Multi-objective collaborative optimization pipe network automatic transmission and distribution control method

Abstract

The invention discloses a multi-objective collaborative optimization pipe network automatic transmission and distribution control method, and belongs to the technical field of process control and system automation, and the method comprises the steps: obtaining pipe network operation data, generating robust features through adversarial domain adaptation training, monitoring environment mutation in real time, building a dynamic multi-objective optimization model through the robust features, and carrying out the optimization of the dynamic multi-objective optimization model. And generating a Pareto solution set by adopting a dynamic multi-objective evolutionary algorithm, and finally selecting an optimal control solution in combination with a collaborative balance index and a preference function. According to the method, data field differences are eliminated by adopting an antagonistic domain adaptation technology, robust features insensitive to working condition changes are obtained, and an optimal equilibrium solution among energy consumption, pressure safety and service quality can be tracked and determined in real time when a pipe network load and an operating environment dynamically change in combination with a dynamic multi-objective evolutionary algorithm responding to environment mutation, so that the service quality of the pipe network is improved. And the real-time performance, the robustness and the multi-target collaborative optimization level of automatic transmission and distribution control of the pipe network are improved.

Description

TECHNICAL FIELD

[0001] The present application relates to the field of process control and system automation, in particular to a multi-objective collaborative optimization method for automatic distribution control of pipe network. BACKGROUND

[0002] The automatic distribution control system of pipe network is an important part of modern urban infrastructure, widely used in water supply, heat supply, gas supply and other fields. This kind of system can accurately control the pump station, valve and other actuators in the pipe network, so as to effectively manage the key operation parameters such as pressure and flow, and ensure the safe, efficient and economic operation of the system. The core is to generate and execute the optimal control strategy according to the real-time state of the pipe network, which belongs to the technical category of adjusting and controlling non-electric variables.

[0003] In the prior art, the automatic distribution control of pipe network usually relies on optimization algorithm based on hydraulic model. The control system first calibrates the hydraulic model by using historical operation data, and then combines the real-time collected operation data to calculate the best setting value of the control device by setting the optimization target such as the lowest energy consumption or the most stable pressure, using mathematical optimization algorithm. Some advanced methods will use multi-objective optimization technology to consider multiple targets such as energy consumption and pressure, and provide a set of compromise control schemes for decision makers.

[0004] However, the existing control method has some inherent technical defects. First, the model trained based on historical data will have performance degradation when facing real-time changing working conditions due to inconsistent data distribution, and the model has insufficient generalization ability and robustness. Secondly, most of the traditional control strategies are static, which is difficult to quickly and effectively respond to sudden events occurring in the pipe network, such as pipe burst or sudden increase of user demand, and lack of adaptability to dynamic environment. Finally, when dealing with multi-objective problems, the existing method often has difficulty in selecting a single optimal solution from a large number of compromise solutions that best meet the current actual operation demand and are balanced among various targets, making the decision-making process subjective and blind. SUMMARY

[0005] To solve the above problems, the present application provides a multi-objective collaborative optimization method for automatic distribution control of pipe network, which adopts the technical means of combining adversarial domain adaptation training, dynamic multi-objective evolutionary algorithm and collaborative balance evaluation, and can overcome the difference in data distribution, quickly respond to environmental mutations, and realize dynamic collaboration and balance control of multiple operation targets.

[0006] To achieve the above purpose, the present application adopts the following technical scheme: In a first aspect, a multi-objective collaborative optimization method for automatic distribution control of pipe network is provided, comprising: S1, acquire real-time running parameters, device states and external environment data of the pipe network as real-time state data, define the real-time state data as target domain data, acquire historical running data, historical energy consumption data and historical load data of the pipe network and define as source domain data; S2, based on the source domain data and the target domain data, eliminate the difference between the source domain data and the target domain data by performing adversarial domain adaptation training, generate robust features capable of representing the essential law of pipe network operation, and acquire domain adaptation error in the training process; S3, real-time monitor the change rate of the target domain data and the domain adaptation error, and generate an environmental mutation judgment result; S4, use the robust features to construct a dynamic multi-objective optimization model containing decision variables and multi-objective functions; S5, based on the dynamic multi-objective optimization model, respond to the environmental mutation judgment result, and use a dynamic multi-objective evolutionary algorithm to generate a Pareto solution set containing a plurality of non-inferior solutions; S6, perform target balance evaluation on the non-inferior solutions in the Pareto solution set, and calculate the synergistic balance index between the solutions in the multi-objective functions; S7, acquire a preference function representing the current operation preference, select an optimal control solution from the Pareto solution set in combination with the synergistic balance index, and generate a control instruction according to the optimal control solution and issue it to the execution mechanism.

[0007] Based on the above technical solution, in the pipe network automatic delivery and control method provided in the application, the technical means of combining adversarial domain adaptation training, dynamic multi-objective evolutionary algorithm and synergistic balance evaluation can overcome data distribution differences, quickly respond to environmental mutations, and realize dynamic synergistic and balanced control of multiple operation targets.

[0008] In combination with the first aspect, in a possible implementation manner, the generating robust features capable of representing the essential law of pipe network operation includes: performing feature extraction on the source domain data and the target domain data to generate control features, and performing domain classification judgment on the data source of the control features to generate domain classification loss; by alternately optimizing the main task loss of the feature extraction and the domain classification loss of the domain classification, the feature extraction is trained so that the feature extraction can optimize the domain classification error; the feature extraction after training convergence is applied to real-time data, and the domain-invariant features output as robust features.

[0009] In a possible implementation manner of the first aspect, the real-time monitoring of the change rate and the domain adaptation error of the target domain data to generate the environment mutation judgment result comprises: calculating a change amount of the target domain data in consecutive time steps to generate a change rate index; comparing the change rate index and the domain adaptation error with a threshold value representing a normal operation fluctuation range to generate a comparison result; and according to the comparison result, generating an environment mutation judgment result identifying that the environment has mutated when the change rate index or the domain adaptation error exceeds the threshold value.

[0010] In a possible implementation manner of the first aspect, the constructing of the dynamic multi-objective optimization model containing the decision variable and the multi-objective function comprises: defining a decision variable set containing a control device operation parameter; establishing a multi-objective function for quantifying energy consumption cost, pipe network pressure safety and service quality tracking respectively; and integrating the robust feature as a dynamic input parameter into the multi-objective function for real-time adjustment of parameters and model constraints of the multi-objective function to form the dynamic multi-objective optimization model.

[0011] In a possible implementation manner of the first aspect, based on the dynamic multi-objective optimization model, in response to the environment mutation judgment result, the dynamic multi-objective evolutionary algorithm is used to generate a Pareto solution set containing a plurality of non-inferior solutions, which comprises: generating an initial population by using a Pareto solution set of a previous control period or a prediction result based on the robust feature; if the environment mutation judgment result is yes, adjusting search parameters of evolutionary operation to generate a child population by performing evolutionary operation on the initial population; and evaluating the initial population and the child population by using the dynamic multi-objective optimization model, and generating the Pareto solution set by non-dominated sorting.

[0012] In a possible implementation manner of the first aspect, the target balance evaluation of the non-inferior solutions in the Pareto solution set to calculate a coordination balance index of each solution between the multi-objective functions comprises: determining ideal points of the Pareto solution set on each target dimension, and calculating target deviation values of each non-inferior solution relative to the ideal points; performing normalization processing on the target deviation values to eliminate dimensional influence to generate normalized deviation values; and calculating a dispersion degree of the normalized deviation values of each non-inferior solution, and taking the dispersion degree as the coordination balance index.

[0013] In a possible implementation manner of the first aspect, the acquiring the preference function for representing the current operation preference, and selecting the optimal control solution from the Pareto solution set in combination with the coordination balance degree indicator comprises: acquiring working condition information reflecting a current working condition of the pipe network, and determining target weights in the preference function according to the working condition information; applying the preference function, calculating a comprehensive evaluation value of each non-inferior solution in the Pareto solution set in combination with the target weights and the coordination balance degree indicator; and selecting a non-inferior solution with an optimal comprehensive evaluation value as the optimal control solution.

[0014] In a possible implementation manner of the first aspect, the method further comprises: after the control instruction is issued, collecting control-after-pipe-network-operation data generated by the control instruction; updating the source domain data by using the control-after-pipe-network-operation data; and adjusting the adversarial domain adaptation training based on the updated source domain data, to realize continuous optimization of the model.

[0015] In a possible implementation manner of the first aspect, the method further comprises: analyzing an optimal control solution and a corresponding pipe network operation result in a historical control period, to generate a performance stability degree indicator; comparing the performance stability degree indicator with a threshold value for defining a performance stability interval, to generate a period adjustment decision; and dynamically adjusting a running period length of the dynamic multi-objective evolutionary algorithm according to the period adjustment decision.

[0016] The second aspect provides a multi-objective coordinated optimization pipe network automatic delivery and distribution control system, comprising: a data acquisition module, a domain adaptation training module, an environment monitoring module, an optimization model construction module, an optimization decision module, a balance evaluation module, and an optimal solution selection and instruction generation module; wherein: The data acquisition module is configured to acquire real-time running parameters, device states, and external environment data of the pipe network as real-time state data, define the real-time state data as target domain data, and acquire historical running data, historical energy consumption data, and historical load data of the pipe network and define the historical running data, the historical energy consumption data, and the historical load data as source domain data; The domain adaptation training module is configured to generate robust features capable of representing the nature law of the pipe network operation by performing adversarial domain adaptation training based on the source domain data and the target domain data, and acquire a domain adaptation error of the training process; The environment monitoring module is configured to monitor a change rate of the target domain data and the domain adaptation error in real time, and generate an environment mutation determination result; The optimization model construction module is configured to construct a dynamic multi-objective optimization model comprising decision variables and multi-objective functions by using the robust features; The optimization decision module is configured to generate a Pareto solution set comprising a plurality of non-inferior solutions by using a dynamic multi-objective evolutionary algorithm based on the dynamic multi-objective optimization model and in response to the environment mutation determination result; The balance evaluation module is configured to evaluate the target balance of the non-inferior solutions in the Pareto solution set, and calculate a synergistic balance index of each solution between multiple objective functions. The optimal solution selection and instruction generation module is configured to obtain a preference function representing a current operation preference, select an optimal control solution from the Pareto solution set in combination with the synergistic balance index, and generate a control instruction for the execution mechanism according to the optimal control solution.

[0017] Compared with the prior art, the application has the following advantages: The application effectively eliminates the distribution difference between historical data and real-time data by introducing adversarial domain adaptation training, and generates robust features that can represent the essential laws of pipe network operation. This makes the control model not affected by data drift caused by pipe network aging, seasonal replacement and other factors, improves the generalization ability and prediction accuracy of the model under complex and variable working conditions, and ensures the long-term effectiveness and reliability of the control strategy.

[0018] The application establishes a sensitive environmental mutation monitoring and rapid response mechanism, which can accurately identify sudden changes in the operation state of the pipe network by monitoring the data change rate and model adaptation error in real time. In the event of environmental mutation, the search strategy of the dynamic multi-objective evolutionary algorithm can be automatically adjusted to quickly adapt to the new operating environment and find a new optimal control interval, ensuring the stability and adaptability of the control system in dynamic and uncertain environments.

[0019] The application proposes a multi-dimensional decision-making method that takes into account operation preference and target balance. Not only does it generate a series of non-inferior solutions, but it also quantifies the balance of each solution between multiple objectives through the calculation of a synergistic balance index, and conducts comprehensive evaluation in combination with a preference function that changes dynamically with the working condition. This makes the final selected optimal control solution not only meet the current operation focus, but also avoid sacrificing other key performance due to excessive emphasis on a certain target, achieving true synergy and balance between objectives.

[0020] The application designs a closed-loop self-learning and adaptive optimization mechanism. By feeding back the actual operation data after control for model updating, the application realizes continuous self-evolution of the control method, enabling it to automatically adapt to long-term slow changes in pipe network characteristics. At the same time, by analyzing the performance stability to dynamically adjust the optimization period, the application realizes intelligent allocation of computing resources, reduces system operating costs on the premise of ensuring control accuracy, and achieves an optimal balance between performance and efficiency.

[0021] It should be understood that the descriptions of technical features, technical solutions, beneficial effects, or similar language in this application do not imply that all features and advantages can be achieved in any single embodiment. Rather, it is understood that the description of a feature or beneficial effect means that a specific technical feature, technical solution, or beneficial effect is included in at least one embodiment. Therefore, the descriptions of technical features, technical solutions, or beneficial effects in this specification do not necessarily refer to the same embodiment. Furthermore, the technical features, technical solutions, and beneficial effects described in this embodiment can be combined in any suitable manner. Those skilled in the art will understand that embodiments can be implemented without one or more specific technical features, technical solutions, or beneficial effects of a particular embodiment. In other embodiments, additional technical features and beneficial effects may be identified in specific embodiments that do not embody all embodiments. Attached Figure Description

[0022] To more clearly illustrate the technical solutions in the embodiments of this application or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are some embodiments of this application. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0023] Figure 1 A system architecture diagram of a multi-objective collaborative optimization pipeline automatic transmission and distribution control system provided in this application embodiment; Figure 2 A flowchart illustrating a multi-objective collaborative optimization automatic pipeline distribution control method provided in this application embodiment; Figure 3 This is a diagram illustrating the effect of adversarial domain adaptation training provided in an embodiment of this application.

[0024] Figure 4 This is a schematic diagram of the dynamic Pareto solution set responding to environmental mutations provided in the embodiments of this application. Detailed Implementation

[0025] In the description of this application, unless otherwise stated, " / " means "or," for example, A / B can mean A or B. The "and / or" in this document is merely a description of the relationship between related objects, indicating that three relationships can exist. For example, A and / or B can represent: A alone, A and B simultaneously, and B alone. Furthermore, "at least one" means one or more, and "multiple" means two or more. The terms "first," "second," etc., do not limit the quantity or order of execution, and "first," "second," etc., do not necessarily imply differences.

[0026] It should be noted that, in this application, the terms "exemplary" or "for example" are used to indicate that something is being described as an example, illustration, or illustration. Any embodiment or design described as "exemplary" or "for example" in this application should not be construed as being more preferred or advantageous than other embodiments or design solutions. Specifically, the use of terms such as "exemplary" or "for example" is intended to present the relevant concepts in a concrete manner.

[0027] The multi-objective collaborative optimization pipeline automatic transmission and distribution control method provided in this application embodiment can be applied to, for example... Figure 1 In the multi-objective collaborative optimization pipeline automatic transmission and distribution control system 100 shown, such as Figure 1 As shown, the system includes: a data acquisition module, a domain adaptation training module, an environmental monitoring module, an optimization model building module, an optimization decision-making module, a balance evaluation module, and an optimal solution selection and instruction generation module; Among them, the data acquisition module is used to acquire real-time operating parameters, equipment status and external environment data of the pipeline network as real-time status data, define the real-time status data as target domain data, and acquire historical operating data, historical energy consumption data and historical load data of the pipeline network and define them as source domain data. The domain adaptation training module is used to generate robust features that can characterize the essential laws of pipeline network operation by performing adversarial domain adaptation training based on the source domain data and target domain data, and to obtain the domain adaptation error of the training process. The environmental monitoring module is used to monitor the rate of change and domain adaptation error of the target domain data in real time and generate environmental mutation judgment results. The optimization model building module is used to utilize the robust features to build a dynamic multi-objective optimization model that includes decision variables and multiple objective functions; The optimization decision module is used to generate a Pareto solution set containing several non-dominated solutions based on the dynamic multi-objective optimization model and in response to the environmental mutation determination result, using a dynamic multi-objective evolutionary algorithm. The balance evaluation module is used to evaluate the objective balance of the non-dominated solutions in the Pareto solution set and calculate the cooperative balance index of each solution among multiple objective functions. The optimal solution selection and instruction generation module is used to obtain a preference function that characterizes the current operational preference, select the optimal control solution from the Pareto solution set in combination with the cooperative balance index, and generate control instructions to be issued to the actuators based on the optimal control solution.

[0028] like Figure 2 As shown in the figure, this application provides a multi-objective collaborative optimization method for automatic pipeline distribution control, including: S1, acquire real-time running parameters, device states and external environment data of the pipe network as real-time state data, define the real-time state data as target domain data, acquire historical running data, historical energy consumption data and historical load data of the pipe network and define as source domain data; S2, based on the source domain data and the target domain data, eliminate the difference between the source domain data and the target domain data by performing adversarial domain adaptation training, generate robust features capable of representing the essential law of pipe network operation, and acquire domain adaptation error in the training process; S3, real-time monitor the change rate of the target domain data and the domain adaptation error, and generate an environment mutation judgment result; S4, use the robust features to construct a dynamic multi-objective optimization model containing decision variables and multi-objective functions; S5, based on the dynamic multi-objective optimization model, respond to the environment mutation judgment result, and use a dynamic multi-objective evolutionary algorithm to generate a Pareto solution set containing a plurality of non-inferior solutions; S6, perform target balance evaluation on the non-inferior solutions in the Pareto solution set, and calculate the synergistic balance index between the solutions in the multi-objective functions; S7, acquire a preference function representing the current operation preference, select an optimal control solution from the Pareto solution set in combination with the synergistic balance index, and generate a control instruction according to the optimal control solution and issue it to the execution mechanism.

[0029] It should be noted that the application uses adversarial domain adaptation training technology to extract robust features capable of reflecting the inherent hydraulic characteristics of the pipe network from the historical source domain data and real-time target domain data with distribution difference by letting the feature extractor and the domain classifier compete with each other. The change rate of real-time data and the error in the domain adaptation training process are used as double indicators to sensitively monitor the mutation of the environment and as a trigger signal. The extracted robust features are used to construct a dynamic multi-objective optimization model that can reflect the dynamic changes of the pipe network in real time. When the environment mutates, the model can be efficiently solved by a dynamic multi-objective evolutionary algorithm to generate a Pareto solution set containing a plurality of optimal compromise schemes. Finally, by introducing a synergistic balance index to evaluate the balance of each scheme among the multi-objectives, and combining a preference function representing the current operation focus, a unique optimal control solution is selected from the plurality of optimal schemes, and a control instruction is issued.

[0030] In a possible implementation manner of the embodiment of the application, in combination Figure 2 The above S2 can be implemented through the following S21, S22 and S23, which will be described in detail as follows: S21, feature extraction is performed on the source domain data and the target domain data to generate control features, and domain classification loss is generated by judging the data source of the control features; In some implementations, an adversarial network consisting of a feature extractor and a domain classifier is constructed and trained. The role of the feature extractor is to receive the input source domain data or target domain data and map it to a high-dimensional feature space, generating control features. The task of the domain classifier is to receive the control features output by the feature extractor and determine whether the features originate from source domain data or target domain data. During the training process, the objectives of the two modules are adversarial. On the one hand, the domain classifier is optimized to minimize its domain classification loss, striving to accurately distinguish the data source. On the other hand, the feature extractor is trained to maximize the classification loss of the domain classifier while optimizing the main task loss, i.e., to confuse the domain classifier by generating control features that are difficult to distinguish. The main task loss is usually defined based on the source domain data, for example, the loss of predicting pipe network energy consumption or pressure, which ensures that the extracted features are meaningful for pipe network control tasks. As Figure 3 The curve shows the accuracy of the domain classifier during the adversarial domain adaptation training process. The curve gradually decreases from high accuracy and stabilizes around 0.5, indicating that the feature extraction operation successfully confuses the domain source of the data, thereby achieving the extraction of domain-invariant robust features.

[0031] For example, in a pipe network transmission and distribution system, the feature extraction network uses a network structure containing 4 layers of 1D-CNN, with each layer having a convolution kernel size of , and the activation function is LeakyReLU. The input 120-dimensional real-time running state vector is converted into a 64-dimensional control feature vector, which contains the pressure, flow, pool liquid level, pump station frequency, etc. of 30 key nodes. The domain classification network uses a classifier containing 3 fully connected layers, with the number of neurons being , and after input, outputs two probability values of the probability of coming from the source domain or the target domain. For example, if the probability output by the classifier is , it indicates that it is highly confident to be judged as source domain data.

[0032] S22, training the feature extraction by alternately optimizing the main task loss of the feature extraction and the domain classification loss of the domain classification, so that the feature extraction can optimize the domain classification error; In some implementations, the adversarial training process can be achieved through a unified optimization objective, in which the parameter update of the feature extractor aims to minimize the main task loss while maximizing the domain classification loss. This is usually achieved by introducing a gradient reversal layer, which multiplies the gradient from the domain classification loss by a negative constant during backpropagation, thereby converting the minimization objective of the domain classifier into the maximization objective of the feature extractor. The optimization objective function of the entire network is represented as: ; wherein, represent the overall optimization objective; , , represent the network parameters of the feature extractor, the main task predictor and the domain classifier respectively; is the prediction error of the main task predictor on the source domain data, i.e., the main task loss; is the domain classification loss of the domain classifier, measuring its ability to distinguish the features of the source domain and the target domain; is a hyper-parameter balancing the weights of the main task loss and the domain classification loss in the optimization process of the feature extractor. Through optimization algorithms such as gradient descent, the parameters and of the feature extractor and the main task predictor are updated towards the direction of minimizing , while the parameters of the domain classifier are updated towards the direction of maximizing , forming a dynamic minimax game process.

[0033] For example, the main task loss of the feature extraction operation is defined as the mean square error of predicting the total energy consumption of the pipe network in the next 6 time steps, i.e., 30 minutes. The domain classification loss is defined as the binary cross-entropy loss. The training uses optimizer with a batch size of 128. In each iteration, the parameter update target of the feature extraction operation is: , and the parameter update target of the domain classification operation is . For example, in a certain iteration, if is , and is , the feature extraction operation will mainly adjust towards the direction of maximizing , generating more "ambiguous" features.

[0034] S23, apply the trained and converged feature extraction to real-time data, and output the domain-invariant features as robust features.

[0035] In some implementations, when the training process converges, it means that the feature extractor has learned a feature representation that can effectively support the prediction of the main task and successfully confuse the differences between the source domain and the target domain, so that the classification accuracy of the domain classifier is close to random guessing. At this time, the trained feature extractor is fixed, and when new target domain data, i.e., real-time state data, is input, the features output by the feature extractor are domain-invariant features, which capture the internal laws of pipe network operation that do not change with operating conditions, and are therefore defined as robust features for subsequent dynamic multi-objective optimization modeling.

[0036] For example, the convergence criterion is set as follows: the classification accuracy of the domain classifier is continuously increasing on the independent test set. Round iterations remain Within the range, and the energy consumption error predicted by the main task is less than At this point, training convergence is determined. In online applications, the converged feature extraction network is solidified and integrated into the real-time data processing flow of the control system. For example, given a target domain vector contaminated by sensor noise as real-time input, the robust features output by this network are significantly better than those obtained directly using... Its signal-to-noise ratio is improved This 64-dimensional vector is directly used as the initial input for the next stage of the algorithm to determine the optimal control solution.

[0037] In one possible implementation, combining Figure 2 The above-mentioned S3 can be implemented through the following S31, S32 and S33, which are explained in detail below: S31. Calculate the change in the target domain data within a continuous time step and generate a rate of change index; In some implementations, the rate of change of the target domain data over a continuous time step is calculated. The target domain data is a real-time status data vector containing multiple dimensions, such as pressure, flow rate, and equipment operating parameters at various monitoring points in the pipeline network. To eliminate the influence of different physical dimensions and numerical ranges, this data vector is first normalized. The difference between the normalized data vector of the current time step and the previous time step is calculated. The rate of change index is the norm of this difference vector, typically the L2 norm, and its calculation method is as follows: ; in, For the current moment The rate of change index; and These are the current times. and the previous time step The target domain data vector after normalization; This represents the L2 norm of the calculated vector, i.e., the Euclidean distance, which quantifies the magnitude of change in the overall operating status of the pipeline network over a short period of time.

[0038] For example, control cycle setting Minutes. Target domain data. It is dimensional vector, containing Pressure at key nodes Flow rate of the main pipeline The power of each pumping station, etc., and after Interval normalization. Rate of change index. Using the current time the L2 norm of the normalized vector difference of the previous time step. For example, under normal operating conditions, When a pipe burst or an important equipment trips, the change rate index can rapidly increase to or more within .

[0039] S32, comparing the change rate index and the domain adaptation error with a threshold value representing the normal operating fluctuation range, to generate a comparison result; In some implementations, the domain adaptation error defined in the previous stage of adversarial domain adaptation training is obtained in real time. This error is essentially a measure of the ability of the pre-trained domain classifier to misclassify the robust features extracted from the current real-time state data as source domain data. A higher domain adaptation error value indicates that the current target domain data distribution has deviated significantly from the historical source domain data distribution used for training, suggesting that the underlying patterns of the pipe network operation may have changed.

[0040] For example, the threshold value is set according to the confidence interval of the stable operating data in the past year: the change rate index threshold value is set to ; and the domain adaptation error threshold value, i.e., the cross-entropy loss of the domain classifier on real-time data is set to . At a certain time step , if the calculated , and the domain adaptation error at the same time step , then the following comparison results are generated: the change rate index comparison result is "exceeded", i.e., , and the domain adaptation error comparison result is "normal", i.e., .

[0041] S33, according to the comparison result, when the change rate index or the domain adaptation error exceeds the threshold value, generating an environmental mutation judgment result identifying an environmental mutation.

[0042] In some implementations, the calculated change rate index and the real-time obtained domain adaptation error are compared with their respective threshold values. These two threshold values represent the reasonable fluctuation range of the pipe network under normal operating conditions, and their values are set according to statistical analysis of a large amount of historical stable operating data, for example, they can be set to the normal fluctuation mean plus three times the standard deviation. When the change rate index exceeds its threshold value, or the domain adaptation error exceeds its threshold value, it is determined that an environmental mutation has occurred, and the corresponding environmental mutation judgment result is generated. This result is usually a Boolean signal, which is used to trigger subsequent dynamic optimization adjustment.

[0043] ​For example, the mutation decision logic is set as an "or" relationship: any indicator exceeding the threshold value is determined as a mutation. Since the change rate indicator exceeds the threshold value , an environmental mutation decision result True is immediately generated. This result subsequently triggers the dynamic multi-objective evolutionary algorithm, for example, switches the population initialization strategy from "memory-based" to "random-based", and immediately adjusts the mutation rate of the evolutionary operation from to , to accelerate the search for a new Pareto frontier and quickly adapt to the new operating conditions.

[0044] In one possible implementation, in combination with Figure 2 , the above S4 can be implemented by the following S41, S42 and S43, which are described in detail below: S41, define a decision variable set containing the operating parameters of the control equipment; In some implementations, a decision variable set is defined. Decision variables are physical quantities that can be directly adjusted by the control system, constituting the solution space of the optimization problem. The decision variable set mainly contains the operating parameters of all controllable equipment in the pipe network. The set consists of the operating frequencies of each frequency conversion pump and the valve openings of each electrically operated regulating valve. These variables together constitute a decision vector, each element of which corresponds to the control setting value of a specific device, whose value range is strictly constrained by the physical performance of the device and the safety operation procedures.

[0045] For example, in a medium-sized water supply pipe network, the decision variable set is defined as a vector containing 15 elements. The vector consists of the real-time operating frequencies of 5 main frequency conversion pump stations and the openings of 10 key electrically operated regulating valves. For example, the value range of the pump station frequency is limited to Hz to Hz; the value range of the valve opening is limited to to , which is pre-set according to the safety and operation limits of the equipment.

[0046] S42, establish multi-objective functions that respectively quantify the energy consumption cost, pipe network pressure safety and service quality tracking; In some implementations, a target function is established to quantify multiple operating objectives. To achieve multi-objective collaborative optimization, the operating level objectives, i.e. energy consumption cost, pipe network pressure safety and service quality tracking, need to be converted into specific mathematical functions. The energy consumption cost objective function , whose goal is to minimize the power consumption of the entire transmission and distribution system. This function is mainly based on the power consumption model of the pump set, and the total power consumption is the sum of the power consumption of all operating pump sets, while the power consumption of a single pump is related to its flow rate and head, and further related to the frequency of the frequency conversion and the valve opening in the decision variable. Therefore, is a function of the decision vector. The pressure safety objective function , which aims to ensure that the pressure of all critical nodes in the network is maintained within the pre-set safety upper and lower limits. This function is usually constructed as a penalty function, which grows non-linearly with the degree of deviation and the length of time that any node's predicted pressure is outside its safety range. The pressure values themselves are calculated from the hydraulic model, based on the decision vector, so is also a function of the decision vector. The service quality tracking objective function , which aims to minimize the deviation between the predicted service pressure and the desired pressure setting at the critical user nodes. This function calculates the weighted sum of the difference between the predicted pressure and the target pressure at all critical service points. Again, the point pressures are a function of the decision vector, which makes also dependent on the decision vector.

[0047] For example, the multi-objective functions include: the energy cost objective, which aims to minimize the total electricity bill in the next control period. This function is calculated based on the real-time power curve of each pump station and the local peak and valley electricity price table. The pressure safety objective: ensures that the pressure of 20 key monitoring points must be maintained within the safe pressure upper and lower limits. This function is constructed as a penalty term, which will impose a quadratic penalty once the predicted pressure exceeds this range. The service quality tracking objective: minimizes the percentage deviation between the predicted water supply pressure and the dispatch expected pressure of 10 main user service areas.

[0048] S43, integrate the robust features as dynamic input parameters into the multi-objective function, for real-time adjustment of the parameters and model constraints of the multi-objective function, to form a dynamic multi-objective optimization model.

[0049] In some implementations, the robust features are integrated to form a dynamic multi-objective optimization model. The key to achieving model dynamic adaptability. The hydraulic model relied on by each objective function is static in traditional methods, and its parameters such as pipe roughness and local head loss coefficient are fixed. The robust features generated by the adversarial domain adaptation training are integrated into the hydraulic model and the objective function as dynamic input parameters. The robust features contain deep information such as the current operation mode of the network and unmodeled changes in water demand, which cannot be directly measured. By integrating as one of the inputs to the hydraulic model, for example , the pressure prediction can reflect the actual hydraulic characteristics of the network in real time. In addition, can also be used to dynamically adjust the model constraints, such as adjusting the minimum service pressure standard according to the current demand pattern. Finally, the optimization model is expressed as finding the optimal decision vector to minimize a set of functions that depend on and dynamic parameters of the objective function, i.e., .

[0050] Exemplarily, the 64-dimensional robust feature vector output by the adversarial domain adaptation module is used as the dynamic calibration input of the model. For example, the vector is input into a parameter adaptive network for real-time adjustment of the Hessian-William coefficient (H-W roughness) in the pipe network hydraulic model to correct the distortion of the hydraulic model caused by pipe network aging or sudden failure. In addition, the robust feature vector is also used to dynamically adjust the minimum safety pressure threshold in the pressure safety target, allowing the minimum pressure threshold to be safely reduced slightly to save energy consumption when extreme low load is predicted.

[0051] In a possible implementation manner, in combination with Figure 2 , the S5 can be implemented by the following S51, S52 and S53, which are specifically described as follows: S51, generating an initial population by using the Pareto solution set of the last control period or the prediction result based on the robust feature; In some implementation manners, the algorithm needs an initial population to start the evolutionary search of the current control period. The generation strategy of the initial population is closely related to the running state of the pipe network. In the case that the pipe network runs smoothly, i.e., no environmental mutation is detected, the algorithm directly uses the Pareto solution set calculated in the last control period as the initial population of the current period. This “hot start” mode inherits the high-quality solution at the last time, so that the algorithm can fine-tune and optimize on the basis of the existing good foundation, thereby improving the convergence speed. Conversely, if the environmental monitoring module issues a determination result of environmental mutation, it means that the original optimal solution region may have failed, at this time, the algorithm will generate a new set of solutions that are more likely to adapt to the current new environment based on the latest robust feature as the initial population, or introduce random solutions to increase the diversity of the population, in order to prepare for searching for a new optimal region.

[0052] Exemplarily, a dynamic variant of the NSGA-II algorithm is adopted. Normally, the algorithm directly uses the Pareto solution set containing 100 non-inferior solutions obtained in the last control period as the initial population of the current period, which is called “hot start” based on memory. If the environmental monitoring module determines that mutation occurs, the algorithm activates the prediction initialization based on the robust feature: input the current 64-dimensional robust feature vector into a pre-trained prediction network to generate 5 predicted optimal control solutions, and mix the 5 solutions with 95 random solutions as a new initial population.

[0053] S52, if the environmental mutation determination result is yes, adjusting the search parameters of the evolutionary operation to generate a child population by performing the evolutionary operation on the initial population; In some implementations, the core of the algorithm lies in the response mechanism to environmental mutation. When receiving the environmental mutation decision result identifying that the environment has mutated, the algorithm will immediately adjust the internal evolutionary operation search parameters. For example, it will automatically increase the strength and probability of the mutation operator, or introduce a special population diversity maintenance mechanism. The purpose of this adjustment is to enhance the exploration ability of the algorithm, so that the population can break away from the constraints of the previous optimal solution area and explore the wider unknown areas in the solution space, quickly positioning the new Pareto front formed by the environmental mutation. Without environmental mutation, the algorithm will maintain normal search parameters, focusing on fine search of the current optimal area and enhancing its convergence ability. The algorithm performs evolutionary operations on the population initialized by strategy, including selection, crossover and mutation, to generate a child population. Then, the initial population and the child population are merged, and the dynamic multi-objective optimization model is used to evaluate each individual in the merged population, i.e., each candidate control strategy, to calculate its corresponding energy consumption cost, pipe network pressure safety and service quality tracking three objective function values. As shown in the following table, the initial population has 100 individuals, and the child population also has 100 individuals. The total number of individuals in the merged population is 200. Figure 4 The following table shows the changes in the Pareto solution set before and after the dynamic multi-objective evolutionary algorithm responds to environmental mutation. After the environmental mutation, the optimal solution set as a whole shifts to the high cost and high penalty direction. The algorithm can quickly transition from the old Pareto front to the new Pareto front by adjusting the search parameters, demonstrating dynamic adaptability.

[0054] For example, when the environmental mutation decision result is "yes", the algorithm immediately reduces the crossover probability from the normal to to protect excellent individuals, while significantly increasing the mutation rate from to and switching the mutation operator from polynomial mutation to non-uniform mutation to enhance the global exploration ability of the population and quickly jump out of the old optimal area. If the decision result is "no", the mutation rate remains at , focusing on local fine search. Subsequently, the algorithm performs selection, crossover and mutation operations on the initial population to generate a child population of 100 individuals.

[0055] S53, evaluate the initial population and the child population using the dynamic multi-objective optimization model, and generate a Pareto solution set through non-dominated sorting.

[0056] In some implementations, the algorithm employs a non-dominated sorting technique to stratify the entire population after evaluation. The principle of non-dominated sorting is that if individual A does not have any objective function value worse than individual B and at least one objective function value better than individual B, then A dominates B. All individuals that are not dominated by any other individual in the population are classified into the first layer, i.e. constitute the Pareto solution set obtained in this round of optimization, which are all non-inferior solutions and represent the optimal trade-off relationship between different objectives under the current operating condition.

[0057] Exemplarily, the initial population and the offspring population are combined into one mixed population. The three objective function values of the 200 candidate solutions are evaluated using the dynamic multi-objective optimization model. Subsequently, the NSGA-II non-dominated sorting and crowding distance calculation are performed on the 200 individuals: all individuals determined to be in the first layer constitute the Pareto solution set of this period, for example, 112 non-inferior solutions are obtained in this period, which are used for subsequent evaluation of the degree of synergy balance.

[0058] In one possible implementation, in combination Figure 2 , S6 can be implemented by S61, S62 and S63, which are specifically described as follows: S61, determining ideal points of the Pareto solution set in each objective dimension, and calculating objective deviation values of each non-inferior solution relative to the ideal points; In some implementations, the ideal points are determined and the objective deviation values are calculated. After obtaining the Pareto solution set containing several non-inferior solutions, a reference benchmark, i.e. the ideal point, needs to be determined. The ideal point is a virtual solution composed of the optimal values of each objective function in the current Pareto solution set. For the three minimization objectives of energy consumption cost, pipe network pressure safety and service quality tracking, the coordinates of the ideal point are composed of the minimum values of each objective dimension in the Pareto solution set. For each non-inferior solution, the objective deviation value in each objective dimension is the difference between its objective function value and the corresponding dimension value of the ideal point.

[0059] Exemplarily, in the analysis of a Pareto solution set containing 100 non-inferior solutions, the ideal points of the three objectives are determined: the energy consumption cost of the ideal point takes the lowest value among the 100 solutions, for example, 1500 yuan; the pressure safety penalty takes the lowest value; the service quality deviation takes the lowest value. For any non-inferior solution, its objective deviation value is the absolute difference between its objective function value and the ideal point value. For example, if the energy consumption cost of the non-inferior solution is 1800 yuan, then its energy consumption deviation value is yuan.

[0060] S62, performing normalization processing on the objective deviation values to eliminate the dimensional influence and generate normalized deviation values; In some implementations, the target deviation values are normalized. Since the three objective functions have different physical meanings and numerical ranges, it is not reasonable to directly compare the target deviation values. In order to eliminate the dimension effect, the deviation values in each objective dimension need to be normalized. Normalization usually adopts the maximum-minimum normalization method, which maps the deviation value of each non-inferior solution in a certain objective dimension to the interval of 0 to 1. The calculation formula of the normalized deviation value is: ; wherein, is the normalized deviation value of the jth non-inferior solution in the ith objective dimension; is the original target deviation value of the jth non-inferior solution in the ith objective dimension relative to the ideal point; and are the minimum value and the maximum value of the target deviation values of all solutions in the Pareto solution set in the ith objective dimension, respectively. Through this step, the importance of all objective dimensions is numerically leveled, and subsequent horizontal comparison can be performed.

[0061] For example, in order to eliminate the dimensional differences between "yuan", "point" and "percentage", the maximum-minimum normalization method is adopted to process the deviation values of all non-inferior solutions. For example, the energy consumption deviation value is the largest in all solutions, which is 600 yuan, and the smallest is 100 yuan. For the solution with a deviation value of 300 yuan, the calculation logic of the normalized deviation value is: After this processing, the deviation values of all solutions in the three objectives are mapped to a dimensionless interval of 0 to 1.

[0062] S63, calculate the dispersion degree of the normalized deviation value of each non-inferior solution, and take the dispersion degree as the coordination balance degree index.

[0063] In some implementations, the dispersion degree of the normalized deviation value is calculated and taken as the coordination balance degree index. For each non-inferior solution, it now has a set of normalized deviation values, which reflect its relative distance to the "absolute optimum" in each objective. A solution that performs well in each objective will have a relatively close normalized deviation value, and a solution that performs well in some objectives and poorly in others will have a very dispersed normalized deviation value. Therefore, the dispersion degree of the set of normalized deviation values can be calculated to quantify the balance of the solution. The dispersion degree can be measured by calculating its standard deviation or variance. The calculation of the coordination balance degree index is: ​​​​​​ wherein, is the degree of synergy balance of the th non-inferior solution; is the number of objective functions, generally taking the value of 3; is the average of all normalized deviation values of the th solution. The smaller the value of the degree of synergy balance, the more balanced the solution performs on multiple objectives, and the better the synergy.

[0064] Exemplarily, the degree of synergy balance adopts the standard deviation of the three normalized deviation values of each non-inferior solution. A good balanced solution will have three normalized deviation values very close to each other, resulting in a standard deviation close to zero. On the contrary, the normalized deviation values of an extreme solution will be quite different, resulting in a larger standard deviation. The scheduling system will take the value as a key indicator to measure the synergy balance of the control strategy among multiple objectives.

[0065] In a possible implementation, in combination with Figure 2 , the above S7 can be implemented by the following S71, S72 and S73, which are specifically described as follows: S71, obtaining working condition information reflecting the current working condition of the pipe network, and determining the objective weights in the preference function according to the working condition information; In some implementations, the working condition information of the current pipe network is obtained, and the objective weights in the preference function are dynamically determined according to the working condition information. The working condition information refers to real-time information that can reflect the current running state and external demand of the pipe network, for example, whether it is currently in a peak or off-peak water period, whether an emergency water event such as fire fighting has occurred, or whether a specific area has a repair pressure deficiency, etc. According to these specific working condition information, a preset rule base or model is called to assign different weights to the three objective functions of energy consumption cost, pipe network pressure safety and service quality tracking. For example, during the night off-peak period, energy saving and consumption reduction is the primary task, so the weight of the energy consumption cost objective will be correspondingly increased; while during the daytime peak period, guaranteeing the service quality of users is more critical, and the weight of the service quality tracking objective will be adjusted higher. These weights constitute the preference function representing the current operation preference, which is a quantitative embodiment of the decision maker's intention.

[0066] Exemplarily, it is detected in real time that the current period is “night off-peak period” and no emergency event occurs. According to the preset operation rules, the preference function is called to dynamically assign the three objective weights: the weight of the energy consumption cost objective is set to 0.6; the weight of the pipe network pressure safety objective is set to 0.3; and the weight of the service quality tracking objective is set to 0.1. This set of weight vectors is the quantitative embodiment of the current operation preference.

[0067] ​S72, applying the preference function to calculate a comprehensive evaluation value of each non-inferior solution in the Pareto solution set in combination with the target weight and the synergy balance index; In some implementations, a preference function is applied to calculate a comprehensive evaluation value of each non-inferior solution in the Pareto solution set in combination with the synergy balance index calculated in advance. The comprehensive evaluation value aims to give each solution a comprehensive score, which takes into account two dimensions, one is whether the solution fits the current operational preference, and the other is whether a good balance is achieved among multiple objectives. The calculation formula is: ; wherein, is the comprehensive evaluation value of the i-th non-inferior solution; is a preset hyperparameter, taking a value between 0 and 1, used to adjust the relative importance of operational preference and synergy balance in the final evaluation; is the weight of the i-th objective function determined according to the current operating condition information; is the normalized objective value of the i-th non-inferior solution on the i-th objective function, and the normalization process aims to eliminate the dimensional differences among different objective functions, so that the weighted sum has physical meaning. The first part of the formula measures the degree of satisfaction of the solution to the current operational preference, and the second part evaluates the intrinsic balance of the solution itself.

[0068] For example, the comprehensive evaluation value is calculated by combining weighted sum and balance penalty. For each non-inferior solution in the Pareto solution set, first calculate the weighted objective value: multiply the three normalized objective values of the solution by the target weight of the current operating condition and then sum. Then, add a penalty term proportional to the synergy balance index of the solution to the weighted sum. For example, the final comprehensive evaluation value of one solution is , and the comprehensive evaluation value of another solution is .

[0069] S73, selecting the non-inferior solution with the optimal comprehensive evaluation value as the optimal control solution.

[0070] In some implementations, the non-inferior solution with the optimal comprehensive evaluation value is selected as the final optimal control solution. Since each objective function and the synergy balance index are defined as the smaller the better, the non-inferior solution with the minimum comprehensive evaluation value ​​​​The non-dominated solution is considered the most ideal control scheme under the current operating conditions. This solution not only largely satisfies the current operational priorities, but its decision-making scheme itself also avoids excessive bias between different objectives, exhibiting good overall performance. Once the optimal control solution is determined, its corresponding decision variables, namely the frequencies of each frequency converter and the valve opening degree, constitute the specific control instructions issued to the field actuators.

[0071] For example, the overall evaluation values ​​of all 100 non-dominated solutions in the Pareto solution set are sorted. For instance, among all solutions, the overall evaluation value of a particular solution is... This is the minimum value. This solution is then selected as the optimal control solution for the current control cycle. This solution corresponds to a specific set of decision variables, such as the operating frequency of pump station P3. Hertz, control valve Opening These specific values ​​are then used as control commands and sent to the field actuators through the real-time automated control system.

[0072] In one possible implementation, after S73, the multi-objective cooperative optimization pipeline automatic transmission and distribution control method provided in this application embodiment further includes: After the control command is issued, the post-control pipeline operation data generated by the control command is collected; In some implementations, after generating and issuing control commands based on the optimal control solution, operational data of the pipeline network is continuously collected. This data, known as post-control pipeline network operational data, represents the actual response after the optimization decision was implemented, comprehensively recording the actual performance of various parameters such as pressure, flow rate, and equipment energy consumption under the control command. This data is the most direct and reliable basis for evaluating control effectiveness and identifying model deviations.

[0073] For example, during the 5-minute control cycle following the issuance of the optimal control solution, key data is continuously collected. This data includes: real-time power and head of all 5 pump stations, actual pressure readings at 20 key monitoring points, and actual energy consumption. For instance, if the instruction requires pump station P3 to operate at... Then what is collected is P3 at The actual energy consumption during operation and the resulting downstream pressure data form a control network operation record containing 100-dimensional data.

[0074] The source domain data is updated using the control network operation data. In some implementations, the newly collected post-control network operation data is used to update the source domain data. The source domain data is the basis for the adversarial domain adaptation training, which represents the "historical experience" known by the model. By supplementing the operation data containing the latest control effect into the source domain database, the dynamic expansion of the knowledge base is realized. The update operation can adopt various strategies, for example, the new data can be directly appended to the existing database, or the sliding window mechanism is used to replace the oldest data with new data, so that the source domain data can always reflect the recent operation characteristics of the network.

[0075] For example, the sliding window strategy is used to update the source domain data. The latest 10,000 historical operation records are always maintained in the source domain database. Every time a new post-control network operation record is obtained, it is appended to the end of the source domain database, and the oldest record in the database is deleted. For example, after the 10,001st data is added, the 1st data is removed, ensuring that the source domain data always reflects the operation characteristics of the network in the last 35 days.

[0076] Based on the updated source domain data, the adversarial domain adaptation training is adjusted to realize the continuous optimization of the model.

[0077] In some implementations, the original adversarial domain adaptation training process is adjusted and retrained based on the updated source domain data. As the source domain data changes, the original feature extractor may no longer be able to optimally extract domain-invariant robust features. Therefore, it is necessary to use the updated, more abundant and realistic source domain data to re-execute the adversarial domain adaptation training process. This can be in the form of online incremental learning, that is, fine-tuning based on the original model parameters to adapt to the distribution changes brought by new data; or periodic complete retraining to completely reconstruct the feature extractor and domain classifier. In this way, the adversarial domain adaptation model can continuously learn from new practice and correct its understanding of the nature of the network operation.

[0078] For example, the training adjustment adopts an incremental learning strategy. A small-scale incremental training is performed every 24 hours: use the 288 new data added on the same day as the training batch, and fine-tune the original feature extractor and domain classifier parameters for 1000 iteration cycles. This enables the model to quickly adapt to slow changes such as sensor drift or slight aging of the network, without the need to start from scratch.

[0079] Analyze the optimal control solution and its corresponding network operation results in the historical control period to generate a performance stability index; In some implementations, a performance stability index is generated by analyzing the optimal control solution and its corresponding network operation results in the history control periods. The optimal control solution selected in each control period and the network operation results produced by the solution in practical application are recorded and analyzed in a past period of time. The performance stability index aims to quantify the consistency and volatility of the recent control effect. A simple performance stability index can be the variation range of the optimal control solution in the decision space in a plurality of consecutive control periods, or the variance of the corresponding objective function value sequence. For example, the trace or determinant of the covariance matrix of the objective function value vector sequence of the optimal control solution in the last N periods can be calculated, and the smaller the value, the more stable the performance.

[0080] For example, the average variation range of the optimal control solution in the decision variable space in the last consecutive control periods is recorded. For example, the periodic variation standard deviation of the pump station frequency and valve opening in 10 periods is calculated. If the sum of the standard deviations is less than , it is considered that the control strategy is very stable, and the performance stability index is set to low volatility. If the sum is greater than , it indicates that the optimal solution jumps frequently and dramatically, and the performance stability index is set to high volatility.

[0081] The performance stability index is compared with a threshold value for defining a performance stability interval to generate a period adjustment decision; In some implementations, the calculated performance stability index is compared with a preset threshold value to generate a period adjustment decision. One or more threshold values are defined in advance, which constitute a performance stability interval. When the performance stability index falls within the stability interval, it indicates that the recent network operation state is stable, and the current control strategy performs well and is stable, and frequent global optimization is not required. On the contrary, when the index exceeds the stability interval, it means that the network operation state may be changing, or the current control strategy has a large fluctuation, and more frequent optimization calculation is required to respond quickly. According to the comparison result, a period adjustment decision is generated, which clearly indicates whether the length of the next control period should be extended, shortened, or kept unchanged.

[0082] For example, two threshold values are preset: a high stability threshold of 0.08 and a low stability threshold of 0.25. The current control period length is set to 5 minutes. The calculated performance stability index is compared with the threshold value: if the index is less than the high stability threshold of 0.08, it indicates that it is in an over-stable state, and a "extend period" period adjustment decision is generated. If the index is between the high stability threshold of 0.08 and the low stability threshold of 0.25, it indicates that the state is normal, and a "keep unchanged" decision is generated. If the index is greater than the low stability threshold of 0.25, it indicates that it is unstable, and a "shorten period" decision is generated.

[0083] According to the cycle adjustment decision, the running cycle length of the dynamic multi-objective evolutionary algorithm is dynamically adjusted.

[0084] In some implementations, the running cycle length of the dynamic multi-objective evolutionary algorithm is dynamically adjusted according to a cycle adjustment decision. Upon receiving the cycle adjustment decision, the control system will modify its internal timer or scheduler settings accordingly. If the decision is to extend the cycle, the dynamic multi-objective evolutionary algorithm will wait longer before being triggered next time. During this period, the current optimal control solution can be continued to be executed, or a simpler local adjustment strategy can be adopted. If the decision is to shorten the cycle, the algorithm will be executed more frequently to closely track the changes in the pipe network state. If the decision is to keep the cycle unchanged, the running cycle length will remain as it is.

[0085] For example, upon receiving a decision of "extend cycle", the running cycle length of the algorithm is dynamically adjusted from the current 5 minutes to 7.5 minutes. This results in a decrease in the execution frequency of the algorithm , thus saving computational resources in the steady state. On the contrary, if a decision of "shorten cycle" is received, the cycle length will be adjusted to 3 minutes to ensure that the algorithm can respond more quickly to potential system instability.

[0086] It should be noted that the electrical connection between the various units described above does not necessarily mean a direct connection, and an indirect connection mode can also be applicable to the embodiments of the present application as long as the purpose of the present application is achieved. The above described is only an exemplary embodiment of the present application, and cannot limit the scope of the present application.

[0087] That is, any equivalent changes and modifications made according to the teachings of the present application are still within the scope of the present application. Other embodiments of the present application will be readily apparent to those skilled in the art upon considering the specification and practice of the true principles disclosed herein. The present application is intended to cover any variations, uses, or adaptations of the application following the general principles thereof and including such departures from the present disclosure as come within known or customary practice in the art to which the present application pertains.

Claims

1. A multi-objective collaborative optimization method for automatic pipeline transmission and distribution control, characterized in that, The method includes: The real-time operating parameters, equipment status, and external environment data of the pipeline network are acquired as real-time status data. The real-time status data is defined as target domain data. The historical operating data, historical energy consumption data, and historical load data of the pipeline network are acquired and defined as source domain data. Based on the source domain data and target domain data, adversarial domain adaptation training is performed to eliminate the differences between the source domain data and the target domain data, generate robust features that can characterize the essential laws of pipeline network operation, and obtain the domain adaptation error of the training process. Real-time monitoring of the rate of change and domain adaptation error of the target domain data, generating environmental mutation judgment results; Using the robust features, a dynamic multi-objective optimization model containing decision variables and multiple objective functions is constructed; Based on the dynamic multi-objective optimization model, and in response to the environmental mutation determination result, a Pareto solution set containing several non-dominated solutions is generated using a dynamic multi-objective evolutionary algorithm. The non-dominated solutions in the Pareto solution set are evaluated for objective balance, and the synergistic balance index of each solution among multiple objective functions is calculated. Obtain a preference function to characterize the current operational preferences, select the optimal control solution from the Pareto solution set in combination with the cooperative balance index, and generate control instructions to be issued to the actuators based on the optimal control solution.

2. The multi-objective collaborative optimization method for automatic pipeline transmission and distribution control according to claim 1, characterized in that, The robust features generated that characterize the essential laws governing pipeline network operation include: Feature extraction is performed on the source domain data and target domain data to generate control features, and the data source is determined by domain classification of the control features to generate domain classification loss; The feature extraction is trained by alternately optimizing the main task loss of feature extraction and the domain classification loss of domain classification, so that feature extraction can optimize the domain classification error. The features extracted after training convergence are applied to real-time data, and the output domain-invariant features are used as robust features.

3. The multi-objective collaborative optimization method for automatic pipeline transmission and distribution control according to claim 1, characterized in that, The real-time monitoring of the change rate and domain adaptation error of the target domain data, and the generation of environmental mutation judgment results, include: Calculate the change in the target domain data over a continuous time step and generate a rate of change index; The rate of change index and the domain adaptation error are compared with a threshold representing the normal operating fluctuation range to generate a comparison result; Based on the comparison results, when the rate of change index or domain adaptation error exceeds the threshold, an environmental mutation determination result is generated to identify a sudden change in the environment.

4. The multi-objective collaborative optimization pipeline automatic transmission and distribution control method according to claim 1, characterized in that, The construction of the dynamic multi-objective optimization model, which includes decision variables and multiple objective functions, includes: Define a set of decision variables that include the operating parameters of the control equipment; Establish multiple objective functions to quantify transmission and distribution energy consumption costs, pipeline pressure safety, and service quality tracking; The robust features are integrated as dynamic input parameters into a multi-objective function to adjust the parameters and model constraints of the multi-objective function in real time, forming a dynamic multi-objective optimization model.

5. The multi-objective collaborative optimization method for automatic pipeline transmission and distribution control according to claim 1, characterized in that, Based on the dynamic multi-objective optimization model, and in response to the environmental mutation determination result, a Pareto solution set containing several non-dominated solutions is generated using a dynamic multi-objective evolutionary algorithm, including: The initial population is generated using the Pareto solution set from the previous control period or the prediction results based on robust features. If the environmental mutation determination result is yes, then the search parameters of the evolutionary operation are adjusted, and the evolutionary operation is performed on the initial population to generate the offspring population; The initial population and offspring population are evaluated using the dynamic multi-objective optimization model, and a Pareto solution set is generated through non-dominated sorting.

6. The multi-objective collaborative optimization method for automatic pipeline transmission and distribution control according to claim 1, characterized in that, The objective balance of the non-dominated solutions in the Pareto solution set is evaluated, and the cooperative balance index of each solution among the multiple objective functions is calculated, including: Determine the ideal points of the Pareto solution set in each objective dimension, and calculate the objective deviation value of each non-dominated solution relative to the ideal point; The target deviation value is normalized to eliminate the influence of dimensions, and a normalized deviation value is generated. Calculate the degree of dispersion of the normalized deviation value for each non-dominated solution, and use the degree of dispersion as an index of cooperative balance.

7. The multi-objective collaborative optimization method for automatic pipeline transmission and distribution control according to claim 1, characterized in that, The step of obtaining the preference function to characterize the current operational preference, and selecting the optimal control solution from the Pareto solution set in conjunction with the collaborative balance index, includes: Obtain operating condition information reflecting the current pipeline network operating conditions, and determine the target weight in the preference function based on the operating condition information; The preference function is applied, and the target weight and the collaborative balance index are combined to calculate the comprehensive evaluation value of each non-dominated solution in the Pareto solution set; The non-dominated solution with the best comprehensive evaluation value is selected as the optimal control solution.

8. The multi-objective collaborative optimization method for automatic pipeline transmission and distribution control according to claim 1, characterized in that, The method further includes: After the control command is issued, the post-control pipeline operation data generated by the control command is collected; The source domain data is updated using the control network operation data. Based on the updated source domain data, the adversarial domain adaptation training is adjusted to achieve continuous model optimization.

9. The multi-objective collaborative optimization method for automatic pipeline transmission and distribution control according to claim 1, characterized in that, The method further includes: Analyze the optimal control solution and its corresponding pipeline operation results within the historical control cycle to generate performance stability index; The performance stability index is compared with a threshold used to define the performance stability range to generate a periodic adjustment decision; Based on the aforementioned periodic adjustment decision, the running period length of the dynamic multi-objective evolutionary algorithm is dynamically adjusted.

10. A multi-objective collaborative optimization automatic pipeline transportation and distribution control system, characterized in that, The system is used in a multi-objective collaborative optimization pipeline automatic transmission and distribution control method as described in any one of claims 1-9, the system comprising: The data acquisition module is used to acquire real-time operating parameters, equipment status and external environment data of the pipeline network as real-time status data, define the real-time status data as target domain data, and acquire historical operating data, historical energy consumption data and historical load data of the pipeline network and define them as source domain data. The domain adaptation training module is used to generate robust features that can characterize the essential laws of pipeline network operation by performing adversarial domain adaptation training based on the source domain data and target domain data, and to obtain the domain adaptation error of the training process. The environmental monitoring module is used to monitor the rate of change and domain adaptation error of the target domain data in real time and generate environmental mutation judgment results. The optimization model building module is used to utilize the robust features to build a dynamic multi-objective optimization model that includes decision variables and multiple objective functions; The optimization decision module is used to generate a Pareto solution set containing several non-dominated solutions based on the dynamic multi-objective optimization model and in response to the environmental mutation determination result, using a dynamic multi-objective evolutionary algorithm. The balance evaluation module is used to evaluate the objective balance of the non-dominated solutions in the Pareto solution set and calculate the cooperative balance index of each solution among multiple objective functions. The optimal solution selection and instruction generation module is used to obtain a preference function that characterizes the current operational preference, select the optimal control solution from the Pareto solution set in combination with the cooperative balance index, and generate control instructions to be issued to the actuators based on the optimal control solution.

Images