Centrifugal compressor unit load intelligent distribution system and method

Abstract

The invention discloses an intelligent load distribution system and method for a centrifugal compressor unit. The system comprises a data sensing module, an intelligent decision-making module, an instruction execution module and a model updating module. The method comprises the steps that operation parameters of multiple parallel centrifugal compressors are synchronously collected in a high-precision mode through an industrial real-time network supporting hardware timestamps, and time-aligned system state observation data are obtained; constructing a state vector based on the data and a load prediction result, inputting a deep reinforcement learning agent fused with compressor physical characteristic constraints, and generating a load distribution instruction and a unit start-stop suggestion; after safety boundary verification is conducted on the original action, the guide vane opening degree is adjusted through a feedforward speed planning and fuzzy self-adaption PID composite servo control algorithm, and starting and stopping operation is executed; meanwhile, empirical data are collected online, and model parameters are periodically and finely adjusted. According to the invention, efficient, safe, accurate and self-evolution compressor group control is realized.

Description

TECHNICAL FIELD

[0001] The application belongs to the technical field of industrial automation and fluid machinery control, and relates to a centrifugal compressor unit load intelligent distribution system and method. BACKGROUND

[0002] In the industrial field, such as petrochemical, air separation and large refrigeration system, multiple centrifugal compressors are often used in parallel operation to meet different load requirements. The traditional control method usually relies on artificial experience or PLC logic based on fixed rules to start and stop and distribute the load of each compressor.

[0003] Such methods have obvious deficiencies: first, the dynamic coupling relationship between multiple units cannot be accurately coordinated, which easily causes uneven load distribution and inefficient operation of some units; second, data acquisition generally uses asynchronous mode, and the sampling time of the operating parameters of each compressor is inconsistent, which leads to distorted system state observation and makes it difficult to support high-precision optimization control; third, the control strategy does not fully consider the physical characteristic limitations of the compressor, such as surge boundary and guide vane adjustment rate, which may cause safety risks or frequent triggering of protection shutdown in actual operation.

[0004] In recent years, some research has attempted to use deep reinforcement learning for compressor group control, but most of the schemes are only trained in a simulation environment without embedding real physical constraints, and the stability is poor after deployment; and lack of safety verification mechanism for control instructions and high-precision servo control at the execution level, making it difficult to be reliably applied in industrial sites. At the same time, the existing system usually does not have online learning ability and cannot adapt to long-term dynamics such as equipment aging and working condition changes.

[0005] Therefore, a centrifugal compressor unit load intelligent distribution method is needed, which integrates high-precision synchronous perception, safe intelligent decision-making, precise composite execution and continuous online optimization, to realize efficient, stable and safe operation control. SUMMARY

[0006] To solve the problems in the background art, the application provides a centrifugal compressor unit load intelligent distribution system and method.

[0007] To achieve the above purpose, the technical scheme adopted by the application is as follows: a centrifugal compressor unit load intelligent distribution system, comprising:

[0008] A data perception module configured to perform high-precision synchronous acquisition of the operating parameters of multiple centrifugal compressors operating in parallel through an industrial real-time network supporting hardware timestamps to obtain time-aligned system state observation data;

[0009] An intelligent decision-making module in communication connection with the data perception module and configured to: construct a system state representation based on the system state observation data;

[0010] inputting the system state representation into a pre-trained artificial intelligence decision model, the artificial intelligence decision model being a deep reinforcement learning agent trained based on a proximal policy optimization algorithm and fused with compressor physical property constraints to ensure that the decision process meets dynamic constraints;

[0011] obtaining and outputting load distribution instructions and unit start-stop suggestions generated by the model;

[0012] The instruction execution module is in communication connection with the intelligent decision module and is configured to: parse the load distribution instructions, and drive the corresponding guide vane actuator through a composite servo control algorithm including a feedforward speed planner and a fuzzy adaptive PID controller;

[0013] and parse the unit start-stop suggestions to control the start-stop operation of the corresponding compressor.

[0014] Specifically, the industrial real-time network in the data perception module adopts a master clock device supporting hardware time stamping and a plurality of slave station devices, and based on a master-slave clock synchronization mechanism of a precision time protocol, high-precision alignment of sampling moments at all sites is realized to ensure that the collected operating parameters are consistent in time and meet the needs of the deep reinforcement learning agent for accurate modeling of the dynamic coupling state of multiple units.

[0015] Specifically, the system state representation at least includes real-time isentropic efficiency and load rate of each compressor calculated from the operating parameters, and a system future load demand prediction value output by a load prediction model based on a long short-term memory network;

[0016] The reward function of the deep reinforcement learning agent is configured to comprehensively evaluate the overall energy efficiency, total energy consumption and load balance of the system, and introduce a penalty term when the system outlet pressure deviates from the set value or the guide vane opening approaches the surge boundary, so as to guide the agent to learn an energy-efficient and safe operation strategy.

[0017] Specifically, the composite servo control algorithm includes:

[0018] a feedforward control channel configured to generate a smooth speed planning instruction according to the target guide vane opening in the load distribution instruction;

[0019] a feedback control channel including a fuzzy adaptive PID controller configured to dynamically adjust PID control parameters by querying a pre-set fuzzy rule base according to the real-time deviation and its rate of change between the target guide vane opening and the actual guide vane opening, and generate a feedback control instruction;

[0020] The outputs of the feedforward control channel and the feedback control channel are combined to drive the guide vane actuator to achieve rapid, accurate and smooth positioning.

[0021] Specifically, the fuzzy adaptive PID controller is configured to map the real-time deviation and deviation rate to preset fuzzy subsets, and output real-time adjustment amounts of proportional, integral and differential parameters according to fuzzy inference rules.

[0022] Specifically, the intelligent decision module is further configured with a safety verification unit for verifying whether the guide vane opening adjustment amount exceeds the single-step maximum allowable variation amplitude and whether the target guide vane opening is located in the anti-surge safety operating area determined based on the current flow from the compressor characteristic curve before the instruction output by the model is executed.

[0023] Specifically, the centrifugal compressor unit load intelligent allocation system further comprises a model updating module configured to continuously collect experience data containing system state, executed action, immediate reward and subsequent state during online operation of the system, and periodically fine-tune parameters of the deep reinforcement learning intelligent agent using the collected experience data.

[0024] Based on the centrifugal compressor unit load intelligent allocation system, the application further provides a centrifugal compressor unit load intelligent allocation method, comprising:

[0025] Synchronous acquisition step: high-precision synchronous acquisition of operating parameters of multiple centrifugal compressors operating in parallel is performed through an industrial real-time network supporting hardware time stamping to obtain time-aligned system state observation data;

[0026] Intelligent decision step: based on the system state observation data, a pre-trained artificial intelligence decision model is used to calculate load allocation instructions and unit start-stop suggestions, wherein the model is a deep reinforcement learning intelligent agent trained based on a proximal policy optimization algorithm and fused with compressor physical property constraints;

[0027] Instruction execution step: a compound servo control algorithm fused with feedforward speed planning and fuzzy adaptive PID feedback is used to execute the load allocation instructions to adjust the guide vane opening of the corresponding compressor; and based on the unit start-stop suggestions, the start-stop operation of the corresponding compressor is controlled.

[0028] Specifically, the intelligent decision step comprises:

[0029] Based on the system state observation data and load prediction results, a state vector containing energy efficiency states of each unit and future demand information is constructed;

[0030] The state vector is input into the deep reinforcement learning intelligent agent to obtain an original action output;

[0031] The original action output is subjected to safety boundary verification, and a final control instruction is generated based on the verified load allocation instruction.

[0032] Specifically, the centrifugal compressor unit load intelligent distribution method further comprises an online learning step of monitoring system states and performance indicators after instruction execution to form experience data, and periodically updating parameters of the deep reinforcement learning intelligent agent using the accumulated experience data

[0033] Compared with the prior art, the application has the following beneficial effects: through the support of the hardware time-stamped industrial network, the microsecond-level synchronous acquisition of the operating parameters of multiple compressors is realized, all data are strictly aligned at the same time, and control deviation caused by asynchronous sampling is avoided.

[0034] The deep reinforcement learning model fusing the physical characteristics of the compressor is used for decision-making, which can comprehensively consider energy efficiency, energy consumption and load balance, meet the process pressure requirements while reducing the total energy consumption of the system.

[0035] The safety check is added before the instruction is issued, the guide vane adjustment range is ensured to be not more than 5°, and the opening is always away from the surge boundary, so that the dangerous working conditions such as surge and overload are effectively prevented.

[0036] The compound servo control algorithm using the feedforward speed planning and the fuzzy adaptive PID makes the guide vane action smooth, rapid in response, high in positioning accuracy, reduces pressure fluctuation and mechanical impact.

[0037] The online learning capability is provided, the model parameters can be continuously fine-tuned using actual operation data, the equipment aging or working condition changes are automatically adapted, and long-term efficient and stable operation is maintained. BRIEF DESCRIPTION OF DRAWINGS

[0038] Fig. 1 It is a centrifugal compressor unit load intelligent distribution system architecture diagram of the application;

[0039] Fig. 2 It is a centrifugal compressor unit load intelligent distribution method flow chart of the application. DETAILED DESCRIPTION

[0040] The technical solutions in the embodiments of the application will be clearly and completely described below with reference to the drawings in the embodiments of the application. Obviously, the described embodiments are only part of the embodiments of the application, not all the embodiments. Based on the embodiments in the application, all other embodiments obtained by those skilled in the art without creative labor are within the protection scope of the application.

[0041] As Figs. 1-2 shown, the technical solutions adopted by the application are as follows: a centrifugal compressor unit load intelligent distribution system, comprising:

[0042] The data perception module is configured to synchronously collect high-precision operation parameters of a plurality of centrifugal compressors operating in parallel through an industrial real-time network supporting hardware timestamps, so as to obtain time-aligned system state observation data.

[0043] The data perception module is a functional unit for obtaining original operation information of a controlled object, and the output thereof serves as state input of a deep reinforcement learning intelligent agent.

[0044] The industrial real-time network supporting hardware timestamps refers to a deterministic communication network that adopts IEEE 802.1AS standard and records the sending and receiving time of a message in a physical layer or a dedicated timestamp unit (TSU), and the timestamp generation thereof is independent of an operating system software interrupt, thereby avoiding non-deterministic delay.

[0045] The plurality of centrifugal compressors operating in parallel refers to a centrifugal compressor unit with a number of (wherein ) centrifugal compressors, the outlet pipelines of which converge into the same process pipeline network, forming a pneumatic coupling relationship, and any change in the operating state of one compressor unit will affect the working point of another compressor unit.

[0046] The operation parameters include the inlet volume flow rate , the inlet absolute pressure , the outlet absolute pressure , the inlet temperature , the shaft rotational speed , the actual opening angle of the guide vane , and the motor current of the first compressor at the sampling time . , .

[0047] The high-precision synchronous collection refers to that the analog-to-digital conversion (ADC) actions of all operation parameters are triggered by a local clock after global synchronization, and the deviation of the sampling start time of each slave station device relative to the master clock satisfies , wherein represents the deviation of the actual sampling time of the slave station where the first compressor is located from the ideal sampling time , is the set sampling period.

[0048] The time-aligned system state observation data refers to a state vector composed of the operation parameters of all compressors in each sampling period .

[0049] ;

[0050] Among them, dimensions Each element in this vector corresponds to the same physical moment. This ensures that there are no timing misalignments in state observation.

[0051] Specifically, the industrial real-time network in the data sensing module adopts a master clock device that supports hardware timestamps and multiple slave devices, and achieves high-precision alignment of sampling times across all network sites based on a master-slave clock synchronization mechanism with a precise time protocol. This ensures that the collected operating parameters are consistent in time, meeting the requirements of the deep reinforcement learning agent to accurately model the dynamic coupling state of multiple units.

[0052] The master clock device is an industrial switch that is elected as the Grandmaster Clock in the network topology. It has an IEEE 802.1AS compliant timestamp hardware unit and uses a temperature-compensated crystal oscillator (TCXO) with a frequency stability better than ±50ppb as its local clock source.

[0053] Multiple slave devices include distributed I / O modules, embedded controllers, and smart sensor nodes. Each slave device integrates an Ethernet controller (such as the TISitara AM64x series) that supports the gPTP protocol stack and implements hardware timestamp capture of Sync messages at the PHY layer.

[0054] Precision time protocol specifically refers to the Generalized Precision Time Protocol (gPTP) defined by the IEEE 802.1AS-2011 standard. It completes the master-slave clock offset by periodically transmitting four types of messages: Sync, Follow_Up, Pdelay_Req, and Pdelay_Resp. With link propagation delay The joint estimate.

[0055] The formula for calculating the synchronization error of the master-slave clock synchronization mechanism is as follows:

[0056] ;in The hardware timestamp for sending Sync messages from the master clock. For the station The hardware timestamp of the received message. For the station The hardware timestamp used to send the Pdelay_Req message. The master clock returns the hardware timestamp of the Pdelay_Req message. For the station The one-way link delay to the master clock can be reduced through the above mechanism. Controlled Within.

[0057] High-precision alignment of sampling times across all sites refers to the process where, after clock synchronization is completed, each slave device generates a period based on its local synchronization clock. The sampling enable signal drives the ADC in each The conversion is initiated simultaneously at all times, thereby ensuring that the acquisition of all operating parameters is strictly aligned.

[0058] The collected operating parameters are consistent over time for any two compressors. and The state components satisfy This makes the state vector It can truly reflect the system's performance. The coupling dynamics at any given moment.

[0059] The requirement for the deep reinforcement learning agent to accurately model the dynamic coupling state of multiple units is met because the deep reinforcement learning agent needs to be modeled based on a Markov decision process (MDP), and its state transition probabilities... Depends on the current state The integrity and authenticity. If asynchronous sampling leads to... If data from different physical times is included, the state transition function will introduce spurious correlations, compromising the unbiasedness of policy gradient estimation and leading to convergence failure of the control policy. Therefore, microsecond-level time alignment is a necessary technical prerequisite for ensuring the effectiveness of MDP modeling.

[0060] The intelligent decision-making module is communicatively connected to the data perception module and configured to construct a system state representation based on the system state observation data.

[0061] The system state representation is input into a pre-trained artificial intelligence decision-making model, which is a deep reinforcement learning agent trained based on a proximal policy optimization algorithm and incorporating compressor physical characteristic constraints to ensure that the decision-making process meets dynamic constraints.

[0062] The system acquires and outputs load allocation instructions and unit start-up / shutdown suggestions generated by the artificial intelligence decision-making model.

[0063] The intelligent decision-making module is a dedicated functional unit for generating compressor group control strategies. It is implemented as an embedded software module deployed on an industrial edge computing device. Communication with the data sensing module indicates that this module receives data from the data sensing module via an industrial real-time network that supports hardware timestamps. System state observation data output in each sampling period System state observation data Defined at the sampling time ,in , The original running parameter vector collected synchronously has a dimension of ,Include The inlet volume flow rate of each of the two parallel centrifugal compressors Absolute pressure on imports Absolute pressure on exports Inlet temperature Shaft speed Actual opening angle of the guide vane and motor current ,in , .

[0064] Constructing a system state representation refers to... Converted into feature vectors that can be processed by deep reinforcement learning agents This transformation process involves the fusion of physical derivative calculations and external prediction information to ensure... It can fully characterize the current energy efficiency status and future load trends of the system.

[0065] System state characterization The feature vector defined in the previous sentence serves as the sole input to the AI ​​decision-making model. A pre-trained AI decision-making model refers to a policy model that has undergone offline training and has fixed weights before deployment; its implementation is a neural network inference engine. Training based on the proximal policy optimization algorithm shows that the model uses the Proximal Policy Optimization (PPO) algorithm for policy learning, with the objective function being:

[0066] ;

[0067] Among them, expectations This indicates the sampling step in the experience playback buffer. On average, For pruning hyperparameters; Defined as the policy probability ratio.

[0068] ;in For the current policy network, For old policy networks, and These are the corresponding neural network parameters; For the first The generalized advantage estimate of the step is derived from the value network. The calculation yielded the result.

[0069] Integrating compressor physical characteristic constraints refers to embedding engineering constraints such as surge boundary, guide vane change rate limit, and power limit as soft constraints into the reward function or as hard constraints to limit the action space during the training phase. .

[0070] Ensuring that the decision-making process meets dynamic constraints means that the generated control commands must conform to the time response characteristics of the compressor pneumatic system, so as to avoid surge, overload or mechanical shock caused by sudden changes in commands.

[0071] Deep reinforcement learning agents consist of policy networks and value network Both are composed of multilayer perceptrons (MLPs), with the input being... The outputs are the action distribution and the state value estimate, respectively.

[0072] The load distribution instructions generated by the model refer to the continuous control signals output by the policy network: ,in Indicates the first The target load rate of the compressor in the next control cycle. The unit start-up / shutdown recommendation refers to the binary decision signal output by the strategy network. ,in It is suggested to start the first Taiwan compressor, This indicates a recommendation to keep the system offline or shut down.

[0073] The intelligent decision-making module extracts complete action vectors from the output layer of the policy network and outputs the representation. It is then transmitted to the lower-level execution controller via an industrial real-time network, with the transmission occurring during each sampling cycle. The moment the decision is completed.

[0074] Specifically, the system state characterization includes at least the real-time isentropic efficiency and load rate of each compressor calculated from the operating parameters, and the predicted future load demand of the system output by the load prediction model based on long short-term memory network.

[0075] System state characterization In this sentence, it specifically refers to its content composition, and its dimensions are: ,in .

[0076] The real-time isentropic efficiency of each compressor is indicated by the first... Taiwan compressor isentropic efficiency at time step The calculation formula is as follows:

[0077] ;

[0078] in; The specific heat capacity of air at constant pressure. The air insulation index. Shaft power is determined by motor current. Rated voltage Power factor and motor efficiency Conversion: .

[0079] Load factor is defined as: ;in For the first The rated inlet volumetric flow rate of the compressor, in units of It is determined by the parameters on the equipment nameplate.

[0080] The load prediction model based on the Long Short-Term Memory network is an independently trained time-series prediction model. Its implementation is a single-layer LSTM neural network, with the input being past data. The measured flow sequence of the main pipe for each sampling period Output for the future The total system load demand forecast for each cycle, i.e., the entire forecast sequence output by the LSTM model. The unit is As After Dimensional components.

[0081] The reward function of the deep reinforcement learning agent is configured to comprehensively evaluate the overall energy efficiency, total energy consumption and load balance of the system, and to introduce a penalty term when the system outlet pressure deviates from the set value or the guide vane opening approaches the surge boundary, so as to guide the agent to learn the energy-efficient and safe operating strategy.

[0082] reward function Defined as:

[0083] ;

[0084] Among them, the weighting coefficient , , , , These are pre-set positive real numbers used to balance the importance of various indicators.

[0085] Overall system energy efficiency Defined as:

[0086] Unit is This reflects the system's work capacity per unit of energy consumption.

[0087] Total energy consumption Defined as: The unit is joule (J).

[0088] Load balance Defined as:

[0089] The smaller the value, the more evenly the load distribution among the units.

[0090] The penalty for the system outlet pressure deviating from the set value is as follows:

[0091] ;in The measured pressure in the main pipe (unit: MPa) is obtained by a pressure sensor. Set values ​​for the process. This is the allowable deviation threshold.

[0092] The penalty term for guide vane opening approaching the surge boundary is:

[0093] ;in For the first The compressor is in flow The surge boundary guide vane opening (unit: degrees) is provided by the manufacturer. The characteristic curve was obtained by cubic spline interpolation. This is to pre-set a safety margin.

[0094] Guiding an agent to learn an energy-efficient and safe operating strategy means using the aforementioned multi-objective reward function to simultaneously optimize economic indicators (energy efficiency, energy consumption, and equilibrium) and safety indicators (pressure stability and anti-surge) during strategy gradient updates, ensuring that the learned strategy is both energy-efficient and reliable in actual operation.

[0095] Specifically, the intelligent decision-making module is also equipped with a safety verification unit, which is used to verify whether the guide vane opening adjustment exceeds the maximum allowable change range in a single step before the instruction output by the model is executed, and to verify whether the target guide vane opening is within the anti-surge safe operating range determined from the compressor characteristic curve based on the current flow rate.

[0096] The security verification unit is an independent logic verification submodule within the intelligent decision-making module. It is implemented as a conditional judgment function in embedded C code and runs in the last processing stage before the instruction is issued.

[0097] The instructions output by the model refer to the complete action vector. Load distribution section After mapping, the target guide vane opening of each compressor is obtained. (Calculated by looking up a table or using an inverse function).

[0098] The guide vane opening adjustment amount is defined as: The unit is degrees; the maximum allowable change range per step is a preset constant. This value is determined based on experiments on the maximum angular acceleration and mechanical strength of the guide vane drive mechanism. Then The limit is: ;in It is a symbolic function.

[0099] The anti-surge safe operating range determined from the compressor characteristic curve based on the current flow rate refers to the range of the compressor's current flow rate. A compressor, with known current inlet flow rate Under the premise of this, the minimum safe guide vane opening is found from its compressor characteristic curve:

[0100] The unit is degrees.

[0101] Verifying whether the target guide vane opening is within the anti-surge safe operating range determined from the compressor characteristic curve based on the current flow rate refers to determining whether the Boolean condition is met: If the condition is not met, then a forced correction will be made: This security verification mechanism constitutes the last physical security barrier, ensuring that even if the deep reinforcement learning agent outputs dangerous commands due to generalization errors or abnormal inputs, the actual control actions remain strictly within the safe operating range of the equipment.

[0102] The instruction execution module is communicatively connected to the intelligent decision-making module and is configured to: parse the load allocation instruction and drive the corresponding guide vane actuator through a composite servo control algorithm that includes a feedforward speed planner and a fuzzy adaptive PID controller.

[0103] It also analyzes the unit start-up and shutdown suggestions and controls the start-up and shutdown operations of the corresponding compressors.

[0104] The instruction execution module is a functional unit used to translate high-level decisions into low-level physical actions. It is implemented as firmware deployed in a programmable logic controller (PLC) or embedded motion control card. Communication with the intelligent decision module indicates that this module receives instructions from the intelligent decision module during the sampling period via an industrial real-time network that supports hardware timestamps. Output action vector .

[0105] Load distribution command specifically refers to ,in Indicates the first The target load rate of the compressor.

[0106] Parsing the load allocation instruction means... Mapped to target guide vane opening The mapping relationship is achieved through a lookup table method. The lookup table data comes from the load rate-guide vane opening calibration curve provided by the compressor manufacturer. This curve is stored in local non-volatile memory in steps of 1%.

[0107] The feedforward speed planner is the open-loop control submodule in the composite servo control algorithm. Its function is to... and the current actual opening of the guide vane Generate smooth velocity profiles .

[0108] The fuzzy adaptive PID controller is a closed-loop control submodule whose parameters are adjusted online according to the system's dynamic characteristics. The composite servo control algorithm is a hybrid control structure consisting of a feedforward control channel and a feedback control channel connected in parallel; its total output is the sum of the two. The corresponding guide vane actuator drives the conversion of the synthesized control signal into a 4–20mA analog signal or a CANopen PDO message, which is then sent to the... The compressor has electric and pneumatic guide vane actuators, model Rotork IQ3, with a response bandwidth of not less than 5Hz.

[0109] Unit start-up and shutdown recommendations specifically refer to the binary decision signals output by the intelligent decision module. The analysis of the unit start-up and shutdown recommendations refers to the instruction execution module reading bit by bit. ,like If the compressor is currently stopped, the start-up sequence is triggered. If the compressor is currently running, a shutdown sequence will be triggered.

[0110] Controlling the start and stop of the corresponding compressor refers to sending commands to the first compressor via hard-wired relay output or Modbus TCP write command. The main control PLC of the compressor sends a digital signal to enable or stop the start. The start process follows a preset sequence: first, the lubricating oil pump is turned on (delay 30s), then the motor is started (soft start time 15s); during the stop process, the motor, guide vanes (closed to 0°), and cooling water valve are turned off in sequence. The entire process is constrained by safety interlock logic to ensure no surge or overpressure risk.

[0111] Specifically, the composite servo control algorithm includes:

[0112] The feedforward control channel is configured to generate a smooth speed planning command based on the target guide vane opening in the load distribution command.

[0113] The feedforward control channel is the open-loop path of the composite servo control algorithm, implemented as an S-shaped speed planner. The target guide vane opening is... The unit is degrees (°). Generating a smooth velocity planning command means calculating the guide vane's speed from its current position. To the target location The optimal trajectory, its angular velocity command Defined as:

[0114] ;

[0115] in: For maximum angular acceleration, For the maximum angular velocity, The total travel time is determined by the displacement. The decision was made to ensure that the guide vane movement is shock-free and overshoot-free, avoiding mechanical wear and flow field disturbance.

[0116] The feedback control channel includes a fuzzy adaptive PID controller, configured to dynamically adjust the PID control parameters and generate feedback control commands by querying a preset fuzzy rule base based on the real-time deviation and rate of change between the target guide vane opening and the actual guide vane opening.

[0117] The feedback control channel is the closed-loop path of the composite servo control algorithm. The target guide vane opening remains at... The actual opening of the guide vane is recorded as... The value is obtained in real time by an absolute encoder (resolution 0.1°) mounted on the guide vane shaft.

[0118] Real-time deviation is defined as: The unit is degrees.

[0119] The rate of change of deviation is defined as: ; Calculated using the first-order backward difference approximation:

[0120] ;in .

[0121] The pre-built fuzzy rule base is an expert knowledge base containing 49 If-Then rules, with the input variable being... and The output is the proportional gain adjustment amount. Integral gain adjustment Differential gain adjustment .

[0122] Dynamically adjusting PID control parameters refers to adjusting the basic parameters. , , Add the adjustment amount to obtain the real-time parameters:

[0123] All parameters are limited to positive real numbers.

[0124] Generating feedback control instructions refers to calculating the standard PID output:

[0125] The instruction is in voltage (V) and, after D / A conversion, drives the actuator servo amplifier.

[0126] The outputs of the feedforward control channel and the feedback control channel are combined to drive the guide vane actuator to achieve fast, accurate and stable positioning.

[0127] The output of the feedforward control channel, i.e., the speed planning command. After integration, it is converted into a position feedforward term. The output of the feedback control channel, i.e. "Being synthesized" refers to the algebraic addition of the two to form the overall control command.

[0128] The instruction is output after being limited (range 0–10 V).

[0129] Driving the guide vane actuator means... The setpoint of the servo amplifier controls the torque of the DC servo motor, thereby adjusting the guide vane angle.

[0130] Fast, accurate, and stable positioning refers to the positioning time under a typical step response. steady-state error overshoot This performance index is achieved through a combination of feedforward compensation to eliminate system inertial delay and feedback to suppress external disturbances.

[0131] Specifically, the fuzzy adaptive PID controller is configured to map the real-time deviation and the rate of change of deviation to a preset fuzzy subset, and output the real-time adjustment of the proportional, integral and derivative parameters according to the fuzzy inference rules.

[0132] Real-time deviation and rate of change of deviation These are the two input variables for the fuzzy controller. The preset fuzzy subset pair... Define 7 linguistic variables: NB (negative large), NM (negative medium), NS (negative small), ZO (zero), PS (positive small), PM (positive medium), and PB (positive large), with the universe of discourse being... The trigonometric membership function is used.

[0133] right Similarly, we define 7 subsets, with the domain being... .

[0134] Mapping to a predefined fuzzy subset refers to calculating the membership degree of each input to each linguistic variable using a membership function, for example:

[0135] .

[0136] Other subsets are defined similarly; the fuzzy inference rules adopt Mamdani-type max-min inference, with the rule form being: "If..." for and for ,but for , for , for ".

[0137] The real-time adjustment of the proportional, integral, and derivative parameters is achieved by aggregating the outputs of each activation rule (taking the maximum membership degree), and then using the centroid method for defuzzification to obtain precise values. The values ​​range from [−1.0, 1.0], [−0.3, 0.3], to [−0.2, 0.2], ensuring that the parameter adjustment is both sensitive and stable.

[0138] Specifically, a centrifugal compressor unit load intelligent allocation system also includes a model update module, configured to continuously collect experience data including system status, executed actions, immediate rewards and subsequent status during system online operation, and periodically use the collected experience data to fine-tune the parameters of the deep reinforcement learning agent.

[0139] The model update module is a functional unit used to implement online adaptive optimization of deep reinforcement learning agents. It is implemented as a background daemon deployed on an industrial edge server, running in an independent thread without interfering with the main control loop.

[0140] During system online operation, the period refers to the entire continuous operating time from the system's initial commissioning to the current moment, during which the sampling period is... Continuously increasing, among which , .

[0141] Continuous collection means in each sampling period At the end, the quadruple empirical data is automatically stored in a circular buffer. The buffer capacity is Each record is managed using a first-in, first-out (FIFO) strategy.

[0142] System status refers to the first The period is a system state representation output by the data sensing module and processed by the intelligent decision-making module. ,in As defined above.

[0143] The execution action refers to the intelligent decision-making module in Output the complete motion vector after correction by the security verification unit. .

[0144] Instant rewards refer to The scalar value calculated by the reward function at each time step Its mathematical expression has been clearly defined above.

[0145] Subsequent state refers to the state after an action is performed. Then, in the next sampling period Observed system state characterization Therefore, each empirical data point is a quadruple. Its data structure is a fixed-length array of floating-point and Boolean data, which is passed from the main control thread to the model update module through shared memory or IPC mechanism.

[0146] Regularly utilizing the collected empirical data refers to the use of circular buffer zones. The number of valid samples in the data reaches the threshold. When the sample size is insufficient, a fine-tuning process is triggered. This process is executed at least once every 24 hours and is forcibly started even if the sample size is insufficient to prevent long-term drift.

[0147] Fine-tuning the parameters of the deep reinforcement learning agent refers to using an online fine-tuning variant of the PPO algorithm to adjust the policy network parameters. and value network parameters The objective function for performing gradient updates is:

[0148] ;

[0149] in, To optimize the loss of the aforementioned near-end strategy, The mean squared error loss of the value function. for Return estimate (take) Discount factor ), This is the policy entropy term, used to encourage exploration, with a coefficient of... , For fixed hyperparameters.

[0150] The fine-tuning process is performed on a dedicated GPU accelerator, with each iteration using a randomly sampled batch size B=256, and a maximum number of iterations. Wheel, updated new parameters , After integrity verification, old parameters are replaced atomically to ensure uninterrupted inference. This mechanism enables deep reinforcement learning agents to adapt to long-term dynamics such as compressor performance degradation, changes in environmental conditions, or drift in pipeline characteristics, maintaining the optimality and robustness of the control strategy.

[0151] A method for intelligent load distribution of a centrifugal compressor unit includes:

[0152] Synchronous acquisition steps: Through an industrial real-time network that supports hardware timestamps, the operating parameters of multiple centrifugal compressors operating in parallel are acquired with high precision to obtain time-aligned system status observation data.

[0153] The synchronous acquisition step is the first stage, used to acquire raw sensor information. Industrial real-time networks supporting hardware timestamps refer to deterministic communication networks that conform to the IEEE 802.1AS standard and have physical layer timestamp capabilities. Multiple centrifugal compressors operating in parallel refer to... The outlet manifold compressor unit. Operating parameters include the inlet flow rate of each compressor. Import and export pressure , ,temperature Rotation speed Guide vane opening and motor current ,in , , .

[0154] High-precision synchronous acquisition requires that the sampling time deviation of each slave station be... Time-aligned system state observation data, i.e., state vectors. All its components correspond to the same physical moment. .

[0155] Intelligent decision-making steps: Based on the system status observation data, load allocation instructions and unit start-up and shutdown suggestions are calculated through a pre-trained artificial intelligence decision-making model, wherein the model is a deep reinforcement learning agent trained based on a near-end strategy optimization algorithm and incorporating compressor physical characteristic constraints.

[0156] The intelligent decision-making step is the second stage, in each Execute, input is The pre-trained AI decision-making model is a neural network inference engine with fixed weights. Load allocation instructions are continuous vectors. This represents the target load rate. The unit start-up and shutdown recommendations are binary vectors. .

[0157] The model is trained based on the proximal policy optimization algorithm and then optimized using the PPO algorithm. The objective function is: ,in .

[0158] Integrating compressor physical characteristic constraints refers to embedding engineering constraints such as surge boundaries and guide vane speed limits during training. Deep reinforcement learning agents consist of policy networks. With value network It is incorporated into AI acceleration chips.

[0159] Instruction execution steps: By using a composite servo control algorithm that integrates feedforward speed planning and fuzzy adaptive PID feedback, the load distribution instruction is executed to adjust the guide vane opening of the corresponding compressor; and based on the unit start-up and shutdown suggestions, the start-up and shutdown operations of the corresponding compressor are controlled.

[0160] The instruction execution step is the third stage, where the load allocation instruction is mapped to the target guide vane opening. .

[0161] The composite servo control algorithm includes a feedforward channel (S-shaped speed planning). , ) and feedback channel (fuzzy adaptive PID, input is deviation) (and its rate of change).

[0162] General control command Drive the actuator to achieve positioning error Response time .

[0163] Recommended start-up and shutdown sequence for the unit: When starting, first start the oil pump (30s delay) and then soft start the motor (15s). When shutting down, turn off the motor, guide vanes and cooling water in sequence. The entire process is protected by safety interlocks.

[0164] Specifically, the intelligent decision-making steps include:

[0165] Based on the system status observation data and load forecast results, a state vector containing the energy efficiency status and future demand information of each unit is constructed.

[0166] The state vector is input into the deep reinforcement learning agent to obtain the original action output.

[0167] The original action output is subjected to safety boundary verification, and the final control command is generated based on the verified load distribution command.

[0168] This sentence defines the internal sub-processes of the intelligent decision-making step. System state observation data is... Load forecasting results are based on the future output of a Long Short-Term Memory (LSTM) network. Step-by-step load forecast sequence .

[0169] The energy efficiency status of each unit includes real-time isentropic efficiency. With load factor The state vector is denoted as... , before Weiwei ,back The value represents the load forecast.

[0170] The state vector is input into the deep reinforcement learning agent to obtain the original action output, which is the state vector. Deep reinforcement learning agents, i.e., the aforementioned policy networks. The original action output is an action vector that has not undergone security checks. ;

[0171] in , It is generated directly from the policy network.

[0172] Security boundary verification is performed by a security verification unit, which is implemented as an embedded conditional judgment function; the verification content includes:

[0173] (1) Single-step adjustment amount of guide vane If the limit is exceeded, the amplitude will be limited;

[0174] (2) The target guide vane opening meets the requirements. Otherwise, correct to the safety boundary.

[0175] The verified load distribution instruction is denoted as The final control command is the complete motion vector after verification and correction. , used for subsequent execution.

[0176] Specifically, a method for intelligent load allocation of centrifugal compressor units also includes:

[0177] Online learning steps: Monitor the system status and performance indicators after the command is executed to form experience data; and periodically use the accumulated experience data to fine-tune and update the parameters of the deep reinforcement learning agent.

[0178] The online learning process is the fourth stage, running continuously in the background. The system status after the monitoring command is executed refers to... Obtaining the new state representation Performance metrics include immediate rewards. Its definition is: ;

[0179] Among them, weight .

[0180] Empirical data is quadruples Storage capacity A circular buffer. Periodic use refers to the use of a buffer when the number of samples reaches a certain threshold. Or it can be forcibly triggered once every 24 hours.

[0181] Parameter fine-tuning updates use the PPO fine-tuning objective function:

[0182] ;

[0183] in, Using a batch size of 256 and a maximum of 3 iterations, the network parameters are atomically replaced after the update to ensure that the strategy continuously adapts to the dynamic changes of the system.

[0184] In one specific embodiment, the present invention is applied to an air compressor station in a chemical plant. This station has three centrifugal compressors operating in parallel, providing 0.8 MPa compressed air to the entire plant. Traditional control methods are inefficient and pose safety risks. After applying this system, safe, efficient, and automated load distribution is achieved.

[0185] The system consists of four modules: a data perception module, an intelligent decision-making module, and an instruction execution module, and also includes a model update module for continuous optimization.

[0186] The data sensing module collects data in real time through a high-precision synchronized industrial network. This network uses one master clock switch and three slave controllers, employing a precise time protocol to achieve clock synchronization with microsecond-level accuracy. Every millisecond... Three slave devices simultaneously collect operating parameters of their respective compressors, including 21 parameters such as flow rate, pressure, temperature, speed, actual guide vane opening angle, and motor current. These parameters are collected at the sampling time. Strictly aligned data constitutes a system state observation data vector reflecting the instantaneous state of the entire system. .

[0187] The intelligent decision-making module is deployed on an industrial edge computing device in the field. It first processes... Constructing system state representation This includes calculating the real-time isentropic efficiency and load rate of each machine, and fusing the total system load demand forecasts for the next 10 sampling periods from the output of an independent long short-term memory network load forecast model.

[0188] Then, The input is a pre-trained artificial intelligence decision-making model (a deep reinforcement learning agent trained based on a proximal policy optimization algorithm). The model outputs raw action outputs, including target load rates for each compressor and unit start-up and shutdown recommendations.

[0189] Before the command is issued, the safety verification unit within the intelligent decision-making module immediately intervenes. It performs a double verification: first, it ensures that the calculated guide vane opening adjustment does not exceed the maximum permissible change range per step (set as...). Secondly, based on the current flow rate, query the compressor characteristic curve to ensure that the target guide vane opening is within the anti-surge safe operating range (at least [distance from the surge boundary]). Any unsafe instructions will be automatically corrected. Instructions that pass verification will be issued as final load allocation instructions and unit start-up / shutdown recommendations.

[0190] The agent learns through a comprehensive reward function that simultaneously optimizes the overall system energy efficiency, total energy consumption, and load balance, while penalizing pressure deviations and behaviors near surge limits, thereby guiding it to learn efficient and safe operating strategies.

[0191] The instruction execution module is deployed within the programmable logic controller (PLC). Upon receiving a load allocation instruction, it executes it precisely using a composite servo control algorithm. This algorithm combines a feedforward control channel (for planning smooth movements) and a feedback control channel (using a fuzzy adaptive PID controller for real-time correction), synthesizing a control signal to drive the guide vane actuator, achieving fast, accurate, and stable positioning.

[0192] For unit start-up and shutdown recommendations, this module executes the compressor start-up and shutdown operations by controlling relays according to the preset safety logic sequence (such as starting the lubricating oil pump first, and then soft-starting the motor).

[0193] During system operation, the model update module acts as a background process, continuously collecting experience data (including states, actions, rewards, and subsequent states). This data is stored in a circular buffer. The system periodically (e.g., every 24 hours) uses this data to fine-tune the parameters of the deep reinforcement learning agent, enabling it to adapt to changes in equipment performance and operating conditions, and continuously maintain optimal performance.

[0194] Although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art can still modify the technical solutions described in the foregoing embodiments or make equivalent substitutions for some of the technical features. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the protection scope of the present invention.

Claims

1. A centrifugal compressor unit load intelligent distribution system, characterized in that, include: The data sensing module is configured to collect operating parameters of multiple centrifugal compressors operating in parallel with high precision through an industrial real-time network that supports hardware timestamps, so as to obtain time-aligned system status observation data. The intelligent decision-making module is communicatively connected to the data sensing module and configured to: construct a system state representation based on the system state observation data; The system state representation is input into a pre-trained artificial intelligence decision-making model, which is a deep reinforcement learning agent trained based on a proximal policy optimization algorithm and incorporates compressor physical characteristic constraints to ensure that the decision-making process meets dynamic constraints. Obtain and output the load allocation instructions and unit start-up / shutdown suggestions generated by the model; The instruction execution module is communicatively connected to the intelligent decision-making module and is configured to: parse the load allocation instruction and drive the corresponding guide vane actuator through a composite servo control algorithm that includes a feedforward speed planner and a fuzzy adaptive PID controller; It also analyzes the unit start-up and shutdown suggestions and controls the start-up and shutdown operations of the corresponding compressors.

2. The intelligent load distribution system for a centrifugal compressor unit according to claim 1, characterized in that, The industrial real-time network in the data sensing module adopts a master clock device that supports hardware timestamps and multiple slave devices. Based on the master-slave clock synchronization mechanism of the precise time protocol, it achieves high-precision alignment of the sampling time of all network sites to ensure that the collected operating parameters are consistent in time, thus meeting the needs of the deep reinforcement learning agent to accurately model the dynamic coupling state of multiple units.

3. The intelligent load distribution system for a centrifugal compressor unit according to claim 1, characterized in that, The system state characterization includes at least the real-time isentropic efficiency and load rate of each compressor calculated from the operating parameters, and the predicted future load demand of the system output by the load prediction model based on long short-term memory network. The reward function of the deep reinforcement learning agent is configured to comprehensively evaluate the overall energy efficiency, total energy consumption and load balance of the system, and to introduce a penalty term when the system outlet pressure deviates from the set value or the guide vane opening approaches the surge boundary, so as to guide the agent to learn the energy-efficient and safe operating strategy.

4. The intelligent load distribution system for a centrifugal compressor unit according to claim 1, characterized in that, The composite servo control algorithm includes: The feedforward control channel is configured to generate a smooth speed planning command based on the target guide vane opening in the load distribution command; The feedback control channel includes a fuzzy adaptive PID controller, configured to dynamically adjust the PID control parameters and generate feedback control commands by querying a preset fuzzy rule base based on the real-time deviation and rate of change between the target guide vane opening and the actual guide vane opening. The outputs of the feedforward control channel and the feedback control channel are combined to drive the guide vane actuator to achieve fast, accurate and stable positioning.

5. The intelligent load distribution system for a centrifugal compressor unit according to claim 4, characterized in that, The fuzzy adaptive PID controller is configured to map the real-time deviation and the rate of change of deviation to a preset fuzzy subset, and output the real-time adjustment of the proportional, integral and derivative parameters according to the fuzzy inference rules.

6. The intelligent load distribution system for a centrifugal compressor unit according to claim 1, characterized in that, The intelligent decision-making module is also equipped with a safety verification unit, which is used to verify whether the guide vane opening adjustment exceeds the maximum allowable change range in a single step before the instructions output by the model are executed, and to verify whether the target guide vane opening is within the anti-surge safe operating range determined from the compressor characteristic curve based on the current flow rate.

7. The intelligent load distribution system for a centrifugal compressor unit according to claim 1, characterized in that, It also includes a model update module, configured to continuously collect experience data including system status, executed actions, immediate rewards and subsequent status during system online operation, and periodically use the collected experience data to fine-tune the parameters of the deep reinforcement learning agent.

8. A method for intelligent load distribution in a centrifugal compressor unit, characterized in that, include: Synchronous acquisition steps: Through an industrial real-time network that supports hardware timestamps, the operating parameters of multiple centrifugal compressors operating in parallel are acquired with high precision to obtain time-aligned system status observation data. Intelligent decision-making steps: Based on the system status observation data, load allocation instructions and unit start-up and shutdown suggestions are calculated through a pre-trained artificial intelligence decision-making model, wherein the model is a deep reinforcement learning agent trained based on a near-end strategy optimization algorithm and incorporating compressor physical characteristic constraints; Instruction execution steps: By using a composite servo control algorithm that integrates feedforward speed planning and fuzzy adaptive PID feedback, the load distribution instruction is executed to adjust the guide vane opening of the corresponding compressor; and based on the unit start-up and shutdown suggestions, the start-up and shutdown operations of the corresponding compressor are controlled.

9. The intelligent load allocation method for a centrifugal compressor unit according to claim 8, characterized in that, The intelligent decision-making steps include: Based on the system status observation data and load forecast results, a state vector containing the energy efficiency status and future demand information of each unit is constructed. The state vector is input into the deep reinforcement learning agent to obtain the original action output; The original action output is subjected to safety boundary verification, and the final control command is generated based on the verified load distribution command.

10. The intelligent load allocation method for a centrifugal compressor unit according to claim 8, characterized in that, Also includes: Online learning steps: Monitor the system status and performance indicators after command execution to generate empirical data; The parameters of the deep reinforcement learning agent are periodically fine-tuned and updated using accumulated experience data.

Images