchronous dynamic test data obtained in S1. Let the change sequence of a certain normalized actual control action of the i-th level be... The normalized response sequence of the adjacent j-th level corresponding pressure or mass flow rate is as follows: Calculate the normalized cross-correlation coefficient and take the number of lag points corresponding to the largest absolute value of the cross-correlation coefficient:
[0022] ;
[0023] The propagation delay is determined by the following formula:
[0024] ;
[0025] The feedforward gain between the i-th level, q-th type actual action and the j-th level, r-th type compensation actuator is determined by the following formula:
[0026] ;
[0027] The above calibrations were performed for different corrected speeds, corrected mass flow rates, pressure ratios, and interstage cooling conditions. The resulting propagation delays and feedforward gains were grouped and stored according to the corresponding operating conditions. When the actual operating conditions were between adjacent calibration conditions, interpolation was performed based on the adjacent calibration parameters.
[0028] In the preferred implementation, further, in S3, when two or more machine levels share the same shaft system, each relevant machine level Actor outputs its own speed control requirements for the shared shaft system, but only one shared shaft system speed action variable is set in the actual joint action; when multiple machine levels share the same recirculation valve, each relevant machine level outputs its own opening control requirements, but only one shared recirculation valve opening action variable is set in the actual joint action; after all machine level actions are processed by shared action constraints, they form the joint action actually executed in the training environment, and the centralized Critic receives the global state and joint action, and outputs the corresponding evaluation value.
[0029] In the preferred implementation, further, in S6, multiple representative operating conditions are selected based on the full-condition dynamic test data of S1, and the response time required from the start of local anti-surge vibration operation to the emergence of identifiable improvement in machine-level surge margin or risk state under each representative operating condition is recorded. The i-th level preset time window is determined according to the following formula:
[0030] ;
[0031] In the formula: This represents the time window for evaluating the effectiveness of the i-th level local control. This represents the set of representative operating conditions used for time window calibration. This represents the response time required from the start of local control action to the emergence of an identifiable and improveable risk state under the q-th representative operating condition; This indicates the time margin for covering the detection fluctuations and the discrete control cycle setting; within the preset time window, when the i-th level decreases from the surge state to the warning or safe state, or from the warning state to the safe state, it is determined that the local control has effectively mitigated the surge risk; after the preset time window ends, when the i-th level still maintains the risk state at the beginning of the local control, and the surge margin does not show an increasing trend, it is determined that the local control has not effectively mitigated the surge risk.
[0032] In the preferred implementation, further, in S3, a hierarchical reward function for the i-th level is established:
[0033] ;
[0034] In the formula: This represents the reward for the current control step size at level i; This represents the minimum surge margin required for level i; Indicates the current surge margin at level i; This represents the minimum surge margin required for the adjacent j-th level; Indicates the current surge margin of the adjacent j-th level; This indicates the current efficiency of the i-th stage; Indicates the normalized recirculation valve opening; This represents the current control step size action vector; This represents the action vector of the previous control step. Represents the set of directly adjacent machine levels; , , , and These represent the non-negative weights corresponding to insufficient surge margin, efficiency, recycle, action variation, and insufficient surge margin of adjacent stages, respectively; the sum of rewards for each stage forms the global reward for the current control step size.
[0035] ;
[0036] In the formula: This indicates the global return for the current control step size. This indicates the total number of machine classes.
[0037] In the preferred implementation, further, in S1, while establishing the physical surge boundary, dynamic state features are extracted from the synchronous dynamic signals corresponding to each full-condition test working point. For the i-th level pressure dynamic signal, the average pressure component and low-frequency slow change trend within the analysis time window are removed to obtain the pressure pulsation component, and wavelet packet decomposition is performed. For the i-th level mass flow dynamic signal, after removing the average mass flow within the analysis time window, empirical mode decomposition is performed on the remaining dynamic signal. For the i-th level vibration dynamic signal, spectral analysis is performed on the vibration acceleration signals obtained from the bearing and / or housing positions. The pressure pulsation characteristics, mass flow rate low-frequency oscillation characteristics, and vibration characteristics corresponding to the same test operating point are combined to form a dynamic state feature vector. The dynamic features of different dimensions are normalized, and three categories of fuzzy clustering are performed on the i-th level normalized dynamic state features. The test samples with safety, warning, and surge labels are remapped to the mass flow rate-pressure ratio characteristic curves of the corresponding corrected rotation speed to form the i-th level independent safety zone, warning zone, and surge zone. The i-th level support vector machine classification model is trained with the normalized dynamic state feature vector as input and the safety, warning, and surge state labels as output.
[0038] On the other hand, the present invention also provides a multi-stage centrifugal compressor graded independent monitoring anti-surge protection system, which is applied to a multi-stage centrifugal compressor formed by multiple centrifugal compressor stages connected in series along the working fluid flow direction. The system includes a parameter detection component, a multi-channel synchronous data acquisition component, multiple graded independent anti-surge controllers respectively set one-to-one with each stage, a communication component, a coordination calculation component and an execution component.
[0039] The parameter detection component is used to acquire the inlet pressure, inlet temperature, outlet pressure, outlet temperature, mass flow rate, rotational speed, and vibration signal of each stage.
[0040] The multi-channel synchronous data acquisition component is connected to the parameter detection component, each hierarchical independent anti-surge controller and the coordination calculation component respectively. It is used to collect the operating data of each machine level according to a unified time reference and send the operating data of each machine level to the corresponding hierarchical independent anti-surge controller and the coordination calculation component respectively.
[0041] Each hierarchical independent anti-surge controller is connected to the hierarchical independent anti-surge controller corresponding to the directly adjacent machine level through a communication component, and is also connected to the coordination computing component; each hierarchical independent anti-surge controller also includes an anti-surge agent running inside it, which includes an Actor;
[0042] When the i-th level is determined to be a risk level, the i-th level Actor generates candidate control actions based on the current local observations, and the i-th level hierarchical independent anti-surge controller determines the actual local control action of the i-th level based on the candidate control actions, the baseline control requirements of the recirculation valve, and the coordination results of the inlet guide vane and / or speed.
[0043] After the i-th level independent anti-surge controller forms the actual local control action, it uses the actual executed local action as the calculation input for the feedforward compensation of the adjacent machine level. Based on the corresponding inter-level feedforward gain and propagation delay, it forms the feedforward compensation requirement of the directly adjacent j-th level and sends the feedforward compensation requirement to the j-th level independent anti-surge controller through the communication component.
[0044] When adjustment is required via a shared shaft system or a shared recirculation valve, the j-th level independent anti-surge controller generates control requirements for the corresponding shared shaft system or shared recirculation valve and sends the control requirements to the coordination computing component;
[0045] The coordination computing component includes a processor, a memory, and a communication interface. The memory stores the parameters of the trained centralized Critic network, state normalization parameters, the allowable range of motion of each actuator, and the shared action constraints corresponding to the shared shaft system and the shared recirculation valve.
[0046] After implementing actual local control actions, the i-th level independent anti-surge controller continuously determines the i-th level surge risk status within a preset time window; when the i-th level surge risk is effectively mitigated, it maintains local independent control; when the i-th level surge risk is not effectively mitigated or further aggravated, the i-th level independent anti-surge controller sends a cooperative protection trigger signal to the coordination calculation component and enters the RL-CP cooperative protection mode.
[0047] After entering the RL-CP cooperative protection mode, the risk level and the directly adjacent levels that are affected by the actual local control actions of the risk level and need to be fed forward compensation are identified as the relevant levels for this cooperative operation. Each relevant level Actor generates candidate control actions based on the current local observations.
[0048] The coordination calculation component forms an executable joint candidate action set based on independent action candidate values and shared action candidate values. It inputs the current global state and each executable joint candidate action into the centralized Critic, and selects the joint candidate action with the largest evaluation value as the coordination joint action for the current control cycle based on the evaluation value output by the centralized Critic.
[0049] The execution components include a valve actuator connected to the recirculation valve, a guide vane actuator connected to the inlet guide vane, and a drive device connected to the rotating shaft; each stage of independent anti-surge controller sends the corresponding control quantity in the actual local control action, feedforward compensation action, or coordinated joint action to the corresponding execution component through the actuator control interface, and the execution component adjusts the opening degree of the corresponding recirculation valve, the angle of the inlet guide vane, and / or the rotation speed according to the received control quantity.
[0050] The beneficial effects of this invention are:
[0051] First, the multi-stage centrifugal compressor graded independent monitoring anti-surge protection method of this application establishes the mass flow rate-pressure ratio characteristic relationship and the safe zone, early warning zone and surge zone by using independent sensor data of each stage, full-condition test data and combining one-dimensional thermodynamic model and three-dimensional CFD results. This can avoid substituting the state of the whole machine for the state of a single stage and improve the pertinence of local surge risk identification.
[0052] Secondly, in the method of this application, each stage performs online surge identification and implements recirculation valve, inlet guide vane and / or speed adjustment according to the status of the stage, so that the anti-surge action corresponds to the stage where the risk actually occurs, reducing the ineffective adjustment and energy loss caused by uniform recirculation.
[0053] Third, in the method of this application, after the risk stage operates, compensation is formed to the adjacent stages based on the inter-stage feedforward gain and propagation delay, which can reduce the propagation of pressure and flow disturbances between series stages in advance and reduce the possibility of adjacent stages re-entering an unstable state.
[0054] Fourth, in the method of this application, RL-CP adopts a combination of local intelligent agents at each machine level and centralized evaluation. When the control at this level cannot effectively mitigate the risk, it can coordinate the adjustment of relevant machine levels, take into account the surge margin, operating efficiency and smoothness of each machine level, and improve the adaptability to changing operating conditions.
[0055] Fifth, the multi-stage centrifugal compressor graded independent monitoring anti-surge protection system of this application forms a control structure that combines graded execution and unified coordination through graded independent anti-surge controllers, synchronous data acquisition, inter-stage communication and coordination calculation components. It can be applied to different unit configurations such as independent actuators, shared shaft systems and shared recirculation valves, thereby improving the applicability and reliability of protection control. Attached Figure Description
[0056] Figure 1 This is a flowchart of the multi-stage centrifugal compressor staged independent monitoring anti-surge protection method according to Embodiment 1 of the present invention;
[0057] Figure 2 This is a diagram of the anti-surge controller architecture of the multi-stage series centrifugal compressor graded monitoring and anti-surge protection system according to Embodiment 2 of the present invention;
[0058] Figure 3 This is a diagram of the architecture of a multi-stage series centrifugal compressor graded monitoring and anti-surge protection system according to Embodiment 2 of the present invention. Detailed Implementation
[0059] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.
[0060] This invention aims to provide a method and system for graded independent monitoring and surge protection of multi-stage centrifugal compressors. It establishes a surge boundary identification model and a graded independent surge protection controller corresponding to the aerodynamic characteristics of each stage. The surge risk state of each stage is determined by combining operating signals such as pressure, mass flow rate, and vibration. Recirculation valves, inlet guide vanes, and / or speed regulation are implemented according to the actuator configuration of each stage. Simultaneously, inter-stage compensation is performed using feedforward gain and propagation delay between adjacent stages. If the surge risk cannot be alleviated within a preset time by the control of the current stage, candidate control actions of related stages are coordinated through RL-CP collaborative protection. This achieves independent monitoring of each stage, priority control of the current stage, feedforward compensation of adjacent stages, and multi-stage collaborative protection when necessary.
[0061] Example 1
[0062] This embodiment is based on... The implementation focuses on a multi-stage centrifugal compressor formed by connecting multiple centrifugal compressors in series along the direction of working fluid flow. The value is an integer greater than or equal to 2. Each stage is sequentially designated as stage 1, stage 2, and so on along the direction of the working fluid flow. Level, No. Level represents level 1 to level 2. Any machine level in the class, the first Level representation and the first The upstream or downstream stages that are directly adjacent to the stage.
[0063] Based on the actual mechanical structure of a multi-stage centrifugal compressor, each stage can be equipped with its own inlet guide vanes, recirculation branch, and corresponding recirculation valve, or only some stages can be equipped with the corresponding actuators. Each stage can have an independent shaft, or two or more stages can share the same shaft system. Each stage can be equipped with an independent recirculation valve, or two or more stages can share the same recirculation valve, or a recirculation regulating mechanism shared by the entire unit can be used.
[0064] The hierarchical independent monitoring in this embodiment refers to level 1 to level 2. Each stage establishes a mass flow rate-pressure ratio characteristic relationship, surge boundary, and operating state identification model corresponding to its own aerodynamic performance based on its own operating data, and determines the surge margin and surge risk state of each stage. Hierarchical independent control means that for control actions with independent actuators, the corresponding stage forms and implements its own stage control based on its own surge risk. For shared actuators such as shared shafts and shared recirculation valves, each stage still forms control requirements for the shared actuators based on its own surge state. However, the control requirements formed by each stage are not considered as multiple actual control commands that can simultaneously act on the same shared actuator; instead, the higher-level coordinator forms the actual control quantity according to the actual control degrees of freedom of the shared actuator.
[0065] To clarify the logical relationship between model establishment, control rule establishment, and actual online control, in this embodiment, S1 is used to establish the independent performance relationship, surge boundary, and state identification model of each machine level; S2 is used to pre-establish the deterministic local anti-surge control rules of each machine level and the decoupling control parameters such as inter-level feedforward gain and propagation delay; and S3 is used to establish and train the RL-CP cooperative protection model. In this embodiment, RL-CP (Reinforcement Learning-Coordinated Protection) represents a cooperative protection model based on reinforcement learning. It is not a single reinforcement learning algorithm, but is composed of anti-surge agents corresponding to each machine level, local action networks in each anti-surge agent, a centralized evaluation network, local observations of each machine level, global state, action space, shared action constraints, and reward function. It is used to coordinate the control actions of multiple related machine levels when the local control of a single risk machine level cannot effectively alleviate the surge risk or when there is a need for shared actuator control. S4 to S6 constitute the actual operation phase. S4 performs real-time risk identification, S5 invokes the rules, parameters, and models established in S2 and S3 to implement local control and inter-stage feedforward, and S6 evaluates the effectiveness of local control and initiates collaborative protection if necessary. Therefore, the parameter determination and rule establishment in S2 are not considered as an actual anti-surge control action; the actual control action is invoked and executed in S5 based on the real-time status.
[0066] This embodiment provides a method for staged independent monitoring and surge protection of a multi-stage centrifugal compressor, including the following steps:
[0067] S1: Collect full-condition operating data of each stage of the multi-stage centrifugal compressor, construct the mass flow rate-pressure ratio characteristic curve of each stage by combining the one-dimensional thermodynamic model and the three-dimensional CFD simulation model, determine the surge boundary of each stage according to the surge state, extract pressure pulsation, low-frequency oscillation of mass flow rate and vibration characteristics, and classify the safe zone, early warning zone and surge zone by combining fuzzy clustering and support vector machine, and establish a graded surge boundary identification model.
[0068] Specifically, for Level 1 to Level 2 Each stage will undergo steady-state performance tests and dynamic performance tests covering the entire operating range to obtain the data needed to establish the performance characteristics, surge boundary, and state identification models of each stage. The full-condition operating data includes steady-state operating data as well as dynamic time-series data acquired synchronously with each steady-state operating point.
[0069] No. Steady-state operating condition data should include at least the inlet absolute pressure. Absolute temperature at the inlet Absolute pressure on exports Absolute temperature at the outlet mass flow rate Actual rotor speed and inlet guide vane angle . No. When setting the recirculation regulating mechanism, the opening degree of the recirculation valve is also collected. When a multi-stage centrifugal compressor is equipped with an interstage cooler, the interstage cooling flow rate and cooling temperature related to the inlet state of the corresponding stage are also collected.
[0070] No. Level 1 dynamic time series data should include at least the data from the 1st level. Pressure dynamic signals continuously acquired at the inlet and / or outlet positions 、 and the The corresponding mass flow dynamic signal of the level and from the Vibration acceleration signals obtained from the detection locations of the bearings and / or housing .
[0071] During the test, the inlet mass flow rate, inlet pressure, inlet temperature, inlet guide vane angle, and rotational speed were varied. When interstage coolers were installed, the interstage cooling flow rate and cooling temperature were also changed to create multiple different operating points for each stage within its permissible operating range. For coaxial units, the common shaft speed was changed. For stages with independent shafts, the actual rotor speed of the corresponding stage was changed. The test operating points for each stage covered the stable operating state, the near-surge state gradually approaching surge, and the surge state confirmed by testing.
[0072] Pressure, temperature, mass flow rate, rotational speed, and vibration data are acquired synchronously using a unified time base, ensuring that the steady-state operating parameters at the same operating point correspond to the pressure pulsations, mass flow oscillations, and vibration dynamic signals at that point according to the sampling time. When using real-time industrial Ethernet for data acquisition and transmission, different data channels are kept synchronized at the sub-millisecond level according to the acquisition hardware configuration to reduce the impact of multi-channel sampling time differences on subsequent dynamic feature extraction and inter-stage propagation delay analysis.
[0073] To enable comparison of performance data obtained under different inlet pressures and temperatures under a unified benchmark, the measured mass flow rate and actual rotational speed were converted to a unified reference pressure. and reference temperature Down.
[0074] No. The corrected mass flow rate is determined by the following formula:
[0075]
[0076] In the formula: Indicates the first The corrected mass flow rate of the stage under a unified reference condition; Indicates the first The measured mass flow rate at the current operating point; Indicates a unified reference absolute pressure; Indicates the first The absolute pressure at the current working point entrance; Indicates the first The absolute temperature at the current working point inlet; This indicates a unified reference absolute temperature.
[0077] No. The stage correction speed is determined according to the following formula:
[0078]
[0079] In the formula: Indicates the first Level correction speed; Indicates the first The actual rotor speed at the current operating point of the stage; Indicates a unified reference absolute temperature; Indicates the first The absolute temperature at the current working point inlet.
[0080] No. The stage pressure ratio is determined by the following formula:
[0081]
[0082] In the formula: Indicates the first Current operating point pressure ratio; Indicates the first The absolute pressure at the current working point of the level; Indicates the first The absolute pressure at the current working point entry level.
[0083] After completing the unified operating condition correction for the measured data, a one-dimensional thermodynamic compressible flow model was established for each turbine stage. A first-order one-dimensional thermodynamic model is used to describe the state of a given first-order thermodynamic model. When considering the inlet state, corrected mass flow rate, corrected rotational speed, inlet guide vane state, and corresponding interstage boundary conditions, starting from the first stage... The average thermal state change from stage inlet to outlet.
[0084] No. The first-stage one-dimensional thermodynamic model first calculates the pressure ratio and compression efficiency corresponding to this operating condition based on the corrected mass flow rate, corrected rotational speed, and inlet guide vane angle. The relationship is expressed as follows:
[0085]
[0086]
[0087] In the formula: Indicates the first Pressure ratio calculated using a one-dimensional thermodynamic model; Indicates the first Compression efficiency calculated using a one-dimensional thermodynamic model; Indicates the first Relationship for calculating stage pressure ratio; Indicates the first Efficiency calculation relationship; Indicates the first Level entry absolute pressure; Indicates the first Level inlet absolute temperature; Indicates the first Level correction mass flow rate; Indicates the first Level correction speed; Indicates the first Stage inlet guide vane angle; Indicates the impact of the first Interstage cooling status and corresponding inlet and outlet boundary conditions during stage operation.
[0088] Based on the pressure ratio obtained from the one-dimensional thermodynamic model, calculate the first... Level model export pressure:
[0089]
[0090] In the formula: Indicates the first The outlet absolute pressure calculated using a one-dimensional thermodynamic model; Indicates the first Pressure ratio calculated using a one-dimensional thermodynamic model; Indicates the first Level entry absolute pressure.
[0091] Based on the thermophysical properties of the working fluid, the inlet specific enthalpy and inlet specific entropy are determined by the inlet pressure and inlet temperature:
[0092]
[0093]
[0094] In the formula: Indicates the first Specific enthalpy of the inlet working fluid; Indicates the first Entropy of the inlet working fluid; This indicates the thermophysical relationship of specific enthalpy obtained based on the working fluid pressure and temperature; This indicates the thermophysical property relationship between specific entropy and working fluid pressure and temperature. Indicates the first Level entry absolute pressure; Indicates the first Level inlet absolute temperature.
[0095] Using the inlet specific entropy as the isentropic compression condition, and with As the target pressure after isentropic compression, the theoretical isentropic outlet enthalpy is obtained. The actual export enthalpy is calculated using the following formula:
[0096]
[0097] In the formula: Indicates the first The actual outlet specific enthalpy calculated using a one-dimensional thermodynamic model; Indicates the first Enthalpy at the entrance; Indicates the first The inlet working fluid is compressed to a certain extent while maintaining a constant inlet specific entropy. The corresponding theoretical isentropic exit enthalpy; Indicates the first Compression efficiency calculated using a one-dimensional thermodynamic model.
[0098] Then according to and By querying the corresponding thermophysical property relationships of the working fluid, the outlet temperature of the one-dimensional thermodynamic model can be obtained. Therefore, the first The complete inputs to a one-dimensional thermodynamic model include inlet pressure, inlet temperature, corrected mass flow rate, corrected rotational speed, inlet guide vane angle, and corresponding interstage boundary conditions. Its outputs include at least outlet pressure, outlet temperature, pressure ratio, and compression efficiency. The one-dimensional thermodynamic model is used to continuously describe the relationship between different experimental operating points. The relationship between the overall thermodynamic performance of the stage and the changes in the corrected mass flow rate, corrected rotational speed and inlet conditions.
[0099] For operating points with obtained measured data, the measured outlet pressure, outlet temperature, pressure ratio, and efficiency are compared with the calculation results of a one-dimensional thermodynamic model under the same inlet conditions, mass flow rate, rotational speed, and inlet guide vane conditions. The corresponding performance relationships are then corrected based on the measured data to ensure that the model output aligns with the measured performance trends across all operating conditions. Simultaneously, according to the... A corresponding three-dimensional CFD simulation model was established for the actual flow channel structure. The level 3D CFD computational domain includes at least the first The impeller flow path includes the inlet flow path, diffuser flow path, and corresponding interstage connecting flow path directly related to the impeller flow state. The three-dimensional CFD model calculates the pressure, temperature, and velocity distribution of the working fluid within the three-dimensional flow path based on the conservation of mass, momentum, and energy under compressible flow conditions.
[0100] For a specific test operating point, the corresponding inlet pressure, inlet temperature, rotor speed, inlet guide vane angle, and mass flow rate or outlet pressure are set as boundary conditions for the three-dimensional CFD model, and the interstage cooling state is made consistent with the corresponding test conditions. After the calculation, according to the... The average state of the inlet and outlet sections of the stage is used to obtain the CFD calculation pressure ratio and efficiency, while the pressure and velocity distribution inside the impeller and diffuser channel are also obtained.
[0101] One-dimensional thermodynamic models are used to establish continuous overall performance relationships over a wide operating range, while three-dimensional CFD models are used to verify typical operating points and low-mass-flow operating regions approaching surge. For operating points with available measured data, the measured results are used as the basis for performance data. For regions not continuously covered by adjacent measured operating points, a one-dimensional thermodynamic model corrected for measured data is used to supplement the operating points, and the corresponding three-dimensional CFD results are used to check the pressure ratio, efficiency, and flow variation trends in that region. When the variation trend of the one-dimensional model in a certain local region is inconsistent with the measured data and CFD results, the model parameters for that region are readjusted based on the corresponding measured points and CFD results.
[0102] After the above processing, in the given number Corrected speed At that time, the first The stage forms multiple discrete performance operating points:
[0103] In the formula: Indicates the first Level in A set of performance operating points at a modified speed; Indicates the first Level 1 One corrected rotational speed; This indicates the first time under the corrected speed. Corrected quality flow rate at each operating point; Indicates the first Pressure ratio at each working point; Indicates the first Efficiency of each work point; Indicates the first The total number of operating points at each corrected rotational speed; Indicates the work point number.
[0104] The operating points at the same corrected rotational speed are arranged according to the corrected mass flow rate, and interpolation is performed between adjacent operating points to form the mass flow rate-pressure ratio characteristic curve at the corresponding corrected rotational speed. This process is repeated for multiple corrected rotational speeds to form the... Family of mass flow rate-pressure ratio characteristic curves:
[0105]
[0106] In the formula: Indicates the first Mass flow rate-pressure ratio characteristic relationship; Indicates the first Stage pressure ratio; Indicates the first Level correction mass flow rate; Indicates the first Level correction speed.
[0107] Using the efficiency data corresponding to each operating point, connect operating points with the same or similar efficiency within the family of characteristic curves to form the first... Efficiency lines of different grades.
[0108] Further define the independent surge boundary for each stage. For the first stage... Level, in the Corrected speed While maintaining within the corresponding control allowable deviation range, make the first The operating point gradually changes from the stable operating region along the direction of decreasing mass flow rate, and pressure, mass flow rate and vibration dynamic signals are collected simultaneously.
[0109] As the mass flow rate decreases, when the first When a speed range transitions from a stable operating state through a near-surge state to a confirmed surge state, the corresponding critical operating region is recorded. For operating conditions where the surge initiation critical operating point can be directly obtained, this operating point is used as the surge boundary point at the current corrected speed. For operating conditions where the discrete test point cannot be exactly located at the state transition position, the last stable operating point and the first confirmed surge operating point form a boundary interval, and the surge critical operating point is obtained by interpolation between the two.
[0110] No. Level in The surge boundary point at the corrected rotational speed is represented as follows:
[0111]
[0112] In the formula: Indicates the first Level in Surge boundary point at a corrected rotational speed; Indicates the first Level 1 One corrected rotational speed; This indicates the corresponding surge correction mass flow rate; This indicates the pressure ratio at the corresponding surge boundary point.
[0113] Connect or interpolate the surge boundary points at multiple corrected speeds to form the first... Level surge boundary relationship:
[0114]
[0115] In the formula: Indicates the first Level surge boundary function; Indicates the first Level correction speed; This indicates the surge correction mass flow rate corresponding to the corrected rotational speed.
[0116] No. Level at The dimensionless surge margin at time t is determined by the following formula:
[0117]
[0118] In the formula: Indicates the first Level at The dimensionless surge margin at any given moment; Indicates the first Level at Real-time quality flow rate correction; Indicates the first Level at The corrected rotation speed at any time; Indicates according to the first The current corrected rotational speed, determined by the surge boundary function, corresponds to the surge corrected mass flow rate.
[0119] While establishing the physical surge boundary, dynamic state features are extracted from the synchronous dynamic signals corresponding to each full-condition test operating point. For the first... The pressure dynamic signal is analyzed by removing the average pressure component and low-frequency slow change trend within the analysis time window to obtain the pressure pulsation component, which is then subjected to wavelet packet decomposition. The energy of each pressure frequency band is determined according to the following formula:
[0120]
[0121] In the formula: Indicates the first Level 1 Pressure pulsation energy in a pressure frequency band; Indicates the first Within each frequency band Wavelet packet coefficients; Indicates the first The number of wavelet packet coefficients per frequency band; Indicates the pressure frequency band number; This indicates the wavelet packet coefficient index.
[0122] Compare the energy of each frequency band under stable, near-surge, and surge conditions. Select one or more frequency bands that exhibit repetitive variation patterns during the near-surge process at the operating point, and use the corresponding frequency band energy as the time-frequency characteristic of pressure pulsation.
[0123] For the After removing the average mass flow rate within the analysis time window from the dynamic signal of the mass flow rate, empirical mode decomposition is performed on the remaining dynamic signal:
[0124]
[0125] In the formula: Indicates the first The dynamic signal of mass flow rate after removing the average component; Indicates the first One intrinsic mode component; Indicates the total number of intrinsic modal components; Represents the residual trend term after empirical mode decomposition; Indicates the intrinsic modal component number; Indicates the sampling time.
[0126] By comparing the frequency and energy changes of each intrinsic mode under stable, near-surge, and surge conditions, one or more intrinsic modes that oscillate more strongly as the operating point approaches surge in the low-frequency range are selected as the low-frequency oscillation modes of mass flow.
[0127] When using the root mean square value of the selected low-frequency intrinsic mode to describe the amplitude of the low-frequency oscillation, it is determined according to the following formula:
[0128]
[0129] In the formula: Indicates the first Low-frequency oscillation amplitude of mass flow rate; This indicates that the selected low-frequency intrinsic mode is in the th... The values of each sampling point; This represents the total number of sampling points within the analysis time window; Indicates the sampling point number.
[0130] Low-frequency oscillation main frequency Take the frequency corresponding to the position of maximum amplitude in the spectrum of the selected low-frequency intrinsic mode.
[0131] For the Level 1 vibration dynamic signal: Spectral analysis is performed on the vibration acceleration signals obtained from the bearing and / or housing positions. The energy of each vibration characteristic frequency band is determined according to the following formula:
[0132]
[0133] In the formula: Indicates the first Level 1 Vibration energy in each characteristic frequency band; Indicates the first Power spectral density of vibration acceleration signal; Indicates frequency; Indicates the first Lower limit frequency of each characteristic frequency band; Indicates the first The upper limit frequency of each characteristic frequency band; This indicates the frequency band number of the vibration characteristic band.
[0134] The pressure pulsation characteristics, low-frequency oscillation characteristics of mass flow rate, and vibration characteristics corresponding to the same test operating point are combined to form a dynamic state feature vector:
[0135]
[0136] In the formula: Indicates the first Level 1 The dynamic state feature vector of each test sample; This indicates the number of selected pressure characteristic frequency bands; This indicates the number of selected vibration characteristic frequency bands; This indicates the low-frequency oscillation amplitude of the mass flow rate; This indicates the low-frequency oscillation frequency of the mass flow rate; Indicates vector transpose; Indicates the test sample number.
[0137] Furthermore, the dynamic characteristics of different dimensions are normalized:
[0138]
[0139] In the formula: Indicates the first Level 1 The first sample Normalization characteristics; Indicates the corresponding original features; Indicates the first The first in the full-condition sample Minimum of dimensional features; Indicates the first The first in the full-condition sample Maximum value of dimensional feature; Indicates the dynamic feature dimension number; Indicates the sample number.
[0140] For the The normalized dynamic state features are subjected to three categories of fuzzy clustering. In one implementation, fuzzy C-means clustering is used, with the three categories corresponding to the safety, warning, and surge states, respectively. The fuzzy clustering objective function is:
[0141]
[0142] In the formula: Indicates the first Level fuzzy clustering objective function; Indicates the number of participants in the clustering Number of samples under all operating conditions; Indicates the cluster category number; Indicates the first The sample belongs to the first Membership degree of each category; This represents a fuzzy index greater than 1; Indicates the first A normalized dynamic state feature vector; Indicates the first Cluster centers; This represents the L2 norm.
[0143] After clustering, the physical state of each cluster is determined based on the actual test conditions, surge margin, and operating point location in the mass flow rate-pressure ratio characteristic curve. Categories mainly composed of stable operating points with large surge margins are marked as safe states; categories mainly composed of operating points that have already entered surge are marked as surge states; and categories between safe and surge states, where the operating point is approaching the surge boundary but has not yet entered full surge, are marked as warning states.
[0144] Test samples labeled with safety, warning, and surge are remapped to the corresponding corrected speed mass flow rate-pressure ratio characteristic curve. The boundaries of the safety and warning zones are determined based on the transition position between the safety and warning samples, and the boundaries of the warning and surge zones are determined based on the transition position between the warning and surge samples. When no direct state transition point is acquired between adjacent test points in different states, interpolation is performed on the two operating points to determine the state transition position.
[0145] Connect or interpolate the state transition positions obtained at different corrected speeds to form the first... The system defines independent safety zones, warning zones, and surge zones, and determines the first-level safety zone based on the state transition boundary between the safety zone and the warning zone. Minimum surge margin required for level .
[0146] After completing the sample state labeling, the normalized dynamic state feature vector is used as input, and the safety, warning, and surge state labels are used as output to train the first... Level 1 Support Vector Machine Classification Model This enables the system to output safety, early warning, or surge classification results based on real-time pressure pulsation, low-frequency oscillation of mass flow rate, and vibration dynamic characteristics during actual operation.
[0147] Therefore, the first The graded surge boundary identification model is defined as follows:
[0148]
[0149] In the formula: Indicates the first A graded surge boundary identification model; Indicates the first Mass flow rate-pressure ratio characteristic relationship; Indicates the first Level surge boundary relationship; Indicates the first A set of state region parameters for the safety zone, early warning zone, and surge zone; Indicates the first A set of normalized parameters for dynamic state characteristics at each level; Indicates the first Level 1 Support Vector Machine State Classification Model.
[0150] Level 1 to Level 2 Each level establishes an independent graded surge boundary identification model.
[0151] S2: Establish a graded independent anti-surge controller for each stage, which has online surge identification, recirculation valve control, inlet guide vane adjustment and / or speed regulation, as well as inter-stage decoupling coordination rules.
[0152] It should be noted that this step is used to pre-establish the deterministic local anti-surge control rules, actuator constraints, and inter-stage feedforward compensation parameters to be called by the subsequent online control of each stage. This step does not actually output the anti-surge action to the actuator.
[0153] Specifically, for Level 1 to Level 2 Each stage is equipped with a separate anti-surge controller, and each controller has at least the functions of online surge identification, recirculation valve control, inlet guide vane / speed coordination, and inter-stage decoupling coordination.
[0154] In order to make the first In actual operation, the level controller can directly generate control requirements based on the local level status and pre-define the level requirements. The level controller calls the first established by S1. A graded surge boundary identification model is established, and the machine-level surge margin, the state region where the operating point is located, and the dynamic state classification results are obtained based on real-time operating data. The calculation relationships and control mappings corresponding to the above state variables are established and stored in this step, and called in S4 and S5 based on real-time data.
[0155] No. The rate of change of surge margin is determined by the following formula:
[0156]
[0157] In the formula: Indicates the first Level at The rate of change of surge margin at time; Indicates the surge margin of the current control cycle; Indicates the surge margin in the previous control cycle; This indicates the time interval between adjacent control cycles.
[0158] Definition of the first Insufficient surge margin:
[0159]
[0160] In the formula: Indicates the first The amount by which the current surge margin is insufficient relative to the required minimum surge margin; Indicates the first Minimum surge margin required for the level; This indicates the current actual surge margin.
[0161] In this embodiment, the baseline control requirement of this stage refers to the basic control requirement of the recirculation valve determined by the current surge margin deficiency of the local stage and the inter-stage feedforward compensation already applied to the local stage, without relying on the RL-CP collaborative control result.
[0162] No. The baseline opening increment of the recirculation valve caused by insufficient surge margin is determined by the following formula:
[0163]
[0164] In the formula: Indicates the first The normalized recirculation valve baseline opening increment is formed based on its insufficient surge margin. Indicates according to the first The recirculation valve control gain calibrated by the full-condition anti-surge test; This indicates that the current surge margin is insufficient; This indicates that the amplitude is limited according to the allowable opening range of the corresponding recirculation valve and the maximum allowable change in a single control cycle.
[0165] When the Level 1 receives adjacent 1 When the feedforward compensation requirement is at level 1, the first Baseline control requirements for a deterministic recirculation valve are determined by the following formula:
[0166]
[0167] In the formula: Indicates the first Baseline control requirements for deterministic recirculation valves; Indicates the first The baseline opening increment is based on the insufficient surge margin of the level; Indicates the relationship with the first The set of machine levels that are directly adjacent to each other; Indicates the number of adjacent digits Level 1 control action triggers and needs to act on Level 2 Feedforward compensation amount of stage recirculation valve; This represents the recirculation valve's action limiting function. When there is no adjacent stage feedforward compensation, the feedforward compensation is zero.
[0168] No. When a stage has an independent recirculation valve, the aforementioned deterministic recirculation valve baseline control requirement serves as a safety baseline when S5 forms the actual recirculation valve control quantity for that stage, and is not directly output to the recirculation valve in S2. When multiple stages share the same recirculation valve, the first stage... The system only establishes rules for generating local-level opening requirements for shared recirculation valves; the actual shared opening is uniformly coordinated in S6.
[0169] No. When the stage is in a warning state but has not yet entered a surge state, the inlet guide vane and / or speed coordination function determines the change in the operating point position after the inlet guide vane angle or speed change based on the mass flow rate-pressure ratio characteristic curve and isoefficiency line established by S1, and selects the inlet guide vane or speed adjustment direction that can increase the surge margin of the stage. For the common shaft system, the first The stage only addresses the speed control requirements for the shared shaft system.
[0170] Furthermore, to unify the action relationship between recirculation rapid protection and guide vane and / or active speed adjustment, an action selection rule based on the degree of risk urgency is established: when the first When the stage is in a warning state, the current surge margin has not yet fallen below the required minimum surge margin, and the rate of change of surge margin does not indicate that the operating point is rapidly approaching the surge boundary, adjusting the inlet guide vane and / or rotational speed is the preferred active adjustment method to reduce recirculation losses. If the surge margin is lower than the required minimum surge margin, or if the surge margin continues to decrease and is accompanied by pressure pulsation, low-frequency oscillation of mass flow rate, or vibration characteristics developing towards a surge state, or if a surge state has already been entered, the recirculation valve baseline control will be used as a safety protection bottom line. No subsequent Actor candidate action in S5 must cause the actual recirculation volume to fall below this safety baseline. If this safety baseline is met, inlet guide vanes and / or speed adjustments that can increase the machine-level surge margin can be added.
[0171] To enable the inter-stage decoupling coordination function to directly generate feedforward compensation based on actual local actions in S5, this step also uses the full-condition synchronous dynamic test data obtained in S1 to pre-calibrate the propagation delay and feedforward gain between the control actions of each machine stage and the dynamic response of the directly adjacent machine stage.
[0172] Let the first The normalized actual control action change sequence of a certain level is as follows , adjacent The normalized response sequence of the corresponding pressure or mass flow rate is as follows: Calculate the normalized cross-correlation coefficient:
[0173]
[0174] In the formula: Indicates the first Level control actions and the first Level dynamic response with hysteresis points Cross-correlation coefficients at different times; Indicates the first Level 1 Normalized control action variation at each sampling point; Indicates the first time within the corresponding analysis time window Average value of level control actions; Indicates the first Level 1 Normalized pressure or mass flow response at each sampling point; This represents the average value of the corresponding response; Indicates the sampling point number; This indicates the number of lagging sampling points.
[0175] Take the number of lag points corresponding to the largest absolute value of the cross-correlation coefficient:
[0176]
[0177] In the formula: Indicates the first Level control actions and the first The number of hysteresis sampling points when the correlation of the dynamic response level is maximized; This represents the cross-correlation coefficient corresponding to different lag points.
[0178] The propagation delay is determined by the following formula:
[0179]
[0180] In the formula: Indicates the first Pressure or mass flow disturbances caused by stage-level control actions propagate to the next stage. Level of propagation delay; This represents the number of lag points corresponding to the highest degree of correlation. This indicates the sampling time interval for dynamic test data.
[0181] No. Level 1 Class of actual actions and the first Level 1 The feedforward gain between the compensation actuators is determined by the following formula:
[0182]
[0183] In the formula: Indicates the first Level 1 Class control action and adjacent first Level 1 Feedforward gain between similar compensation actions; Indicates the first Level 1 Class control action in the first Normalized action increment for each calibrated sampling point; Indicates the first level used to offset the corresponding inter-level disturbance. Level 1 The normalized compensation action increment of the type of execution agency; Indicates the risk level action type; Indicates the type of adjacent machine-level compensation execution mechanism; This indicates the calibration sampling point number.
[0184] The above calibration was performed for different corrected rotational speeds, corrected mass flow rates, pressure ratios, and interstage cooling conditions, and the resulting propagation delays were then calculated. and feedforward gain The parameters are stored in groups according to their corresponding operating conditions. When the actual operating condition is between adjacent calibration conditions, interpolation can be performed based on the adjacent calibration parameters. Thus, S2 completes the pre-determination of the inter-stage feedforward parameters, and S5 only calls the corresponding parameters based on the actual executed action to calculate the feedforward compensation requirement for this control cycle.
[0185] After completing S2, the first Each level of independent anti-surge controller should at least retain the following: the calculation rules for the surge margin change rate of the machine level, the calculation rules for insufficient surge margin, the baseline control gain and limiting parameters of the recirculation valve, the judgment rules for the inlet guide vane and / or speed adjustment direction, the rules for the risk urgency and actuator action selection, the feedforward gain and propagation delay parameters corresponding to the directly adjacent machine level, and the rules for forming shared control requirements for shared shaft systems or shared recirculation valves.
[0186] S3: Using the local states of the current machine level and adjacent machine levels as local observations for each anti-surge agent, the corresponding action space is determined according to the actuator configuration of each machine level. For cases where two or more machine levels share the same shaft system or the same recirculation valve, the corresponding speed regulation or recirculation valve opening regulation is used as a shared action constraint. Each machine level agent generates control requirements for the shared shaft system or the shared recirculation valve. The MADDPG algorithm is used to train each machine level agent, and a centralized evaluation network is used to evaluate based on the global state and joint actions to establish an RL-CP collaborative protection model.
[0187] It should be noted that in this embodiment, the anti-surge agent is a reinforcement learning decision unit corresponding to a centrifugal compressor stage. Each anti-surge agent corresponds to a local observation space, an action space, and a local action network (Actor). The anti-surge agent uses its own stage state and the local states of directly adjacent stages as local observations, and generates local stage candidate control actions based on these local observations through the Actor. Therefore, the Actor in this embodiment is not the anti-surge agent itself, but rather the action network within the anti-surge agent used to implement control strategy mapping.
[0188] The Actor (Action Network, also known as Policy Network) is used to establish the mapping relationship between local observations and candidate control actions. For the i-th level anti-surge agent, its local observations are input into the corresponding i-th level Actor, which outputs candidate control quantities for the i-th level adjustable recirculation valve, inlet guide vane, and / or rotational speed. When the corresponding actuator is a shared actuator of multiple levels, the i-th level Actor outputs the control requirements for that shared actuator, rather than directly forming the unique actual control quantity for that shared actuator.
[0189] The Critic is a centralized evaluation network used to evaluate the control effects of the control actions generated by each stage Actor under the current overall operating state of the multi-stage centrifugal compressor. The Critic takes the global state composed of the states of each stage and the joint actions after processing shared action constraints as input, and outputs an evaluation value representing the long-term control benefits of the corresponding joint actions. During training, the Critic guides each stage Actor to update network parameters. In actual collaborative protection, it compares multiple executable joint candidate actions and determines the joint control action with the higher evaluation value.
[0190] The MADDPG (Multi-Agent Deep Deterministic Policy Gradient) algorithm is a multi-agent reinforcement learning algorithm used to train the aforementioned multiple anti-surge agents. This embodiment employs a centralized training and decentralized decision-making approach: During training, each agent generates candidate actions based on its local observations. All agent actions are constrained by a shared execution mechanism to form a joint action. A centralized Critic evaluates the joint action based on the global state and the evaluation results, updating the Critic network parameters and the network parameters of each agent using the evaluation results. After training, each agent can generate candidate control actions based on its local observations. When coordination by the shared execution mechanism is required or when entering the RL-CP collaborative protection mode, the centralized Critic is invoked to evaluate the joint candidate actions.
[0191] Specifically, the RL-CP collaborative protection model in this embodiment ranges from level 1 to level 2. The network consists of the local action network (Actor) corresponding to each level, the centralized evaluation network (Critic), the local observation definition for each level, the global state definition, the action space for each level, the shared action constraints, the reward function parameters, the state normalization parameters, and the network parameters after training.
[0192] First, establish the state space for each machine level. The local state vector is represented as follows:
[0193]
[0194] In the formula: Indicates the first Current state vector; Indicates inlet pressure; Indicates the inlet temperature; This indicates export pressure; Indicates the outlet temperature; Indicates mass flow rate; Indicates the rotor speed; Indicates the inlet guide vane angle; Indicates the opening degree of the recirculation valve; Indicates surge margin; Indicates the rate of change of surge margin; This represents the transpose of a vector.
[0195] No. When a certain actuator is not configured at the local level, the state of the corresponding actuator is not used as a valid control state variable at the local level.
[0196] The range of values for each state variable in the full-condition test data was used to evaluate the first... Normalize the level state vector to obtain .
[0197] No. Local observations of a machine-level agent are defined as a combination of the machine-level normalized state and the local states of its directly neighboring machine-level agents.
[0198]
[0199] In the formula: Indicates the first The current local observation of the intelligent agent; Indicates the first Level-normalized state vector; Indicates the relationship with the first The set of local state variables of directly adjacent machine levels.
[0200] In one implementation, At least including the surge margin of directly adjacent turbine stages, the rate of change of surge margin, and the first stage. Pressure and mass flow rates at interstage interfaces between a stage and adjacent stages. For intermediate stages, local observations include the aforementioned local states of the upstream and downstream directly adjacent stages. For the first and last stages, local observations include the aforementioned local states of their only directly adjacent stage. The normalized states of all stages constitute the global state:
[0201]
[0202] In the formula: This indicates the current global status of the multistage centrifugal compressor; to They represent levels 1 to 10 respectively. Level-normalized state vector; This indicates the total number of machine classes.
[0203] Then, the motion space is established based on the actual actuators of each stage. When the stage has an independent recirculation valve, an independent inlet guide vane, and an independent shaft, the action vector of this stage is expressed as:
[0204]
[0205] In the formula: Indicates the first Level action vector; This represents the normalized increment of the recirculation valve opening. This represents the normalized increment of the inlet guide vane angle; This represents the normalized speed increment; This indicates the inter-stage feedforward compensation requirement for directly adjacent stages; This represents the transpose of a vector.
[0206] No. When there is no corresponding independent executive body at the level, from the 1st level... The corresponding action is deleted in the independent action dimension. When two or more machine stages share the same shaft system, each relevant machine stage actor outputs its own speed control requirements for the shared shaft system, but in the actual joint action, only one shared shaft system speed action variable is set. When multiple machine stages share the same recirculation valve, each relevant machine stage outputs its own opening control requirements, but in the actual joint action, only one shared recirculation valve opening action variable is set.
[0207] After defining the state space, observation space, and action space, establish the first... Tiered reward function:
[0208] (2)
[0209] In the formula: Indicates the first Current control step size reward; Indicates the first Minimum surge margin required for the level; Indicates the first Current surge margin; Indicates the adjacent first Minimum surge margin required for the level; Indicates the adjacent first Current surge margin; Indicates the first Current efficiency level; Indicates the normalized recirculation valve opening; This represents the current control step size action vector; This represents the action vector of the previous control step. Represents the set of directly adjacent machine levels; , , , and These represent the non-negative weights corresponding to insufficient surge margin, efficiency, recycle, motion variation, and insufficient surge margin of adjacent machine levels, respectively.
[0210] The summation of rewards at each machine level forms the global reward for the current control step size:
[0211]
[0212] In the formula: This indicates the global return for the current control step size; Indicates the first Level rewards; Indicates the total number of machine classes; Indicates the machine class number.
[0213] Then, a dynamic training environment for MADDPG training is established. The training environment uses the full-condition steady-state operating point, full-condition dynamic response data, corrected one-dimensional thermodynamic model, and the mass flow rate-pressure ratio characteristic relationship verified by three-dimensional CFD, as well as the surge boundary of each stage, obtained from S1. For a training control step, the training environment reads the current inlet state, mass flow rate, speed, and actuator state of each stage, and changes the corresponding recirculation valve opening, inlet guide vane angle, and / or speed according to the current action. The first step is adopted. After the stage performance relationship calculation is performed, the stage pressure ratio and efficiency are determined. Then, the outlet pressure is determined based on the current inlet pressure and the calculated pressure ratio, and the outlet temperature is determined based on the one-dimensional thermodynamic relationship.
[0214] For adjacent stages connected in series, the outlet state of the upstream stage combined with the interstage cooling state serves as the inlet boundary condition for the downstream stage. The flow state of the next stage is updated according to the continuous relationship of interstage mass flow. Then, the surge margin of the next control step is calculated based on the updated corrected mass flow, corrected speed, and surge boundaries of each stage. The actuator response and dynamic response data of pressure and mass flow obtained from the full-condition dynamic test are used to correct the rate of state change between adjacent control steps.
[0215] Through the above processing, in the current global state and actual joint actions Given the current state, the training environment obtains the global state for the next control step. .
[0216] No. Establish the current Actor network at each level. and target Actor network All machine levels share a single centralized Critic network. A target Critic network . No. The current Actor outputs an action based on local observations:
[0217]
[0218] In the formula: Indicates the first The current Actor outputs the action; Indicates the first The current Actor network; Indicates the first Level local observation; Indicates the first The current Actor network parameters.
[0219] All machine-level actions, after being processed by shared motion constraints, constitute the combined actions actually executed in the training environment. The centralized Critic receives the global state and joint actions, and outputs the corresponding evaluation value:
[0220]
[0221] In the formula: This represents the current centralized Critic network; Indicates the current Critic network parameters; Indicates the current global state; This indicates a current joint action that satisfies the constraints of the actual implementing agency.
[0222] After each training step is completed, the experience samples are saved as:
[0223]
[0224] In the formula: express Momentary experience samples; Indicates the current global state; Represents the entire set of local observations at the machine level; Indicates the current joint action; This indicates the current global return. Indicates the global state of the next control step; This represents the set of all machine-level local observations for the next control step.
[0225] Experience samples are stored in an experience sample pool, and random samples are drawn from the experience sample pool. The sample is updated in the network.
[0226] The target Critic rating is calculated using the following formula:
[0227]
[0228] In the formula: Indicates the first The target evaluation value of each training sample; Indicates the first Global reward for each training sample; Indicates the discount factor; Represents the target Critic network; Indicates the first The next global state for each training sample; This represents the joint target action generated by the target Actor based on the next local observation and processed by shared action constraints; This indicates the sample number of the training batch.
[0229] The current Critic loss is determined by the following formula:
[0230]
[0231] In the formula: Indicates the current Critic loss; Indicates the current Critic network parameters; Indicates the number of samples in the training batch; Indicates the first The global state of each sample; Indicates the first Joint action of samples; This indicates the current Critic rating. This represents the corresponding target evaluation value.
[0232] By reducing Update the current Critic network parameters.
[0233] No. The level actor is updated according to the evaluation gradient of the centralized Critic regarding the machine-level action:
[0234]
[0235] In the formula: Indicates the first Level Actor policy gradient; Indicates the first Level Actor parameters; Indicates the first The optimization objective for a Level Actor; Indicates the number of samples in the training batch; This indicates the centralized Critic rating value with respect to the first... The gradient of the action level; Indicates the first The level Actor outputs the gradient with respect to the Actor parameters; Indicates the first The first of the samples Level local observation.
[0236] After completing one round of current network parameter updates, perform a soft update on the target Actor:
[0237]
[0238] In the formula: Indicates the first The first round of updates Level target Actor parameters; Indicates the first The current Actor parameters after the round of updates; Indicates the first Wheel target Actor parameters; Indicates the soft update coefficient; This indicates the number of times the network parameters are updated.
[0239] The target Critic performs a soft update according to the following formula:
[0240]
[0241] In the formula: Indicates the first Wheel target Critic parameters; Indicates the first The current Critic parameters; Indicates the first Wheel target Critic parameters; Indicates the soft update coefficient; This indicates the number of times the network parameters are updated.
[0242] In one implementation, the Actor learning rate is adopted. Critic learning rate adopts Discount factor Using 0.99, soft update coefficient Using 0.005, for each training batch 256 samples were used.
[0243] Training conditions include stable operation, near-surge, surge, inlet pressure changes, inlet temperature changes, load changes, speed changes, and start-stop processes. During training, the average global reward across multiple training rounds is continuously calculated, and validation conditions not involved in the current network update are used to check the surge margin and Actor output of each stage. Offline training is completed when the average global reward stabilizes, the surge margin insufficiency events in each stage during validation conditions no longer increase significantly with continued training, and there is no sustained divergence or continuous large oscillations in the executed actions.
[0244] After training is complete, save levels 1 through 2. The parameters include the current Actor parameters, the current Critic parameters, the state normalization parameters, the action range of each independent actuator, the shared action constraints, and the reward function parameters.
[0245] The trained RL-CP collaborative protection model is represented as follows:
[0246]
[0247] In the formula: This represents the RL-CP collaborative protection model; to Indicates level 1 to level 2. The Actor network that has been trained to the required level; This represents a centralized Critic network that has completed training. Represents the set of state normalization parameters; This represents the set of parameters representing the range of motion of an independent actuator. This represents the set of shared motion constraint parameters corresponding to the shared shaft system and shared recirculation valve. This represents the set of parameters for the reward function.
[0248] In actual operation, each actor generates local candidate actions based on the current local observations; when the shared execution mechanism needs to coordinate or enter the RL-CP collaborative protection mode, the centralized Critic evaluates different executable joint actions in the current global state.
[0249] Furthermore, in actual operation, this embodiment also utilizes an online fine-tuning strategy to continuously optimize the parameters of each level of the intelligent agent. To ensure that this online fine-tuning has a feasible execution path, during actual operation, after each control cycle is completed, the current global state can be adjusted according to the same data structure as the offline training experience samples. All machine-level local observation sets Joint actions actually carried out The actual global return is calculated based on the actual surge margin, efficiency, recycle usage, and motion variations of each engine class. Global state in the next control cycle and the local observation set for the next control cycle The online operational experience samples are compiled and stored in the online experience sample pool.
[0250] It should be noted that online fine-tuning does not necessarily require immediate updates to network parameters in every control cycle. In one possible implementation, when the online experience sample pool accumulates to a preset number of samples, and the unit is not currently in a state requiring immediate emergency anti-surge protection, samples are extracted from the online experience sample pool. The trained Actors and Critics are then incrementally updated using the same Critic loss, Actor policy gradient, and target network soft update relationship as in the offline MADDPG training described above. The learning rate used in online fine-tuning can be set to be no greater than the corresponding offline training learning rate to reduce sudden changes in policy parameters during actual operation.
[0251] To avoid weakening deterministic safety protection with unverified online update results, online fine-tuning is only used to update the intelligent candidate action generation and centralized evaluation parameters in S3. It does not change the physical surge boundary established in S1, nor does it change the deterministic recirculation valve safety baseline and actuator limiting rules established in S2. The updated Actor candidate actions still need to be compared with the deterministic safety baseline and constrained by the actuator in S5 before they can form actual control actions. The updated network parameters can also be evaluated using preset verification conditions or the most recent actual operating samples. If the updated model does not cause a deterioration in the surge margin evaluation and the output action meets the actuator constraints, then the current online model parameters will be replaced; otherwise, the model parameters before the update will continue to be used.
[0252] S4: Real-time acquisition of operating data for each turbine level; determination of surge margin and its changing trend based on the corresponding surge boundary; and determination of surge risk status and risk level for each turbine level by combining pressure pulsation, low-frequency oscillation of mass flow, vibration characteristics, fuzzy logic reasoning and support vector machine classification results.
[0253] Specifically, during actual operation, operating data of each machine level are collected synchronously according to the control cycle, and the correction method of S1 is used to calculate the first... The current corrected mass flow rate, corrected rotational speed, and pressure ratio are determined.
[0254] Using the first The surge boundary function is used to calculate the surge correction mass flow rate corresponding to the current correction speed, and then the current surge margin is calculated according to equation (1), and the surge margin change rate is calculated according to S2.
[0255] At the same time, the current operating point is determined based on the current corrected mass flow rate, corrected rotational speed, and pressure ratio. The position of the mass flow rate-pressure ratio characteristic curve and its location in the safe zone, warning zone, or surge zone.
[0256] Pressure pulsation features are extracted from the real-time pressure dynamic signal using the same method as in S1. Low-frequency oscillation amplitude and dominant frequency are extracted from the mass flow dynamic signal. Vibration features are extracted from the vibration dynamic signal. After processing using the normalized parameters saved in S1, the results are input into the first... Level 1 Support Vector Machine (SVM) classification model.
[0257] Support Vector Machine outputs the current dynamic state category .
[0258] Fuzzy logic reasoning uses surge margin and surge margin change rate as two fuzzy inputs. For surge margin, three fuzzy sets—safety margin, warning margin, and surge margin—are established using the surge margin distributions corresponding to the safe zone, warning zone, and surge zone samples in S1, respectively. These three fuzzy sets overlap in the state boundary region, allowing operating points near the boundaries of the safe zone and warning zone to simultaneously possess both safe and warning membership, and vice versa.
[0259] Get the first Current surge margin for state category Membership degree:
[0260]
[0261] In the formula: Indicates the first Current surge margin for state category Membership degree; Indicates the category of safety, warning, or surge status; Indicates the current control moment.
[0262] For the surge margin change rate, based on the dynamic test data of the S1 full-condition test, three types of change rate samples were obtained: the operating point continuously approaches the surge boundary, the state changes are small, and the operating point recovers in the safe direction. Three fuzzy sets were established based on the range of the three types of samples: approaching surge, basically stable, and far from surge.
[0263] Obtain the surge margin change trend for state category Risk membership degree:
[0264]
[0265] In the formula: Indicates the first The trend of surge margin change on state category Membership degree; Indicates the category of safety, warning, or surge status; Indicates the current control moment.
[0266] Fuzzy logic rules adopt the following relationship:
[0267] When the surge margin is within the safety margin and the trend of the surge margin change is basically stable or far from the surge, the fuzzy logic outputs a safe state. When the surge margin is within the safety margin, but the trend of the surge margin change is approaching the surge, the fuzzy logic outputs a warning state. When the surge margin is within the warning margin, and the trend of the surge margin change is far from the surge or basically stable, the fuzzy logic outputs a warning state. When the surge margin is within the warning margin, and the trend of the surge margin change is approaching the surge, the fuzzy logic outputs a surge state. When the surge margin is within the surge margin, regardless of which of the above fuzzy states the surge margin change trend belongs to, the fuzzy logic outputs a surge state.
[0268] In cases where multiple rules are triggered simultaneously within an overlapping region of a fuzzy set, the smaller value of the corresponding input membership degree is used as the activation degree of the corresponding rule, and the larger value of the activation degree of all rules in the same output state is taken as the fuzzy inference result of that output state.
[0269] Let the activation degrees corresponding to the safety, warning, and surge states obtained by fuzzy logic be respectively... , and .
[0270] Support Vector Machine classification results are converted into state indicators:
[0271]
[0272] In the formula: Indicates the first Level for state category Support vector machine state indicators; Indicates the output state of the support vector machine; It indicates a safe, warning, or surge condition.
[0273] The fuzzy logic output is fused with the SVM classification result:
[0274]
[0275] In the formula: Indicates the first Level for state category The final state evaluation quantity; This represents fuzzy logic for state categories. The output activation level; This represents the state indication of a support vector machine; This represents the fusion weights of the fuzzy logic results; This represents the weights used to fuse the classification results of the support vector machine, and .
[0276] The fusion weights are calibrated using the test samples in S1 that already have actual state labels, and the weight combinations that make the final classification results of the validation samples highly consistent with the test states are selected.
[0277] No. The current risk level is determined by the following formula:
[0278]
[0279] In the formula: Indicates the first Current surge risk status; Indicates state category The corresponding final state evaluation quantity; Indicates a safe, warning, or surge condition; This indicates that the state category with the largest final state evaluation value is selected.
[0280] When two state categories have the same final evaluation value, the state category with the higher risk level is adopted as the current state.
[0281] when In case of early warning or surge, the first The risk level has been determined to be level 3.
[0282] S5: The anti-surge agent corresponding to the risk level generates candidate control actions based on local observations. It combines the baseline control requirements of the recirculation valve formed by the corresponding hierarchical independent anti-surge controller in S2 with the coordination results of the inlet guide vane and / or speed to determine the actual local control action of the risk level. After the actual local control action is formed, it forms feedforward compensation requirements to adjacent levels based on the inter-level feedforward gain and propagation delay pre-calibrated in S2. When a shared shaft system or a shared recirculation valve is involved, each relevant level forms control requirements for the corresponding shared actuator and sends the shared control requirements to S6 for coordination.
[0283] Specifically, the first Once the risk level is determined to be a risk level, the RL-CP collaborative protection model established by S3 is called from the first level. Level Actor, and read the first Local observation at level .
[0284] Local observations include the first The inlet pressure, inlet temperature, outlet pressure, outlet temperature, mass flow rate, rotor speed, inlet guide vane angle, recirculation valve opening, surge margin and surge margin change rate of the stage itself, as well as the surge margin, surge margin change rate of the directly adjacent stage and the corresponding interstage interface pressure and mass flow rate, are all processed according to the normalized parameters saved in S3.
[0285] No. Level Actor outputs local-level candidate control actions:
[0286]
[0287] In the formula: Indicates the first The current candidate control action for the level Actor; This indicates the completion of the training. Actor of the Class; Indicates the first Current local observations at level [level].
[0288] Then, based on the risk status determined by S4, the current surge margin, and the rate of change of surge margin, the action selection rule established by S2 is invoked. The level is in a warning state, and the current surge margin is not lower than [previous level]. If the risk does not deteriorate rapidly, and an independent inlet guide vane and / or an independent rotating shaft are configured, the candidate actions of the inlet guide vane and / or rotation speed given by the Actor will be checked first. Based on the machine-level mass flow rate-pressure ratio characteristic relationship and isoefficiency line established by S1, the direction of the working point movement after the action will be determined. Only actions that can increase the machine-level surge margin and are within the allowable range of the actuator will be retained.
[0289] When the Level surge margin is lower than If the surge margin continues to decrease and the dynamic characteristics develop towards a surge state, or if the system is already in a surge state, the recirculation valve baseline control rule established in S2 is invoked. For the first... In the case where the stage has an independent recirculation valve, read the deterministic recirculation valve baseline control requirements calculated by S2 based on the state of this control cycle. This serves as the safety baseline that actual recirculation control must not fall below.
[0290] No. The normalized recirculation valve candidate opening increment given by the level Actor is denoted as When the local stage is in a state of insufficient surge margin, the control quantity of the local independent recirculation valve is determined according to the following formula:
[0291]
[0292] In the formula: This represents the normalized opening increment of the actual output to the independent recirculation valve in this stage; This indicates the deterministic baseline control requirements obtained from S2; This represents the increment of the candidate opening of the recirculation valve obtained by the Actor; This indicates that the amplitude is limited according to the actual operating range of the recirculation valve.
[0293] When the candidate Actor recycle opening is lower than the deterministic baseline requirement, or when a reduction in the necessary recycle amount is required under conditions of insufficient surge margin, the deterministic baseline requirement is adopted.
[0294] For situations where the risk level is in a warning state but has not yet entered a surge state, if the risk urgency assessment established by S2 indicates that the current situation is a non-emergency warning condition, then inlet guide vane and / or independent speed candidate actions that meet the above conditions can be implemented first. If the surge margin is still lower than the required value after implementing guide vane / speed adjustment, or if the risk further deteriorates, then recirculation protection should be implemented according to the above recirculation valve safety baseline. For situations where the recirculation valve safety baseline has already been triggered, inlet guide vane and / or speed candidate actions can still be superimposed as auxiliary actions, but not at the expense of reducing the necessary recirculation amount.
[0295] No. After a level actually generates a local control action, the actual executed local action is used as the input for calculating inter-level disturbances.
[0296] It should be noted that S5 no longer re-identifies the propagation delay and feedforward gain. Instead, it reads or interpolates the propagation delay between the i-th stage and the adjacent j-th stage from the parameters already calibrated and stored in S2, based on the current corrected rotational speed, corrected mass flow rate, pressure ratio, and interstage cooling status. And the feedforward gain corresponding to the current action type q and the compensation actuator type r .
[0297] In actual operation, the first The feedforward compensation requirement is determined according to the following formula:
[0298]
[0299] In the formula: Indicates by the first Level local actions are generated and sent to the first Level 1 Feedforward compensation requirements of similar execution agencies; This indicates the feedforward gain corresponding to the current operating condition; Indicates the first The current actual execution level is the first Class-based normalization control actions.
[0300] According to the propagation delay Arrange adjacent The execution time of the compensation action at the first level, so that the compensation action is performed in the first... Pressure or mass flow disturbances caused by local control actions reach the first level It begins to play a regulatory role at or before the time of the event.
[0301] Adjacent When a level has a corresponding independent executive body, it shall be handled by the first level. The first-level controller performs the corresponding feedforward compensation. When compensation requires the use of a shared shaft system or a shared recirculation valve, the second-level controller... The level only forms the shared execution mechanism control requirements.
[0302] When the risk level itself needs to adjust the shared actuator, it will only generate the corresponding shared control requirements.
[0303] The set of shared control requirements for the current control cycle is represented as follows:
[0304]
[0305] In the formula: This indicates the set of shared control requirements for the current control cycle; Indicates the first Item sharing control requirements; Indicates the current number of shared control requirements; Indicates the control requirement number.
[0306] By the end of S5, a set of local independent actions, adjacent machine-level feedforward compensation requirements, and shared control requirements that have been actually implemented at the risk level have been identified. The above results will proceed to S6.
[0307] S6: Continuously monitor the surge risk after local control of the risk level. If the risk is effectively mitigated within the preset time window, maintain independent control of this level. If it is not effectively mitigated or further aggravated, switch to RL-CP collaborative protection mode. The upper-level coordinator coordinates the recirculation valve, inlet guide vane and / or speed of the relevant level according to the candidate actions of each agent and the global state evaluation results.
[0308] Specifically, since No. From the moment the S5 local control action is initiated, data will be continuously collected. The operating data of the stage and its directly adjacent stages are used, and the surge margin, surge margin change rate and risk status are recalculated according to S4.
[0309] No. The start time of the local control effect evaluation is recorded as follows: .
[0310] Based on the dynamic test data of S1 under all operating conditions, several representative operating conditions were selected, and the response time required from the start of local anti-surge vibration operation to the emergence of identifiable improvement of machine-level surge margin or risk status was recorded under each representative operating condition.
[0311] No. The preset time window is determined according to the following formula:
[0312]
[0313] In the formula: Indicates the first Evaluation time window for local control effectiveness; This represents the set of representative operating conditions used for time window calibration. Indicates the first The response time required from the start of local control actions to the emergence of an identifiable and achievable risk condition under a representative operating condition; This indicates the time margin for covering detection fluctuations and discrete control cycle settings; This indicates the representative operating condition number.
[0314] Within the preset time window, the first When the surge level is reduced from a state of surge to a state of warning or safety, or from a state of warning to a state of safety, it is determined that the local control has effectively mitigated the surge risk.
[0315] The risk level remains unchanged, but the surge margin is relative to... When the surge margin change rate no longer remains less than zero, and the pressure pulsation, mass flow rate low-frequency oscillation, vibration characteristics, and fusion state results obtained in S4 do not continue to change towards a surge state, it is also determined that the surge risk has been effectively mitigated.
[0316] After the preset time window ends, the first If the risk level remains the same as when local control began, and the surge margin does not show an increasing trend, it is determined that local control has not effectively mitigated the surge risk.
[0317] The preset time window has not yet ended, but the first... When the risk level changes from a warning state to a surge state, or when the surge margin continues to decrease and the dynamic characteristics continue to change towards a surge state, the surge risk is determined to have further increased.
[0318] S6 simultaneously receives the set of shared control requirements generated by S5.
[0319] For a shared execution agency ,Depend on The control requirements formed by each related machine level are expressed as follows:
[0320]
[0321] In the formula: Indicates targeting shared execution agencies The resulting set of control requirements; to They represent Each relevant level is for shared execution agencies The proposed control requirements; Indicates the number of machine levels that use the shared actuator; Indicates the shared executive agency number.
[0322] When local controls have effectively mitigated the risk, but shared actuator control needs still exist, the local actions that the risk-level machine can already execute independently are not changed. The higher-level coordinator restricts the control needs of each machine-level machine belonging to the same shared actuator to the actuator's allowable range, and forms a shared action candidate set based on the values of each demand value and the values between adjacent demand values according to the actuator's minimum controllable increment.
[0323] Each shared action candidate value is combined with the currently determined independent actions to form a finite number of joint candidate actions, and a centralized Critic evaluation is performed using S3 training. The shared action with the highest evaluation value is selected as the actual control quantity for that shared actuator. This coordination only addresses the problem that the same physical actuator cannot simultaneously execute multiple different requirements; it does not reallocate other control actions that can already be executed independently at the local level.
[0324] When the risk level is not effectively mitigated within the preset time window or the risk further intensifies, switch to RL-CP collaborative protection mode.
[0325] After entering the RL-CP cooperative protection mode, the risky machine level and its directly adjacent machine levels affected by the local actions of the risky machine level and requiring inter-level compensation are identified as the relevant machine levels for this cooperative operation. Each relevant machine level actor generates new candidate control actions based on the current local observations.
[0326] For actions that can be executed independently, the candidate actions output by the Actor are used as the corresponding candidate center values, and a limited number of neighboring candidate values are obtained within the allowable range of the actuator based on the minimum controllable change. For shared shaft systems or shared recirculation valves, shared action candidate values are obtained based on the relevant shared control requirements formed in S5, and it is ensured that the same shared actuator has only one actual action variable within a set of joint candidate actions.
[0327] Form a set of executable joint candidate actions:
[0328]
[0329] In the formula: This indicates the set of joint candidate actions that can be executed in the current control cycle; Indicates the first Groups can execute joint candidate actions; Indicates the total number of joint candidate actions; Indicates the joint candidate action number.
[0330] Each set of joint candidate actions satisfies the allowable action range of the independent actuators and the shared action constraint that the shared shaft system and the shared recirculation valve have only one actual control degree of freedom.
[0331] Current global state Each joint candidate action is input into the centralized Critic trained by S3.
[0332] The final coordinated joint actions are determined according to the following formula:
[0333]
[0334] In the formula: This indicates the coordinated joint actions determined in the current control cycle; Indicates the set of executable joint candidate actions; Represents any joint candidate action in the set; This indicates a centralized Critic that has completed S3 training; Indicates the current global state; Indicates the execution of a combined action in the current global state. The corresponding evaluation value; This indicates that the joint candidate action with the highest evaluation value is selected.
[0335] For independent recirculation valves, independent inlet guide vanes, and independent rotating shafts, The corresponding actual control quantity is sent to the corresponding actuator.
[0336] For a shared shaft system It only includes the actual speed control value of a single common shaft system; for a common recirculation valve, It contains only the actual opening control value of a single shared recirculation valve.
[0337] After completing the current coordinated action, S4 is re-executed in the next control cycle. If the risk has been effectively mitigated, the current RL-CP cooperative protection mode is exited, and hierarchical independent control is restored. If the risk has not been effectively mitigated, the corresponding control procedures in S5 and S6 are re-executed based on new local observations, global status, and shared control requirements.
[0338] Furthermore, after completing an actual control cycle and obtaining the operating status of the next control cycle after execution, corresponding online operating experience samples are formed according to the online fine-tuning mechanism established in S3. If the current unit is in a risky state that requires immediate execution of anti-surge protection, only the sample is recorded while the updates of Actor and Critic parameters are frozen. When the risk has been mitigated and the online update conditions are met, the accumulated actual operating samples are used to incrementally fine-tune Actor and Critic. Thus, online fine-tuning does not participate in the immediate closed-loop calculation of the current emergency protection action, but rather adaptively updates the RL-CP strategy parameters used in subsequent control cycles without weakening the S1 physical surge boundary and the S2 deterministic safety baseline.
[0339] For example, in a three-stage series centrifugal compressor, stages 1, 2, and 3 each have their own graded surge boundary identification model and graded independent anti-surge controller. When stage 2 enters a warning state, stage 2 Actor reads its own state and the local states of stages 1 and 3 to generate candidate control actions. When the current surge margin of stage 2 is not lower than the required minimum surge margin and the risk does not deteriorate rapidly, candidate actions for increasing the surge margin of stage 2, such as inlet guide vanes and / or rotational speed, are first checked. When the surge margin of stage 2 is insufficient or the risk continues to worsen, the baseline control rule for the recirculation valve established by S2 is invoked based on the insufficient surge margin of stage 2, and the actual local recirculation valve opening increment is determined by comparing it with the candidate recirculation valve actions of the Actor.
[0340] After the second stage implements local control actions, it generates feedforward compensation requirements for adjacent stages based on the pre-calibrated feedforward gain and propagation delay between the second and first / third stages. When the third stage has independent inlet guide vanes, it adjusts the inlet guide vanes in advance based on the feedforward compensation requirements received by the third stage. When the three stages share the same shaft system, the first, second, and third stages generate their own speed control requirements for the shared shaft system, but the shared shaft system only executes one actual speed control quantity.
[0341] Starting from the implementation of local control actions at Level 2, the surge margin, surge margin change rate, and risk status are continuously monitored within the corresponding preset time window for Level 2. If Level 2 recovers to a safe state or shows a clear risk mitigation trend within the preset time window, Level 2 local control continues. If coaxial speed requirements still exist, shared action coordination is only performed on the shared shaft speed.
[0342] If Level 2 fails to provide effective relief within the preset time window or further enters a surge state, the RL-CP collaborative protection mode is activated. Levels 1, 2, and 3 actors generate candidate actions based on their local observations. The higher-level coordinator constructs joint candidate actions that satisfy the single speed freedom requirement of the coaxial system based on the allowable range of each independent actuator and the speed requirements of the three machine levels for the shared shaft system. The centralized Critic evaluates the current global state of the three machine levels and each joint candidate action, and then determines the corresponding control quantity for the independent actuator and the unique actual speed control quantity for the shared shaft system.
[0343] Therefore, in this embodiment, S1 is responsible for establishing the independent performance relationship, surge boundary, and dynamic state identification model of each machine level; S2 is responsible for pre-establishing the deterministic anti-surge control rules, action selection rules, and decoupling parameters such as inter-level feedforward gain and propagation delay for each machine level; S3 establishes and trains the RL-CP collaborative protection model composed of each level of Actor, centralized Critic, state space, action space, shared action constraints, and reward parameters, and performs online fine-tuning according to safety constraints after the accumulation of actual operating data; S4 determines the risk machine level according to the real-time surge margin, change trend, dynamic characteristics, fuzzy logic reasoning, and support vector machine classification results; S5 no longer re-establishes the control laws or identifies feedforward parameters in S2, but directly calls the baseline rules, feedforward gain, and propagation delay already determined in S2, and combines the Actor candidate actions in S3 to form the actual local actions of the risk machine level, the feedforward compensation requirements of adjacent machine levels, and the control requirements of shared actuators; S6 determines whether to maintain hierarchical independent control, only coordinate shared actuators, or enter RL-CP collaborative protection based on the actual risk changes after local control, thereby forming a hierarchical anti-surge control process for multi-stage centrifugal compressors.
[0344] Example 2
[0345] This embodiment provides a multi-stage centrifugal compressor staged independent monitoring anti-surge protection system. This system is applied to a multi-stage centrifugal compressor consisting of n-stage centrifugal compressors connected in series along the working fluid flow direction, where n is an integer greater than or equal to 2. Each stage is sequentially designated as stage 1, stage 2, to stage n along the working fluid flow direction. Each stage is equipped with inlet guide vanes, recirculation branches, and corresponding recirculation valves according to its actual structure; alternatively, only some stages may have corresponding actuators. Each stage can have independent rotating shafts, or two or more stages can share the same shaft system. Each stage can have its own independent recirculation valve, or two or more stages can share the same recirculation valve.
[0346] The graded independent monitoring anti-surge protection system includes a parameter detection component, a multi-channel synchronous data acquisition component, multiple graded independent anti-surge controllers that correspond one-to-one with each machine level, a communication component, a coordination calculation component, and an execution component.
[0347] The parameter detection component is used to acquire the inlet pressure, inlet temperature, outlet pressure, outlet temperature, mass flow rate, rotational speed, and vibration signals of each stage.
[0348] Specifically, the parameter detection components include inlet pressure and inlet temperature detection devices respectively installed at the inlet of each stage, outlet pressure and outlet temperature detection devices respectively installed at the outlet of each stage, a mass flow detection device for obtaining the corresponding stage mass flow rate, a speed detection device corresponding to the stage shaft, and a vibration detection device installed on the bearing housing and / or casing of each stage. Stages with inlet guide vanes are also equipped with guide vane position feedback devices, and stages with recirculation valves are also equipped with valve position feedback devices. When an interstage cooler is installed, flow detection devices and temperature detection devices for obtaining the corresponding interstage cooling flow rate and cooling temperature are also provided.
[0349] In one embodiment, both the inlet pressure detection device and the outlet pressure detection device are pressure sensors used to convert the working fluid pressure at the corresponding location into a pressure detection signal. Both the inlet temperature detection device and the outlet temperature detection device are temperature sensors used to acquire the working fluid temperature at the corresponding location. The mass flow detection device is a mass flow meter used to acquire the working fluid mass flow rate at the corresponding stage. The speed detection device is a speed sensor used to acquire the actual speed of the corresponding independent shaft or shared shaft system. The vibration detection device is a vibration acceleration sensor used to acquire the vibration acceleration signal at the corresponding bearing housing and / or housing location. The guide vane position feedback device is an angle encoder or displacement sensor used to acquire the actual angle of the inlet guide vane. The valve position feedback device is a valve position transmitter or displacement sensor used to acquire the actual opening degree of the recirculation valve. The flow detection device for acquiring the interstage cooling flow rate is a flow meter, and the temperature detection device for acquiring the interstage cooling temperature is a temperature sensor.
[0350] For a machine stage with an independent shaft, the speed detection device acquires the actual rotor speed of the corresponding machine stage. For two or more machine stages sharing the same shaft system, the speed detection device corresponding to the shared shaft system acquires the actual shaft system speed, and provides the actual shaft system speed to the respective graded independent anti-surge controllers for each machine stage.
[0351] The mass flow detection device is positioned to obtain the actual mass flow rate of the corresponding stage. When there is no branch between two directly adjacent stages that causes the mainstream mass flow rate to split or merge, the corresponding mass flow rate can be obtained by a mass flow detection device positioned between the two adjacent stages.
[0352] The multi-channel synchronous data acquisition component is connected to the parameter detection component, each level of independent anti-surge controller, and the coordination calculation component, respectively. It is used to collect the operating data of each machine level according to a unified time reference, and send the operating data of each machine level to the corresponding level of independent anti-surge controller and coordination calculation component.
[0353] The multi-channel synchronous data acquisition component includes an acquisition interface, a unified clock unit, a synchronous sampling unit, and a real-time communication interface. Each pressure detection device, temperature detection device, mass flow detection device, speed detection device, vibration detection device, guide vane position feedback device, and valve position feedback device is connected to its corresponding acquisition interface. Pressure dynamic signals, mass flow dynamic signals, and vibration dynamic signals are continuously acquired through their respective acquisition channels, while temperature, speed, inlet guide vane position, and recirculation valve position are acquired according to the control cycle. The unified clock unit provides a unified time reference to different acquisition channels, and the synchronous sampling unit performs time mapping on each acquired data according to the unified time reference, ensuring that different operating data at the same sampling time are correlated.
[0354] Each tiered independent anti-surge controller is configured with its corresponding machine level. Each tiered independent anti-surge controller communicates with its directly adjacent tiered independent anti-surge controller via a communication component, and also communicates with the coordination computing component. Tiered independent anti-surge controllers in directly adjacent machine levels transmit feedforward compensation requirements to each other. Each tiered independent anti-surge controller also sends its own machine-level operating status, candidate control actions generated by the Actor, and control requirements for shared shaft systems or shared recirculation valves to the coordination computing component.
[0355] Each level of independent anti-surge controller includes a processor, memory, data acquisition interface, communication interface, and actuator control interface. The memory of the i-th level of independent anti-surge controller stores the mass flow rate-pressure ratio characteristic relationship corresponding to the i-th level, surge boundary relationship, regional parameters of the safe zone, warning zone, and surge zone, dynamic state feature normalization parameters, support vector machine classification model, minimum surge margin required for this level, recirculation valve control gain and limiting parameters, inlet guide vane and / or speed regulation parameters, inter-stage feedforward gain and propagation delay parameters corresponding to the directly adjacent level, preset time window, and parameters of the trained i-th level Actor network.
[0356] As per the instruction manual Figure 2 As shown, each stage of independent anti-surge controller includes an online surge identification module, a recirculation valve control module, an inlet guide vane adjustment and / or speed regulation module, and an adjacent stage feedforward compensation module. Furthermore, each stage of independent anti-surge controller also includes an anti-surge intelligent agent operating within it, which includes an Actor.
[0357] The surge online identification module determines the surge margin, surge margin change rate, and surge risk status of the machine stage based on the machine stage's operating data and stored surge boundary relationships and dynamic state identification parameters. The recirculation valve control module generates baseline control requirements for the recirculation valve based on insufficient surge margin at the machine stage and generates corresponding recirculation valve control quantities when recirculation adjustment is required at the machine stage. The inlet guide vane adjustment and / or speed adjustment module generates inlet guide vane and / or speed control requirements based on the machine stage's surge risk status, mass flow rate-pressure ratio characteristics, and the allowable operating range of the corresponding actuator. The adjacent machine stage feedforward compensation module generates feedforward compensation requirements for directly adjacent machine stages based on stored inter-stage feedforward gain and propagation delay after the machine stage generates actual local control actions.
[0358] As per the instruction manual Figure 3 As shown, each machine level is equipped with an anti-surge agent, and each anti-surge agent runs in a corresponding hierarchical independent anti-surge controller. The i-th level anti-surge agent uses the i-th level state and the local state of the directly adjacent machine level as local observations, and the local action network Actor is the policy network in the anti-surge agent that generates candidate control actions based on local observations.
[0359] When stage i is identified as a risky stage, stage i Actor generates candidate control actions based on current local observations. Stage i independent anti-surge controller determines the actual local control action of stage i based on the candidate control actions, the baseline control requirements of the recirculation valve, and the coordination results of the inlet guide vane and / or speed. When stage i has insufficient surge margin, the actual recirculation valve control quantity of stage i is not lower than the baseline control requirements of the recirculation valve.
[0360] After the i-th level independent anti-surge controller forms the actual local control action, it uses the actual executed local action as the calculation input for the feedforward compensation of the adjacent machine level. Based on the corresponding inter-level feedforward gain and propagation delay, it forms the feedforward compensation requirement of the directly adjacent j-th level and sends the feedforward compensation requirement to the j-th level independent anti-surge controller through the communication component.
[0361] When the j-th stage has a corresponding actuator capable of independent adjustment, the j-th stage independent anti-surge controller generates the corresponding control quantity based on the received feedforward compensation requirements. When adjustment is required via a shared shaft system or a shared recirculation valve, the j-th stage independent anti-surge controller generates control requirements for the corresponding shared shaft system or shared recirculation valve and sends the control requirements to the coordination calculation component.
[0362] When two or more machine stages share the same shaft system, each relevant machine stage actor generates speed control requirements for the shared shaft system. When multiple machine stages share the same recirculation valve, each relevant machine stage generates opening control requirements for the shared recirculation valve. Each relevant independent anti-surge controller sends the control requirements for the same shared shaft system or the same shared recirculation valve to the coordination calculation component.
[0363] The coordinating computing component includes a processor, memory, and a communication interface. The memory stores the trained centralized Critic network parameters, state normalization parameters, the permissible range of motion for each actuator, and the shared motion constraints corresponding to the shared shaft system and shared recirculation valve.
[0364] The centralized Critic receives the global state formed by the current operating status of each machine level and the joint actions that satisfy the allowable action range of the actuators and shared action constraints, and outputs the corresponding evaluation value. Each machine level has its own corresponding Actor, and all machine levels share the centralized Critic.
[0365] When machine-level control has effectively mitigated surge risk, but control requirements still exist for shared shaft systems or shared recirculation valves, the coordination calculation component maintains the already determined independent actions, limiting the control requirements for the same shared shaft system or the same shared recirculation valve to the allowable range of the corresponding actuator, and forming a set of shared action candidate values. Each shared action candidate value is combined with the already determined independent actions to form joint candidate actions, which are then evaluated by a centralized Critic. The shared action with the highest evaluation value is selected as the actual control quantity for the corresponding shared actuator.
[0366] For a shared shaft system, the actual control quantity of the shared actuator is the actual speed control quantity of the shared shaft system. For a shared recirculation valve, the actual control quantity of the shared actuator is the actual opening control quantity of the shared recirculation valve.
[0367] After implementing actual local control actions, the i-th level independent anti-surge controller continuously determines the i-th level surge risk status within a preset time window. When the i-th level surge risk is effectively mitigated, local independent control is maintained; when the i-th level surge risk is not effectively mitigated or further aggravated, the i-th level independent anti-surge controller sends a cooperative protection trigger signal to the coordination calculation component and enters the RL-CP cooperative protection mode.
[0368] After entering the RL-CP cooperative protection mode, the risk level and the directly adjacent levels that are affected by the actual local control actions of the risk level and require feedforward compensation are identified as the relevant levels for this cooperative operation. Each relevant level Actor generates candidate control actions based on the current local observations.
[0369] For actions that can be executed independently, the candidate control actions output by the Actor are used as the corresponding candidate center values, and the corresponding candidate values are obtained within the allowable action range of the corresponding actuator according to the minimum controllable change of the actuator. For shared shaft systems or shared recirculation valves, shared action candidate values are obtained according to the control requirements formed by each relevant stage, and the same shared shaft system or the same shared recirculation valve has only one actual action variable in a set of joint candidate actions.
[0370] The coordination calculation component forms an executable joint candidate action set based on independent action candidate values and shared action candidate values. It inputs the current global state and each executable joint candidate action into the centralized Critic, and selects the joint candidate action with the largest evaluation value as the coordination joint action for the current control cycle based on the evaluation value output by the centralized Critic.
[0371] The coordinated joint action satisfies the permissible operating range of each independent actuator, as well as the shared action constraints corresponding to the shared shaft system and shared recirculation valve. For independent recirculation valves, independent inlet guide vanes, and independent rotating shafts, the corresponding control quantities in the coordinated joint action are sent to the respective actuators. For the shared shaft system, only one actual speed control quantity for the shared shaft system is set in the coordinated joint action; for the shared recirculation valve, only one actual opening degree control quantity for the shared recirculation valve is set in the coordinated joint action.
[0372] The actuation components include a valve actuator connected to the recirculation valve, a guide vane actuator connected to the inlet guide vane, and a drive unit connected to the rotating shaft. Each stage of independent anti-surge controller sends the corresponding control quantity from the actual local control action, feedforward compensation action, or coordinated joint action to the corresponding actuation component through the actuator control interface. The actuation component adjusts the opening degree of the corresponding recirculation valve, the angle of the inlet guide vane, and / or the rotational speed according to the received control quantity.
[0373] After the execution component completes the current control action, the parameter detection component continues to acquire the operating data of each machine level, and the multi-channel synchronous data acquisition component sends the new operating data to the corresponding hierarchical independent anti-surge controller and coordination calculation component according to a unified time base.
[0374] Once the surge risk at the risk level is effectively mitigated, the current RL-CP collaborative protection mode is exited, and hierarchical independent control is restored. If the surge risk is still not effectively mitigated, each hierarchical independent anti-surge controller and coordination calculation component determines the control action for the next control cycle based on new local observations, global status, and control requirements arising from shared shaft systems or shared recirculation valves.
[0375] The above are merely embodiments of the present invention, and common knowledge regarding specific structures and characteristics in the solutions has not been described in detail here. It will be apparent to those skilled in the art that the present invention is not limited to the details of the above exemplary embodiments, and that the present invention can be implemented in other specific forms without departing from the spirit or essential characteristics of the invention. Therefore, the embodiments should be considered exemplary and non-limiting in all respects, and the scope of the present invention is defined by the appended claims rather than the foregoing description. Therefore, it is intended that all variations falling within the meaning and scope of equivalents of the claims be included within the present invention. No reference numerals in the claims should be construed as limiting the scope of the claims.
Claims
1. A method for staged independent monitoring and surge protection of a multi-stage centrifugal compressor, characterized in that, The method includes: S1: Collect full-condition operating data of each stage of the multi-stage centrifugal compressor, construct the mass flow rate-pressure ratio characteristic curve of each stage by combining the one-dimensional thermodynamic model and the three-dimensional CFD simulation model, determine the surge boundary of each stage according to the surge state, extract pressure pulsation, low-frequency oscillation of mass flow rate and vibration characteristics, and classify the safe zone, warning zone and surge zone by combining fuzzy clustering and support vector machine to establish a hierarchical surge boundary identification model; S2: Establish a hierarchical independent anti-surge controller for each stage, which has online surge identification, recirculation valve control, inlet guide vane adjustment and / or speed regulation, as well as inter-stage decoupling coordination rules; S3: The local states of the machine level and adjacent machine levels are used as local observations of each anti-surge agent. The corresponding action space is determined according to the configuration of the actuators of each machine level. For cases where two or more machine levels share the same shaft system or the same recirculation valve, the corresponding speed regulation amount or recirculation valve opening regulation amount is used as a shared action constraint. Each machine level agent generates control requirements for the shared shaft system or the shared recirculation valve. The MADDPG algorithm is used to train each machine level agent, and the centralized evaluation network is used to evaluate based on the global state and joint actions to establish an RL-CP collaborative protection model. S4: Real-time acquisition of operating data for each turbine level; determination of surge margin and its changing trend based on the corresponding surge boundary; and determination of surge risk status and risk level for each turbine level by combining pressure pulsation, low-frequency oscillation of mass flow, vibration characteristics, fuzzy logic reasoning and support vector machine classification results. S5: The anti-surge agent corresponding to the risk level generates candidate control actions based on local observations, and combines the recirculation valve baseline control requirements formed by the corresponding hierarchical independent anti-surge controller in S2 with the inlet guide vane and / or speed coordination results to determine the actual local control actions of the risk level. After forming the actual local control actions, it forms feedforward compensation requirements to adjacent levels based on the inter-level feedforward gain and propagation delay pre-calibrated in S2. When a shared shaft system or a shared recirculation valve is involved, each relevant level forms control requirements for the corresponding shared actuator and sends the shared control requirements to S6 for coordination. S6: Continuously monitor the surge risk after local control of the risk level. If the risk is effectively mitigated within the preset time window, maintain independent control of this level. If it is not effectively mitigated or further aggravated, switch to RL-CP collaborative protection mode. The upper-level coordinator coordinates the recirculation valve, inlet guide vane and / or speed of the relevant level according to the candidate actions of each agent and the global state evaluation results.
2. The multi-stage centrifugal compressor staged independent monitoring anti-surge protection method according to claim 1, characterized in that, In S5, after the i-th stage actually forms a local control action, the actual executed local action is used as the input for inter-stage disturbance calculation. Based on the current corrected speed, corrected mass flow rate, pressure ratio, and inter-stage cooling state, the propagation delay between the i-th stage and the adjacent j-th stage, as well as the feedforward gain corresponding to the current action type q and the compensation actuator type r, are read or interpolated from the parameters calibrated and stored in S2. The feedforward compensation requirement of the j-th stage is determined according to the following formula: ; In the formula: This represents the feedforward compensation requirement generated by the i-th level local action and sent to the j-th level r-type actuator; This indicates the feedforward gain corresponding to the current operating condition; This indicates the type q normalized control action currently being executed at level i; the execution time of adjacent level j compensation actions is arranged according to the propagation delay, so that the compensation actions begin to play a regulating role when or before the pressure or mass flow disturbance caused by the local control action at level i reaches level j.
3. The multi-stage centrifugal compressor staged independent monitoring anti-surge protection method according to claim 1, characterized in that, In S6, when local control has effectively mitigated the risk but there is still a need for shared actuator control, the local actions that the risk level can already execute independently are not changed. The control needs of each level belonging to the same shared actuator are restricted to the range allowed by the actuator. A shared action candidate set is formed by the values of each demand value and the values obtained by the actuator according to the minimum controllable increment of the actuator. Each shared action candidate value is combined with the currently determined independent actions to form a limited joint candidate action. The centralized Critic evaluation completed in S3 is used to select the shared action with the largest evaluation value as the actual control quantity of the shared actuator. When the risk level is not effectively mitigated or the risk is further aggravated within the preset time window, the system switches to RL-CP collaborative protection mode.
4. The multi-stage centrifugal compressor staged independent monitoring anti-surge protection method according to claim 1, characterized in that, In S5, the normalized candidate opening increment of the recirculation valve given by the i-th stage Actor is denoted as... When the local stage is in a state of insufficient surge margin, the control quantity of the local independent recirculation valve is determined according to the following formula: ; In the formula: This represents the normalized opening increment of the actual output to the independent recirculation valve in this stage; This indicates the baseline control requirements for the recirculation valve obtained from S2; This represents the increment of the candidate opening of the recirculation valve obtained by the Actor; This indicates that the amplitude is limited according to the actual operating range of the recirculation valve; when the candidate recirculation opening of the Actor is lower than the baseline control requirement of the recirculation valve, or when it is required to reduce the necessary recirculation amount under insufficient surge margin, the baseline control requirement of the recirculation valve is adopted.
5. The multi-stage centrifugal compressor staged independent monitoring anti-surge protection method according to claim 1, characterized in that, In S2, the propagation delay and feedforward gain between the control actions of each machine level and the dynamic response of the directly adjacent machine level are pre-calibrated using the full-condition synchronous dynamic test data obtained in S1. Let the normalized actual control action change sequence of the i-th level be... The normalized response sequence of the adjacent j-th level corresponding pressure or mass flow rate is as follows: Calculate the normalized cross-correlation coefficient and take the number of lag points corresponding to the largest absolute value of the cross-correlation coefficient: ; The propagation delay is determined by the following formula: ; The feedforward gain between the i-th level, q-th type actual action and the j-th level, r-th type compensation actuator is determined by the following formula: ; The above calibrations were performed for different corrected speeds, corrected mass flow rates, pressure ratios, and interstage cooling conditions. The resulting propagation delays and feedforward gains were grouped and stored according to the corresponding operating conditions. When the actual operating conditions were between adjacent calibration conditions, interpolation was performed based on the adjacent calibration parameters.
6. The multi-stage centrifugal compressor staged independent monitoring anti-surge protection method according to claim 1, characterized in that, In S3, when two or more machine classes share the same shaft system, each relevant machine class actor outputs its own speed control requirements for the shared shaft system, but in the actual joint action, only one shared shaft system speed action variable is set; when multiple machine classes share the same recirculation valve, each relevant machine class outputs its own opening control requirements, but in the actual joint action, only one shared recirculation valve opening action variable is set; after all machine class actions are processed by shared action constraints, they form the joint action actually executed in the training environment. The centralized Critic receives the global state and joint action and outputs the corresponding evaluation value.
7. The multi-stage centrifugal compressor staged independent monitoring anti-surge protection method according to claim 3, characterized in that, In S6, based on the dynamic test data of the full-condition S1, multiple representative operating conditions are selected, and the response time required from the start of local anti-surge vibration operation to the emergence of identifiable improvement in machine-level surge margin or risk state under each representative operating condition is recorded. The i-th level preset time window is determined according to the following formula: ; In the formula: This represents the time window for evaluating the effectiveness of the i-th level local control. This represents the set of representative operating conditions used for time window calibration. This represents the response time required from the start of local control action to the emergence of an identifiable and improveable risk state under the q-th representative operating condition; This indicates the time margin for covering the detection fluctuations and the discrete control cycle setting; within the preset time window, when the i-th level decreases from the surge state to the warning or safe state, or from the warning state to the safe state, it is determined that the local control has effectively mitigated the surge risk; If, after the preset time window ends, the i-th level still maintains the risk state at the start of local control, and the surge margin does not show an increasing trend, it is determined that the local control has not effectively mitigated the surge risk.
8. The multi-stage centrifugal compressor staged independent monitoring anti-surge protection method according to claim 1, characterized in that, In S3, the hierarchical reward function for the i-th level is established: ; In the formula: This represents the reward for the current control step size at level i; This represents the minimum surge margin required for level i; Indicates the current surge margin at level i; This represents the minimum surge margin required for the adjacent j-th level; Indicates the current surge margin of the adjacent j-th level; Indicates the current efficiency of the i-th level; Indicates the normalized recirculation valve opening; This represents the current control step size action vector; This represents the action vector of the previous control step. Represents the set of directly adjacent machine levels; , , , and These represent the non-negative weights corresponding to insufficient surge margin, efficiency, recycle, action variation, and insufficient surge margin of adjacent stages, respectively; the sum of rewards for each stage forms the global reward for the current control step size. ; In the formula: This indicates the global return for the current control step size. This indicates the total number of machine classes.
9. The multi-stage centrifugal compressor staged independent monitoring anti-surge protection method according to claim 1, characterized in that, In S1, while establishing the physical surge boundary, dynamic state features are extracted from the synchronous dynamic signals corresponding to each full-condition test working point. For the i-th level pressure dynamic signal, the average pressure component and low-frequency slow change trend within the analysis time window are removed to obtain the pressure pulsation component, and wavelet packet decomposition is performed. For the i-th level mass flow dynamic signal, after removing the average mass flow within the analysis time window, empirical mode decomposition is performed on the remaining dynamic signal. For the i-th level vibration dynamic signal, spectral analysis is performed on the vibration acceleration signals obtained from the bearing and / or housing positions. The pressure pulsation features, mass flow rate low-frequency oscillation features, and vibration features corresponding to the test point are used to form a dynamic state feature vector. The dynamic features of different dimensions are normalized, and three-category fuzzy clustering is performed on the i-th level normalized dynamic state features. The test samples with safety, warning, and surge labels are remapped to the mass flow rate-pressure ratio characteristic curves of the corresponding corrected rotation speed to form the i-th level independent safety zone, warning zone, and surge zone. The i-th level support vector machine classification model is trained with the normalized dynamic state feature vector as input and the safety, warning, and surge state labels as output.
10. A multi-stage centrifugal compressor staged independent monitoring anti-surge protection system, characterized in that, Applied to a multi-stage centrifugal compressor consisting of multiple centrifugal compressor stages connected in series along the direction of working fluid flow, the system includes a parameter detection component, a multi-channel synchronous data acquisition component, multiple graded independent anti-surge controllers respectively set one-to-one with each stage, a communication component, a coordination calculation component, and an execution component; The parameter detection component is used to acquire the inlet pressure, inlet temperature, outlet pressure, outlet temperature, mass flow rate, rotational speed, and vibration signal of each stage. The multi-channel synchronous data acquisition component is connected to the parameter detection component, each hierarchical independent anti-surge controller and the coordination calculation component respectively. It is used to collect the operating data of each machine level according to a unified time reference and send the operating data of each machine level to the corresponding hierarchical independent anti-surge controller and the coordination calculation component respectively. Each hierarchical independent anti-surge controller is connected to the hierarchical independent anti-surge controller corresponding to the directly adjacent machine level through a communication component, and is also connected to the coordination computing component; each hierarchical independent anti-surge controller also includes an anti-surge agent running inside it, which includes an Actor; When the i-th level is determined to be a risk level, the i-th level Actor generates candidate control actions based on the current local observations, and the i-th level hierarchical independent anti-surge controller determines the actual local control action of the i-th level based on the candidate control actions, the baseline control requirements of the recirculation valve, and the coordination results of the inlet guide vane and / or speed. After the i-th level independent anti-surge controller forms the actual local control action, it uses the actual executed local action as the calculation input for the feedforward compensation of the adjacent machine level. Based on the corresponding inter-level feedforward gain and propagation delay, it forms the feedforward compensation requirement of the directly adjacent j-th level and sends the feedforward compensation requirement to the j-th level independent anti-surge controller through the communication component. When adjustment is required via a shared shaft system or a shared recirculation valve, the j-th level independent anti-surge controller generates control requirements for the corresponding shared shaft system or shared recirculation valve and sends the control requirements to the coordination computing component; The coordination computing component includes a processor, a memory, and a communication interface. The memory stores the parameters of the trained centralized Critic network, state normalization parameters, the allowable range of motion of each actuator, and the shared action constraints corresponding to the shared shaft system and the shared recirculation valve. After implementing the actual local control action, the i-th level independent anti-surge controller continuously determines the i-th level surge risk status within a preset time window; When the risk of surge at level i is effectively mitigated, the local independent control is maintained; when the risk of surge at level i is not effectively mitigated or is further aggravated, the level i-level independent anti-surge controller sends a cooperative protection trigger signal to the coordination computing component and enters the RL-CP cooperative protection mode. After entering the RL-CP cooperative protection mode, the risk level and the directly adjacent levels that are affected by the actual local control actions of the risk level and need to be fed forward compensation are identified as the relevant levels for this cooperative operation. Each relevant level Actor generates candidate control actions based on the current local observations. The coordination calculation component forms an executable joint candidate action set based on independent action candidate values and shared action candidate values. It inputs the current global state and each executable joint candidate action into the centralized Critic, and selects the joint candidate action with the largest evaluation value as the coordination joint action for the current control cycle based on the evaluation value output by the centralized Critic. The execution components include a valve actuator connected to the recirculation valve, a guide vane actuator connected to the inlet guide vane, and a drive device connected to the rotating shaft; each stage of independent anti-surge controller sends the corresponding control quantity in the actual local control action, feedforward compensation action, or coordinated joint action to the corresponding execution component through the actuator control interface, and the execution component adjusts the opening degree of the corresponding recirculation valve, the angle of the inlet guide vane, and / or the rotation speed according to the received control quantity.
