Single-pump modeling and pump group collaborative optimization method and system of a pump station system

Abstract

The application belongs to the technical field of hydraulic machinery and intelligent control, and discloses a single-pump modeling and pump group collaborative optimization method and system of a pump station system, which comprises the following steps: obtaining single-pump operation parameters, constructing an exponential physical formula and a neural network model, embedding the exponential physical formula into the neural network model, and obtaining an E-PINN model; based on the E-PINN model, constructing a TS-PINN model comprising structural knowledge and trend knowledge and a pump group level loss function; taking the minimum total power as an optimization objective, determining a flow constraint condition and an efficiency constraint condition; and optimizing the E-PINN model, the TS-PINN model and a pump group control strategy by using a double-layer optimization architecture. The application realizes collaborative optimization control of multiple pump stations based on accurate characteristic curves and reduces the total energy consumption of the system.

Description

TECHNICAL FIELD

[0001] The present application relates to the field of hydraulic machinery and intelligent control technology, in particular to a single-pump modeling and pump group collaborative optimization method and system for a pump station system. BACKGROUND

[0002] In view of the core needs of pump station modeling and operation optimization in the scenes of water supply pump stations, industrial circulating pumps, central air conditioning cooling pumps, and urban water supply and distribution systems, the traditional methods have limitations in single-pump modeling accuracy and pump group collaborative energy efficiency, and there is an urgent need in the field to provide key technical support for intelligent operation of pump station systems.

[0003] Traditional single-pump modeling mainly adopts two methods: one is least squares fitting, which performs regression analysis on the collected Q-H data through linear or nonlinear fitting algorithms; the other is artificial neural network, which relies on data-driven capability to fit the Q-H characteristic curve without pre-setting physical formula structure.

[0004] The multi-pump parallel system often uses equal flow control or fixed start-stop control strategy, i.e., the system flow is evenly distributed according to the number of pumps, or the start-stop state of the pump is fixed according to the pre-set load threshold, lacking dynamic adaptation to the differences in single-pump characteristics and real-time working conditions.

[0005] In the prior art, physical information neural network (PINN) has been used for modeling of some industrial equipment, which improves modeling accuracy by embedding device-specific physical constraints, but its application in the field of pump stations is still limited to single-device modeling and has not been extended to pump group collaborative optimization scenarios.

[0006] In summary, the defects in the prior art are as follows: First, the single-pump modeling accuracy and physical consistency are insufficient. Although least squares fitting is simple to operate, it is prone to produce results that do not conform to physical laws, such as "head increase" and "curve passing through zero point". Artificial neural network has strong fitting capability, but it is easily disturbed by data noise and prone to characteristic distortion in data sparse or extrapolation scenarios, which cannot guarantee the physical rationality of the Q-H curve.

[0007] Secondly, the equal flow or fixed start-stop control strategy does not consider the differences in single-pump characteristics (such as different efficient regions of different pumps), which easily leads to some pumps running in the low-efficiency region for a long time, cannot achieve the optimal energy consumption of the whole system, and causes energy waste.

[0008] Finally, although PINN can improve generalization by physical constraints in device modeling, the existing methods do not fully utilize the exclusive physical knowledge such as "pump head-flow monotone decreasing, convex upward", and lack of technical extension from single-pump modeling to pump group collaborative optimization, which limits the modeling accuracy and system energy efficiency to the single-device level, and cannot support the optimization needs of pump station system level. SUMMARY

[0009] To solve the above problems, the present application provides a single pump modeling and pump group collaborative optimization method and system of a pump station system, the present application constructs a precise pump station (Q-H) characteristic curve fitting model based on an exponential physical formula, improves the fitting accuracy and physical consistency, realizes collaborative optimization control of multiple pump stations based on the precise characteristic curve, and reduces the total energy consumption of the system.

[0010] To achieve the above purpose, the present application adopts the following technical solutions:

[0011] In the first aspect, the present application provides a single pump modeling and pump group collaborative optimization method of a pump station system, comprising the following steps:

[0012] Obtain single pump operating parameters and preprocess the single pump operating parameters;

[0013] Based on the preprocessed single pump operating parameters, construct an exponential physical formula and a neural network model, and embed the exponential physical formula into the neural network model to obtain an E-PINN model, and train the E-PINN model by fusing a loss function of physical constraints and a Gaussian negative log-likelihood loss function;

[0014] Based on the E-PINN model, construct a TS-PINN model including structural knowledge and trend knowledge and a pump group level loss function;

[0015] Take the minimum total power as the optimization objective, determine the flow constraint condition and the efficiency constraint condition; optimize the E-PINN model, the TS-PINN model and the pump group control strategy by using a double-layer optimization architecture.

[0016] As an optional implementation, the single pump operating parameters include flow parameters, head parameters, power parameters and frequency parameters, and the preprocessing includes local anomaly filtering based on a confidence interval, standardization processing and incremental clustering processing.

[0017] As an optional implementation, the exponential physical formula constructed is expressed as:

[0018] ; wherein, is the head parameter in the single pump operating parameters, is the flow parameter in the single pump operating parameters, , , is a to-be-determined parameter.

[0019] As an optional implementation, the loss function fusing physical constraints includes a mean square error loss function, a monotonicity loss function and a curvature loss function; the loss function fusing physical constraints is expressed as: ​

[0020] ;in, To incorporate the loss function based on physical constraints, Let the mean squared error loss function be used. It is a monotonic loss function. Let curvature loss function be used. , This is a weighting factor.

[0021] As an alternative implementation, the Gaussian negative log-likelihood loss function is expressed as:

[0022] ;in, Let Gaussian negative log-likelihood loss function be used. This represents the total number of samples. For indexing, For the first The true value of each sample For the first The output prediction mean of each sample For the first The output prediction variance for each sample.

[0023] As an alternative implementation, the structural knowledge in the TS-PINN model includes parallel flow conservation, head consistency, and total power consumption, specifically expressed as follows:

[0024] ; ; ;

[0025] in, For total flow, For the first Flow parameters of the pump For the first The head parameters of the trolley pump Total power consumption, For the density of the medium, It is the acceleration due to gravity. For the first The efficiency of the pump;

[0026] Trend-based knowledge includes the monotonic increase in power consumption as flow rate and head increase, and the increase in total energy consumption when the load difference between adjacent pumps exceeds a threshold.

[0027] As an alternative implementation, the bottom layer of the dual-layer optimization architecture adopts the Adam-GC optimization algorithm and introduces gradient pruning and momentum adaptation. The upper layer of the dual-layer optimization architecture constructs a distributed cooperative control framework based on multi-agent reinforcement learning.

[0028] In a second aspect, the present application provides a single-pump modeling and pump group collaborative optimization system of a pump station system, comprising the following modules:

[0029] A parameter acquisition module configured to acquire single-pump operation parameters and pre-process the single-pump operation parameters;

[0030] A single-pump modeling module configured to construct an exponential physical formula and a neural network model based on the pre-processed single-pump operation parameters, embed the exponential physical formula into the neural network model, obtain an E-PINN model, and train the E-PINN model by fusing a loss function of a physical constraint and a Gaussian negative log-likelihood loss function;

[0031] A pump group modeling module configured to construct a TS-PINN model comprising structural knowledge and trend knowledge and a pump group level loss function based on the E-PINN model;

[0032] An optimization module configured to determine flow constraint conditions and efficiency constraint conditions with the minimum total power as an optimization objective, and optimize the E-PINN model, the TS-PINN model and a pump group control strategy by using a double-layer optimization architecture.

[0033] In a third aspect, the present application provides an electronic device comprising a memory and a processor, and computer instructions stored in the memory and running on the processor, when the computer instructions are run by the processor, the method of the first aspect is completed.

[0034] In a fourth aspect, the present application provides a computer readable storage medium for storing computer instructions, when the computer instructions are executed by a processor, the method of the first aspect is completed.

[0035] Compared with the prior art, the present application has the following beneficial effects:

[0036] The present application provides high-quality samples for modeling by the process of "pre-processing single-pump operation parameters, constructing and embedding an exponential physical formula, and training an E-PINN model by double loss functions", guarantees the physical reasonableness of the model from the structure by embedding the exponential physical formula into the neural network, and finally forces the model to meet the monotonicity and curvature constraints by fusing the loss function of the physical constraint, and improves the robustness of the model to noisy and sparse data by automatically learning the data noise characteristics through the modeling of the Gaussian distribution by the Gaussian negative log-likelihood loss function. The E-PINN model (single-pump model) finally trained can output physically consistent Q-H characteristic curves even in the data sparse and noisy scene, and the extrapolation can reach 0.97, completely avoids the unreasonable prediction of the traditional method, and significantly improves the single-pump modeling accuracy and physical consistency.

[0037] The present application is based on the accurate single pump characteristics output by the E-PINN model, and through "constructing a TS-PINN model, determining the optimization target and constraint, and optimizing the pump group control strategy by a double-layer optimization architecture", the TS-PINN model includes structural knowledge and trend knowledge, and accurately calculates the total power consumption of the pump group under different load distribution; the double-layer optimization architecture is used to optimize the E-PINN model, the TS-PINN model and the pump group control strategy, to dynamically distribute the load and adjust the frequency, and to avoid the long-term low-efficiency operation of part of the pumps. The total energy consumption of the pump group is reduced by 18%-22%, and the pressure fluctuation is reduced by 25%, solving the problem of low energy efficiency of the traditional control strategy, and realizing the energy efficiency optimization of the collaborative operation of the pump group.

[0038] Through the progressive design of "E-PINN single pump modeling - TS-PINN pump group modeling - double-layer optimization architecture linkage", the present application breaks through the expansion limitation, forms a whole-process technical chain of "single pump modeling - pump group modeling - system optimization", and completely solves the expansion problem of the traditional scheme "single device modeling and system optimization disconnection", providing complete technical support for the pump station system from single pump credible modeling to global optimal scheduling.

[0039] The advantages of the additional aspects of the present application will be partially given in the following description, partially will become obvious from the following description, or will be known by the practice of the present application. BRIEF DESCRIPTION OF DRAWINGS

[0040] The drawings accompanying the specification integrated into part of the present application are used to provide further understanding of the present application, the illustrative embodiments of the present application and the description thereof are used to explain the present application, and do not constitute improper limitation to the present application.

[0041] Figure 1 A flowchart of a single pump modeling and pump group collaborative optimization method of a pump station system provided for the present application embodiment 1;

[0042] Figure 2 A neural network structure diagram for the present application embodiment 1. DETAILED DESCRIPTION

[0043] The present application will be further described below in combination with the drawings and embodiments.

[0044] It should be pointed out that the following detailed description is exemplary and is intended to provide further explanation of the present application. Unless otherwise specified, all technical and scientific terms used herein have the same meaning as that generally understood by those skilled in the art to which the present application belongs.

[0045] It should be noted that the terminology used herein is for the purpose of describing particular embodiments only and is not intended to limit the exemplary embodiments of the invention. As used herein, unless the context clearly indicates otherwise, the singular form is intended to include the plural form as well. Furthermore, it should be understood that the terms "comprising" and "having," and any variations thereof, are intended to cover a non-exclusive inclusion, for example, a process, method, system, product, or apparatus that comprises a series of steps or units is not necessarily limited to those steps or units explicitly listed, but includes other steps or units not explicitly listed or inherent to such processes, methods, products, or apparatus.

[0046] Where there is no conflict, the embodiments and features in the embodiments of the present invention can be combined with each other.

[0047] Example 1

[0048] like Figure 1 As shown, this embodiment provides a method for single-pump modeling and pump group collaborative optimization of a pumping station system, including the following steps:

[0049] S1: Obtain the single pump operating parameters and preprocess them.

[0050] In step S1, parameter acquisition is online data acquisition and preprocessing, which is the core supporting link for realizing real-time modeling and dynamic optimization of the pump station system. Its essence is to replace the batch static processing in the traditional offline mode with an "incremental processing architecture driven by real-time data stream" to build a sample update mechanism that adapts to the dynamic changes in the pump station's operating conditions.

[0051] Obtaining single pump operating parameters involves the following steps:

[0052] S11. Real-time parameter acquisition: The edge-driven streaming access mechanism breaks through the limitations of the offline "full storage-batch processing" mode, and builds an online collection architecture of "nearby deployment of edge nodes + sliding window caching" to solve the problems of remote transmission latency and massive storage.

[0053] S12, Edge Localized Reception: Based on industrial protocols such as Modbus TCP / IP and Profinet, it realizes direct connection between sensors and edge nodes, and obtains flow parameters (Q), head parameters (H), power parameters (P), and frequency parameters (f) at a sampling frequency of 1Hz. The end-to-end delay is ≤50ms, which meets the industrial-grade real-time requirements and avoids the network jitter and bandwidth limitation problems of cloud transmission.

[0054] S13, sliding window cache: a fixed time window (one data point per second) is used to construct an "first-in first-out" cache pool. When new data is accessed, the oldest historical data is automatically removed to ensure that the cache data is a "recent working condition sample set", which not only reduces the total storage resource consumption, but also provides a data source that fits the real-time state for subsequent processing.

[0055] In step S1, preprocessing of single-pump operating parameters includes local anomaly filtering based on confidence interval, standardization processing, and incremental clustering processing, specifically:

[0056] Based on local anomaly filtering based on confidence interval: the mean and standard deviation of parameters in the sliding window are used to construct a confidence interval. If the new data exceeds the interval, it is determined to be abnormal, such as sensor pulse interference. Relying on "recent working condition local distribution" judgment, historical abnormal interference is avoided to ensure data effectiveness.

[0057] Standardization processing: when the window is updated, the mean and standard deviation of Q and H are recalculated simultaneously. This is used as a basis for dimensionless, real-time adaptation to working condition changes, eliminating the influence of dimension differences on clustering and improving sample consistency.

[0058] Incremental clustering: iterative optimization sample update mechanism, three-level incremental architecture + low complexity calculation, breaking the "one-time generation of fixed samples" mode of offline full K-means, constructing a "initial template-single sample iteration-periodic calibration" three-level incremental architecture to balance efficiency and accuracy: in the startup phase, 100 cluster centers (covering low / rated / high load working conditions) generated offline are used as the initial template to avoid deviations caused by insufficient real-time data; sample-model linkage breaks through the limitations of offline "one-way sample input", and a "sample update-model parameter adjustment" real-time closed-loop interaction mechanism is constructed: after updating the cluster center, 100 representative samples (flow parameter (Q), head parameter (H), power parameter (P), frequency parameter (f) parameter) are pushed to the E-PINN model in real time. The model adjusts the parameters in small batches online to avoid full retraining to ensure synchronization with working conditions. At the same time, the original data, preprocessed data, and cluster center records are stored in InfluxDB / TimescaleDB according to the timestamp. Relying on the columnar storage and time indexing of the time series database, it supports working condition backtracking and fault troubleshooting.

[0059] S2: based on the preprocessed single-pump operating parameters, construct an exponential physical formula and a neural network model, and embed the exponential physical formula in the neural network model to obtain an E-PINN model. The E-PINN model is trained by fusing a loss function with physical constraints and a Gaussian negative log-likelihood loss function.

[0060] In step S2, as Figure 2The constructed exponential physical formula is embedded into a neural network (PINN), combining data-driven and physical constraints to achieve accurate fitting. The constructed exponential physical formula is expressed as:

[0061] ; wherein, is the head parameter in the single pump operating parameter, is the flow parameter in the single pump operating parameter, , , is the to-be-determined parameter.

[0062] The constructed neural network is used to approximate the three to-be-determined parameters. The neural network takes the flow Q and head H in the pump station operating parameter as input, and outputs the values of , , . In this way, the structure of the physical formula is retained, and the neural network can adapt to different working conditions.

[0063] In step S2, the E-PINN model is based on the ordinary PINN, and adds the "monotonicity + curvature" physical constraint and noise adaptive estimation, so that the model can accurately fit the Q-H characteristics of the pump in the case of sparse data and noise. Traditional PINN only uses differential equation residuals and does not use the engineering common sense that "pump head monotonically decreases with flow and is convex upward";

[0064] First-order derivative constraint:

[0065] ;

[0066] Second-order derivative constraint:

[0067] ;

[0068] The loss function that fuses physical constraints includes the mean square error loss function, the monotonicity loss function, and the curvature loss function. The loss function that fuses physical constraints is expressed as:

[0069] ;

[0070] wherein, is the loss function that fuses physical constraints, is the mean square error loss function, which is used to fit data points and ensure that the model can "learn data"; is the monotonicity loss function, which ensures that the head monotonically decreases with the increase of flow; is the curvature loss function, which ensures that the curve is convex overall; , are weighting coefficients.

[0071] In this embodiment, and can be activated step by step to avoid oscillation at the beginning of training. At the beginning of training = 0, and after stabilization > 0, the physical term is activated step by step to realize a two-stage learning strategy of "fitting first and then correcting" to ensure stable training.

[0072] In step S2, at each training epoch, a batch of fake data is generated. To generate fake data, first, random values are sampled from the effective domain of other input features to fix them, and then a target feature sequence is generated.

[0073] Considering data noise, when generating the sequence, the interval or distance between two adjacent points should be greater than the sensor noise. After generating the fake data, the fake data and the historical data are fed into the neural network to minimize the loss function. The historical data is different from the standard neural network for this T-PINN (trend (T)). The conditional output is modeled as a Gaussian distribution ( ), rather than a single-point prediction. Therefore, there are two output neurons in the neural network, representing the mean and the mean square error.

[0074] To learn this distribution, Gaussian negative likelihood (GNLL) is used as the data loss. The Gaussian negative log-likelihood loss function is represented as:

[0075] ; wherein, is the Gaussian negative log-likelihood loss function, is the total number of samples, is the index, is the true value of the th sample, is the output predicted mean of the th sample, is the output predicted variance of the th sample. The neural network can automatically learn the mean and the mean square error from the historical data, fit the observed values through the Gaussian negative log-likelihood loss function, and improve the robustness under abnormal data and drift data.

[0076] In training, the historical data and the fake data are given complementary statistical-physical roles to jointly form a "dirty-clean collaborative" sample space: the former is the noisy observation collected by the sensor, carrying the likelihood information of the real working condition, used to estimate the physical parameters and heteroscedasticity; the latter is a disturbance-free sequence generated according to the principle of "adjacent interval > noise amplitude" based on the historical distribution edge and the upper limit of the sensor noise, which is specifically used to strengthen the monotonicity and curvature of the macroscopic physical constraints. Under mixed training, the network automatically amplifies the In the sparse domain covered by pseudo-data, clean samples are used to correct the error. This achieves statistical-physical dual optimality of "data-driven fitting" and "physical law conservation", significantly improving the interpretability and extrapolation robustness of the model outside the sparse domain.

[0077] In this embodiment, K-means clustering is also used to extract representative data points, reducing the training scale by more than 80%; gradient clipping and early stopping mechanisms are set to improve convergence speed and stability, specifically:

[0078] Gradient clipping and early stopping control:

[0079] 1. Before each backpropagation, calculate the gradient norm of the loss function with respect to the network parameters; if the norm exceeds a preset threshold, scale the gradient to the threshold range according to the norm ratio to achieve gradient clipping and prevent gradient explosion.

[0080] 2. Synchronous monitoring of validation set loss: When the loss value does not decrease within 200 consecutive iterations, training is immediately terminated and the weights are rolled back to the historical optimal weights, thus completing early stopping.

[0081] Gradient clipping and early stopping control improve the model convergence speed by 30%–40%, effectively suppress curve oscillations caused by overtraining, and ensure physical consistency and numerical stability in the extrapolation region.

[0082] S3: Based on the E-PINN model, construct the TS-PINN model, which includes structural knowledge and trend knowledge, and the pump group-level loss function.

[0083] In step S3, the TS-PINN (Trend+Structure Physics-Informed Neural Network) model is a type of physical information neural network that combines structural knowledge (S) and trend knowledge (T), enabling the model to both "abide by physical laws" and "conform to engineering experience," thereby exhibiting higher stability and extrapolation ability in actual engineering modeling and optimization.

[0084] The structural knowledge in the TS-PINN model includes parallel flow conservation, head consistency, and total power consumption, specifically expressed as follows:

[0085] ; ; ;

[0086] in, For total flow, For the first Flow parameters of the pump For the first The head parameters of the trolley pump Total power consumption For the density of the medium, It is the acceleration due to gravity. For the first The efficiency of the pump;

[0087] Trend-based knowledge includes the monotonic increase in power consumption as flow rate and head increase, and the increase in total energy consumption when the load difference between adjacent pumps exceeds a threshold.

[0088] In addition to the data fitting term, the designed pump-stage loss function also incorporates a trend constraint term, expressed as:

[0089] ;

[0090] in, For the pump set level loss function, This is the data loss function, used to measure the error between the model's predicted values ​​and the actual sampled data; This is a structural physical constraint loss function used to constrain the model to satisfy the basic physical laws of the pumping station. This is a trend-based constraint loss function used to embed engineering empirical patterns (monotonicity / trend). , These are weighting coefficients used to balance the effects of the three types of losses. Used to adjust the importance of structural constraints. The importance of using trend-based constraints.

[0091] In actual training, , The constraints can be gradually increased; that is, first let the model fit the data, and then gradually introduce physical constraints to avoid initial oscillations.

[0092] S4: Using the minimum total power as the optimization objective, determine the flow constraints and efficiency constraints; use a two-layer optimization architecture to optimize the E-PINN model, TS-PINN model, and pump control strategy.

[0093] In step S4, the optimization objective is to minimize the total power while meeting the operating requirements.

[0094] The constraints include flow constraints and efficiency constraints, specifically:

[0095] Flow constraints: This means ensuring that the total output traffic of the system is not less than the traffic demanded by the user.

[0096] Efficiency constraints: This means that the operating efficiency of each pump must not be lower than the set lower limit (80%), in order to avoid long-term inefficient operation of the pump.

[0097] In step S4, the pump set control strategy is a specific operation scheme for actually controlling the operation of the pump set based on the energy consumption distribution (modeling results of the TS-PINN model), which clearly specifies "which pump to start / stop and each pump to run at what frequency (or flow rate)" to achieve the goal of "minimum total power".

[0098] In the intelligent process of the pump station system, two core challenges are faced: first, the physical information neural network is extremely unstable in high-dimensional input training due to its complex multi-objective loss function, and there are problems of gradient explosion and convergence shock, which seriously restrict the modeling efficiency; second, the multi-pump parallel system lacks coordination during operation, and the traditional independent control strategy is difficult to achieve global energy consumption optimization while meeting the flow demand, which easily leads to low equipment efficiency and system fluctuation.

[0099] Therefore, in step S4, a double-layer optimization architecture is designed to realize the whole-process closed loop from accurate and rapid modeling to distributed collaborative control, wherein the bottom layer of the double-layer optimization architecture adopts the Adam-GC optimization algorithm, and gradient clipping and momentum adaptation are introduced, and the upper layer of the double-layer optimization architecture builds a distributed collaborative control framework based on multi-agent reinforcement learning.

[0100] The bottom layer of the double-layer optimization architecture adopts an improved Adam-GC optimization algorithm, which introduces dual mechanisms of gradient clipping (Gradient Clipping) and momentum adaptation (Momentum Adaptation) on the basis of the traditional Adam-GC algorithm. Gradient clipping is: before each back propagation, the gradient norm of the corresponding loss function relative to the parameters of the E-PINN model and the TS-PINN model is calculated; if the norm exceeds the preset threshold (set according to the engineering scene, such as 10 or 20), the gradient is scaled to the threshold range in proportion to the norm, effectively suppressing the gradient mutation caused by "structural constraint terms (such as flow conservation)", and avoiding the divergence of model parameters caused by gradient explosion.

[0101] Momentum adaptation dynamically adjusts the hyperparameters (reduces momentum to accelerate convergence at the beginning of training, and increases momentum to stabilize optimization in the middle and later stages), balancing training speed and stability.

[0102] The training speed of the E-PINN model and the TS-PINN model is improved by about 35%, and the convergence curve is smooth, providing a high-precision, physically consistent reliable model for the upper layer control.

[0103] The upper layer of the double-layer optimization architecture builds a distributed collaborative control framework based on multi-agent reinforcement learning (MARL). In this framework, each pump station is regarded as an agent. The state space of the agent integrates real-time monitoring data and predicted states output by the bottom model. The action space is the start-stop and frequency adjustment of the pump. The decision goal is guided by a well-designed reward function, with the core of minimizing the total power consumption, while considering the flow tracking accuracy, single-pump efficiency constraints and pipe network pressure stability. In this embodiment, the MADDPG algorithm is used to make each agent learn to cooperate by relying only on local observations through centralized training and distributed execution, and finally output the globally optimal control instructions.

[0104] The bottom improved Adam-GC optimization algorithm ensures the efficient and stable training of the E-PINN model and the TS-PINN model, providing an accurate digital twin for the system. The upper MARL collaborative control strategy utilizes the model to achieve dynamic load distribution and energy efficiency optimization of the pump group. The test shows that this scheme not only improves the modeling and training efficiency by 35%, but also reduces the total energy consumption of the system by 18%-22% and the pressure fluctuation by 25% at the control level, completely solving the chain problem from "modeling difficulty" to "control sub-optimization" and laying the key technical foundation for the intelligentization of the pump station system.

[0105] Embodiment 2

[0106] The embodiment provides a single-pump modeling and pump group collaborative optimization system for a pump station system, comprising the following modules:

[0107] The parameter acquisition module is configured to acquire single-pump operating parameters and pre-process the single-pump operating parameters;

[0108] The single-pump modeling module is configured to construct an exponential physical formula and a neural network model based on the pre-processed single-pump operating parameters, embed the exponential physical formula into the neural network model to obtain an E-PINN model, and train the E-PINN model by fusing a loss function of physical constraints and a Gaussian negative log-likelihood loss function;

[0109] The pump group modeling module is configured to construct a TS-PINN model including structural knowledge and trend knowledge and a pump group level loss function based on the E-PINN model;

[0110] The optimization module is configured to determine flow constraint conditions and efficiency constraint conditions with the minimum total power as the optimization target, and optimize the E-PINN model, the TS-PINN model and the pump group control strategy by using the double-layer optimization architecture.

[0111] It should be noted that the above modules correspond to the steps in Embodiment 1, and the above modules have the same examples and application scenarios as the steps they correspond to, but are not limited to the content disclosed in Embodiment 1. It should be noted that the above modules can be executed in a computer system as part of a system.

[0112] In more embodiments, there are also provided:

[0113] An electronic device comprising a memory and a processor and computer instructions stored on the memory and running on the processor, when the computer instructions are run by the processor, the method in Embodiment 1 is completed. For brevity, it will not be repeated here.

[0114] It should be understood that in the embodiments, the processor can be a central processing unit CPU, and the processor can also be other general-purpose processors, digital signal processors DSPs, application-specific integrated circuits ASICs, field programmable gate arrays FPGA or other programmable logic devices, discrete gates or transistor logic devices, discrete hardware components, etc. The general-purpose processor can be a microprocessor or any conventional processor, etc.

[0115] A computer-readable storage medium for storing computer instructions, when the computer instructions are executed by a processor, the method in Embodiment 1 is completed.

[0116] The method in Embodiment 1 can be directly executed by a hardware processor, or executed by a combination of hardware and software modules in the processor. The software modules can be located in random access memory, flash memory, read-only memory, programmable read-only memory, or electrically erasable programmable memory, registers, etc. The storage medium is located in the memory, and the processor reads the information in the memory and combines the hardware to complete the steps of the above method. To avoid repetition, it will not be described in detail here.

[0117] A computer program product comprising a computer program, the computer program being executed by a processor to implement the method in Embodiment 1.

[0118] The present application also provides at least one computer program product tangibly stored on a non-transitory computer readable storage medium. The computer program product includes computer executable instructions, such as instructions included in program modules, which are executed in devices on a target real or virtual processor to perform processes / methods as described above. Generally, program modules include routines, programs, libraries, objects, classes, components, data structures, etc. that perform particular tasks or implement particular abstract data types. In various embodiments, the functions of the program modules can be combined or divided as desired. Machine executable instructions for program modules can be executed within a local or distributed device. In a distributed device, program modules can be located in local and remote storage media.

[0119] Computer program code for carrying out operations of the present application can be written in any combination of one or more programming languages. The computer program code can execute entirely on a computer, a special purpose computer, or other programmable apparatus to produce the functions / acts specified in the flow diagrams and / or block diagrams. The program code can execute entirely on a computer, a special purpose computer, or other programmable apparatus, as a stand-alone software package, partly on the computer and partly on a remote computer, or entirely on the remote computer or server.

[0120] In the context of the present application, the computer program code or related data can be carried by any suitable carrier, to enable the device, apparatus or processor to perform the various processes and operations described above. Examples of carriers include signals, computer readable media, and the like. Examples of signals can include electrical, optical, radio, sound or other forms of propagated signals, such as carrier waves, infrared signals, and the like.

[0121] Those skilled in the art can realize that the units and algorithm steps of the examples described in conjunction with the present embodiments can be realized in electronic hardware or a combination of computer software and electronic hardware. Whether the functions are realized in hardware or software depends on the specific application and design constraints of the technical solution. Those skilled in the art can use different methods to realize the described functions for each specific application, but such implementation should not be considered beyond the scope of the present embodiments.

[0122] The above describes the specific embodiments of the present application in conjunction with the accompanying drawings, but is not a limitation on the scope of protection of the present application. Those skilled in the art should understand that various modifications or variations made by those skilled in the art on the basis of the technical solutions of the present application without inventive labor are still within the scope of protection of the present application.

Claims

1. A method for single pump modeling and pump group collaborative optimization of a pump station system, characterized in that, The method comprises the following steps: acquiring single-pump operation parameters, and preprocessing the single-pump operation parameters, wherein the single-pump operation parameters comprise flow parameters, head parameters, power parameters and frequency parameters; Based on the pretreated single-pump operation parameters, an exponential physical formula and a neural network model are constructed, the exponential physical formula is embedded into the neural network model, an E-PINN model is obtained, and the E-PINN model is trained through a loss function of physical constraints and a Gaussian negative log-likelihood loss function, and the pretreatment includes local anomaly filtering, standardization processing and incremental clustering processing based on confidence intervals constructing a TS-PINN model and a pump group level loss function comprising structural knowledge and trend knowledge based on an E-PINN model; determining flow constraint conditions and efficiency constraint conditions with the minimum total power as an optimization objective; optimizing the E-PINN model, the TS-PINN model and the pump group control strategy by using a double-layer optimization architecture.

2. The single pump modeling and pump group collaborative optimization method of a pump station system according to claim 1, wherein, The exponential physical formula is constructed as: ; wherein, is the head parameter in the single pump operating parameter, is the flow parameter in the single pump operating parameter, , , is the parameter to be determined.

3. The single pump modeling and pump group collaborative optimization method of a pump station system of claim 1, wherein, The loss function fused with physical constraints comprises a mean square error loss function, a monotonicity loss function and a curvature loss function; and the loss function fused with physical constraints is expressed as: ; wherein, is a loss function incorporating physical constraints, is a mean squared error loss function, is a monotonicity loss function, is a curvature loss function, , is a weighting coefficient.

4. The single pump modeling and pump group collaborative optimization method of a pump station system of claim 1, wherein, The Gaussian negative log-likelihood loss function is expressed as: ; wherein, is a Gaussian negative log-likelihood loss function, is the total number of samples, is an index, is a true value of the th sample, is an output prediction mean of the th sample, is an output prediction variance of the th sample.

5. The single pump modeling and pump group collaborative optimization method of a pump station system of claim 1, wherein, The structural knowledge in the TS-PINN model comprises parallel flow conservation, consistent head and total power consumption, and is specifically expressed as: ； ； ； wherein, Qtotal is the total flow, Q1 is the flow parameter of the first pump, Qn is the flow parameter of the nth pump, H1 is the head parameter of the first pump, Hn is the head parameter of the nth pump, Ptotal is the total power consumption, p is the medium density, g is the gravitational acceleration, η1 is the efficiency of the first pump, ηn is the efficiency of the nth pump. The trend knowledge comprises that the power consumption monotonically increases when the flow and head increase, and the total energy consumption rises when the load difference of adjacent pumps exceeds a threshold.

6. The single pump modeling and pump group collaborative optimization method of a pump station system of claim 1, wherein, The bottom layer of the double-layer optimization architecture adopts an Adam-GC optimization algorithm, and gradient clipping and momentum self-adaptation are introduced; and the top layer of the double-layer optimization architecture constructs a distributed collaborative control framework based on multi-agent reinforcement learning.

7. A single pump modeling and pump group collaborative optimization system of a pump station system, characterized in that, The method comprises the following modules: a parameter acquisition module configured to acquire single-pump operation parameters, and preprocess the single-pump operation parameters, wherein the single-pump operation parameters comprise flow parameters, head parameters, power parameters and frequency parameters; The single-pump modeling module is configured to: based on the preprocessed single-pump operation parameters, construct an exponential physical formula and construct a neural network model, embed the exponential physical formula in the neural network model to obtain an E-PINN model, and train the E-PINN model by fusing a loss function of a physical constraint and a Gaussian negative log-likelihood loss function, wherein the preprocessing includes local anomaly filtering, standardization processing and incremental clustering processing based on confidence intervals. a pump group modeling module configured to construct a TS-PINN model and a pump group level loss function comprising structural knowledge and trend knowledge based on an E-PINN model; an optimization module configured to determine flow constraint conditions and efficiency constraint conditions with the minimum total power as an optimization objective; optimizing the E-PINN model, the TS-PINN model and the pump group control strategy by using a double-layer optimization architecture.

8. An electronic device, comprising: A computer readable storage medium storing computer instructions, wherein the computer instructions are executed by a processor to implement the single-pump modeling and pump group collaborative optimization method of the pump station system according to any one of claims 1-6.

9. A computer-readable storage medium, characterized in that, A computer readable storage medium storing computer instructions, wherein the computer instructions are executed by a processor to implement the single-pump modeling and pump group collaborative optimization method of the pump station system according to any one of claims 1-6.

Images