Intelligent control method for multi-working-condition super-high-lift single-stage centrifugal pump

Abstract

The application relates to the technical field of centrifugal pump control and discloses a kind of intelligent regulation and control multi-working condition superhigh-lift single-stage centrifugal pump control methods, the method includes: the attention of multi-source working condition data is guided and the dimensionality reduction fusion and multi-scale feature extraction, obtain multi-scale operation feature, using the time clustering algorithm based on density to the multi-scale operation feature is time evolution and clustering processing, obtain the current working condition of superhigh-lift single-stage centrifugal pump;The current working condition, performance parameter and performance score of superhigh-lift single-stage centrifugal pump are regarded as state vector, and the optimal control vector is selected from the control space using the reinforcement learning model.The application realizes intelligent perception and adaptive control of superhigh-lift single-stage centrifugal pump under multi-working condition and full life cycle by fusing multi-source data, introducing time evolution modeling and reinforcement learning optimization strategy, breaks through the problem that traditional rule-driven control is slow in response and has regulation lag under complex operating state, and improves energy efficiency ratio, operating stability and equipment life.

Description

TECHNICAL FIELD

[0001] The present application relates to the field of centrifugal pump control, and in particular to a method for intelligently regulating a multi-working-condition super-high-lift single-stage centrifugal pump. BACKGROUND

[0002] A super-high-lift single-stage centrifugal pump is a typical fluid conveying device that transports liquid from a low-pressure area to a high-pressure area through impeller rotation. It has the advantages of simple structure, low cost, and convenient installation and maintenance, and is widely used in industrial cooling, urban water supply, farmland irrigation, mine drainage, and chemical industries. In particular, in applications that require continuous liquid transport and stable pressure output, super-high-lift single-stage centrifugal pumps are preferred due to their high reliability and low failure rate. However, the lift of traditional super-high-lift single-stage centrifugal pumps is limited, and they cannot provide the required working pressure in high-drop or high-resistance conveying environments. Therefore, super-high-lift single-stage centrifugal pumps have emerged, which achieve higher lift in a single-stage structure through impeller structure optimization, speed increase, and hydraulic parameter adjustment, and are widely used in high-rise building water supply, mountain water transport, and remote pipe network pressure compensation.

[0003] With the increasing complexity of conveying conditions, super-high-lift single-stage centrifugal pumps face frequent load fluctuations, multi-working-condition switching, changes in liquid physical properties, and equipment aging during operation. These factors not only pose higher requirements on the structural strength and hydraulic performance of the pump body, but also make it difficult for traditional control methods based on fixed control logic or static empirical models to adapt. Such control methods cannot dynamically respond to complex working condition changes and are difficult to achieve comprehensive optimal regulation of operating efficiency, system stability, and equipment life.

[0004] Current research has attempted to introduce artificial intelligence and model prediction methods to improve centrifugal pump control. For example, patent CN113901710B proposes a centrifugal pump operation optimization control method, which establishes a system mathematical model by constructing an offline database and corrects the improved BPNN prediction model using actual operation data. The binary tree search algorithm is used to achieve online rapid control rate matching, thereby optimizing the operating parameters of lift, efficiency, and flow. This method realizes fast adjustment and adaptive compensation based on empirical data, to some extent, alleviating the operating error problem caused by model deviation. However, this method is based on pre-constructed databases and offline models for regulation, lacks real-time recognition and dynamic adaptation to complex and variable working conditions (such as load disturbance, liquid property change, frequent start-stop, etc.), and mainly uses static models. Although a historical data correction mechanism is introduced, it does not model the performance degradation trend caused by equipment aging and running phase migration over time.

[0005] To solve the problem, the application provides a kind of intelligent control multi-working condition super-high lift single-stage centrifugal pump control method, the operation optimization and regulation control of pump equipment in multiple working conditions are realized by artificial intelligence technology, and the overall performance and service life of super-high lift single-stage centrifugal pump are significantly improved. SUMMARY

[0006] The application provides a kind of intelligent control multi-working condition super-high lift single-stage centrifugal pump control method, since traditional method relies on static sensing parameter, it is difficult to accurately identify the running state of super-high lift single-stage centrifugal pump in variable environment, the application introduces attention mechanism to adaptively weight fusion of multi-source working condition data, and adopts density-based time clustering to process time evolution of multi-scale running features, realizes high-robustness identification and group perception of variable working conditions, solves the problem of inaccurate working condition perception under complex conditions, and uses neural regression forest to fit and predict the running features under current working condition, accurately outputs multiple performance parameters such as actual lift and hydraulic efficiency, and comprehensively quantifies through nonlinear performance scoring function, solves the problems of performance multi-index modeling difficulty and insufficient dynamic response capability, takes the current working condition, performance parameter and performance score as state vector, dynamically selects the optimal control vector in control space, realizes joint regulation of frequency converter frequency, main shaft speed and valve opening, so as to solve the problems of control parameter adjustment lag, control target conflict and poor system stability.

[0007] To achieve the above purpose, the application provides a kind of intelligent control multi-working condition super-high lift single-stage centrifugal pump control method, including the following steps:

[0008] S1: periodically collect multi-source working condition data in the running process of super-high lift single-stage centrifugal pump by using acquisition device, and perform attention-guided dimension reduction fusion on multi-source working condition data to obtain multi-source working condition fusion tensor;

[0009] S2: multi-scale feature extraction is performed on the multi-source working condition fusion tensor to obtain multi-scale running features of super-high lift single-stage centrifugal pump, density-based time clustering algorithm is used to process time evolution of multi-scale running features, and the multi-scale running features after time evolution processing are clustered into clustering clusters;

[0010] S3: the working condition corresponding to the clustering cluster to which the multi-scale running features belong is taken as the current working condition of super-high lift single-stage centrifugal pump, the performance parameters of super-high lift single-stage centrifugal pump under current working condition are output based on current working condition and multi-scale running features by using neural regression forest, and the performance parameters are converted into performance score by using comprehensive performance scoring function;

[0011] S4: taking the current working condition, performance parameter and performance score of the super-high-lift single-stage centrifugal pump as a state vector, selecting an optimal control vector corresponding to the state vector from a control space by using a reinforcement learning model, and adjusting the variable frequency, main shaft speed and valve opening of the super-high-lift single-stage centrifugal pump by using the optimal control vector.

[0012] As a further improved method of the application:

[0013] Optionally, the acquisition device comprises a pressure sensor, an electromagnetic flowmeter, a current sensor, a power meter, a rotating speed / frequency sensor, a temperature sensor and an acceleration sensor; the multi-source working condition data in the running process of the super-high-lift single-stage centrifugal pump is periodically acquired by using the acquisition device, comprising:

[0014] The multi-source working condition data comprises a plurality of working condition data, which are, in sequence, inlet instantaneous pressure data, outlet instantaneous pressure data, rotating speed data, frequency data, temperature data, acceleration data, current data, power data and instantaneous flow data of the super-high-lift single-stage centrifugal pump;

[0015] The working condition data in the multi-source working condition data are all sequence data with a length of N, N represents the number of sampling time points in the process of acquiring the multi-source working condition data, the time interval of adjacent sampling time points is set to 100 milliseconds, and N is set to 50;

[0016] The missing values in the multi-source working condition data are interpolated and completed by using a linear weighted interpolation method to obtain the completed multi-source working condition data;

[0017] The data values of all the working condition data in the completed multi-source working condition data are differentially filtered by using an adaptive weighted difference filtering method.

[0018] Optionally, the multi-source working condition data after the differential filtering is subjected to attention-guided dimension reduction fusion, comprising:

[0019] The multi-source working condition data after the differential filtering is converted into a matrix form:

[0020] ;

[0021] Wherein, is a matrix with N rows and M columns, the first n rows and the first m columns in the matrix , the nth row and the mth column in the matrix , represents the nth working condition data in the mth column, ​The differential filtering result of the data values ​​at each sampling time point, where M represents the number of data categories in the multi-source operating condition data. The first to Mth types of operating condition data are, respectively, the inlet instantaneous pressure data, outlet instantaneous pressure data, speed data, frequency data, temperature data, acceleration data, current data, power data, and instantaneous flow rate data of the ultra-high head single-stage centrifugal pump. M is 9. This indicates the sequence data length of the operating condition data in the multi-source operating condition data after differential filtering.

[0022] The first in The row is represented as This indicates that the ultra-high head single-stage centrifugal pump is in the first stage. A vector composed of the data values ​​of M types of working conditions at each sampling time;

[0023] Calculate the temporal attention at each sampling time in multi-source working condition data in matrix form, where the temporal attention at each sampling time is calculated as follows:

[0024] ;

[0025] Where T represents transpose. Indicates the first The temporal attention at each sampling time is a scalar between 0 and 1;

[0026] Let represent an exponential function with base to the natural constant, and Q represent a trainable query vector. Represents a trainable linear transformation matrix;

[0027] matrix Convert to time attention weighted covariance matrix ;

[0028] Time attention weighted covariance matrix Perform eigenvalue decomposition, select the eigenvectors corresponding to the K largest eigenvalues ​​to form the principal component matrix, and use the principal component matrix to transform the matrix... Projected onto a low-dimensional space, this becomes a multi-source operating condition fusion tensor, which is an N-1 row, K column matrix. Behavior No. Multi-source dimensionality reduction projection vector at each sampling time.

[0029] Optionally, multi-scale feature extraction is performed on the multi-source condition fusion tensor, including:

[0030] Local similarity and global similarity of the multi-source dimension reduction projection vectors of any two sampling time points in the multi-source working condition fusion tensor are calculated, and the local similarity and global similarity are taken as matrix elements to construct a local adjacency graph and a global adjacency graph of N-1 sampling time points, both of which are in the form of N-1 rows and N-1 columns.

[0031] The multi-source working condition fusion tensor, the local adjacency graph and the global adjacency graph are taken as inputs of the graph neural network.

[0032] The graph neural network includes a plurality of graph convolution layers nested in a hierarchical structure.

[0033] A residual channel enhancement module is introduced between each graph convolution layer to enhance the stability and nonlinear expression ability of feature transmission.

[0034] The residual channel enhancement module includes an SE channel attention mechanism and a residual connection path.

[0035] After the graph neural network is processed by multiple layers of graph convolution, the embedded features of the last layer of graph convolution are output as multi-scale running features F representing the running state of the super-high lift single-stage centrifugal pump.

[0036] Optionally, a density-based time clustering algorithm is used to perform time evolution processing on the multi-scale running features, and the multi-scale running features after time evolution processing are clustered into a clustering cluster, including:

[0037] The density-based time clustering algorithm is used to generate a time evolution factor of the multi-scale running features F .

[0038] The time evolution factor is used to perform time evolution processing on the multi-scale running features F to obtain the multi-scale running features after time evolution processing , wherein the control coefficient is set to 0.1.

[0039] The Euclidean distance between the multi-scale running features after time evolution processing and the core points of different clustering clusters is calculated, and the multi-scale running features after time evolution processing are clustered into the clustering cluster with the closest Euclidean distance, and the working condition label of the clustered clustering cluster is taken as the current working condition of the super-high lift single-stage centrifugal pump.

[0040] Optionally, based on the current working condition of the super-high lift single-stage centrifugal pump and the multi-scale running features, a neural regression forest is used to output performance parameters of the super-high lift single-stage centrifugal pump under the current working condition, including:

[0041] ​The neural regression forest comprises a plurality of parallel regression tree structures, a split function of each regression tree is parameterized and modeled by a neural network, so that the split function has stronger nonlinear fitting capability; a current operating condition and multi-scale operating characteristics of the super-high lift single-stage centrifugal pump are jointly constructed as a high-dimensional feature vector, the high-dimensional feature vector is taken as an input of a regression tree in the neural regression forest, each regression tree performs nonlinear regression mapping on the high-dimensional feature vector through an embedded shallow neural network to obtain a regression output of each regression tree, wherein the regression output is a mapping result of the high-dimensional feature vector to a performance parameter dimension; and a performance parameter of the super-high lift single-stage centrifugal pump under the current operating condition is obtained by fusing the regression outputs of the plurality of regression trees, the performance parameter comprising an actual lift, a hydraulic efficiency, a theoretical flow deviation, a shaft power and a running resistance change rate.

[0042] Optionally, the performance parameters are converted into performance scores by using a comprehensive performance score function, comprising:

[0043] The performance parameters are normalized;

[0044] The normalized performance parameters are taken as inputs of the comprehensive performance score function to obtain performance scores corresponding to the performance parameters:

[0045] ;

[0046] wherein, the normalized performance parameter vector comprises a normalized actual lift , a normalized hydraulic efficiency , a normalized theoretical flow deviation , a normalized shaft power and a normalized running resistance change rate , the performance score corresponding to the normalized performance parameter vector , are score coefficients, and are set to be equal and 0.2.

[0047] Optionally, an optimal control vector corresponding to a state vector is selected from a control space by using a reinforcement learning model, comprising:

[0048] constructing a state vector as an input of the reinforcement learning model: [current operating condition of the super-high lift single-stage centrifugal pump, performance parameter, performance score];

[0049] constructing a control space C of the super-high lift single-stage centrifugal pump, the control space C comprising the following three continuous control dimensions:

[0050] a frequency conversion frequency of a frequency converter in the super-high lift single-stage centrifugal pump ;​

[0051] Spindle speed ;

[0052] Valve opening at the outlet of ultra-high head single-stage centrifugal pump ;

[0053] in, To control the frequency conversion in space C The output range, This is the minimum output value of the frequency converter. This represents the maximum output value of the frequency converter. To control the spindle speed in space C The output range, This indicates the minimum output value of the spindle speed. This indicates the maximum output value of the spindle speed. To control the valve opening in space C The output range;

[0054] A policy network in a reinforcement learning model is constructed using a policy gradient-based reinforcement learning method. The policy network takes a state vector as input and selects the optimal frequency, spindle speed, and valve opening from three continuous control dimensions corresponding to the state vector in the control space C to form an optimal control vector. The optimal control vector is then used to adjust the frequency, spindle speed, and valve opening of an ultra-high head single-stage centrifugal pump.

[0055] Compared with existing technologies, this invention proposes a smart control method for ultra-high head single-stage centrifugal pumps under multiple operating conditions, which has the following beneficial effects:

[0056] First, this invention significantly enhances the expressive power of complex operating conditions by combining temporal attention weighting and covariance principal component decomposition. This attention-guided dimensionality reduction fusion method not only considers the statistical synergistic relationships between features but also dynamically identifies anomalous time points with key diagnostic value in the time-series data (such as sudden vibrations, drastic pressure changes, etc.) and assigns them higher participation weights, thereby better preserving representative change information during the dimensionality reduction process. Compared with traditional PCA, attention-guided dimensionality reduction fusion effectively solves the problem of important detail submersion caused by global averaging and improves the sensitivity to non-stationary and abrupt operating conditions, especially suitable for ultra-high head centrifugal pump systems with high-speed rotation and strong nonlinear responses. Specifically, the dimensionality reduction results have higher discriminativeness and compressibility, providing more compact and dynamically semantic input features for subsequent state identification, strategy optimization, and other modules, significantly improving the response accuracy and adaptive capability of the entire control system.

[0057] Meanwhile, the neural regression forest model is used for accurately predicting key performance parameters of the super-high-lift single-stage centrifugal pump, such as actual lift, hydraulic efficiency, theoretical flow deviation, shaft power and running resistance change rate, so that the multi-dimensional performance of the super-high-lift single-stage centrifugal pump can be comprehensively described, the comprehensive performance scoring function can quantitatively evaluate multiple heterogeneous performance indexes, and the comparability and sensitivity of decision-making are enhanced. Further, the reinforcement learning model is used for dynamically deciding an optimal control strategy based on a state vector, so that adaptive adjustment of a variable frequency frequency, a main shaft speed and a valve opening can be realized under varying working conditions, and thus the energy efficiency ratio, stability and service life of the super-high-lift single-stage centrifugal pump are improved. BRIEF DESCRIPTION OF DRAWINGS

[0058] Figure 1 A flowchart of a control method for intelligently regulating a super-high-lift single-stage centrifugal pump under multiple working conditions is provided for an embodiment of the present application.

[0059] Figure 2 A structural diagram of a super-high-lift single-stage centrifugal pump is provided for an embodiment of the present application. DETAILED DESCRIPTION

[0060] The implementation, functional features and advantages of the present application will be further described with reference to the embodiments and the accompanying drawings. It should be understood that the specific embodiments described herein are only used to explain the present application and do not limit the present application.

[0061] An embodiment of the present application provides a control method for intelligently regulating a super-high-lift single-stage centrifugal pump under multiple working conditions. The execution subject of the control method for intelligently regulating a super-high-lift single-stage centrifugal pump under multiple working conditions includes but is not limited to at least one of electronic devices such as a server, a terminal and the like which can be configured to execute the method provided by the embodiment of the present application. In other words, the control method for intelligently regulating a super-high-lift single-stage centrifugal pump under multiple working conditions can be executed by software or hardware installed in a terminal device or a server device, and the software can be a blockchain platform. The server includes but is not limited to a single server, a server cluster, a cloud server or a cloud server cluster and the like.

[0062] Referring to Figure 1 and Figure 2 , embodiment 1 of the present application is:

[0063] S1: periodically collecting multi-source working condition data in the running process of the super-high-lift single-stage centrifugal pump by using a collection device, and performing attention-guided dimension reduction fusion on the multi-source working condition data to obtain a multi-source working condition fusion tensor.

[0064] The acquisition device comprises a pressure sensor, an electromagnetic flowmeter, a current sensor, a power meter, a rotating speed / frequency sensor, a temperature sensor and an acceleration sensor; the acquisition device is used to periodically acquire multi-source working condition data in the operation process of the single-stage centrifugal pump with super-high lift, comprising:

[0065] The multi-source working condition data comprises various working condition data, which are in sequence respectively the inlet instantaneous pressure data, the outlet instantaneous pressure data, the rotating speed data, the frequency data, the temperature data, the acceleration data, the current data, the power data of the single-stage centrifugal pump with super-high lift and the instantaneous flow data of the single-stage centrifugal pump with super-high lift;

[0066] In the embodiment of the application, the structure of the single-stage centrifugal pump with super-high lift comprises a high specific speed closed impeller, a double volute, a main shaft and bearing assembly, a multi-stage sealed pump body and a variable frequency drive motor system; referring to Figure 2 The figure is a structural schematic diagram of the single-stage centrifugal pump with super-high lift provided by an embodiment of the application;

[0067] Specifically, the high specific speed closed impeller adopts a long blade closed structure, has a large impeller diameter, realizes a larger fluid tangential velocity to obtain a higher lift, and adopts a corrosion-resistant or high-strength alloy material to withstand the stress generated by high-speed rotation; the double volute is used for balancing the radial force, reducing the flow velocity, recovering the kinetic energy and improving the outlet pressure; the main shaft adopts a heat-treated stainless steel or high-strength steel material to adapt to high rotating speed and large load; the multi-stage sealed pump body adopts a combination of mechanical sealing, labyrinth sealing and balance pipe sealing to prevent high-pressure side liquid leakage; the variable frequency drive motor system is matched with a frequency converter and is used to adjust the variable frequency, the rotating speed of the main shaft and the valve opening degree of the single-stage centrifugal pump with super-high lift;

[0068] It should be noted that the pressure sensor is arranged at the outlet side and the inlet side of the single-stage centrifugal pump with super-high lift to acquire the inlet instantaneous pressure data and the outlet instantaneous pressure data of the single-stage centrifugal pump with super-high lift; the rotating speed / frequency sensor is used to perceive the feedback result of the frequency converter in the single-stage centrifugal pump with super-high lift to obtain the rotating speed data and the frequency data; the temperature sensor and the acceleration sensor are arranged in the main shaft area of the single-stage centrifugal pump with super-high lift to perceive the main shaft temperature and the acceleration of the main shaft in the horizontal direction to obtain the temperature data and the acceleration data; the electromagnetic flowmeter is arranged inside the pump body of the single-stage centrifugal pump with super-high lift to acquire the instantaneous flow through the pump body as the instantaneous flow data through the single-stage centrifugal pump with super-high lift; the current sensor and the power meter are used to acquire the current data and the power data of the single-stage centrifugal pump with super-high lift;

[0069] Specifically, the instantaneous pressure data are used to evaluate the head and impedance changes, the instantaneous flow data are used to characterize the load of the super-high head single-stage centrifugal pump, the rotation speed data and the frequency data are used to characterize the motion state of the frequency converter in the super-high head single-stage centrifugal pump, the current data and the power data are used to characterize the energy consumption state of the super-high head single-stage centrifugal pump, and the temperature data and the acceleration data are used to characterize the structural stability of the super-high head single-stage centrifugal pump.

[0070] The working condition data in the multi-source working condition data are sequence data with a length of N, N represents the number of sampling time points in the process of collecting the multi-source working condition data, the time interval of adjacent sampling time points is set to 100 milliseconds, and N is set to 50;

[0071] The missing values in the multi-source working condition data are interpolated and completed by using a linear weighted interpolation method to obtain the completed multi-source working condition data.

[0072] The data values of all working condition data in the completed multi-source working condition data are differentially filtered by using an adaptive weighted difference filtering method. Specifically, the differential filtering process of the adaptive weighted difference filtering method is as follows:

[0073] The working condition data are extracted from the completed multi-source working condition data , wherein, are the data values collected at the 1st to Nth sampling time points in the working condition data x, respectively, ;

[0074] The local standard deviation of each data value in the working condition data x is calculated, and the local standard deviation of the data value is:

[0075] ;

[0076] , wherein, represents the local standard deviation of the data value , L represents the length of the local time window, and L is set to 5, ;

[0077] The local standard deviation of the data value is converted into an adaptive weighting coefficient, and the local standard deviation is converted into an adaptive weighting coefficient according to the following formula:

[0078] ;

[0079] , wherein, represents the local standard deviation of the (n-1)th data value in the working condition data x, represents the adaptive weighting coefficient converted from the local standard deviation ;

[0080] Adaptive weighted differential filtering is performed on the data values ​​using adaptive weighting coefficients, where the data values The adaptive weighted differential filter formula is:

[0081] ;

[0082] in, Represents data value The adaptive weighted differential filtering results;

[0083] Compared to the traditional fixed-coefficient first-order difference method, the adaptive weighted difference filtering method adaptively allocates difference weights according to the local variation trend of the input data. Specifically, when the fluctuation is small (significant low-frequency drift), the difference effect is enhanced. When the fluctuation is large (abrupt changes exist), the weights are smoothly adjusted, reducing the sensitivity of the difference to high-frequency noise, thereby enhancing the response capability to small changes. While preserving key structural features, it suppresses high-frequency noise interference, which can effectively alleviate the information distortion problem caused by the original difference when processing non-stationary and noise-sensitive data.

[0084] Attention-guided dimensionality reduction fusion of multi-source operating condition data after differential filtering includes:

[0085] Convert the multi-source operating condition data after differential filtering into matrix form:

[0086] ;

[0087] in, for A matrix of M rows and M columns. Located in the matrix The first in Row m, Indicates the first Data on various working conditions in the first The differential filtering result of the data values ​​at each sampling time point, where M represents the number of data categories in the multi-source operating condition data. The first to Mth types of operating condition data are, respectively, the inlet instantaneous pressure data, outlet instantaneous pressure data, speed data, frequency data, temperature data, acceleration data, current data, power data, and instantaneous flow rate data of the ultra-high head single-stage centrifugal pump. M is 9. This indicates the sequence data length of the operating condition data in the multi-source operating condition data after differential filtering.

[0088] The row is represented as This indicates that the ultra-high head single-stage centrifugal pump is in the first stage. A vector composed of the data values ​​of M types of working conditions at each sampling time;

[0089] This indicates the sequence data length of the operating condition data in the multi-source operating condition data after differential filtering.

[0090] In multi-source operating condition data in matrix form, the temporal attention at each sampling time is calculated, where the temporal attention at each sampling time is calculated as follows:

[0091] ;

[0092] Where T represents transpose. Indicates the first The temporal attention at each sampling time is a scalar between 0 and 1;

[0093] Let represent an exponential function with base to the natural constant, and Q represent a trainable query vector. Represents a trainable linear transformation matrix;

[0094] It is a scalar between 0 and 1; in this embodiment of the invention, a linear transformation matrix is ​​used to transform... Projecting dimension M onto the hidden space and using query vectors to generate the importance of sampling moments, the trainable query vectors are used to automatically identify key moments during the operation of ultra-high head single-stage centrifugal pumps, such as abnormal vibration, energy consumption peaks, and sudden changes in operating load. These key moments are assigned higher weights to ensure that their influence is maximized in subsequent feature extraction.

[0095] matrix Convert to time attention weighted covariance matrix :

[0096] ;

[0097] ;

[0098] in, Representation matrix The time attention weighted mean vector; specifically, the time attention weighted covariance matrix is ​​used to amplify the contribution of key moments in the statistical space, focusing on fluctuation features with actual diagnostic value, so that abnormal and drastic time segments have a stronger deterministic effect on the main direction of features.

[0099] Time attention weighted covariance matrix Perform eigenvalue decomposition, select the eigenvectors corresponding to the K largest eigenvalues ​​to form the principal component matrix, and use the principal component matrix to transform the matrix... Projected onto a low-dimensional space, this becomes a multi-source operating condition fusion tensor, which is an N-1 row, K column matrix. Behavioral The multi-source dimension reduction projection vectors of the N-1 sampling time points are obtained by performing the multi-source dimension reduction projection on the multi-source fusion tensor.

[0100] S2: Multi-scale feature extraction is performed on the multi-source fusion tensor to obtain multi-scale operation features of the super-high lift single-stage centrifugal pump, a time evolution process is performed on the multi-scale operation features by using a density-based time clustering algorithm, and the multi-scale operation features after the time evolution process are clustered into clustering clusters.

[0101] The multi-scale feature extraction on the multi-source fusion tensor includes:

[0102] The local similarity and the global similarity of the multi-source dimension reduction projection vectors of any two sampling time points in the multi-source fusion tensor are calculated, and the local similarity and the global similarity are taken as matrix elements to construct a local adjacency graph and a global adjacency graph of N-1 sampling time points, respectively, wherein the local adjacency graph and the global adjacency graph are both in the form of an N-1 row and N-1 column matrix.

[0103] As an embodiment of the present application, the multi-source dimension reduction projection vector of the first sampling time point and the multi-source dimension reduction projection vector of the jth sampling time point The local similarity and the global similarity calculation method are as follows:

[0104] ;

[0105] ;

[0106] wherein, indicates the local similarity of the multi-source dimension reduction projection vector and the multi-source dimension reduction projection vector indicates the global similarity of the multi-source dimension reduction projection vector and the multi-source dimension reduction projection vector indicates a similarity scale coefficient, which is set to 3, indicates a local sampling time point threshold, which is set to 5, L1 norm;

[0107] indicates a global similarity fusion coefficient (for example, set to 0.5), indicates the multi-source dimension reduction projection vector and the multi-source dimension reduction projection vector ​​​​​cosine similarity, a trainable K by K projection matrix;

[0108] The representation of the local adjacency graph is: , The representation of the global adjacency graph is: , The global adjacency graph is represented by:

[0109] Specifically, by introducing local similarity and global similarity, the feature extraction capability of the working condition evolution pattern under different time scales is significantly enhanced, wherein the local adjacency graph accurately captures the sensitive changes of the super-high-lift single-stage centrifugal pump in a short time, effectively identifying transient anomalies; the global adjacency graph integrates the static geometric structure and dynamic projection information of the super-high-lift single-stage centrifugal pump, and can learn the nonlinear dependence between distant time steps. Specifically, the cosine similarity measures the direction consistency of the multi-source dimension reduction projection vector, emphasizes the trend commonality, and improves the perception ability of complex dynamic behaviors such as periodic instability and hysteresis effect.

[0110] The multi-source working condition fusion tensor, the local adjacency graph and the global adjacency graph are used as inputs of the graph neural network;

[0111] The graph neural network includes a plurality of graph convolution layers nested in a hierarchical structure;

[0112] A residual channel enhancement module is introduced between each graph convolution layer to enhance the stability and nonlinear expression capability of feature transmission;

[0113] The residual channel enhancement module includes an SE channel attention mechanism and a residual connection path;

[0114] After the graph neural network is processed by multiple layers of graph convolution, the embedding features of the output of the last layer of graph convolution are output as multi-scale running features F representing the running state of the super-high-lift single-stage centrifugal pump.

[0115] The graph convolution processing procedure based on the local adjacency graph and the global adjacency graph is:

[0116] ;

[0117] ;

[0118] ;

[0119] wherein, represents the embedding features output by the y+1th graph convolution layer in the graph neural network, represents the embedding features output by the yth graph convolution layer in the graph neural network, the local adjacency graph is represented by: a degree matrix of the local adjacency graph, a degree matrix of the global adjacency graph, a degree matrix of the global adjacency graph, a normalized result of the local adjacency graph, a normalized result of the global adjacency graph;

[0120] an activation function, and the activation function is set as a ReLU function;

[0121] a trainable local adjacency convolution matrix,

[0122] The double-graph fusion graph convolution operation sufficiently integrates local and global information, and improves the robust modeling capability of the pump group under multiple working conditions on non-stationary sequences.

[0123] The time evolution processing is performed on the multi-scale running features by using a density-based time clustering algorithm, and the multi-scale running features after the time evolution processing are clustered into clustering clusters, including:

[0124] The time evolution factor of the multi-scale running features F is generated by using a density-based time clustering algorithm:

[0125] ;

[0126] wherein, the time evolution factor of the multi-scale running features F, an exponential function with a natural constant as a base, a current collection number of the multi-source working condition data, a preset maximum collection number, time represents a total running time of the super-high lift single-stage centrifugal pump, and TIME represents a maximum running time of the super-high lift single-stage centrifugal pump, all are time evolution coefficients, and are set to 0.4 and 0.6 respectively;

[0127] The time evolution factor is used to perform time evolution processing on the multi-scale running features F, and the multi-scale running features after the time evolution processing are obtained. a control coefficient, and is set to 0.1;

[0128] ​​​​​​​The time-evolution-processed multi-scale operating feature and the core points of different clustering clusters, the time-evolution-processed multi-scale operating feature is clustered into the clustering cluster closest to the Euclidean distance.

[0129] S3: taking the working condition corresponding to the clustering cluster to which the multi-scale operating feature belongs as the current working condition of the super-high lift single-stage centrifugal pump, outputting the performance parameter of the super-high lift single-stage centrifugal pump under the current working condition by using the neural regression forest based on the current working condition and the multi-scale operating feature, and converting the performance parameter into a performance score by using a comprehensive performance scoring function.

[0130] taking the working condition label of the clustered clustering cluster as the current working condition of the super-high lift single-stage centrifugal pump;

[0131] It should be noted that the multi-scale operating features under multiple working conditions are collected in advance, and the DBSCAN algorithm is used for clustering to obtain multiple clustering clusters and core points of the clustering clusters, and the proportion of the working condition label of each clustering cluster is counted, and the working condition label with the highest proportion is selected as the working condition label of the clustering cluster.

[0132] Based on the current working condition of the super-high lift single-stage centrifugal pump and the multi-scale operating feature, the neural regression forest outputs the performance parameter of the super-high lift single-stage centrifugal pump under the current working condition, including:

[0133] The neural regression forest includes multiple parallel regression tree structures, and the split function of each regression tree is parameterized modeled by a neural network, so that it has stronger nonlinear fitting capability; the current working condition of the super-high lift single-stage centrifugal pump and the multi-scale operating feature are jointly constructed into a high-dimensional feature vector, and the high-dimensional feature vector is taken as the input of the regression tree in the neural regression forest; each regression tree performs nonlinear regression mapping on the high-dimensional feature vector through an embedded shallow neural network to obtain the regression output of each regression tree, wherein the regression output is the mapping result of the high-dimensional feature vector to the performance parameter dimension; by fusing the regression outputs of multiple regression trees, the performance parameter of the super-high lift single-stage centrifugal pump under the current working condition is obtained, and the performance parameter includes actual lift, hydraulic efficiency, theoretical flow deviation, shaft power and running resistance change rate, which can realize adaptive modeling of the relationship between non-stationary and nonlinear operating features and performance responses.

[0134] As an embodiment of the present application, the neural regression forest is trained offline using a historical operation data set of the super-high-lift centrifugal pump to obtain an initial performance prediction model, and in the real-time operation process of the super-high-lift centrifugal pump, new working condition parameters and performance feedback data are continuously collected as online learning samples to incrementally update part of the regression trees in the neural regression forest, so as to ensure that the neural regression forest can adapt to changes in the running environment in a timely manner. For example, when the super-high-lift centrifugal pump causes the hydraulic efficiency to decrease due to changes in water quality, the updated neural regression forest can automatically correct the performance prediction and output a shaft power and a running resistance change rate that are closer to the actual running conditions. In this way, the neural regression forest can maintain prediction accuracy and generalization ability in the face of non-stationary working conditions and long-term running states, thereby improving the intelligent operation control level of the super-high-lift centrifugal pump.

[0135] The performance parameters are converted into performance scores by using a comprehensive performance score function, including:

[0136] The performance parameters are normalized; optionally, the normalization method is a maximum-minimum normalization method;

[0137] The normalized performance parameters are used as inputs of the comprehensive performance score function to obtain performance scores corresponding to the performance parameters:

[0138] ;

[0139] wherein, represents a normalized performance parameter vector, the normalized performance parameter vector including a normalized actual head , a hydraulic efficiency , a theoretical flow deviation , a shaft power , and a running resistance change rate , represents a performance score corresponding to the normalized performance parameter vector .

[0140] are score coefficients, and are set to are equal and are 0.2.

[0141] S4: The current working condition, performance parameters, and performance scores of the super-high-lift single-stage centrifugal pump are used as a state vector, an optimal control vector corresponding to the state vector is selected from a control space by using a reinforcement learning model, and the optimal control vector is used to adjust the frequency of the frequency conversion, the speed of the main shaft, and the opening of the valve of the super-high-lift single-stage centrifugal pump.

[0142] A state vector as an input of the reinforcement learning model is constructed: [current working condition of the super-high-lift single-stage centrifugal pump, performance parameters, performance scores].

[0143] A control space C of the super-high-lift single-stage centrifugal pump is constructed, and the control space C includes three continuous control dimensions as follows:

[0144] A variable frequency of the frequency converter in the super-high-lift single-stage centrifugal pump ;

[0145] A rotation speed of the main shaft ;

[0146] An opening degree of a valve at an outlet of the super-high-lift single-stage centrifugal pump ;

[0147] Wherein, is an output range of the variable frequency in the control space C, is a minimum output value (for example, 20 Hz) of the variable frequency, is a maximum output value (for example, 60 Hz) of the variable frequency, is an output range of the rotation speed of the main shaft in the control space C, represents a minimum output value (for example, 30% of rated rotation speed) of the rotation speed of the main shaft, represents a maximum output value (for example, 120% of rated rotation speed) of the rotation speed of the main shaft, is an output range of the opening degree of the valve in the control space C.

[0148] A policy network in a reinforcement learning model is constructed by using a policy gradient type reinforcement learning method, the policy network takes a state vector as an input, and selects a variable frequency, a rotation speed of the main shaft and an opening degree of the valve corresponding to the state vector from three continuous control dimensions in the control space C, to form an optimal control vector, and the optimal control vector is used to adjust the variable frequency, the rotation speed of the main shaft and the opening degree of the valve of the super-high-lift single-stage centrifugal pump.

[0149] As an embodiment of the present application, a training process of the policy network is as follows:

[0150] A simulation environment of the super-high-lift single-stage centrifugal pump is constructed, and specifically, refer to as Figure 2A structural schematic diagram of the super-high-lift single-stage centrifugal pump is shown, a three-dimensional structural model of the super-high-lift single-stage centrifugal pump is generated in a simulation environment, the three-dimensional structural model is subjected to high-quality body grid division, including that a high-specific-speed closed impeller region adopts structured or near-structured grid to ensure blade boundary layer analysis, a gap / seam region is subjected to grid densification, and an unstructured transition grid is adopted for a rotor periphery; fluid mechanics equations (Navier-Stokes equations) are used to model continuity and momentum conservation of fluid in the pump cavity, so as to obtain velocity distribution, pressure distribution and energy transfer law of fluid in the super-high-lift centrifugal pump, the energy transfer process of the high-specific-speed closed impeller is combined, and a theoretical lift model of the super-high-lift centrifugal pump is constructed with reference to Euler turbine equations, for reflecting energy lifting effect provided by the high-specific-speed closed impeller on fluid, and meanwhile, the theoretical model is compared with and corrected according to an empirical curve obtained through actual test, so that the three-dimensional structural model in the simulation environment can accurately reflect real running performance of the super-high-lift centrifugal pump under different control conditions such as flow rate, rotational speed and valve opening degree.

[0151] Initialize trainable parameters in a policy network, the policy network being configured to receive a state vector, calculate a policy selection probability of a control vector meeting a variable frequency, a main shaft rotational speed and a valve opening degree limit in a control space C, and select the control vector with the highest policy selection probability as an optimal control vector corresponding to the state vector;

[0152] Extract a state vector of a current three-dimensional model structure in a simulation environment, take a performance score of a state vector after the control vector is adopted as a reward function, and interactively iterate the state vector and trainable parameters by using a policy optimization algorithm until a preset maximum number of interactive iterations (50 times) is reached, and construct a policy network based on trainable parameters obtained through final iteration, the trainable parameters of the policy network including weight matrix parameters and bias parameters;

[0153] The interactive iteration process is as follows: the policy network generates a policy selection probability of different control vectors under a current state vector, selects a control vector with the highest policy selection probability to optimize control of the three-dimensional model structure, generates a state vector of the three-dimensional model structure in a new state, calculates a reward function value of the selected control vector, and iteratively updates the trainable parameters by using a policy optimization method;

[0154] Optionally, the policy optimization method adopts a PPO (Proximal Policy Optimization) method, a REINFORCE with baseline method or an Advantage Actor-Critic method.

[0155] It should be noted that the neural regression forest model can accurately predict the key performance parameters of the super-high lift single-stage centrifugal pump, such as actual lift, hydraulic efficiency, theoretical flow deviation, shaft power and running resistance change rate, so as to comprehensively characterize the multi-dimensional performance of the super-high lift single-stage centrifugal pump, and the comprehensive performance scoring function can quantitatively evaluate the multiple heterogeneous performance indicators, thereby enhancing the comparability and sensitivity of the decision. Further, the reinforcement learning model can dynamically decide the optimal control strategy based on the state vector, so as to realize the adaptive adjustment of the frequency, shaft speed and valve opening of the variable frequency under the changing conditions, thereby improving the energy efficiency ratio, stability and service life of the super-high lift single-stage centrifugal pump.

[0156] As a preferred embodiment of the present application, the actual lift is the energy height of the actual output of the super-high lift single-stage centrifugal pump, and insufficient lift will lead to substandard fluid delivery, and too high lift will waste energy, both of which will reduce energy efficiency. High-lift operation is often accompanied by high load, which can easily aggravate impeller fatigue and bearing wear. The hydraulic efficiency is the proportion of the input shaft power used for water delivery, and the hydraulic efficiency directly determines the effective delivery capacity under unit power consumption. Low efficiency operation is often accompanied by unstable factors such as turbulent flow and cavitation, which affects the stability of the super-high lift single-stage centrifugal pump. The theoretical flow deviation is the difference between the actual flow and the theoretically designed flow, and a large deviation means that the pump operation deviates from the optimal operating point, and the energy utilization rate decreases. Long-term non-theoretical flow operation can easily lead to component fatigue or local overheating. The shaft power is the total power input required to drive the super-high lift single-stage centrifugal pump to run, and continuous high shaft power operation will overload the motor and bearing, accelerating aging. The running resistance change rate is the change rate of the resistance of the fluid in the pump cavity and the pipeline system of the super-high lift single-stage centrifugal pump. Sudden resistance means that the system is blocked, scaled or cavitated, which is prone to cause vibration. The above performance indicators all obviously affect the energy efficiency ratio, stability and service life of the super-high lift single-stage centrifugal pump.

[0157] It should be understood that the above-mentioned embodiments are only for illustration, and the scope of the patent application is not limited by the structure.

[0158] It should be noted that the above-mentioned embodiment numbers of the present application are only for description, and do not represent the advantages and disadvantages of the embodiments. And the terms "include", "contain" or any other variants in this paper are intended to cover non-exclusive inclusion, so that the process, device, article or method including a series of elements not only includes those elements, but also includes other elements not explicitly listed, or includes elements inherent to such process, device, article or method. Without more limitations, the element defined by the statement "including a" does not exclude the presence of another identical element in the process, device, article or method including the element.

[0159] Those skilled in the art can clearly understand the above-mentioned embodiment method can be realized by means of software and the necessary general hardware platform, of course, also can be through hardware, but in many cases the former is the better embodiment. Based on such understanding, the technical solutions of the present application essentially or say the part of the contribution to the prior art can be embodied in the form of software products, the computer software product is stored in a storage medium (such as ROM / RAM, magnetic disc, optical disc) as described above, including a number of instructions to make a terminal device (may be a mobile phone, computer, server, or network equipment, etc.) executes the method described in various embodiments of the present application.

[0160] The above is only the preferred embodiment of the present application, not therefore limit the patent scope of the present application, any equivalent structure or equivalent flow transformation made by using the content of the present application specification and drawings, or directly or indirectly applied in other related technical fields, are also included in the patent protection scope of the present application.

Claims

1. A method for intelligently controlling a single-stage centrifugal pump with ultra-high head under multiple operating conditions, characterized in that: The method includes: S1: Periodically collect multi-source operating condition data during the operation of an ultra-high head single-stage centrifugal pump using acquisition equipment, and perform attention-guided dimensionality reduction fusion on the multi-source operating condition data to obtain the multi-source operating condition fusion tensor. S2: Multi-scale feature extraction is performed on the multi-source operating condition fusion tensor to obtain the multi-scale operating features of the ultra-high head single-stage centrifugal pump. A density-based time clustering algorithm is used to perform time evolution processing on the multi-scale operating features, and the time-evolved multi-scale operating features are clustered into clusters. S3: The operating conditions corresponding to the clusters to which the multi-scale operating features belong are taken as the current operating conditions of the ultra-high head single-stage centrifugal pump. Based on the current operating conditions and multi-scale operating features, the performance parameters of the ultra-high head single-stage centrifugal pump under the current operating conditions are output using neural regression forest, and the performance parameters are converted into performance scores using a comprehensive performance scoring function. S4: The current operating conditions, performance parameters, and performance score of the ultra-high head single-stage centrifugal pump are used as the state vector. The optimal control vector corresponding to the state vector is selected from the control space using a reinforcement learning model. The frequency conversion frequency, spindle speed, and valve opening of the ultra-high head single-stage centrifugal pump are adjusted using the optimal control vector.

2. The intelligent control method for a multi-condition ultra-high head single-stage centrifugal pump as described in claim 1, characterized in that, The data acquisition equipment includes a pressure sensor, an electromagnetic flowmeter, a current sensor, a power meter, a speed / frequency sensor, a temperature sensor, and an acceleration sensor. The equipment is used to periodically acquire multi-source operating condition data during the operation of the ultra-high head single-stage centrifugal pump, including: The multi-source operating condition data includes various operating condition data, namely, the inlet instantaneous pressure data, outlet instantaneous pressure data, speed data, frequency data, temperature data, acceleration data, current data, power data, and instantaneous flow rate data of the ultra-high head single-stage centrifugal pump. The operating condition data in the multi-source operating condition data are all sequence data of length N, where N represents the number of sampling moments in the process of collecting multi-source operating condition data. The time interval between adjacent sampling moments is set to 100 milliseconds, and N is set to 50. The missing values ​​in the multi-source working condition data are interpolated and filled using a linear weighted interpolation method to obtain the filled multi-source working condition data. An adaptive weighted differential filtering method is used to perform differential filtering on the data values ​​of all operating conditions in the completed multi-source operating condition data.

3. The intelligent control method for a multi-condition ultra-high head single-stage centrifugal pump as described in claim 2, characterized in that, Attention-guided dimensionality reduction fusion of multi-source operating condition data after differential filtering includes: Convert the multi-source operating condition data after differential filtering into matrix form: ； in, for A matrix of M rows and M columns. Located in the matrix The first in Row m, Indicates the first Data on various working conditions in the first The differential filtering result of the data values ​​at each sampling time point, where M represents the number of data categories in the multi-source operating condition data. The first to Mth types of operating condition data are, respectively, the inlet instantaneous pressure data, outlet instantaneous pressure data, speed data, frequency data, temperature data, acceleration data, current data, power data, and instantaneous flow rate data of the ultra-high head single-stage centrifugal pump. M is 9. This indicates the sequence data length of the operating condition data in the multi-source operating condition data after differential filtering. The first in The row is represented as This indicates that the ultra-high head single-stage centrifugal pump is in the first stage. A vector composed of the data values ​​of M types of working conditions at each sampling time; Calculate the temporal attention at each sampling time in multi-source working condition data in matrix form, where the temporal attention at each sampling time is calculated as follows: ； Where T represents transpose. Indicates the first The temporal attention at each sampling time is a scalar between 0 and 1; Let represent an exponential function with base to the natural constant, and Q represent a trainable query vector. Represents a trainable linear transformation matrix; matrix Convert to time attention weighted covariance matrix ; Time attention weighted covariance matrix Perform eigenvalue decomposition, select the eigenvectors corresponding to the K largest eigenvalues ​​to form the principal component matrix, and use the principal component matrix to transform the matrix... Projected onto a low-dimensional space, this becomes a multi-source operating condition fusion tensor, which is an N-1 row, K column matrix. Behavior No. Multi-source dimensionality reduction projection vector at each sampling time.

4. The intelligent control method for a multi-condition ultra-high head single-stage centrifugal pump as described in claim 3, characterized in that, Multi-scale feature extraction is performed on the multi-source working condition fusion tensor, including: Calculate the local similarity and global similarity of the multi-source dimension-reduced projection vectors at any two sampling times in the multi-source working condition fusion tensor. Use the local similarity and global similarity as matrix elements to construct local adjacency graphs and global adjacency graphs at N-1 sampling times, respectively. Both the local adjacency graph and the global adjacency graph are in the form of N-1 rows and N-1 columns. The multi-source working condition fusion tensor, local adjacency graph, and global adjacency graph are used as inputs to the graph neural network; The graph neural network includes multiple graph convolutional layers nested in a hierarchical structure; A residual channel enhancement module is introduced between each graph convolutional layer to enhance the stability of feature propagation and nonlinear expressive power. The residual channel enhancement module includes an SE channel attention mechanism and a residual connection path; After the graph neural network is processed by multiple graph convolutions, the embedding features of the last graph convolutional layer are output as multi-scale operating features F to characterize the operating status of the ultra-high head single-stage centrifugal pump.

5. The intelligent control method for a multi-condition ultra-high head single-stage centrifugal pump as described in claim 4, characterized in that, A density-based temporal clustering algorithm is used to perform temporal evolution processing on multi-scale operational features, and the time-evolved multi-scale operational features are clustered into clusters, including: The time evolution factor of multi-scale running features F is generated using a density-based temporal clustering algorithm. ; Using the time evolution factor The multi-scale operational feature F is subjected to time evolution processing to obtain the time-evolutionized multi-scale operational feature. , Indicates the control coefficient, set It is 0.1; Computational time evolution processing of multi-scale operational characteristics The Euclidean distance between the core points of different clusters represents the multi-scale operational characteristics after time evolution processing. Cluster the pump to the cluster with the closest Euclidean distance, and use the operating condition label of the cluster as the current operating condition of the ultra-high head single-stage centrifugal pump.

6. The intelligent control method for a multi-condition ultra-high head single-stage centrifugal pump as described in claim 1, characterized in that, Based on the current operating conditions and multi-scale operational characteristics of the ultra-high head single-stage centrifugal pump, neural regression forest is used to output the performance parameters of the ultra-high head single-stage centrifugal pump under the current operating conditions, including: The neural regression forest comprises multiple parallel regression tree structures. The splitting function of each regression tree is parameterized and modeled using a neural network, giving it stronger nonlinear fitting capabilities. The current operating conditions and multi-scale operating characteristics of the ultra-high head single-stage centrifugal pump are jointly constructed into a high-dimensional feature vector. This high-dimensional feature vector is used as the input to the regression trees in the neural regression forest. Each regression tree performs a nonlinear regression mapping on the high-dimensional feature vector through an embedded shallow neural network, yielding the regression output of each regression tree. The regression output is the mapping result of the high-dimensional feature vector to the performance parameter dimension. By fusing the regression outputs of multiple regression trees, the performance parameters of the ultra-high head single-stage centrifugal pump under the current operating conditions are obtained. These performance parameters include actual head, hydraulic efficiency, theoretical flow deviation, shaft power, and rate of change of operating resistance.

7. The intelligent control method for a multi-condition ultra-high head single-stage centrifugal pump as described in claim 6, characterized in that, The performance parameters are converted into performance scores using a comprehensive performance scoring function, including: The performance parameters are normalized. The normalized performance parameters are used as input to the comprehensive performance scoring function to obtain the performance score corresponding to the performance parameters: ； in, This represents the normalized performance parameter vector, which includes the normalized actual head. Hydraulic efficiency Theoretical flow deviation Shaft power and the rate of change of operating resistance , Represents the normalized performance parameter vector The corresponding performance score All are rating coefficients.

8. The intelligent control method for a multi-condition ultra-high head single-stage centrifugal pump as described in claim 1, characterized in that, The optimal control vector corresponding to the state vector in the control space is selected using a reinforcement learning model, including: Construct a state vector as input to the reinforcement learning model: [current operating condition, performance parameters, and performance score of the ultra-high head single-stage centrifugal pump]; A control space C is constructed for an ultra-high head single-stage centrifugal pump, and the control space C includes the following three continuous control dimensions: The frequency conversion frequency of the inverter in the ultra-high head single-stage centrifugal pump ; Spindle speed ; Valve opening at the outlet of ultra-high head single-stage centrifugal pump ; in, To control the frequency conversion in space C The output range, This is the minimum output value of the frequency converter. This represents the maximum output value of the frequency converter. To control the spindle speed in space C The output range, This indicates the minimum output value of the spindle speed. This indicates the maximum output value of the spindle speed. To control the valve opening in space C The output range; A policy network in a reinforcement learning model is constructed using a policy gradient-based reinforcement learning method. The policy network takes a state vector as input and selects the optimal frequency, spindle speed, and valve opening from three continuous control dimensions corresponding to the state vector in the control space C to form an optimal control vector. The optimal control vector is then used to adjust the frequency, spindle speed, and valve opening of an ultra-high head single-stage centrifugal pump.

Images