Axial flow pump operation condition intelligent optimization control method and system

Abstract

This invention belongs to the field of axial flow pump control technology, specifically relating to an intelligent optimization control method and system for axial flow pump operation. The method includes: constructing and fusing a Gaussian process regression surrogate model corresponding to the physical mechanism and operating data; the model synchronously outputting performance prediction values ​​and prediction uncertainties; calculating the state entropy corresponding to the current operating condition based on the surrogate model; performing heuristic pruning by using the surrogate model to estimate the change in state entropy; the system selecting a subset of candidate rules and calling the surrogate model to infer and calculate the comprehensive gain value; the inference depth decreasing as the prediction uncertainty increases; establishing a penalty mechanism that applies penalties to the gain value based on historical failure probabilities and the current state entropy value; and the system determining the optimal control rule and issuing commands based on the adjusted gain value. This invention can reduce the underlying computing power consumption, prevent long-term error accumulation, prevent repeated equipment oscillations, and ensure the long-term efficient operation of the axial flow pump.

Description

Technical Field

[0001] This invention relates to the field of axial flow pump control technology. More specifically, this invention relates to an intelligent optimization control method and system for the operating conditions of an axial flow pump. Background Technology

[0002] As core flow-passing devices, axial flow pumps and mixed flow pumps require intelligent optimization of their operating conditions to ensure the stability and economy of systems such as circulating cooling and working fluid transportation. In actual industrial processes, the internal flow field of the pump body exhibits strong nonlinearity, unsteadiness, and multi-parameter coupling characteristics. Its hydraulic conditions are easily affected by fluctuations in pipeline resistance, changes in external water level, and load disturbances, causing the unit to deviate from its optimal efficiency zone. This can lead to a surge in hydraulic losses, increased energy consumption, and may even induce dangerous conditions such as stall and surge.

[0003] Traditional pump operating condition regulation methods mostly rely on static performance curves or simple feedback control logic, which are difficult to accurately characterize transient dynamic characteristics including time variables, rotor inertia and fluid acceleration. When faced with complex and rapid changing operating conditions, such methods often lag in identifying the pump's actual operating state and stall precursors, resulting in insufficient real-time optimization and control capabilities, and failing to effectively solve the problem of balancing control performance and operational stability.

[0004] To address this, the industry has proposed several improved control strategies, such as enhancing system responsiveness by predicting unit load and storing water in advance. Referring to Chinese patent application CN121803483A, a method, system, terminal, and storage medium for controlling the blade angle of a mixed-flow pump are disclosed. This application obtains and analyzes preset real-time and historical load data sets of the unit to determine a confidence-based fusion load. Based on the comparison between this confidence-based fusion load and preset first and second response thresholds, it controls whether the water storage tank supplies water to the condenser and the blade angles of the water storage mixed-flow pump and the water intake mixed-flow pump.

[0005] However, while this application can address the issue of poor load response of the mixed-flow pump due to the lag in incremental water flow to some extent, and balances the unit load through the approach of pre-storage and delayed replenishment, its model construction process does not fully integrate hydraulic loss mechanisms, system dynamic equations, and data-driven surrogate models. Furthermore, it does not explicitly output the uncertainty of the prediction results, thus its generalization ability and prediction reliability across the entire operating range still need improvement. Secondly, the method's assessment of unit status and decision-making regarding replenishment flow primarily revolves around single load data, lacking a comprehensive state evaluation mechanism that incorporates multi-dimensional information such as efficiency margin and stall precursors. More importantly, its reasoning and control decisions employ a deterministic conditional comparison framework, lacking heuristic search and sequential reasoning based on model predictions. When facing new fluctuating operating conditions not fully reflected in historical data, the calculation process lacks sufficient foresight, and the fixed control logic is prone to repeatedly triggering the same adjustment rule, posing a risk of control oscillations. Summary of the Invention

[0006] To address the problems of insufficient reliability in predicting the full-condition performance of axial flow pumps in existing technologies, the single evaluation index of operating status, redundant inference search calculations that are prone to getting trapped in local suboptimal conditions, and the tendency to repeatedly trigger invalid control rules and generate control oscillations under complex fluctuating conditions, this invention provides solutions in the following aspects.

[0007] In a first aspect, the present invention provides an intelligent optimization control method for the operating conditions of an axial flow pump, comprising: constructing a Gaussian process regression surrogate model, which is established by integrating flow field numerical simulation with historical operating data and combining system equations including hydraulic loss mechanisms, time variables, rotor inertia, and fluid acceleration, for simultaneously outputting performance prediction values ​​and prediction uncertainties; calculating the state entropy of the current operating condition based on the surrogate model, wherein the state entropy is weighted by normalizing the nonlinear spatial distance between the operating point and the optimal efficiency curve, the gradient of the efficiency contour lines, and the amplitude of the stall precursor pressure pulsation characteristics; and utilizing the surrogate model during the forward chain inference process of the inference engine. The instantaneous state entropy change after the execution of the predictive rule is estimated and heuristically pruned to select a subset of candidate rules that are expected to reduce state entropy. The proxy model is invoked to perform sequential reasoning on each rule in the subset of candidate rules to calculate the initial comprehensive gain value, wherein the reasoning depth decreases as the prediction uncertainty increases. A penalty mechanism is established: if the execution of a rule within a preset historical time window has led to an increase in state entropy, a penalty is applied to the initial comprehensive gain value based on the historical failure probability and the current state entropy value. Based on the adjusted comprehensive gain value, the optimal control rule is determined in the subset of candidate rules, the target speed and blade angle adjustment amounts are generated, and execution commands are issued.

[0008] This invention overcomes the prediction divergence defect of pure data models in unknown operating conditions by constructing a bottom-level model by integrating hydraulic loss mechanisms and data-driven algorithms. It utilizes prediction uncertainty to assess output confidence levels, and the state entropy index integrates spatial distance, efficiency gradient, and frequency domain pulsation characteristics, avoiding the inability of a single efficiency index to characterize the stall boundary blind zone. Within the optimization stage, heuristic pruning eliminates a large number of invalid computational branches, reducing the system's underlying computing power consumption. A dynamic inference depth mechanism performs deep planning in areas with high model confidence and shortens the inference steps in areas with low model confidence, preventing the long-term prediction error from accumulating and amplifying over time. A historical penalty mechanism extracts equipment failure characteristics and applies negative feedback suppression to high-risk actions; the penalty intensity dynamically increases with the degree of deterioration of the current operating conditions, preventing repeated oscillations of the equipment on the edge of danger and ensuring the long-term efficient operation of the axial flow pump in complex pipeline disturbance environments.

[0009] Preferably, the construction of the Gaussian process regression surrogate model for simultaneously outputting performance prediction values ​​and prediction uncertainties includes: using the prediction results output by the system equations as the prior mean function of Gaussian process regression; calculating the residual between the actual performance and the prior mean function using the flow field numerical simulation data and the historical operating data; training a residual prediction model using Gaussian process regression with the residual as the target; combining the prior mean function and the residual prediction model to form the surrogate model, and outputting the prediction uncertainty through the posterior standard deviation of Gaussian process regression.

[0010] This invention combines prior physical mechanisms with data-driven residual correction, thereby improving the accuracy of model performance prediction under nonlinear flow field conditions.

[0011] Preferably, the system equations include rotor inertia equations, fluid acceleration equations, and hydraulic loss mechanism equations; wherein, the hydraulic loss mechanism equations include formulas for calculating friction loss and local impact loss, used to calculate the total hydraulic loss and correct the theoretical head obtained based on the Euler pump equations, so as to obtain a predicted pump head value that includes physical mechanism constraints.

[0012] This invention comprehensively incorporates the transient characteristics of pipeline networks and the friction features of pipelines, enhancing the model's ability to respond to head decay trends under variable speed conditions.

[0013] Preferably, the step of calculating the state entropy of the current operating condition based on the surrogate model includes: determining the optimal efficiency curve through the surrogate model; calculating the shortest Euclidean distance from the current operating point to the optimal efficiency curve after dimensionless processing as the nonlinear spatial distance; calculating the normal gradient magnitude of the efficiency contour line at the current operating point as the efficiency contour line gradient; extracting the low-frequency broadband amplitude or characteristic peak value in the axial flow pump pressure pulsation signal as the stall precursor pressure pulsation characteristic amplitude; and after normalizing the nonlinear spatial distance, the efficiency contour line gradient, and the stall precursor pressure pulsation characteristic amplitude, performing linear weighted summation using preset weighting coefficients to obtain the state entropy of the current operating condition.

[0014] This invention assesses the operating status from three dimensions: deviation, attenuation sensitivity, and fluid excitation, thereby improving the sensitivity of early identification of potential equipment faults.

[0015] Preferably, the step of using the surrogate model to estimate the instantaneous state entropy change after rule execution and performing heuristic pruning to select a subset of candidate rules expected to reduce state entropy includes: obtaining the target speed change and blade angle change caused by the corresponding control action in the rule base; superimposing the current active control parameters with the changes, and recalculating the flow rate and head as passive response quantities in combination with the pipeline resistance curve; inputting the superimposed active control parameters and the passive response quantities into the surrogate model to calculate the expected state entropy after executing each rule; subtracting the expected state entropy from the state entropy of the current operating condition to obtain the instantaneous state entropy change, eliminating rules with a change greater than or equal to zero, and forming the candidate rule subset from the remaining rules.

[0016] Preferably, the inference depth decreases as the prediction uncertainty increases, including: obtaining the prediction uncertainty output by the surrogate model in the current operating condition region and performing dimensionless processing; When the dimensionless prediction uncertainty is less than a preset lower threshold, the inference depth is set to a preset maximum depth value. When the dimensionless prediction uncertainty is greater than or equal to a preset upper limit threshold, the inference depth is set to a preset minimum depth value. When the dimensionless prediction uncertainty is between the lower threshold and the upper threshold, the inference depth is calculated using the inverse linear mapping function and rounded down.

[0017] This invention dynamically adjusts the time planning span based on the model confidence level, balancing long-term optimization benefits with short-term decision security.

[0018] Preferably, the step of calling the proxy model to perform sequence reasoning on each rule in the subset of candidate rules and calculating the initial comprehensive gain value includes: within each reasoning step, using the proxy model to iteratively predict the expected state entropy at the next moment, and calculating the decrease in state entropy corresponding to each reasoning step as the single-step gain of the reasoning sequence; and performing cumulative discounting calculation on all single-step gains within the reasoning depth to obtain the initial comprehensive gain value corresponding to each rule.

[0019] Preferably, the step of applying a penalty to the initial comprehensive gain value based on the historical failure probability and the current state entropy value includes: counting the total number of times a certain rule is triggered and executed within the preset historical time window and the number of failures that cause an increase in state entropy; calculating the ratio of the number of failures to the total number of failures as the historical failure probability; multiplying the historical failure probability, the normalized current state entropy value, and a preset adjustment factor to obtain a penalty coefficient; and subtracting the penalty term calculated from the penalty coefficient from the initial comprehensive gain value to obtain the adjusted comprehensive gain value.

[0020] Preferably, the step of determining the optimal control rule based on the adjusted comprehensive gain value in the candidate rule subset, generating the target speed and blade angle adjustment amount, and issuing execution instructions includes: sorting the adjusted comprehensive gain values ​​of each rule in the candidate rule subset in descending order; selecting the rule with the largest adjusted comprehensive gain value that is greater than zero as the optimal control rule; parsing and extracting the target speed adjustment amount and blade angle adjustment amount from the optimal control rule, and converting them into control electrical signals to be sent to the frequency converter and hydraulic adjustment mechanism of the axial flow pump.

[0021] This invention ensures that the equipment strictly executes the control commands corresponding to the highest positive returns, thus accelerating the recovery from abnormal operating conditions to a highly efficient and stable state.

[0022] Secondly, the present invention provides an intelligent optimization control system for the operating conditions of an axial flow pump, comprising a processor and a memory, wherein the memory stores computer program instructions, and when the computer program instructions are executed by the processor, the aforementioned intelligent optimization control method for the operating conditions of an axial flow pump is implemented.

[0023] By adopting the above technical solution, a computer program is generated from the above-mentioned intelligent optimization control method for the operating conditions of an axial flow pump, and stored in a memory for loading and execution by a processor. Terminal devices are then created based on the memory and processor for convenient use.

[0024] The beneficial effects of this invention are as follows: This invention constructs a surrogate model by integrating hydraulic loss mechanisms and Gaussian process regression algorithms, which improves the predictive ability of axial flow pumps under all operating conditions and simultaneously represents prediction uncertainty, thus enhancing the reliability of the data base. Utilizing a state entropy index weighted by the spatial distance of efficiency curves, contour gradients, and stall precursor features, a comprehensive quantitative assessment of equipment health and energy efficiency is achieved. During the inference optimization process, heuristic pruning is performed using estimated state entropy changes, reducing the computational load of inference and increasing the search rate; by correlating inference depth with prediction uncertainty, blind searching in unknown operating condition regions is reduced. Furthermore, the constructed historical feedback penalty mechanism can impose restrictions on rules that have previously led to deterioration of operating conditions based on the severity of the current state entropy, promoting the convergence of the control system towards a highly stable target operating condition, thereby improving the overall operating quality and energy-saving effect of the axial flow pump. Attached Figure Description

[0025] Figure 1 A flowchart of an intelligent optimization control method for the operating conditions of an axial flow pump; Figure 2 This is a schematic diagram illustrating the relationship between inference depth and prediction uncertainty. Figure 3 This is a schematic diagram illustrating the expected gain change under the sequence inference step size. Figure 4 This diagram illustrates the comparison of early warning success rates and system efficiency for four different models. Detailed Implementation

[0026] The technical solutions of the present invention will be clearly and completely described below with reference to the accompanying drawings of the embodiments of the present invention. Obviously, the described embodiments are some embodiments of the present invention, but not all embodiments.

[0027] The specific embodiments of the present invention will now be described in detail with reference to the accompanying drawings.

[0028] This invention discloses an intelligent optimization control method for the operating conditions of an axial flow pump, referring to... Figure 1 This includes steps S1-S4: S1. Construct a Gaussian process regression surrogate model.

[0029] A Gaussian process regression surrogate model is constructed. This model is established by integrating flow field numerical simulation with historical operating data and combining system equations that include hydraulic loss mechanism, time variable, rotor inertia and fluid acceleration. It is used to simultaneously output performance prediction values ​​and prediction uncertainty.

[0030] Three-dimensional flow field pressure distribution data and three-dimensional velocity distribution data are obtained through fluid dynamics simulation software as flow field numerical simulation data. Historical sensor operating data, including flow rate parameters, head parameters, speed parameters, and blade angle parameters, are collected through a programmable logic controller. The historical sensor operating data are then processed by sliding window filtering to remove noise using a data processing and analysis library.

[0031] To characterize the transient speed response, a relationship was established between the rotor inertia equation and the motor load parameters, as follows:

[0032] in, This represents the moment of inertia, preferably set between 150-500 kg·m². The angular acceleration is calculated by differentiating the rotational speed from historical sensor data. This represents the motor's output torque, which is obtained by converting the underlying motor current data using a programmable logic controller. This represents the hydraulic load torque of the axial flow pump, which is obtained through numerical simulation data of the flow field.

[0033] To characterize the transient flow rate variation in the pipeline network, a relationship between the fluid acceleration equation and the pipeline network structural parameters was constructed, as follows:

[0034] in, The equivalent pipeline length is indicated by on-site measurements of the pipeline network. This represents gravitational acceleration, which can be obtained by consulting a table of physical constants. Indicates the flow area of ​​the pipeline. The rate of change in flow is represented by calculations based on the internal flow derivative of historical sensor operating data. The pump head is indicated and is obtained by measuring a head sensor. This represents the system head, which is calculated using a pipeline resistance model.

[0035] To calculate the energy dissipation of fluid flowing inside the pipe, a relationship between friction loss along the pipe and fluid dynamic parameters was established, as follows:

[0036] in, This indicates friction loss along the path. The friction factor is obtained by referring to the standard Moody diagram. The equivalent flow channel length is represented and obtained through measurement using a three-dimensional structural model. Indicates the equivalent hydraulic diameter. The average flow velocity is represented by the flow rate divided by the pipe flow area.

[0037] To calculate the energy drop of fluid flowing through valve and pipe fitting structures, a relationship was established between local impact loss and local resistance characteristic parameters. The specific relationship is as follows:

[0038] in, Indicates localized impact loss. This represents the local loss coefficient, which can be obtained by consulting the underlying data manual provided by the valve manufacturer.

[0039] To obtain the total system resistance data, a relationship was constructed between the total hydraulic loss and various independent loss parameters. The specific relationship is as follows:

[0040] in, This indicates the total hydraulic loss.

[0041] To predict the system's head performance, a relationship was established between the predicted pump head and the system's hydraulic loss parameters. The specific relationship is as follows:

[0042] in, This indicates the predicted pump head. The theoretical head is represented by the value, which is calculated using the Euler pump equation.

[0043] To evaluate the system's energy conversion efficiency, a relationship was constructed between the predicted efficiency and mechanical performance parameters, as follows:

[0044] in, This represents the predicted efficiency value. Fluid density is indicated by consulting a standard fluid property table. The flow rate is indicated and obtained through measurement using an ultrasonic flow meter. The shaft power is represented and obtained by measuring a torque-speed composite sensor.

[0045] The predicted values ​​of pump head and efficiency output from the above system equations are encapsulated into a prior mean vector. This prior mean vector is used as the initial prior mean function for the Gaussian process regression calculation stage. Using 5000-10000 sets of flow field numerical simulation data and actual operating performance parameters, the underlying residual between the actual performance and the prior mean function is calculated. The residual prediction model is trained under supervision with the residual as the target. The residual prediction model is constructed using the Matern kernel function, and the Matern kernel function is optimized by maximum likelihood estimation. The underlying length scale of the Matrn kernel function is optimized to a preset range, which in this embodiment is optimized to a preset range of 0.1-2. The underlying signal variance of the Matrn kernel function is optimized to a target value range, which in this embodiment is optimized to a range of 0.01-0.5. The prior mean function and the residual prediction model are fully superimposed and fused. After the superposition and fusion process is completed, a Gaussian process regression surrogate model is obtained. The Gaussian process regression surrogate model outputs the performance prediction value simultaneously during the output of the posterior mean stage. The Gaussian process regression surrogate model uses the diagonal elements of its own posterior covariance matrix to extract the square root data and output the prediction uncertainty.

[0046] In this way, by integrating the prior knowledge of hydraulic loss mechanism and dynamic equation, a Gaussian process regression surrogate model is constructed, which cuts off the prediction failure error link generated by the pure data-driven model in the sparse sample region, reduces the prediction deviation amplitude in the variable operating condition region, and improves the accuracy of equipment performance prediction under all operating conditions and the reliability of system status.

[0047] S2. Calculate the state entropy of the current operating condition.

[0048] The state entropy of the current operating condition is calculated based on the surrogate model. The state entropy is composed of the nonlinear spatial distance between the operating point and the optimal efficiency curve, the gradient of the efficiency contour line, and the amplitude of the stall precursor pressure pulsation characteristics.

[0049] To map the current traffic to the feature space, a relationship between the dimensionless traffic and the current actual traffic was constructed, as follows:

[0050] in, Represents dimensionless flow rate. This indicates the current actual flow rate, obtained through measurement using an ultrasonic flow meter. This indicates the flow rate parameter at the rated optimal operating point, which can be obtained by referring to the factory parameters on the pump nameplate.

[0051] To map the current head into the feature space, a relationship between the dimensionless head and the current actual head was constructed, as follows:

[0052] in, This represents the dimensionless head. The current actual head is indicated by measurement using a pressure sensor. This indicates the head parameter at the rated optimal operating point, which can be obtained by referring to the factory parameters on the pump nameplate.

[0053] The underlying efficiency surface is output using a Gaussian process regression surrogate model to obtain dimensionless flow coordinates and dimensionless head coordinates. The efficiency surface is processed by a polynomial fitting algorithm to obtain the optimal efficiency curve. The shortest Euclidean distance between the current operating condition coordinate point and the nearest point on the optimal efficiency curve is calculated using the tree search algorithm inside the spatial data structure library. The shortest Euclidean distance is defined as a nonlinear spatial distance.

[0054] To assess the sensitivity of efficiency degradation around the operating point, a relationship was constructed between the gradient magnitude of the efficiency contour lines and the partial derivatives, as follows:

[0055] in, This represents the magnitude of the gradient of the efficiency contour lines. The partial derivatives of the efficiency surface in dimensionless flow coordinates are calculated using the automatic differentiation algorithm within the scientific computing library. The partial derivatives of the efficiency surface in dimensionless head coordinates are calculated using the automatic differentiation algorithm within the scientific computing library.

[0056] A high-frequency pressure sensor installed at the impeller inlet is used to acquire real pressure pulsation signals. The sampling rate of the high-frequency pressure sensor is set to 2000 to 5000 Hz. The real pressure pulsation signals are processed by frequency domain transformation using a fast Fourier transform algorithm to extract frequency domain feature amplitudes within the range of 0.2 to 0.8 times the impeller rotation frequency. These frequency domain feature amplitudes are used as the characteristic amplitudes of pressure pulsation indices representing stall precursors. Under the current operating conditions, the pressure pulsation feature amplitudes are extracted from real-time data collected by the sensor. When calculating the expected state entropy in the future, since the pulsation features are relatively stable within a short-term inference window, the pressure pulsation feature amplitudes extracted at the current moment are used as an approximation to reduce the prediction dimensionality of the underlying model.

[0057] The minimum-maximum mapping algorithm is used to map nonlinear spatial distances, to map the gradient magnitude of efficiency contour lines, and to map the amplitude of pressure pulsation characteristics. The minimum-maximum mapping algorithm maps the above three characteristics to the interval between 0 and 1. The minimum-maximum mapping algorithm depends on the bottom maximum and minimum values ​​being pre-calibrated from historical operating data and flow field numerical simulation samples, or the minimum-maximum mapping algorithm depends on the bottom maximum and minimum values ​​being pre-calibrated from preset safety boundaries.

[0058] To characterize the degree of deviation from optimal operating conditions and the potential stall risk of an axial flow pump as a comprehensive scalar, a relationship between state entropy and normalized characteristic quantities was constructed, as follows:

[0059] in, Represents state entropy, Represents normalized spatial distance. This represents the normalized gradient magnitude. This represents the normalized pressure pulsation characteristic amplitude, and the Min-Max algorithm is used uniformly to represent it. , and The three core parameters are mapped to the [0,1] interval for normalization. The maximum and minimum values ​​used in the Min-Max algorithm are pre-calibrated from historical running data, CFD flow field simulation samples, or preset safety boundaries, and remain unchanged within the same control calibration period. The weighting coefficients corresponding to nonlinear spatial distances are obtained through calibration using the analytic hierarchy process (AHP). The weighting coefficients corresponding to the gradient magnitude of the efficiency contour lines are obtained through experimental calibration and experience. The weighting coefficients representing the characteristic amplitude of pressure pulsation are obtained through expert-based parameter settings. These weighting coefficients must satisfy the following conditions: , The value range is from 0.2 to 0.3. The value range is from 0.2 to 0.3. The value range is set to 0.4 to 0.6.

[0060] In this way, the limitations of a single efficiency evaluation index are overcome. By integrating the sensitivity of efficiency decay and the frequency domain characteristics of early stall, a state entropy index is constructed. This cuts off the error links within potentially dangerous areas that pure operating data cannot perceive, improves the accuracy of equipment health assessment, and enables precise quantitative assessment of the overall energy efficiency level and potential stall danger boundaries of the entire system.

[0061] S3. Use a proxy model to filter a subset of candidate rules.

[0062] During the forward chain reasoning process of the inference engine, the surrogate model is used to predict the instantaneous change in state entropy after rule execution and perform heuristic pruning to select a subset of candidate rules that are expected to reduce state entropy. The surrogate model is then invoked to perform sequential reasoning on each rule in the subset of candidate rules and calculate the initial comprehensive gain value, where the reasoning depth decreases as the prediction uncertainty increases.

[0063] The underlying control action corresponding to each rule is extracted from the expert rule base. The extracted control action includes the change in target rotational speed and the change in blade angle.

[0064] To update the active control parameter speed characteristics, a relationship between the target speed and the change in target speed was constructed, as follows:

[0065] in, Indicates the target rotational speed. The initial rotational speed is obtained through historical sensor data. This represents the change in target rotational speed, obtained through a query from the expert rule base.

[0066] To update the active control parameters of the blade, a relationship between the target blade angle and the change in blade angle was constructed, as follows:

[0067] in, Indicates the target blade angle. The initial blade angle is indicated by data obtained from historical sensor operation. This indicates the change in blade angle, obtained through a query from the expert rule base.

[0068] To characterize the passive response boundary conditions of the pipeline network system, a relationship between the system head and flow parameters was constructed, as follows:

[0069] in, Indicates the system head. The pipeline resistance coefficient is determined by calibrating based on flow rate and head units combined with pipeline test data. The flow rate is indicated and obtained through measurement using an ultrasonic flow meter. The static head is determined by measuring the height of the pipeline at the site.

[0070] The supply and demand balance is calculated within the Gaussian process regression surrogate model using the Newton-Raphson iteration method. The iteration stops when the absolute error between the pump head and the system head is less than 0.01 meters. The new flow rate and new head of the passive response are obtained. The active control parameters and the passive response are input into the Gaussian process regression surrogate model. The Gaussian process regression surrogate model outputs the expected state entropy. The instantaneous state entropy change is obtained by subtracting the current state entropy from the expected state entropy. Invalid rules corresponding to instantaneous state entropy changes greater than or equal to zero are eliminated, and the remaining valid rules are retained to form a subset of candidate rules.

[0071] To strongly correlate inference sequence depth with model prediction uncertainty, a relationship between inference sequence depth and dimensionless prediction uncertainty was constructed, as follows:

[0072] in, Indicates the depth of the reasoning sequence. This indicates the preset maximum inference depth, which is obtained through system initialization parameter settings. This indicates the preset minimum inference depth, which is obtained through system initialization parameter settings. This indicates the preset upper limit threshold. This indicates the preset lower threshold. and The lower limit threshold of this uncertainty is pre-set by obtaining parameters based on expert experience. A value of 0.05 indicates that the model has low prediction uncertainty in the current operating condition region, and the upper limit threshold is preset. Set to 0.2. This represents the dimensionless prediction uncertainty. It is obtained in real-time from the posterior variance output of the surrogate model through Gaussian process regression, representing the current operating condition prediction uncertainty. This uncertainty is then normalized according to a preset reference standard deviation or the statistical range of the training samples to obtain the dimensionless prediction uncertainty. .

[0073] To evaluate the asymptotic response characteristics of the actuator's rotational speed, a relationship was constructed between the rotational speed at the next moment and the actuator's rotational speed response coefficient. The specific relationship is as follows:

[0074] in, Indicates the rotational speed at the next moment. The rotational speed at the current moment is obtained through sequence iteration and state propagation. This represents the speed response coefficient of the actuator, which can be obtained by referring to the factory parameters at the bottom layer of the frequency converter. Indicates the target rotational speed.

[0075] To evaluate the asymptotic response characteristics of the actuator blade angle, a relationship was constructed between the blade angle at the next moment and the actuator angle response coefficient. The specific relationship is as follows:

[0076] in, Indicates the blade angle at the next moment. The current blade angle is obtained through sequential iteration and state propagation. This represents the angular response coefficient of the actuator, which can be obtained by referring to the underlying factory parameters of the hydraulic servo drive. Indicates the target blade angle.

[0077] To perform cumulative discounting calculations on all single-step gains within the inference sequence depth, a relationship was established between the initial comprehensive gain value and the attenuation discount factor, as follows:

[0078] in, This represents the initial combined gain value. The depth of the inference sequence is represented by the aforementioned depth-inverse linear mapping function. This represents the attenuation discount factor, obtained through system configuration parameters, and its value is limited to between 0.8 and 0.95. This represents the single-step gain, which is calculated by the expected decrease in state entropy within the inference step.

[0079] By exponentially discounting and summing the gains of each single step within the sequence depth, the output initial comprehensive gain value G not only represents the rule's transient correction capability but also serves to evaluate the rule's long-term expected contribution to maintaining stable system operation within a future time window determined by the inference step size and inference depth. Figure 3 As shown, the contribution of the single-step gain after discounting decreases with the increase of time steps, and the increase of the cumulative discounted gain gradually decreases and tends to stabilize, indicating a reasonable reduction in the long-term control effect and an emphasis on the short-term correction capability.

[0080] For example, a preset maximum inference depth is specified. The process is 10 steps, with a preset minimum reasoning depth. It consists of two steps. In the judgment logic, if This indicates that the model has high prediction reliability in the current operating condition area, allowing for longer sequence inference, with an inference depth of 10 steps assigned; if This indicates that the current operating point may be located in a sparse region of the training samples. To reduce the impact of unreliable long-term predictions on control decisions, the inference depth is limited to a minimum of 2 steps. When the prediction uncertainty is between 0.05 and 0.2, for example, the current... The result is 0.1. The inverse linear mapping function is called for interpolation calculation, and the example numerical values ​​are substituted to obtain the result. The result is 7.33. Rounding down this result determines the inference sequence depth to be 7 steps. Figure 2 As shown, this dynamic inference depth mechanism based on uncertainty can fully leverage the inference optimization capability when the model accuracy is high, and reduce the accumulation of long-term errors when the model prediction uncertainty is high.

[0081] S4. Establish a penalty mechanism and determine the optimal control rules before issuing execution instructions.

[0082] If the execution of a rule within a preset historical time window has led to an increase in state entropy, a penalty is applied to the initial comprehensive gain value based on the historical failure probability and the current state entropy value. The optimal control rule is determined from the candidate rule subset based on the adjusted comprehensive gain value, and the target speed and blade angle adjustment amounts are generated and execution commands are issued.

[0083] The total number of times a rule is triggered and executed within a preset historical time window is counted. The number of failures that lead to an increase in state entropy is then identified, and the historical failure probability is obtained by dividing the number of failures by the total number of triggers and executions. To dynamically amplify the penalty based on the severity of the working conditions, a relationship between the penalty coefficient and the historical failure probability is constructed, as follows:

[0084] in, Indicates the penalty coefficient. The historical failure probability is represented by statistics obtained through an underlying data cache maintained by a specific double-ended queue. This represents the current state entropy value. This represents a preset adjustment factor, set within the range of 1.5 to 3; in this embodiment, it is set to 2. This indicates the preset upper limit value, obtained through a safety threshold setting. .

[0085] To apply negative feedback to weaken the initial composite gain, a relationship was established between the adjusted composite gain and the penalty coefficient, as follows:

[0086] in, This represents the adjusted overall gain value. This represents the initial combined gain value. This represents the penalty coefficient.

[0087] The system uses a quicksort algorithm to sort all the rules in the candidate rule subset in descending order of their adjusted comprehensive gain values. The system selects the rule with the largest adjusted comprehensive gain value that is greater than zero as the optimal control rule. If the adjusted comprehensive gain value of the rule at the head of the queue is less than or equal to zero, the system maintains the current control parameters, or the system directly executes the preset safety protection rule when the gain value is less than or equal to zero.

[0088] The system uses a rule parser to decompose the conclusion text corresponding to the optimal control rule. The system extracts the precise operation command parameters, which include the target speed adjustment and blade angle adjustment. The programmable logic controller (PLC) or distributed control system (DCS) maps the target speed adjustment and blade angle adjustment into standard control electrical signals. For speed adjustment operations, the system generates a 4-20mA standard current signal, or sends a digital frequency command to the drive motor's corresponding frequency converter via an industrial fieldbus protocol. The frequency converter smoothly reduces its output frequency from 45Hz to 43.5Hz. For blade angle adjustment operations, the system outputs a 0-10V proportional voltage signal to the electro-hydraulic proportional valve inside the hydraulic servo system. The electro-hydraulic proportional valve drives the cylinder extension mechanism to rotate the blade connecting rod by the corresponding angle.

[0089] The system controls the electrical signal transmission delay to within 50-200ms, and drives the axial flow pump to smoothly transition to the expected entropy reduction safe operating point through electromechanical-hydraulic coordinated response.

[0090] For example, a specific double-ended queue is used to calculate the total number of execution triggers as 10, the number of failures that cause an increase in state entropy as 2, and the historical failure probability as 0.2. The current state entropy is set to 0.8, the preset adjustment factor is set to 2, and the preset upper limit is set to 1. These parameters are then substituted into the penalty coefficient constraint function to calculate a penalty coefficient of 0.32. The initial overall gain is set to 100. The initial overall gain and the penalty coefficient are then substituted into the formula for the adjusted overall gain to calculate a final adjusted overall gain of 68.

[0091] In this way, the internal computational load of the reasoning and optimal solution search stage is reduced by heuristic pruning. A dynamic reasoning depth mechanism is established in combination with prediction uncertainty to ensure the optimization depth in unknown operating conditions and avoid the accumulation of long-term prediction errors. A historical feedback penalty mechanism is constructed to weaken the weight of actions that have caused deterioration of operating conditions. The internal error link of the equipment repeatedly triggering invalid control commands in the harsh operating conditions area is cut off, preventing the system from experiencing severe control oscillations and ensuring the reliability of the equipment's operation across the entire domain and the efficiency of system regulation.

[0092] To verify the actual effects and performance advantages of each module of the present invention, four comparison groups were set: baseline model A, which adopts a traditional feedback control agentless model and inference mechanism; variant model B, which adopts a pure data-driven agent model and fixed step-size inference; variant model C, which adds mechanism fusion and heuristic rule pruning on the basis of variant model B; and the complete present invention model D, which includes sequence depth and penalty mechanism.

[0093] The experiment was conducted using a large-scale vertical axial flow pump hardware-in-the-loop test platform. The system's rated flow rate was 8.5 m³ / s and the rated head was 6.5 m. The experiment was designed to run continuously for 72 hours under extreme pipeline resistance change and stall conditions. High-frequency pressure signals and low-frequency operating parameters were collected synchronously throughout the experiment. The experiment focused on examining the performance of each group on four key indicators: average state entropy, control response delay time, stall warning success rate, and overall system operating efficiency.

[0094] Data analysis after a full 72-hour test cycle showed that baseline model A had an average state entropy of 0.68, a control response delay of 12.5 seconds, a stall warning success rate of only 45.2%, and an overall operating efficiency of 78.4%.

[0095] The average state entropy of variant model B decreased to 0.55, the control response delay time was shortened to 8.2s, the stall warning success rate was increased to 68.5%, and the overall operating efficiency reached 81.2%.

[0096] The average state entropy of variant model C was further optimized to 0.42, the control response delay time was reduced to 5.4s, the stall warning success rate was 82.6%, and the overall operating efficiency was improved to 83.7%.

[0097] Model D of this invention performs better in various comparison indicators, with an average state entropy value of 0.25, a control response delay time reduced to 1.8s, a stall warning success rate of 98.5%, and an overall operating efficiency of 86.3%.

[0098] like Figure 4 As shown, compared with baseline model A, variant models B and C, model D of this invention has improved both the early warning success rate and system efficiency, verifying the dual advantages of the proposed control method in stall prevention and energy efficiency optimization.

[0099] Further analysis of the above results reveals that variant model C improves upon variant model B in various metrics, indicating that incorporating prior knowledge of equations such as hydraulic loss compensates for the prediction blind spots in sparse sample regions caused by purely data-driven approaches. Model D further improves core performance compared to variant model C, with an average state entropy reduction of 40.5% and a control response delay time reduction of 66.7%. These results demonstrate that a dynamic inference depth mechanism, inversely proportional to prediction uncertainty, can balance optimization depth with the risk of error accumulation. Simultaneously, a penalty mechanism built upon historical failure probabilities reduces the probability of repeated invalid actions under harsh operating conditions, enhancing the operational reliability and control efficiency of large axial flow pumps under complex nonlinear flow field disturbances.

[0100] This invention also discloses an intelligent optimization control system for the operating conditions of an axial flow pump, including a processor and a memory. The memory stores computer program instructions, and when the computer program instructions are executed by the processor, an intelligent optimization control method for the operating conditions of an axial flow pump according to the present invention is implemented.

[0101] The aforementioned intelligent optimization control system for axial flow pump operation also includes other components well known to those skilled in the art, such as communication buses and communication interfaces. Their settings and functions are known in the art and will not be described in detail here.

[0102] In the description of this specification, "multiple" or "several" means at least two, such as two, three or more, unless otherwise expressly and specifically defined.

Claims

1. A method for intelligent optimization control of axial flow pump operating conditions, characterized in that, include: S1. Construct a Gaussian process regression proxy model. This model is established by integrating flow field numerical simulation with historical operating data and combining system equations that include hydraulic loss mechanism, time variable, rotor inertia and fluid acceleration. It is used to simultaneously output performance prediction value and prediction uncertainty. S2. Calculate the state entropy of the current working condition based on the surrogate model. The state entropy is composed of the nonlinear spatial distance between the working point and the optimal efficiency curve, the gradient of the efficiency contour line, and the amplitude of the stall precursor pressure pulsation characteristic after normalization. S3. During the forward chain reasoning process of the inference engine, the surrogate model is used to predict the instantaneous change in state entropy after rule execution and perform heuristic pruning to select a subset of candidate rules that are expected to reduce the state entropy; the surrogate model is called to perform sequential reasoning on each rule in the subset of candidate rules and calculate the initial comprehensive gain value, wherein the reasoning depth decreases as the prediction uncertainty increases; S4. Establish a penalty mechanism. If the execution of a rule within a preset historical time window has led to an increase in state entropy, then apply a penalty to the initial comprehensive gain value based on the historical failure probability and the current state entropy value. Determine the optimal control rule based on the adjusted comprehensive gain value in the candidate rule subset, generate the target speed and blade angle adjustment amount, and issue execution instructions.

2. The intelligent optimization control method for the operating conditions of an axial flow pump according to claim 1, characterized in that, The construction of the Gaussian process regression surrogate model, used to simultaneously output performance predictions and prediction uncertainties, includes: The prediction results output by the system equations are used as the prior mean function of the Gaussian process regression. Using the numerical simulation data of the flow field and the historical operating data, the residual between the actual performance and the prior mean function is calculated. Using the residuals as the target, a residual prediction model is obtained by training with Gaussian process regression; The surrogate model is formed by combining the prior mean function with the residual prediction model, and the prediction uncertainty is output through the posterior standard deviation of Gaussian process regression.

3. The intelligent optimization control method for the operating conditions of an axial flow pump according to claim 2, characterized in that, The system equations include the rotor inertia equation, the fluid acceleration equation, and the hydraulic loss mechanism equation; The hydraulic loss mechanism equation includes formulas for calculating friction loss and local impact loss, which are used to calculate the total hydraulic loss and correct the theoretical head obtained based on the Euler pump equation, so as to obtain a predicted pump head value that includes physical mechanism constraints.

4. The intelligent optimization control method for the operating conditions of an axial flow pump according to claim 1, characterized in that, The calculation of the state entropy of the current operating condition based on the proxy model includes: The optimal efficiency curve is determined by the surrogate model, and the shortest Euclidean distance from the current operating point after dimensionless processing to the optimal efficiency curve is calculated as the nonlinear spatial distance. Calculate the normal gradient magnitude of the efficiency contour lines at the current operating point as the gradient of the efficiency contour lines; Extract the low-frequency broadband amplitude or characteristic peak value from the axial flow pump pressure pulsation signal as the stall precursor pressure pulsation characteristic amplitude; After normalizing the nonlinear spatial distance, the efficiency contour gradient, and the stall precursor pressure pulsation characteristic amplitude, the current state entropy is obtained by linear weighted summation using preset weighting coefficients.

5. The intelligent optimization control method for the operating conditions of an axial flow pump according to claim 1, characterized in that, The process of using the surrogate model to predict the instantaneous change in state entropy after rule execution and performing heuristic pruning to filter out a subset of candidate rules expected to reduce state entropy includes: Obtain the target speed change and blade angle change caused by the corresponding control action from the rule base; The current active control parameters are superimposed with the changes, and the flow rate and head are recalculated as passive response quantities in combination with the pipeline resistance curve. The superimposed active control parameters and the passive response quantity are input into the agent model to calculate the expected state entropy after executing each rule. The instantaneous change in state entropy is obtained by subtracting the expected state entropy from the current state entropy. Rules with a change greater than or equal to zero are eliminated, and the remaining rules form the candidate rule subset.

6. The intelligent optimization control method for the operating conditions of an axial flow pump according to claim 1, characterized in that, The reasoning depth decreases as the prediction uncertainty increases, including: Obtain the prediction uncertainty output by the surrogate model in the current working condition region and perform dimensionless processing; When the dimensionless prediction uncertainty is less than a preset lower threshold, the inference depth is set to a preset maximum depth value. When the dimensionless prediction uncertainty is greater than or equal to a preset upper limit threshold, the inference depth is set to a preset minimum depth value. When the dimensionless prediction uncertainty is between the lower threshold and the upper threshold, the inference depth is calculated using the inverse linear mapping function and rounded down.

7. The intelligent optimization control method for the operating conditions of an axial flow pump according to claim 1, characterized in that, The step of invoking the proxy model to perform sequence reasoning on each rule in the subset of candidate rules and calculating the initial comprehensive gain value includes: Within each inference step, the surrogate model is used to iteratively predict the expected state entropy at the next moment, and the decrease in state entropy corresponding to each inference step is calculated as the single-step gain of the inference sequence. The initial integrated gain value corresponding to each rule is obtained by summing and discounting all single-step gains within the inference depth.

8. The intelligent optimization control method for the operating conditions of an axial flow pump according to claim 1, characterized in that, The penalty applied to the initial synthesis gain value based on the historical failure probability and the current state entropy value includes: The total number of times a certain rule is triggered and executed within the preset historical time window, as well as the number of failures that cause an increase in state entropy, are counted. The ratio of the number of failures to the total number of failures is then calculated as the historical failure probability. The penalty coefficient is obtained by multiplying the historical failure probability, the normalized current state entropy value, and the preset adjustment factor. The adjusted overall gain value is obtained by subtracting the penalty term calculated from the penalty coefficient from the initial overall gain value.

9. The intelligent optimization control method for the operating conditions of an axial flow pump according to claim 1, characterized in that, The process of determining the optimal control rule based on the adjusted comprehensive gain value in the candidate rule subset, generating the target speed and blade angle adjustment amounts, and issuing execution commands includes: The adjusted overall gain values ​​of each rule in the candidate rule subset are sorted in descending order. The rule that has the largest adjusted overall gain value and is greater than zero is selected as the optimal control rule; The target speed adjustment and blade angle adjustment in the optimal control rules are analyzed and extracted, and converted into control electrical signals and sent to the frequency converter and hydraulic adjustment mechanism of the axial flow pump.

10. An intelligent optimization control system for the operating conditions of an axial flow pump, characterized in that, include: A processor and a memory, wherein the memory stores computer program instructions, and when the computer program instructions are executed by the processor, an intelligent optimization control method for the operating conditions of an axial flow pump according to any one of claims 1-9 is implemented.

Images