A multi-objective programming based peak-avoiding method for power consumption management

Abstract

The application relates to the technical field of power system management, and discloses a power consumption management peak-shaving method based on multi-target programming, which comprises the following steps: constructing a heterogeneous graph network structure of a power system, defining an uncertainty representation model on the graph structure, constructing an uncertainty propagation dynamics model, realizing an uncertainty propagation algorithm with graph structure perception, constructing a robust optimization framework with graph structure perception, realizing a multi-level risk partition management strategy, and discovering and utilizing the uncertainty complementary effect; the application effectively solves the problem that a traditional method ignores the influence of a system topological structure, balances the contradiction between the calculation complexity and the modeling accuracy, realizes differentiated processing of different levels of uncertainty, can discover and utilize the uncertainty complementary effect, and improves the effectiveness of power grid peak-shaving management, system adaptability and resource utilization efficiency.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of power system management, more particularly, it relates to a power consumption management peak avoidance method based on multi-objective programming. BACKGROUND

[0002] With the development of power systems towards high proportion of renewable energy, high digitization and intelligence, the uncertainty factors in power grid operation are increasing and complex. Traditional power system uncertainty management methods are mainly based on probability statistics model or scenario analysis. These methods usually regard uncertainty sources as independent or simply correlated random variables, which is difficult to accurately depict the propagation characteristics of uncertainty in complex network structure.

[0003] The current power consumption management peak avoidance technology mainly relies on load forecasting and simple demand response mechanism, and lacks in-depth consideration of system topology structure and uncertainty propagation path. These methods often face the contradiction between computational complexity and modeling accuracy when dealing with multi-source heterogeneous uncertainty, especially in complex scenarios such as high renewable energy penetration or extreme weather events. The system's evaluation of uncertainty is often inaccurate, resulting in poor effect of peak avoidance measures.

[0004] In addition, the existing technology generally adopts a "one-size-fits-all" management strategy, lacks differentiated processing ability for different levels of uncertainty influence, and is also difficult to find and utilize the complementary effects that may exist between uncertainties. These limitations seriously restrict the adaptability and operation efficiency of power systems in the face of complex uncertainty environment, and a power consumption management peak avoidance method is needed that can perceive network topology structure, accurately depict the propagation characteristics of uncertainty and realize multi-objective optimization. SUMMARY

[0005] The present application provides a power consumption management peak avoidance method based on multi-objective programming, which solves the technical problem of ignoring the influence of system topology structure on uncertainty propagation in related art power system uncertainty modeling methods.

[0006] The present application provides a power consumption management peak avoidance method based on multi-objective programming, comprising the following steps:

[0007] Constructing a heterogeneous graph network structure of the power system, the nodes representing entities in the power system, and the edges representing physical connections or data correlations between entities;

[0008] Based on the constructed heterogeneous graph network, defining an uncertainty representation model on the graph structure, and defining uncertainty state spaces for nodes and edges in the graph;

[0009] Using the defined uncertainty representation model, constructing an uncertainty propagation dynamics model to describe the diffusion process of uncertainty in the network;

[0010] Based on the constructed propagation dynamics model, an uncertainty propagation algorithm with graph structure perception is implemented to generate the spatiotemporal evolution trajectory of uncertainty in the network.

[0011] According to the generated spatiotemporal evolution trajectory of uncertainty, a robust optimization framework with graph structure perception is constructed to generate a multi-objective peak avoidance decision scheme that adapts to uncertainty.

[0012] Based on the results of the robust optimization framework, a multi-level risk zoning management strategy is implemented to adopt differentiated management strategies for different risk level areas.

[0013] In the implementation of risk zoning management, the complementary effect of uncertainty is discovered and utilized to reduce the overall risk of the system.

[0014] In a preferred embodiment, the step of constructing a heterogeneous graph network structure of the power system specifically includes:

[0015] Node identification and classification, identifying entities in the power system and classifying them into different types of nodes according to their functional characteristics;

[0016] Edge relationship construction and weighting, based on physical connection relationship and data correlation, constructing the edge connection between nodes;

[0017] Graph structure optimization, through graph sparsification technology and importance sampling method, optimizing the graph structure and removing the edges with minimal impact on uncertainty propagation.

[0018] In a preferred embodiment, the step of defining the uncertainty representation model on the graph structure specifically includes:

[0019] Uncertainty state space construction, defining the uncertainty state space for each node and edge in the graph;

[0020] Initial uncertainty quantification, based on historical data and expert knowledge, assigning initial uncertainty state values to the uncertainty source nodes in the graph;

[0021] Uncertainty evolution function construction, defining the function of the evolution of the uncertainty state of nodes and edges over time.

[0022] In a preferred embodiment, the step of constructing an uncertainty propagation dynamics model specifically includes:

[0023] Propagation function construction, constructing the function of uncertainty propagation between nodes according to node types and edge characteristics;

[0024] Propagation dynamics modeling, integrating the node evolution function and the propagation function to construct the complete uncertainty propagation dynamics equation;

[0025] Parameter learning and model calibration, based on historical data, use parameter learning algorithms to optimize the parameters of the propagation dynamics model.

[0026] In a preferred embodiment, the step of implementing the graph structure-aware uncertainty propagation algorithm specifically includes:

[0027] Spacetime discretization, discretizing continuous propagation dynamics equations into a form suitable for computer solution;

[0028] Graph neural network construction, based on the discretized dynamics equation, construct the graph neural network structure for efficient implementation of uncertainty propagation simulation calculation;

[0029] Parallel computing optimization, for large-scale power system graph structure, use parallel computing and graph partitioning technology to optimize algorithm execution efficiency.

[0030] In a preferred embodiment, the graph neural network includes a message generation function, a message aggregation function, and a node state update function;

[0031] The message generation function generates messages according to the source node state, the target node state, and the edge attribute;

[0032] The message aggregation function aggregates the messages of all neighbor nodes into comprehensive information;

[0033] The node state update function combines the current state of the node with the aggregated message to update the state of the node.

[0034] In a preferred embodiment, the step of constructing a graph structure-aware robust optimization framework specifically includes:

[0035] Uncertainty impact assessment, analyze the impact of uncertainty propagation on key indicators of the power system;

[0036] Robust optimization model construction, construct a multi-objective robust optimization model considering uncertainty propagation;

[0037] Multi-objective weight adaptive adjustment, dynamically adjust the weights of each objective in multi-objective optimization according to decision preferences and risk tolerance under different scenarios.

[0038] In a preferred embodiment, the multi-objective robust optimization model includes:

[0039] Definition of decision variables and uncertain parameters, decision variables represent controllable parameters of the system, and uncertain parameters represent uncertainties existing in the system;

[0040] Construction of multi-objective function, considering multiple optimization objectives, including cost minimization, risk minimization, and user satisfaction maximization;

[0041] Robust constraint setting ensures that the system operates under all preset uncertainties to meet safety and reliability requirements;

[0042] Decision space limitation explicitly defines the value range of decision variables and constraints;

[0043] Based on the construction of uncertain set of graph structure, the possible value range of uncertainty parameters is defined by using the heterogeneous graph network structure characteristics of power system.

[0044] In a preferred embodiment, the step of discovering and utilizing the complementary effect of uncertainty specifically includes:

[0045] Complementary pattern recognition analyzes the correlation and complementarity of different uncertainty sources in time and space dimensions;

[0046] Complementary effect quantification quantitatively evaluates the degree of reduction of the overall uncertainty of the system by complementary effect;

[0047] Complementary resource collaborative scheduling optimizes resource allocation and scheduling strategy to maximize the utilization of complementary effect.

[0048] In a preferred embodiment, a computer readable storage medium for storing computer readable instructions capable of running a multi-objective planning based peak-shaving method for electricity management when read by a computer.

[0049] The beneficial effects of the present application are:

[0050] The uncertainty propagation model perceived by the graph structure effectively solves the problem of ignoring the influence of system topology structure in traditional methods. This model can accurately capture the complex propagation path and diffusion effect of uncertainty in the network, making the system more accurate in evaluating uncertainty under high renewable energy penetration or complex weather events, significantly improving the effectiveness of peak-shaving management of power grid. Practical application shows that the prediction accuracy of this method under complex weather conditions far exceeds that of traditional methods, providing more reliable decision support for power grid dispatching.

[0051] The graph neural network algorithm effectively balances the contradiction between computational complexity and modeling accuracy. Through time and space discretization and parallel computing optimization, this algorithm can significantly reduce the computational complexity while maintaining high accuracy, enabling the complex uncertainty propagation model to be applied to real-time decision-making scenarios. This technical breakthrough enables the system to complete uncertainty evaluation of large-scale power grid in a short time, providing timely support for peak-shaving decision-making.

[0052] The multi-level risk zoning management strategy provides differentiated processing capabilities for different levels of uncertainty impacts. Through the robust optimization framework, the system can develop differentiated response strategies according to the propagation characteristics of different uncertainty sources, avoiding the limitations of traditional "one-size-fits-all" management methods. This refined management approach significantly improves the adaptability and resource utilization efficiency of the system, making the peak-shaving measures more accurate and effective.

[0053] Through the complementary effect recognition module, the complementary effect between uncertainties is successfully discovered and utilized. The system can identify patterns in which different sources of uncertainty cancel out or weaken each other under certain conditions, and maximize the utilization of this complementary effect by optimizing resource allocation and scheduling strategies. This innovation significantly reduces the overall risk of the system, improves the stability and reliability of the power grid, and provides a new technical path for efficient peak-shaving management of the power system. BRIEF DESCRIPTION OF DRAWINGS

[0054] Figure 1 is a flowchart of a power consumption management peak-shaving method based on multi-objective programming of the present application;

[0055] Figure 2 is a line graph of the propagation process of uncertainty in the power system of the present application;

[0056] Figure 3 is a column chart comparing the performance of the traditional method and the graph structure perception method of the present application;

[0057] Figure 4 is a radar chart for multi-dimensional performance comparison of the present application;

[0058] Figure 5 is a line graph of the peak-shaving effect under different renewable energy penetration rates of the present application;

[0059] Figure 6 is a column chart of the technical effect of the graph structure perception method of the present application; DETAILED DESCRIPTION

[0060] The subject matter described herein will now be discussed with reference to example implementations. It should be understood that discussions of these implementations are merely provided to give a more fuller understanding of the subject matter described herein and are not intended to limit the scope of the protection of this specification. The functions and arrangements of the elements discussed can be varied from those described and can be changed to suit the needs of the particular implementation. Various processes or components can be omitted, substituted, or added according to desired configuration. In addition, features described in some examples can be combined in other examples.

[0061] In at least one embodiment of the present application, a power consumption management peak-shaving method based on multi-objective programming is disclosed, as shown in Figure 1 including the following steps:

[0062] Step 1, Constructing the heterogeneous graph network structure of the power system, nodes represent entities in the power system, and edges represent physical connections or data correlations between entities;

[0063] Specifically, the following sub-steps are included:

[0064] Step 1.1, Node identification and classification;

[0065] Identify the entities in the power system (such as power generation units, load centers, substations, sensors, etc.), and classify them into different types of nodes according to their functional characteristics. Each type of node has a specific set of attributes to describe the characteristics of that type of node.

[0066] For example, for power generation nodes, attributes include power generation type (such as thermal power, wind power, photovoltaic, etc.), rated capacity, historical fluctuation characteristics, etc.; for load nodes, attributes include load type (such as industrial, commercial, residential, etc.), power consumption mode, demand response capability, etc.

[0067] Optionally, other types of nodes such as energy storage nodes, transmission nodes, and monitoring nodes can also be included. In some embodiments, nodes can also have time-varying attributes to represent their state changes over time.

[0068] Specifically, the heterogeneous graph network construction algorithm of the present application contains the following implementation details:

[0069] The heterogeneous graph network construction algorithm first extracts structural information from the physical topology and data relationships of the power system.

[0070] For physical topology, the algorithm parses the power grid connection graph, device parameter table, and geographic location data;

[0071] For data relationships, the algorithm analyzes the correlation matrix of historical operation data, power flow distribution, and frequency response characteristics.

[0072] The algorithm adopts a multi-level hierarchical processing flow:

[0073] First, determine the core backbone network, including main power plants, substations, and transmission lines;

[0074] Then gradually add branch networks, including distributed energy, load centers, and distribution networks;

[0075] Finally, integrate auxiliary data sources such as weather stations, power markets, and social and economic data.

[0076] Optionally, in the case of limited computing resources, a simplified two-level processing flow can be used, containing only the backbone network and the key branch network.

[0077] The node classification adopts a clustering method enhanced by domain knowledge, combining power system professional rules and data-driven feature extraction.

[0078] The algorithm first classifies based on device type and function, then subdivides according to operational characteristics (such as fluctuation patterns, response speed, control flexibility), and finally identifies groups of nodes with similar behavior patterns through cluster analysis. For edge construction, the algorithm considers both physical connection strength (such as transmission capacity, electrical distance) and data correlation strength (such as correlation coefficient, mutual information, causal strength), filtering weak connections through an adaptive threshold method and retaining significant edge relationships.

[0079] It should be noted that domain knowledge can be introduced in the form of a rule base or expert system, and data-driven features can be extracted through various statistical learning methods.

[0080] Specific application examples in complex weather conditions:

[0081] In a certain coastal power grid system, it faces multiple uncertainty challenges caused by typhoon weather. The system first constructs a heterogeneous graph network, including meteorological nodes (weather stations, satellite monitoring points), power grid physical nodes (transmission lines, substations, distribution equipment), power generation nodes (conventional power plants, wind farms, distributed photovoltaic), load nodes (industrial parks, commercial centers, residential areas), and social impact nodes (emergency centers, hospitals, transportation hubs).

[0082] Edge relationships include: physical power flow connections (transmission lines, distribution networks), meteorological influence relationships (wind speed and wind power output, cloud cover and photovoltaic power generation), load response relationships (temperature and electricity demand, rainfall and drainage electricity), and emergency plan connections (faulty equipment and backup resources).

[0083] The algorithm automatically identifies high correlation and strong causal relationships by analyzing historical typhoon data, such as specific wind directions and vulnerable lines, rainfall intensity and drainage pump station load, wind farm output and backup power demand.

[0084] Based on the constructed heterogeneous graph network, the system starts monitoring meteorological node data changes 48 hours before the typhoon arrives, propagating meteorological uncertainty to potentially affected power grid components through the graph structure, predicting possible failure points and load changes. For example, when detecting that the wind speed in a certain area will exceed the safety threshold, the system predicts that the relevant transmission lines may be damaged, affecting downstream substations and load areas, while considering the combined effects of wind farm output changes and emergency load increases.

[0085] This refined uncertainty propagation analysis enables the system to take targeted preventive measures, such as adjusting the power grid topology in advance, deploying maintenance teams to high-risk areas, activating backup power sources, etc., significantly improving the resilience and peak avoidance capabilities of the power grid under extreme weather conditions.

[0086] Step 1.2, edge relationship construction and weighting;

[0087] Based on physical connection relationships (such as power transmission lines) and data correlations (such as correlation coefficients of historical data), the edge connections between nodes are constructed.

[0088] The type of edge is determined according to the type of connected nodes and the nature of the connection, and is given different weights to represent the strength and speed of uncertainty propagation.

[0089] For example, for edges representing physical power flow, weights can be determined based on transmission capacity and impedance; for edges representing data correlation, weights can be determined based on correlation coefficients or mutual information.

[0090] Step 1.3, graph structure optimization;

[0091] Through graph sparsification techniques and importance sampling methods, the graph structure is optimized, removing edges that have minimal impact on uncertainty propagation, reducing subsequent computational complexity.

[0092] The remaining edges are reweighted according to their importance to uncertainty propagation, ensuring that the simplified graph structure still accurately represents uncertainty propagation characteristics.

[0093] As shown in Figure 2 , the time evolution process of uncertainty propagation from renewable energy sources (photovoltaic uncertainty) through distribution centers to the load end is shown. It can be observed that there is a time delay in the propagation of uncertainty, and the uncertainty level at the distribution center is between the source and the load.

[0094] Step 2, based on the constructed heterogeneous graph network, define the uncertainty representation model on the graph structure, define the uncertainty state space for nodes and edges in the graph;

[0095] Specifically, the following sub-steps are included:

[0096] Step 2.1, uncertainty state space construction;

[0097] Define the uncertainty state space for each node and edge in the graph, including the type, range, distribution characteristics, etc. of uncertainty. The state space can be in the form of a scalar (such as simple fluctuation amplitude), vector (such as multi-dimensional feature description) or matrix (such as time series correlation description), and the appropriate representation form is selected according to the modeling accuracy requirements and computational resource constraints.

[0098] Optionally, the state space can also contain a set of discrete states to represent certain types of uncertain events, such as equipment failure, extreme weather, etc.

[0099] Step 2.2, initial uncertainty quantification;

[0100] Based on historical data and expert knowledge, the initial uncertainty state value is assigned to the uncertainty source nodes in the graph (such as wind farms, photovoltaic power stations, etc.).

[0101] The initial state can be obtained through statistical analysis, probability distribution fitting or machine learning methods, reflecting the inherent uncertainty characteristics of the node.

[0102] It should be noted that in some embodiments, the quantification of initial uncertainty can also be dynamically updated according to real-time data to adapt to changing environmental conditions.

[0103] Step 2.3, uncertainty evolution function construction;

[0104] Define the function of the uncertainty state of nodes and edges evolving over time, describing the dynamic change law of the uncertainty source itself.

[0105] The evolution function can be constructed based on time series analysis, state space model or deep learning model, capturing the change trend, periodicity, mutation characteristics of uncertainty over time, etc.

[0106] In addition, linear or nonlinear evolution function forms can be selected according to actual needs to balance modeling accuracy and computational complexity.

[0107] In this application, the mathematical expression of the graph structure perception uncertainty propagation theory (GUPT) is:

[0108] ;

[0109] where, represents the uncertainty state on the power system graph structure at time , is a complete uncertainty characterization model that describes the distribution, characteristics and propagation law of uncertainty in the system; is the system graph structure, representing the topology of the entire power system; represents the node set, representing each entity in the system (such as power generation units, load centers, substations, etc.); represents the edge set, representing the connection relationship between nodes (such as power transmission lines, data association, etc.); represents the uncertainty state of the node at time , describing the uncertainty characteristics of each node; represents the uncertainty state of the edge at time , describing the uncertainty characteristics of each edge; The uncertainty state of a node, describing the uncertainty characteristics of the node; The uncertainty evolution function of a node, describing how the uncertainty of a node naturally changes over time; The uncertainty evolution function of an edge, describing how the uncertainty on an edge naturally changes over time; The uncertainty propagation function, describing how uncertainty propagates from one node to another through an edge.

[0110] As Figure 3 shown, the performance of the traditional method and the graph structure-aware method proposed in this invention in predicting error rate and computation time, two key indicators, is compared. The graph structure-aware method significantly reduces computation time while maintaining a low prediction error.

[0111] Step 3, using the defined uncertainty representation model, construct the uncertainty propagation dynamics model, describe the diffusion process of uncertainty in the network;

[0112] Specifically, the following sub-steps are included:

[0113] Step 3.1, propagation function construction;

[0114] According to the node type and edge characteristics, construct the function of uncertainty propagation between nodes , describe how uncertainty propagates from one node to the adjacent node through the edge. The propagation function considers the characteristics of nodes and edges, physical constraints and historical propagation patterns, and can be linear or nonlinear.

[0115] For linear propagation, it can be based on the Laplace operator;

[0116] For nonlinear propagation, it can be based on neural networks or complex system models.

[0117] It should be noted that the selection of the propagation function should be determined according to the specific uncertainty type and network characteristics. Different types of uncertainty may require different forms of propagation function.

[0118] Specifically, the implementation of the propagation function includes the following forms:

[0119] For linear propagation, the propagation function is implemented based on the Laplace operator, and the calculation process is as follows:

[0120] First, get the current uncertainty state values of the source node and the target node;

[0121] Then calculate the difference between the two state values, which represents the uncertainty gradient between the two nodes;

[0122] Then this difference value is multiplied by a propagation coefficient, which reflects the propagation strength of the edge and can be determined by the edge's attributes (such as electrical impedance, physical distance, or data correlation);

[0123] The final result is the amount of uncertainty influence propagated from the source node to the target node.

[0124] The linear propagation model is simple and intuitive, with high computational efficiency, and is suitable for cases where the uncertainty propagation characteristics are relatively simple.

[0125] For nonlinear propagation scenarios, the propagation function is implemented based on a neural network model, and the calculation process is as follows:

[0126] First, the source node state, target node state, and edge state are input as input features into a specially designed neural network;

[0127] Then, through the multi-layer structure of the neural network, feature extraction and nonlinear transformation are performed to capture the complex propagation relationship between the source node and the target node;

[0128] Finally, a numerical value or vector representing the propagation influence is output. The structure of the neural network can be a fully connected network, a convolutional network, or a recurrent network, depending on the spatiotemporal characteristics of the uncertainty propagation.

[0129] The nonlinear propagation model can handle more complex propagation phenomena, such as saturation effects, threshold effects, and multimodal propagation.

[0130] In addition, for specific types of power system uncertainty, the propagation function can also be implemented based on a physical model. For example, for power flow-related uncertainty, the calculation process is as follows:

[0131] First, the phase angle information is extracted from the uncertainty states of the source node and the target node;

[0132] Then, the phase angle difference between the two nodes is calculated, which reflects the direction and size of the power flow;

[0133] Next, this difference value is multiplied by the line susceptance value, which reflects the electrical characteristics of the line;

[0134] The final result is the amount of uncertainty influence propagated through the line.

[0135] This method based on a physical model can accurately reflect the physical laws of the power system and is suitable for uncertainty propagation scenarios with clear physical mechanisms.

[0136] In the embodiments of the present application, an adaptive hybrid propagation function is used to automatically select the appropriate propagation model according to the type of uncertainty and the characteristics of the network, and the calculation process is as follows:

[0137] First, calculate the output values of the linear and nonlinear propagation models respectively;

[0138] Then, determine a mixing coefficient based on the current system state and historical data, which reflects the relative importance of the linear and nonlinear models;

[0139] Next, multiply the output of the linear model by the mixing coefficient and multiply the output of the nonlinear model by the complement of the mixing coefficient;

[0140] Finally, add the two weighted results to obtain the final propagation impact value.

[0141] This hybrid method combines the computational efficiency of the linear model and the expressive power of the nonlinear model, and can adapt to various complex uncertainty propagation scenarios.

[0142] Step 3.2, Propagation dynamics modeling;

[0143] Integrate the node evolution function and the propagation function to build a complete uncertainty propagation dynamics equation, describing the evolution law of uncertainty in the entire system over time and network structure.

[0144] The basic form of the dynamics equation is:

[0145] ;

[0146] Where, represents the uncertainty state of node at time ; represents the rate of change of the uncertainty state of node over time; is the uncertainty evolution function of the node, describing how the node's uncertainty naturally changes over time; represents the neighbor set of node , containing all nodes directly connected to node ; represents a neighbor node of node ; represents the uncertainty state of neighbor node at time ; represents the uncertainty state of the edge connecting node and node at time ; is the propagation function of uncertainty, describing how uncertainty propagates from one node to another through edges.

[0147] The calculation process of the uncertainty propagation dynamics equation is:

[0148] First, the uncertainty evolution term of the node itself is calculated, which is derived from the uncertainty evolution function of the node processing the current uncertainty state, reflecting the natural change trend of the node uncertainty;

[0149] Then, the uncertainty propagation influence from all neighboring nodes is calculated, which requires applying the propagation function to each neighboring node and accumulating all results;

[0150] Finally, the two results are added together to get the rate of change of the node uncertainty state over time.

[0151] Among them, the uncertainty evolution function of the node captures the changes caused by internal factors such as device aging and load fluctuation; the propagation function captures the influence caused by external factors such as the uncertainty transmission of adjacent nodes.

[0152] This equation comprehensively considers both internal evolution and external propagation, and can fully describe the spatio-temporal evolution law of uncertainty in power systems.

[0153] Step 3.3, parameter learning and model calibration;

[0154] Based on historical data, the parameter learning algorithm is used to optimize the parameters of the propagation dynamics model, so that the model prediction results are maximally matched with the historical observation data.

[0155] Parameter learning can be based on gradient descent, evolutionary algorithm or Bayesian optimization method, and the objective function is the minimization of prediction error.

[0156] In addition, in some embodiments, cross-validation method can be used to evaluate the generalization performance of the model to avoid overfitting problem.

[0157] As shown in Figure 4 , the performance of the graph structure perception method and the traditional method is compared from five dimensions of prediction accuracy, calculation efficiency, peak avoidance effect, grid stability and resource utilization. The graph structure perception method shows obvious advantages in all dimensions.

[0158] Step 4, based on the constructed propagation dynamics model, realize the uncertainty propagation algorithm of graph structure perception, and generate the spatio-temporal evolution trajectory of uncertainty in the network;

[0159] Specifically, the following sub-steps are included:

[0160] Step 4.1, spatio-temporal discretization;

[0161] Discretize the continuous propagation dynamics equation into a form suitable for computer solution, and determine the appropriate time step and spatial representation.

[0162] Discretization needs to balance the accuracy and efficiency of calculation, and adaptive time step strategy can be used to use smaller time step in the period of rapid change of uncertainty.

[0163] In addition, in some embodiments, a multi-resolution spatial discretization method can be used to use finer grids for key areas.

[0164] Step 4.2, graph neural network construction;

[0165] Based on the discretized dynamic equation, the structure of the graph neural network is constructed, which is used to efficiently realize the simulation calculation of uncertainty propagation.

[0166] Each layer of the graph neural network corresponds to a time step, and the inter-layer connection reflects the spatio-temporal propagation relationship of uncertainty.

[0167] The input of the network is the initial uncertainty state, and the output is the predicted future uncertainty state evolution trajectory.

[0168] Specifically, the graph neural network model of the present application contains the following structure and implementation details:

[0169] The graph neural network model adopts the message passing framework, which contains three main components: message generation function, message aggregation function and node state update function.

[0170] The message generation function generates messages according to the source node state, target node state and edge attribute, which can be represented as:

[0171] ;

[0172] Where, represents the uncertainty state of the source node , which contains all the feature information of the node; represents the uncertainty state of the target node , which contains all the feature information of the node; represents the state of the edge connecting the source node and the target node , which contains the physical properties and uncertainty information of the edge; represents the message transmitted from the source node to the target node , which contains the information of uncertainty propagation; represents the message generation function, which is used to convert the state of nodes and edges into propagation messages.

[0173] The message aggregation function aggregates all the neighbor node messages into a comprehensive information, which can be represented as:

[0174] ;

[0175] where, denotes the aggregated message of node , containing the comprehensive information collected from all neighbor nodes; denotes the aggregation function used to combine multiple messages into one comprehensive information; denotes the message passed from node to node , containing uncertainty propagation information; denotes the source node sending the message; denotes the neighbor set of node , containing all nodes directly connected to node ; the aggregation operation can be summation, average, maximum, etc., and the appropriate aggregation method is selected according to the specific application scenario.

[0176] The node state update function combines the current node state and the aggregated message to update the node state, which can be represented as:

[0177] ;

[0178] where, denotes the uncertainty state of node at the next time step ; denotes the uncertainty state of node at the current time step ; denotes the aggregated message of node , containing the comprehensive information collected from all neighbor nodes; denotes the node state update function used to calculate the new state of the node according to the current state and the aggregated message; denotes the discretized time step index.

[0179] where, the specific implementation of the message generation function is:

[0180] First, the source node state , the target node state and the edge state are mapped to the same hidden space through feature transformation;

[0181] Then these features are spliced or combined with weights;

[0182] Finally, a multi-layer perceptron is used for nonlinear transformation to generate the message vector.

[0183] Specifically, for different types of nodes and edges, different weight matrices are used for feature transformation, for example:

[0184] ;

[0185] where, is a message generation function, which is used to generate a message vector representing the uncertainty information from the source node to the target node according to the source node state , the target node state and the edge state ; is a learnable weight matrix of the source node features, which is used to transform the feature representation of the source node; is a learnable weight matrix of the target node features, which is used to transform the feature representation of the target node; is a learnable weight matrix of the edge features, which is used to transform the feature representation of the edge; represents the feature vector of the source node, which contains the uncertainty state information of the source node; represents the feature vector of the target node, which contains the uncertainty state information of the target node; represents the feature vector of the edge connecting the source node and the target node, which contains the physical characteristics and uncertainty information of the edge; MLP represents a multi-layer perceptron, which is a kind of feedforward neural network, used to perform nonlinear transformation on the weighted combined features to generate the final message representation.

[0186] The specific implementation of the message aggregation function is to collect messages for all neighbor nodes of the node , and then process them according to the aggregation strategy. Common aggregation strategies include summation, average, maximum, and weighted average. In this application, an attention weighted aggregation mechanism is adopted to calculate the weighted sum of neighbor node messages, and the weight coefficient reflects the influence degree of neighbor nodes on the center node.

[0187] The specific form is:

[0188] ;

[0189] where, represents a message aggregation function, which is used to combine multiple messages into a comprehensive information; represents the message from node to node , which contains uncertainty propagation information; represents the neighbor set of node , which contains all nodes directly connected to node ; is an attention coefficient, which represents the influence weight of node on node ; denotes the summation operation over all neighbor nodes of the node .

[0190] attention coefficient is calculated by the states of source and target nodes and the characteristics of edges, expressed as:

[0191] .

[0192] where, denotes the attention coefficient; denotes the uncertainty state of the source node , containing all feature information of the node; denotes the uncertainty state of the target node , containing all feature information of the node; denotes the state of the edge connecting the source node and the target node , containing the physical characteristics and uncertainty information of the edge; is an attention function used to calculate the correlation between nodes, which can be a multi-layer neural network or a simple similarity calculation; is a normalization function to ensure the sum of all attention coefficients is 1, making the weight distribution reasonable.

[0193] node state update function is implemented as follows: the current state of the node is combined with the aggregated message , and a new state is generated through an update network. The update network can be a gating unit (such as GRU or LSTM), residual connection, or a simple fully connected network.

[0194] In this application, a gating update mechanism is adopted, which can be represented as:

[0195] .

[0196] where, denotes the node state update function, used to update the node state according to the current state and the aggregated message; denotes the uncertainty state of the node at time step ; denotes the aggregated message of the node , containing the comprehensive information collected from all neighbor nodes; is an update gate that controls the proportion of preserving the original state and introducing new information, with a value range of 0 to 1; is a candidate state generated by the aggregated message and the original state ; denotes the proportion of the original state retained; denotes element-wise multiplication (Hadamard product), i.e., multiplying elements at corresponding positions; denotes the retained original state part; denotes the introduced new information part;

[0197] update gate and candidate state The calculation formulas are respectively:

[0198] ;

[0199] ;

[0200] where, denotes the update gate, controlling the proportion of the retained original state and the introduced new information, with a value range of 0 to 1; denotes the candidate state, i.e., new state information that may be used for updating; denotes the current uncertainty state of the node at time step ; denotes the aggregated message of the node , containing comprehensive information collected from all neighboring nodes; denotes concatenating the current state and the aggregated message into a vector; is a sigmoid activation function that maps input values to between 0 and 1, ensuring that the value of the update gate is within a reasonable range; is a hyperbolic tangent activation function that maps input values to between -1 and 1, ensuring that the value of the candidate state is within a normalized range; is a learnable weight matrix of the update gate, used to calculate the update proportion from the concatenated vector; is a learnable weight matrix of the candidate state, used to generate new state information from the concatenated vector.

[0201] In specific implementation, different types of nodes and edges are processed using different parameterized neural networks. For example, for power generation nodes, the message generation function considers power generation type, capacity, and fluctuation characteristics; for load nodes, it considers load type, power consumption mode, and demand response capability; for grid connection edges, it considers physical capacity and impedance characteristics. This type-aware design enables the model to more accurately capture the differentiated behavior of different nodes and edges in uncertainty propagation.

[0202] Optionally, in the case of limited computing resources, a parameter sharing strategy can be adopted, with the same type of nodes and edges using the same parameters to reduce model complexity.

[0203] Specific application example in a power grid system with high renewable energy penetration:

[0204] In a certain regional power grid, there are multiple wind farms, photovoltaic power stations, and traditional generating units, as well as various types of load centers. The output of wind farms and photovoltaic power stations is greatly uncertain due to weather conditions. The system first constructs a heterogeneous graph network, with nodes including power generation units (wind farms, photovoltaic power stations, traditional power plants), load centers (industrial, commercial, residential), and key nodes of the transmission network (substations, distribution centers); edges include physical transmission lines and data correlation connections.

[0205] Through the graph neural network model, the system can simulate how the uncertainty caused by changes in specific weather conditions (such as increased cloud coverage leading to reduced photovoltaic output) propagates in the power grid. For example, when a certain photovoltaic power station experiences a drop in output due to cloud coverage, the uncertainty propagates through the graph structure to the connected distribution centers, and then affects the load centers and traditional power plants that need to compensate for power generation. Through multiple rounds of message passing, the system accurately captures the diffusion path and impact of uncertainty, and predicts the uncertainty state changes of each node in the future period.

[0206] Based on these predictions, the system generates robust peak avoidance strategies, such as scheduling traditional power plants in advance to increase reserve capacity, activating demand response resources in certain areas, adjusting the power grid topology to isolate uncertainty propagation, etc. In practical applications, compared with traditional probabilistic prediction methods, this method reduces the peak load prediction error caused by extreme weather events by 35%, significantly improving the effectiveness and effectiveness of peak avoidance measures.

[0207] Step 4.3, parallel computing optimization;

[0208] For large-scale power system graph structures, parallel computing and graph partitioning techniques are used to optimize algorithm execution efficiency.

[0209] Community detection and partitioning of the graph are performed so that strongly interacting nodes are placed in the same subgraph, reducing communication overhead between subgraphs while ensuring the accuracy of the calculation results.

[0210] As shown in Figure 5 , the trend of peak avoidance effectiveness of traditional methods and graph structure-aware methods is shown as the penetration rate of renewable energy increases. When the penetration rate reaches 60%, the peak avoidance effectiveness of traditional methods drops significantly to 41%, while the graph structure-aware method still maintains high performance of 85%.

[0211] Step 5, based on the generated uncertainty spatiotemporal evolution trajectory, construct a graph structure-aware robust optimization framework, and generate a multi-objective peak avoidance decision scheme that adapts to uncertainty;

[0212] Specifically, the following sub-steps are included:

[0213] Step 5.1, uncertainty impact assessment;

[0214] The impact of uncertainty propagation on key indicators of the power system (such as load balancing, network stability, peak avoidance effect, etc.) is analyzed, and a mapping relationship between the uncertainty state and system performance is established.

[0215] The impact assessment can be based on Monte Carlo simulation, scenario analysis or sensitivity analysis methods.

[0216] It should be noted that different evaluation methods can be selected for different performance indicators, for example, probability distribution method can be used for load balancing, and extreme value analysis method can be used for network stability.

[0217] Step 5.2, robust optimization model construction;

[0218] Based on the impact assessment results, a multi-objective robust optimization model considering uncertainty propagation is constructed. The objective function of the optimization model includes multiple dimensions such as peak avoidance effect, system stability and economic cost, and the constraint conditions include physical constraints, operation constraints and uncertainty constraints.

[0219] The model form can be robust linear programming, robust quadratic programming or robust mixed integer programming, which is selected according to the problem complexity and solution efficiency requirements.

[0220] Specifically, the robust optimization model of the present application contains the following implementation details:

[0221] The robust optimization model uses an adaptive uncertainty set method based on graph structure, which is different from the traditional interval or ellipsoid uncertainty set representation method. The model dynamically constructs the uncertainty set according to the propagation characteristics of uncertainty in the graph structure. The core idea of the model is to represent uncertainty as a multi-level propagation process on the graph. The uncertainty of each node is not only affected by its inherent characteristics, but also affected by the influence of adjacent nodes through the edges.

[0222] The mathematical form of the model is a multi-objective robust optimization problem:

[0223] ;

[0224] ;

[0225] ;

[0226] ;

[0227] where, is the decision variable vector, representing the controllable parameters of the system, such as generation scheduling plan, demand response activation strategy, load transfer amount, etc. is the uncertain parameter vector, representing the uncertain factors in the system, such as renewable energy output, user load changes, etc. is the graph structure based on the constructed uncertainty set, containing all possible combinations of uncertain parameters, reflecting the propagation characteristics of uncertainty in the graph network; , , represent the first , , objective function, respectively, representing different optimization objectives, such as peak load minimization, economic cost minimization, system stability maximization, etc. is the total number of objective functions, representing the number of objectives that need to be optimized simultaneously; is the inequality constraint function, representing the inequality constraint conditions that the system must satisfy, such as load balance constraints, network capacity constraints, etc. is the total number of inequality constraints, representing the number of inequality constraint conditions that the system needs to satisfy; is the equality constraint function, representing the equality conditions that the system must satisfy accurately, such as generator physical limitations, node power balance, etc. is the total number of equality constraints, representing the number of equality constraint conditions that the system needs to satisfy; is the feasible region of decision variables, representing the value range and constraint conditions of decision variables; is the minimax optimization structure, representing the optimal decision scheme under the most unfavorable uncertainty, ensuring the robustness of the decision.

[0228] The solution process of the multi-objective robust optimization problem is as follows:

[0229] First, determine the decision variables, i.e., the controllable parameters of the system, such as generation scheduling plan, demand response activation strategy, etc.

[0230] Then determine multiple optimization objectives, such as peak load minimization, cost minimization, system stability maximization, etc.

[0231] Next, construct the uncertainty set based on the graph structure, representing the possible uncertainty situations in the system;

[0232] Then, under all possible uncertainty situations, find the optimal decision scheme, so that under the most unfavorable uncertainty, the performance of multiple objective functions is still the best.

[0233] This "minimax" optimization approach ensures the robustness of the decision, as the system can maintain good performance even under the most unfavorable uncertainty. At the same time, the decision must satisfy multiple constraints, including uncertainty-related inequality constraints (such as load balancing, network capacity, etc.) and deterministic equality constraints (such as generator physical limitations, etc.).

[0234] This multi-objective robust optimization model based on graph structure can fully consider the propagation characteristics of uncertainty and generate peak avoidance decision schemes that are highly adaptable to uncertainty.

[0235] Uncertainty set The construction method of the uncertainty set is a key innovation of this model:

[0236] First, based on the uncertainty propagation dynamics model of each node in the graph structure, a large number of possible uncertainty scenarios are generated;

[0237] Then, through dimensionality reduction techniques such as principal component analysis or manifold learning, the core features of these scenarios are extracted;

[0238] Finally, a compact uncertainty set representation with geometric and topological significance is constructed.

[0239] This graph-based uncertainty set construction method not only captures the complex correlation between uncertainties, but also avoids the problem of excessive conservatism in traditional methods.

[0240] In some embodiments, the degree of conservatism of the uncertainty set can also be adjusted according to the risk preference of the decision maker, achieving a balance between risk and return.

[0241] The solution algorithm uses a combination of multi-objective decomposition and surrogate optimization:

[0242] First, the multi-objective problem is converted into a series of single-objective problems through weighting or epsilon constraint method;

[0243] Then, for each single-objective problem, the interior point method or branch and bound method is used to solve its robust dual problem;

[0244] Finally, through multiple iterations, the Pareto optimal solution set is generated, and the final solution is selected based on the decision maker's preference.

[0245] In addition, for large-scale problems, heuristic or approximation algorithms can be optionally used to obtain near-optimal solutions within a limited time.

[0246] Specific application examples in large-scale user behavior management scenarios:

[0247] In a certain city power system, containing millions of household users, tens of thousands of commercial users and thousands of industrial users, the system needs to manage the load during peak electricity consumption. The uncertainty of user behavior constitutes a complex network structure: there are behavior correlations between different user groups (such as the opening time of a shopping mall affecting the electricity consumption of surrounding residents), time-dependent behavior in different time periods (such as air conditioner use related to the temperature in the previous period), and spatial correlation between users in different regions (such as similar communities having similar electricity consumption patterns).

[0248] The system first constructs a heterogeneous graph network of user behavior, where nodes represent different types of user groups and edges represent behavior correlations. Based on historical electricity consumption data and socio-economic information, the system quantifies the inherent uncertainty of each node and the propagation strength of the edge. For example, analysis shows that the opening hours of commercial areas will affect the morning and evening peak electricity consumption patterns of surrounding residential areas through employee commuting; production adjustments in industrial parks will affect the electricity consumption behavior of upstream and downstream enterprises through supply chain relationships.

[0249] Using a robust optimization model with graph structure awareness, the system formulates a multi-objective peak avoidance strategy: minimizing peak load, minimizing user inconvenience, while ensuring that the power grid safety constraints are met.

[0250] The optimization result is a series of differentiated demand response schemes, such as implementing peak-shifting operation incentives for key commercial users, implementing tiered electricity prices for high-consumption residents, and providing auxiliary service subsidies for flexible industrial loads.

[0251] These schemes are fine-tuned according to the location and influence of users in the graph network, giving priority to activating key nodes that not only have large electricity consumption themselves, but also have a significant impact on the behavior of other users.

[0252] Step 5.3, multi-objective weight adaptive adjustment;

[0253] According to the decision preferences and risk tolerance under different scenarios, the weights of each objective in the multi-objective optimization are dynamically adjusted to achieve flexibility and adaptability of decision-making.

[0254] Weight adjustment can be achieved based on hierarchical analysis, fuzzy comprehensive evaluation or machine learning methods, reflecting the decision maker's priority considerations under different circumstances.

[0255] It should be noted that the weight adjustment process can be automated or semi-automated, allowing human-machine interaction to incorporate the professional judgment of decision makers.

[0256] For example, Figure 6As shown, the graph structure perception method demonstrates a percentage improvement in multiple technical indicators compared to traditional methods, including a 40% improvement in prediction accuracy, a 60% reduction in model parameters, a 250% improvement in computing efficiency, a 30% improvement in grid stability, a 25% improvement in resource utilization, a 20% reduction in overall system uncertainty, and a 45% reduction in recovery time.

[0257] Step 6, based on the results of the robust optimization framework, implement a multi-level risk zoning management strategy, and adopt differentiated management strategies for different risk level areas;

[0258] Specifically, the following sub-steps are included:

[0259] Step 6.1, uncertainty aggregation point identification;

[0260] Based on the simulation results of the uncertainty propagation model, identify the key nodes in the power grid that are prone to accumulate uncertainty. These nodes are usually located in special topological positions of the network, such as high centrality nodes, key connection points, or highly connected community boundaries.

[0261] The identification algorithm can be based on centrality measures, clustering analysis, or anomaly detection methods. Optionally, during the identification process, the physical characteristics and historical failure data of the nodes can also be considered to improve the accuracy of identification.

[0262] Step 6.2, risk level division;

[0263] According to the influence range and intensity of uncertainty propagation, the power grid is divided into different risk level areas, and differentiated management strategies are adopted for different risk levels.

[0264] The level division can be achieved based on community detection, hierarchical clustering or spectral clustering methods, ensuring that nodes within the same level have similar risk characteristics.

[0265] In addition, in practical applications, risk level division may need to consider geographical location, administrative division and other factors to facilitate the implementation of management measures.

[0266] Step 6.3, differentiated peak avoidance strategy generation;

[0267] For different risk level areas, generate customized peak avoidance management strategies, including resource allocation, early warning mechanisms and emergency response measures, etc.

[0268] High-risk areas adopt more conservative peak avoidance strategies and stricter monitoring measures, while low-risk areas can adopt more flexible strategies to improve economic efficiency.

[0269] In some embodiments, the differentiated strategy can also consider the time dimension, adopting more conservative strategies during periods of high uncertainty and appropriately relaxing constraints during stable periods.

[0270] Step 7. In the implementation of risk zoning management process, the complementary effect of uncertainty is discovered and utilized to reduce the overall risk of the system;

[0271] Specifically, the following sub-steps are included:

[0272] Step 7.1, complementary pattern recognition;

[0273] Analyze the correlation and complementarity of different uncertainty sources (such as wind power, photovoltaic, load, etc.) in time and space dimensions, and identify possible negative correlation patterns.

[0274] Complementary pattern recognition can be based on correlation analysis, principal component analysis or deep learning methods, aiming to find the complementary rules in time and space.

[0275] Optionally, complementary pattern recognition can also consider external factors such as seasonality and weather conditions to discover conditional complementary relationships.

[0276] Step 7.2, complementary effect quantification;

[0277] Quantitative evaluation of the degree of reduction of complementary effect on system overall uncertainty, and establishment of the relationship model between complementary pattern and system risk. Quantitative method can be based on risk measurement, information theory or statistical variance analysis, to provide quantitative basis for decision-making.

[0278] In addition, complementary effect quantification can be static or dynamic, and the evaluation result can be updated according to real-time conditions.

[0279] Step 7.3, complementary resource collaborative scheduling;

[0280] Based on the identified complementary pattern and quantified complementary effect, optimize resource allocation and scheduling strategy, maximize the utilization of complementary effect, and reduce the overall risk of the system.

[0281] Collaborative scheduling can be achieved through model predictive control, reinforcement learning or combinatorial optimization methods, while meeting system constraints and maximizing complementary benefits.

[0282] In some embodiments, collaborative scheduling can be combined with power market mechanisms to guide the collaborative operation of different resources through price signals.

[0283] The above describes embodiments of the present application, but the embodiments are not limited to the specific embodiments described above, which are only illustrative and not limiting. Those skilled in the art can make more forms of equivalent embodiments under the inspiration of the embodiments, which are all within the protection scope of the embodiments.

Claims

1. A peak-shaving method for electricity management based on multi-objective programming, characterized in that, Includes the following steps: Construct a heterogeneous graph network structure for the power system, where nodes represent entities in the power system and edges represent physical connections or data dependencies between entities; Based on the constructed heterogeneous graph network, an uncertainty representation model on the graph structure is defined, and an uncertainty state space is defined for the nodes and edges in the graph; Using the defined uncertainty characterization model, an uncertainty propagation dynamics model is constructed to describe the diffusion process of uncertainty in the network; Based on the constructed propagation dynamics model, an uncertainty propagation algorithm with graph structure awareness is implemented to generate the spatiotemporal evolution trajectory of uncertainty in the network; Based on the generated uncertain spatiotemporal evolution trajectory, a robust optimization framework with graph structure awareness is constructed to generate a multi-objective peak avoidance decision scheme that adapts to uncertainty. Based on the results of the robust optimization framework, a multi-level risk zoning management strategy is implemented, and differentiated management strategies are adopted for regions with different risk levels. In the process of implementing risk zoning management, we can identify and utilize the complementary effects of uncertainty to reduce the overall risk of the system.

2. The peak-shaving method for electricity management based on multi-objective programming according to claim 1, characterized in that, The steps for constructing the heterogeneous graph network structure of the power system specifically include: Node identification and classification involves identifying entities in a power system and classifying them into different types of nodes based on their functional characteristics. Edge relationship construction and weighting: Based on physical connection relationships and data correlation, edge connections between nodes are constructed; Graph structure optimization involves optimizing the graph structure using graph sparsification techniques and importance sampling methods, and removing edges that have the least impact on the propagation of uncertainty.

3. The peak-shaving method for electricity management based on multi-objective programming according to claim 1, characterized in that, The steps of defining the uncertainty representation model on the graph structure specifically include: Uncertainty state space construction: Define the uncertainty state space for each node and edge in the graph; Initial uncertainty quantification involves assigning initial uncertainty state values ​​to the uncertainty source nodes in the graph based on historical data and expert knowledge. Uncertainty evolution function construction: Defines the function that describes the evolution of the uncertain state of nodes and edges over time.

4. The peak-shaving method for electricity management based on multi-objective programming according to claim 1, characterized in that, The specific steps for constructing the uncertainty propagation dynamics model include: The propagation function is constructed by creating a function that allows uncertainty to propagate between nodes, based on node type and edge characteristics. Propagation dynamics modeling integrates node evolution functions and propagation functions to construct a complete uncertainty propagation dynamics equation; Parameter learning and model calibration: Based on historical data, parameter learning algorithms are used to optimize the parameters of the propagation dynamics model.

5. The peak-shaving method for electricity management based on multi-objective programming according to claim 1, characterized in that, The steps for implementing the uncertainty propagation algorithm for graph structure awareness specifically include: Spatiotemporal discretization transforms the continuous propagation dynamics equations into a form suitable for computer solving. Graph neural network construction: Based on discretized dynamic equations, a graph neural network structure is constructed for efficient simulation calculation of uncertainty propagation. Parallel computing optimization is employed for large-scale power system graph structures, using parallel computing and graph partitioning techniques to optimize algorithm execution efficiency.

6. The peak-shaving method for electricity management based on multi-objective programming according to claim 5, characterized in that, Graph neural networks include message generation functions, message aggregation functions, and node state update functions; The message generation function generates a message based on the source node state, the target node state, and edge attributes; The message aggregation function aggregates messages from all neighboring nodes into a single, comprehensive message. The node state update function combines the current node state with the aggregated message to update the node state.

7. The peak-shaving method for electricity management based on multi-objective programming according to claim 1, characterized in that, The specific steps for constructing a graph structure-aware robust optimization framework include: Uncertainty impact assessment, analyzing the impact of uncertainty propagation on key power system indicators; Robust optimization model construction: Construct a multi-objective robust optimization model that considers the propagation of uncertainty; The weights of the multi-objectives are adaptively adjusted, dynamically adjusting the weights of each objective in the multi-objective optimization based on decision preferences and risk tolerance in different scenarios.

8. A peak-shaving method for electricity management based on multi-objective programming according to claim 7, characterized in that, Multi-objective robust optimization models include: Definitions of decision variables and uncertain parameters: Decision variables represent the controllable parameters of the system, while uncertain parameters represent the uncertainties existing in the system; The construction of a multi-objective function considers multiple optimization objectives simultaneously, including cost minimization, risk minimization, and user satisfaction maximization. The setting of robust constraints ensures that the system operates in accordance with safety and reliability requirements under all preset uncertainties; The decision space is defined by clarifying the range of values ​​and constraints of the decision variables; Based on the construction of uncertain sets in graph structures, the possible range of values ​​for uncertain parameters is defined by utilizing the structural characteristics of heterogeneous graph networks in power systems.

9. A peak-shaving method for electricity management based on multi-objective programming according to claim 1, characterized in that, The steps for discovering and utilizing the complementary effect of uncertainty specifically include: Complementary pattern recognition analyzes the correlation and complementarity of different uncertainty sources in time and space dimensions; Quantitative analysis of complementary effects: quantitatively assessing the degree to which complementary effects reduce the overall uncertainty of the system. Complementary resources are coordinated and scheduled to optimize resource allocation and scheduling strategies and maximize the utilization of complementary effects.

10. A computer-readable storage medium, characterized in that, It is used to store computer-readable instructions, which, when read by a computer, enable the execution of a power consumption management peak-shaving method based on multi-objective planning as described in any one of claims 1-9.

Images