Distributed source-load collaborative optimization method based on high-order topology and multi-scale attention

Abstract

The invention relates to a distributed source-load collaborative optimization method based on high-order topology and multi-scale attention, and the method comprises the steps: firstly providing a high-order graph construction method driven by structural interaction, and achieving the structural embedded expression of a physical interaction relation between multi-source equipment through a hyperedge-line graph mapping mechanism and functional attribute coding; secondly, a graph feature extraction method based on a multi-scale joint attention mechanism is designed, topology and state information are fused, and the inter-node adjustment collaboration recognition capability is improved; further constructing a source-load collaborative optimization scheduling model, introducing a particle swarm optimization algorithm to obtain an initial feasible strategy, and establishing a state-action mapping relation based on a deep reinforcement learning framework driven by graph embedding to realize autonomous learning and rolling optimization of a distributed control strategy; and finally, constructing an operation feedback closed loop mechanism, and introducing a graph structure migration and strategy adaptive updating method to enhance the response capability of the system to topological change and dynamic disturbance.

Description

Technical Field

[0001] This invention relates to the field of power system regulation and intelligent management of distributed energy, and in particular to a distributed source-load collaborative optimization method based on high-order topology and multi-scale attention. Background Technology

[0002] With the large-scale integration of distributed renewable energy and the continuous increase in demand for source-load interaction, renewable energy-dominated power systems exhibit operational characteristics of "multi-source heterogeneity, strong coupling, and high dynamism." Source-load collaborative optimization, as a core means to ensure system power balance and stable regulation, has received widespread attention in distributed energy system dispatching, regional energy internet management, and multi-energy complementarity scenarios. Among these challenges, achieving efficient and intelligent collaborative optimization control in the context of strong equipment heterogeneity, complex interaction mechanisms, and dynamically changing system structures has become a key difficulty restricting the operational efficiency and regulatory resilience of renewable energy systems.

[0003] First, current source-load-storage equipment is diverse, with significant differences in operating characteristics, control interfaces, and response mechanisms, encompassing devices with varying response times and control precisions, including photovoltaic, wind power, energy storage, electric loads, and flexible loads. In source-load collaborative modeling, traditional methods often employ centralized or static mathematical models, which struggle to effectively describe the dynamic coupling and interaction relationships between various devices. They also lack structural modeling support for physical mechanisms such as control paths and operational attributes, resulting in poor strategy generalization and low model scalability. Second, the device connection relationships and communication structures in the system frequently evolve. Traditional methods based on fixed topology or global adjacency matrices cannot handle dynamic scenarios such as device online / offline operations and local topology changes. While some graph neural network methods have attempted to incorporate structural information between devices, they are generally based on first-order adjacency graph modeling, making it difficult to capture potential multiple adjustment links, higher-order collaborative relationships, and functional connection patterns between sources and loads. This results in insufficient structural awareness, limiting the ability to express node features and generate global strategies. Furthermore, in real-world systems, the control capabilities and perception granularity of various devices differ significantly, making it difficult to unify adjustment timing and behavior. Centralized synchronous scheduling mechanisms are prone to communication bottlenecks and untimely state updates. While some reinforcement learning methods possess policy self-learning capabilities, they lack a modeling foundation that synergizes with graph structures, resulting in slow policy convergence and poor generalization, making it difficult to support efficient policy transfer and structural adaptation in distributed environments.

[0004] In summary, existing source-load collaborative optimization techniques still have many limitations in structural modeling and representation, control strategy learning, and multi-node adjustment execution. There is an urgent need for a distributed control method that integrates high-order graph structural modeling, structurally aware feature extraction, and reinforcement strategy learning to achieve generalized structural modeling, intelligent strategy generation, and asynchronous collaborative adjustment of multiple devices in complex interactive environments, thereby improving the overall robustness, adaptability, and control efficiency of the system.

[0005] A search revealed Chinese invention patent CN118783486A, which discloses a method and system for joint scheduling of load aggregators and distributed energy storage, relating to the field of energy storage scheduling technology. The method includes establishing a physical model of the load aggregator and constructing a joint scheduling model; constructing a reward function to update operating status information; constructing a distributed near-end strategy optimization algorithm; using historical source-load data to train the joint scheduling model offline to obtain the optimal scheduling decision model; and inputting real-time source-load data into the model for online learning and updating, enabling joint real-time scheduling of load aggregators and distributed energy storage resources. This invention helps to achieve optimal resource utilization, improve the adaptability and efficiency of decision-making, enable the system to respond quickly to changes in supply and demand, and reduce the risk of system imbalance caused by scheduling delays.

[0006] The technical comparison between this application and the aforementioned patent is as follows:

[0007] 1. The patent "A Joint Scheduling Method and System for Load Aggregators and Distributed Energy Storage" is based on Markov decision processes and near-end policy optimization algorithms. It constructs a unified policy network to globally optimize and train the joint scheduling scheme of load aggregators and distributed energy storage, outputting a charging and discharging power policy at the aggregation level to minimize costs and improve demand response efficiency. In contrast, this patent is based on a high-order heterogeneous graph structure and multi-scale attention mechanism, combined with node-level reinforcement learning strategies and graph neural network modeling methods, to construct a source-load collaborative optimization framework with structure awareness and adaptive capabilities. This enables node-level physical action policy learning and distributed collaborative control of multiple types of equipment. The two patents differ fundamentally in their technical approaches.

[0008] 2. The patent "A Joint Scheduling Method and System for Load Aggregators and Distributed Energy Storage" outputs charging and discharging power commands at the load aggregator level through an end-to-end unified strategy network. Its reinforcement training objective is to maximize overall returns under constraints of operating costs and user response. In contrast, this patent treats distributed source, load, and storage nodes as heterogeneous intelligent agents, employing a multi-agent reinforcement learning strategy guided by graph neural networks to jointly learn the local physical control strategies of each node. It then achieves system-level collaborative scheduling and state synchronization through a graph structure perception mechanism. The two patents differ fundamentally in their algorithmic structure and action modeling dimensions.

[0009] A search revealed that Chinese invention patent CN112713618A discloses a method for coordinated optimization of the source-grid-load-storage operation of an active distribution network based on multi-scenario technology. First, based on multi-scenario technology, the active power output of distributed renewable energy generation is divided into scenarios, thus considering the uncertainty of the active power output of distributed renewable energy generation. Second, an active distribution network source-grid-load-storage coordinated optimization model is established with the goal of minimizing the daily comprehensive operating cost of the active distribution network. Building upon the above model that only considers source-load-storage coordinated optimization, the coordinated optimization operation of network reconfiguration, the reactive power output of distributed renewable energy generation inverters and energy storage inverters, and the active power output of distributed renewable energy generation is also considered, thus achieving coordinated optimization of the active distribution network source-grid-load-storage operation. Finally, the active distribution network source-grid-load-storage coordinated optimization model is combined with a particle swarm optimization algorithm to solve a mixed-integer programming problem.

[0010] The technical comparison between this application and the aforementioned patent is as follows:

[0011] The two patents differ fundamentally in their technical approaches, core algorithms, and target objectives.

[0012] In terms of technical modeling, the patent "An Active Distribution Network Source-Grid-Load-Storage Coordinated Optimization Operation Method Based on Multi-Scenario Technology" is based on multi-scenario probabilistic modeling, which transforms the uncertainty of distributed energy output into typical scenarios with probabilities, focusing on scenario quantification and probability weighting; while this patent adopts high-order heterogeneous graph modeling, which uses hyperedge-line graph mapping to structurally embed high-order interaction relationships such as physical connections and control dependencies between devices, breaking through the limitations of traditional first-order topology.

[0013] In terms of core algorithms, the patent "An Active Distribution Network Source-Grid-Load-Storage Coordinated Optimization Operation Method Based on Multi-Scenario Technology" uses an improved particle swarm algorithm to solve mixed integer programming problems and performs static optimization based on a preset mathematical model, which belongs to model-driven mathematical optimization; while this patent mainly uses graph neural networks combined with deep reinforcement learning, and realizes policy autonomous learning and dynamic optimization through state-action mapping, which belongs to data-driven autonomous learning.

[0014] In terms of learning objectives, the patent "An Active Distribution Network Source-Grid-Load-Storage Coordinated Optimization Operation Method Based on Multi-Scenario Technology" aims to minimize the comprehensive daily operating cost, focusing on achieving coordinated economic scheduling of source, grid, load, and storage under uncertainty, and pursuing cost optimization and constraint satisfaction; while this patent focuses on improving the system's adaptability and generalization ability to dynamic topology and complex interactions, pursuing the robustness and autonomy of the control strategy.

[0015] In summary, the patent "An Active Distribution Network Source-Grid-Load-Storage Coordinated Optimization Operation Method Based on Multi-Scenario Technology" is a static scheduling paradigm of "scenario quantification + mathematical optimization," relying on model-driven solutions to find the cost-optimal solution under known scenario distributions. This patent, on the other hand, is a dynamic optimization paradigm of "structure perception + autonomous learning," relying on data-driven learning of optimal strategies from complex interactions. There are significant differences between the two.

[0016] A search revealed Chinese invention patent CN119543275A, which discloses a distribution network reconfiguration method under the collaborative optimization of source-load uncertainty and demand response, incorporating energy storage. The method involves constructing distribution network topology information; establishing a probability model for distributed generation output and load fluctuations, considering time-series fluctuation characteristics and their impact on grid operation; handling uncertainties in distributed generation and loads; constructing a distribution network reconfiguration optimization model that considers demand response and source-load characteristics; and introducing second-order conical convex relaxation techniques and linearization methods into the constructed distribution network reconfiguration optimization model to transform the original model into a form with a convex feasible region for solution. This method aims to address the shortcomings of traditional methods in considering the access and uncertainties of distributed generation. By optimizing the configuration and scheduling of energy storage resources, the flexibility and stability of the distribution network can be effectively improved, thereby enhancing the grid's ability to absorb renewable energy and strengthening the system's adaptability to time-series fluctuations. It compensates for the deficiencies of traditional methods in considering changes in distribution network structure and load differences.

[0017] The technical comparison between this application and the aforementioned patent is as follows:

[0018] 1. The patent "A Distribution Network Reconfiguration Method under the Co-optimization of Source-Load Uncertainty and Demand Response with Energy Storage" is aimed at the reconfiguration of distribution networks with energy storage. Its core is the co-optimization of source-load uncertainty and demand response. It establishes a probability model of distributed power output and load fluctuation, uses Latin hypercube simulation and improved K-means clustering to handle uncertainty, and constructs a distribution network reconfiguration optimization model considering demand response. It introduces a second-order cone-convex relaxation technique for solution, focusing on improving the distribution network's ability to absorb renewable energy and its operational economy. It is suitable for distribution network topology adjustments and load dispatching scenarios with a high proportion of distributed power sources. This patent focuses on distributed source-load co-optimization, with its core being dynamic control achieved through high-order topology modeling and autonomous learning. It constructs a multi-view high-order heterogeneous graph integrating physical connections and functional relationships, extracts cross-dimensional features using a multi-scale joint attention mechanism, and then relies on a graph embedding-driven deep reinforcement learning framework to achieve autonomous learning and rolling optimization of control strategies. It focuses on solving the dynamic coupling and complex interaction problems of multi-source heterogeneous devices, and is suitable for scenarios with frequent device access and disconnection and dynamic topology changes in distributed power systems dominated by new energy sources. The two differ significantly in their core technologies and application scenarios.

[0019] 2. The essential differences between the two in terms of methods and algorithms:

[0020] In terms of uncertainty handling, the patent "A method for reconfiguring a distribution network under the collaborative optimization of source-load uncertainty and demand response with energy storage" is based on a probability model and multi-scenario clustering to transform uncertainty into a deterministic scenario solution; while this patent dynamically responds to topology changes and disturbances through graph structure migration and strategy adaptive update mechanism, without relying on a preset probability distribution.

[0021] In terms of optimization algorithms, the patent "A Distribution Network Restructuring Method under the Co-optimization of Source-Load Uncertainty and Demand Response with Energy Storage" uses a second-order cone-convex relaxation and linearization method to transform the nonlinear model into a convex feasible region, and solves it through mathematical programming, which belongs to model-driven static optimization; while the patent uses graph neural networks to extract topological features and combines deep reinforcement learning to achieve policy autonomous learning, which belongs to data-driven dynamic optimization.

[0022] Regarding control objectives, the patent "A Distribution Network Reconfiguration Method under the Coordinated Optimization of Source-Load Uncertainty and Demand Response with Energy Storage" aims to minimize the overall operating cost of the distribution network and focuses on achieving economic dispatch through demand response and topology reconfiguration; while this patent focuses on improving the system's adaptive ability to dynamic changes and pursues the robustness and synergy of the control strategy. Summary of the Invention

[0023] To address the challenges of complex interactions among heterogeneous multi-source devices in distributed energy systems, insufficient structural modeling capabilities, and a lack of adaptive control strategies, this invention discloses a distributed source-load collaborative optimization method based on high-order topology and multi-scale attention. By constructing a heterogeneous graph structure with high-order interaction expression capabilities, and combining a topology-aware graph attention feature extraction mechanism with a graph embedding-driven reinforcement strategy learning framework, this invention effectively enhances the system's control autonomy and operational coordination in scenarios such as dynamic structural evolution, asynchronous response of multiple devices, and intelligent strategy generation, thereby achieving intelligent control and efficient operation of renewable energy-dominated distributed power systems.

[0024] To achieve the above objectives, the technical solution adopted by the present invention is as follows:

[0025] A distributed source-load collaborative optimization method based on high-order topology and multi-scale attention is characterized by the following steps:

[0026] S1: Based on the physical connection, control dependency and interaction behavior information in the distributed source-load-storage system, construct a multi-view high-order heterogeneous graph that integrates structural topology and functional relationship;

[0027] S2: Based on the multi-view high-order heterogeneous graph constructed by S1, a multi-dimensional joint attention mechanism is designed to integrate multi-scale spatial, temporal dynamics and semantic heterogeneous information.

[0028] S3: Based on the learning results of the multi-view high-order heterogeneous graph constructed by S1 and the multi-dimensional joint attention mechanism of S2, an optimization scheduling model driven by power gap identification, compensation allocation and optimization objectives is constructed.

[0029] S4: Based on the joint power scheduling modeling problem of multi-view high-order heterogeneous graphs in S3, a graph embedding-driven deep reinforcement learning solution framework is designed to construct state, action, and reward triples to obtain distributed source-load control policies.

[0030] S5: Based on the distributed source-load control strategy obtained in S4, a coordination control mechanism for heterogeneous node types is constructed to ensure that each control unit can independently execute strategy instructions, and to achieve system-level consistency and responsiveness through a graph structure;

[0031] S6: Based on the real-time monitoring results of the system's operating status, construct an operation feedback and adaptive adjustment module to dynamically correct the control strategy and graph structure representation.

[0032] As a preferred technical solution of the present invention, step S1 includes the following steps:

[0033] S11, Construction of Hyperedge-Line Graph Dual Mapping Mechanism

[0034] Define the system node set as V = {v1, v2, ..., v...} N}, and introduce a set of superedges. And define the hyperedge mapping function:

[0035]

[0036] Map hyperedges to virtual nodes in the line graph. Constructing a line graph ζ l =(V l ,δ l ),in:

[0037]

[0038] Edge set:

[0039]

[0040] S12, Structural Function Attribute Coding

[0041] For each node v i 、side (v i ,v j Assigning attribute vectors to the nodes and edges, and using the attributes of nodes and edges together as input features of the graph neural network, achieves unified encoding of structure and function;

[0042] S13, Multi-subgraph Modeling

[0043] Based on functional attributes, the high-order heterogeneous graph is split into subgraphs, as follows:

[0044] S131, Source-Load Energy Supply Sub-diagram ζ (1) =(V s ∪V l ,δ (1) ),

[0045] Where the boundary set δ (1) ={(v s ,v l )|v s ∈V s ,v l ∈V l} indicates the power supply path;

[0046] S132, Storage-Load Adjustment Subgraph ζ (2) =(V e ∪V l ,δ (2) ),

[0047] Where the boundary set δ (2) ={(v e ,v l ),(v l ,v e )|v e ∈V e ,v l ∈V l} represents the relationship between energy storage charging / discharging and load regulation;

[0048] S133, Source-Source Collaborative Subgraph ζ (3) =(V s ,δ (3) ),

[0049] edge set δ (3) This indicates the collaborative relationships between the sources;

[0050] S134. Combine the above subgraphs into a multi-view high-order heterogeneous graph:

[0051] ζ={ζ (1) ,ζ (2) ,ζ (3)}

[0052] As a preferred technical solution of the present invention, step S2 includes the following steps:

[0053] S21. Design a multi-scale spatial attention mechanism, as follows:

[0054] The formula for calculating the weight is:

[0055]

[0056] Through hierarchical weight normalization:

[0057]

[0058] S22. Design a multi-scale temporal attention mechanism, as follows:

[0059] By constructing temporal feature descriptions across multiple time windows, the temporal correlation weights are calculated respectively:

[0060] The normalized weights are:

[0061]

[0062] Ultimately, multi-scale temporal attention is achieved through weighted fusion:

[0063]

[0064] Where, β τ These are trainable parameters that reflect the contribution of different time scales to the dynamic dependence of nodes; S23, design a semantic heterogeneous attention mechanism, as follows:

[0065]

[0066] Normalized to semantic attention weights:

[0067]

[0068] S24. Design a cross-dimensional joint attention fusion strategy, as follows:

[0069]

[0070] Where, λ s ,λ t ,λ sem w is the global adjustment coefficient. l These are spatial hierarchy weights, all of which are trainable parameters. As a preferred embodiment of the present invention, step S3 includes the following steps:

[0071] S31. Design power gap detection and dynamic evaluation mechanism:

[0072] The real-time power gap ΔP of each load node is estimated by using the node features output by the graph neural network. i (t):

[0073]

[0074] ΔP(t) is the input of the dynamic compensation strategy, which reflects the power balance state of the system in real time;

[0075] S32. Based on the node functional attributes and system constraints, design a priority-based power compensation allocation strategy:

[0076] S33. Based on real-time power gap and power compensation allocation strategies, construct an optimized scheduling model with the objective function of minimizing the total system operating cost while satisfying power balance and operational constraints.

[0077]

[0078] Among them, P i (t) represents the adjustment power of node i at time t. The model constraints include:

[0079] Power balance constraints:

[0080]

[0081] Among them, P i (t) represents the adjustment power of node i at time t, ΔP net (t)=∑ i ΔP i (t) represents the total power deficit of the system;

[0082] Node power upper and lower limits:

[0083] P i min ≤P i (t)≤P i max ;

[0084] Energy storage state constraints:

[0085]

[0086] Among them, SOC i (t) represents the energy storage state, η i The charging / discharging efficiency is given by Δt, which is the time step.

[0087] S34. Design a heuristic solution strategy based on improved particle swarm optimization, which includes:

[0088] S341, Particle Encoding: The adjustment power of each node is used as the particle position vector;

[0089] S342. Fitness function: defined as the objective function J plus a penalty function for violating constraints;

[0090] S343, Dynamic Weight Adjustment: Combines real-time load and power gap to dynamically adjust algorithm parameters;

[0091] S344, Distributed Computing Architecture: Designing a parallel computing framework for large-scale networks.

[0092] As a preferred technical solution of the present invention, step S4 includes the following steps:

[0093] S41. Encode the node embedding features generated by the multi-view high-order heterogeneous graph in S1 through the multi-dimensional joint attention mechanism in S2. As the system state s t The components, while introducing auxiliary information:

[0094] s t =[H t ,ΔP net (t),SOC t V t ,L t ];

[0095] S42, Action Space A is defined as the set of adjustment operations for all controllable nodes, including photovoltaic output regulation, energy storage charging and discharging power adjustment, and the response amplitude of controllable loads:

[0096]

[0097] Action space A must satisfy the following physical constraints:

[0098] Photovoltaic regulation range:

[0099] Energy storage capacity and power limitations:

[0100] Load regulation rate limit:

[0101] S43. Design a multi-objective reward function that integrates volatility suppression and control cost minimization:

[0102]

[0103] S44. Employ a graph embedding-aware deep reinforcement learning framework, with state s t Input Policy Network π θ (a t |s t Output the optimal action a) t A heterogeneous graph feature residual channel plus state attention fusion structure is introduced, and the training objective is to maximize the cumulative discount reward:

[0104]

[0105] In addition, a strategy stability term J is introduced. stable :

[0106]

[0107] The final output scheduling strategy π * (a|s).

[0108] As a preferred technical solution of the present invention, step S5 includes the following steps:

[0109] S51, Control instruction mapping and node behavior driving, outputting action a from the reinforcement learning policy. t Mapped to physical control commands that can be executed by the node, including:

[0110] S511. Generate output adjustment command for the source node: u pv,i (t)=P pv,i (t);

[0111] S512, for energy storage nodes, decoding is performed to update charging / discharging power and SOC:

[0112] S513. Generate load adjustment commands for load nodes:

[0113] S514, the unified instruction is driven by the embedding results of the node's local state and global strategy, and supports the concurrent execution of heterogeneous control instructions;

[0114] S52. Design a coordination factor constraint mechanism based on graph structure propagation, and define the set of boundary nodes B at the intersection of sub-regions B and C. bd Their power coordination relationship is expressed as:

[0115]

[0116] Where κ is the allowable error tolerance, propagated through the edge weight propagation function φ(e) in the graph. ij Dynamic adjustments are made, and if the error exceeds the tolerance, a local boundary control mechanism is triggered to recalculate the instruction.

[0117] S53, By constructing a "graph propagation coordination graph" ctrl =(V,δ ctrl To achieve synchronous propagation of control states between different nodes, a residual consistency propagation function is used:

[0118]

[0119] Where η ij This represents the propagation coefficient between adjacent nodes.

[0120] As a preferred technical solution of the present invention, step S6 includes the following steps:

[0121] S61. Real-time collection of operational data from each node X t ={V i (t),I i (t),P i (t),SOC i (t)}, through the desired state By comparison, construct the error function:

[0122]

[0123] If the error exceeds the set threshold κ th Then it will enter the S61 retraining phase;

[0124] S62. Based on the error source classification results, design an adaptive adjustment strategy:

[0125] S621. If the error is concentrated in the node power control, fine-tune the adjustment cost term λ2 in the reward function.

[0126] S622. If voltage exceeds the limit frequently, increase the penalty weight in λ3.

[0127] S623. If the strategy has poor stability, increase the regularization coefficient γ of the strategy smoothing term.

[0128] S63. When the system topology changes, the graph structure migration mechanism is triggered. The migration mechanism includes:

[0129] S631, Graph Node Update: Dynamically adding or removing nodes and their connecting edges;

[0130] S632, Node Embedding Initialization: Initialize new nodes using the average embedding vector of nearest neighbor nodes.

[0131] Compared with the prior art, the beneficial effects of the present invention are as follows:

[0132] This invention proposes a source-load collaborative optimization method with high-order structure modeling capability, graph structure perception mechanism and reinforcement strategy learning capability, and constructs a distributed optimization control system with strong structural expressiveness, high level of strategy intelligence and adaptive operation.

[0133] First, addressing the lack of unified characterization of multi-dimensional physical relationships and regulatory functions in existing source-load collaborative modeling, this invention proposes a high-order heterogeneous graph modeling mechanism driven by structural interaction. It employs hyperedge-line graph mapping and multi-type edge construction to embed the multi-view functional interactions of source loads, storage loads, etc., into the graph structure, providing a learnable structural representation foundation for subsequent policy generation. Second, addressing the weak structure perception and limited feature fusion capabilities in traditional graph neural optimization strategies, this invention designs a graph embedding feature extraction method with a joint attention mechanism. It introduces a joint perception mechanism of node topological dependence and state dynamics, effectively improving policy generation. To improve the node collaboration capability and adjustment accuracy in the policy solution stage, a deep reinforcement learning optimization framework based on graph embedding is proposed. By combining state space construction, reward function design, and action generation modules, rolling learning and adaptive optimization of distributed source-load control policies are realized, exhibiting good generalization ability and adjustment response speed. Finally, considering the real-world factors such as frequent structural dynamic evolution and heterogeneous controller capabilities in the operating environment, this invention further introduces an operational feedback and graph structure transfer mechanism. Through error-driven policy retraining and continuous perception of graph topology, the robustness of the system to dynamic disturbances and its online optimization capability are improved. Attached Figure Description

[0134] Figure 1 This is a flowchart of the method of the present invention;

[0135] Figure 2 This is a schematic diagram of the semantic heterogeneous attention mechanism designed in this invention;

[0136] Figure 3 This is a schematic diagram of the graph embedding-driven deep reinforcement learning solution framework designed in this invention. Detailed Implementation

[0137] The present invention will be further described in detail below with reference to the accompanying drawings and specific embodiments:

[0138] like Figure 1 As shown, the distributed source-load collaborative optimization method based on high-order topology and multi-scale attention proposed in this invention includes the following steps:

[0139] S1. Based on the physical connections, control dependencies, and interaction behavior information in the distributed source-load-storage system, a multi-view high-order heterogeneous graph that integrates structural topology and functional relationships is constructed to realize learnable graph structure modeling between multiple types of nodes in the new energy scenario, supporting efficient information transmission and semantic aggregation of subsequent graph neural networks.

[0140] The specific implementation steps are as follows:

[0141] S11. Construction of the hyperedge-line graph dual mapping mechanism, defining the system node set as V = {v1, v2, ..., v...}N} represent various types of nodes, including wind power, photovoltaics, energy storage, and load; a superedge set is also introduced. Each of the super edges Connecting more than two nodes indicates a higher-order group interaction relationship.

[0142] To transform the hypergraph structure into an information propagation structure suitable for graph neural networks, a hyperedge mapping function is defined:

[0143]

[0144] Map hyperedges to virtual nodes in the line graph. Constructing a line graph ζ l =(V l ,δ l ),in

[0145]

[0146] Edge set

[0147]

[0148] This mechanism enables the transformation of high-order relations into learnable structures of graph neural networks, supporting deep representations of complex multi-node interactions;

[0149] S12, Structural action attribute encoding, for each node v i Assigning functional attribute vectors:

[0150] f i =[type i capacity i ,location i ,control-model i ]

[0151] Among them, type i ∈{PV,Wind,Storag,e Loa}d represents the node category; capacity i Rated capacity of the node; control-model i This indicates the adjustment mode of the node, opposite to the edge (v i ,v j Assigning the applied attribute vector:

[0152] s ij =[power-flow ij ,direction ij support-type ij ]

[0153] in, Indicates the direction of power flow; support-type ij ∈{voltage,frequency,energy} represents the functional support type. The attributes of nodes and edges together serve as the input features of the graph neural network, realizing the unified encoding of structure and function.

[0154] S13. Multi-subgraph modeling: Based on functional attributes, high-order heterogeneous graphs are split into multiple subgraphs.

[0155] S131, Source-Load Energy Supply Sub-diagram ζ (1) =(V s ∪V l ,δ (1) ),

[0156] Where the boundary set δ (1) ={(v s ,v l )|v s ∈V s ,v l ∈V l} represents the power supply path.

[0157] S132, Storage-Load Adjustment Subgraph ζ (2) =(V e ∪V l ,δ (2) ),

[0158] Where the boundary set δ (2) ={(v e ,v l ),(v l ,v e )|v e ∈V e ,v l ∈V l} represents the relationship between energy storage charging / discharging and load regulation.

[0159] S133, Source-Source Collaborative Subgraph ζ (3) =(V s ,δ (3) ),

[0160] edge set δ (3) This indicates the synergistic relationship between sources such as wind and solar energy.

[0161] S134. Combine the above subgraphs into a multi-view high-order heterogeneous graph:

[0162] ζ={ζ (1) ,ζ (2) ,ζ (3)}

[0163] It provides a multi-level, multi-functional graph structure foundation for joint optimization and information sharing.

[0164] S2. Based on the multi-view high-order heterogeneous graph constructed by S1, a multi-dimensional joint attention mechanism is designed to integrate multi-scale spatial, temporal dynamics and semantic heterogeneous information to enhance the cross-type and highly dynamic semantic expression capabilities among distributed source-load-storage nodes.

[0165] The specific implementation steps are as follows:

[0166] S21. To address the differences in network topology and neighborhood characteristics at different scales, a multi-scale spatial attention mechanism is designed. Node neighbors are divided into different levels (e.g., first-order neighbors, second-order neighbors, and hyperedge-mapped neighbors), and their spatial attention weights are calculated for each level. The weight calculation formula is as follows:

[0167]

[0168] Where l represents the neighbor level, a (l) W represents the attention vector parameters. (l) Here are the trainable parameters for the corresponding layer, and || denotes vector concatenation, normalized by layer weights:

[0169]

[0170] This achieves adaptive weighting of neighborhoods at different levels, enhancing the graph neural network's ability to express spatial relationships at different scales; where N (l) (i) represents the neighbor set;

[0171] S22. Considering the non-stationary changes of node features over time, a multi-scale temporal attention mechanism is proposed. This mechanism constructs temporal feature descriptions across multiple time windows (short-term, medium-term, and long-term), calculates temporal correlation weights for each window, and specifically defines the node temporal features at each time scale τ. Using multilayer perceptron Time-related scoring:

[0172]

[0173] The function input includes the features of the two nodes on the time scale τ and the time difference Δt = t. j -t i The normalized weights are:

[0174]

[0175] Ultimately, multi-scale temporal attention is achieved through weighted fusion:

[0176]

[0177] Where, β τ These are trainable parameters that reflect the contribution of different time scales to the dynamic dependence of nodes, enabling a comprehensive characterization of temporal dynamics.

[0178] S23. For heterogeneous node types (wind power, photovoltaic, energy storage, load, etc.) and their diverse functional interactions, a semantic heterogeneous attention mechanism is proposed, such as... Figure 2 As shown, a unique embedding vector is defined for each node type. And based on the physical coupling characteristics between types, a type coupling matrix M is constructed, whose elements M pq The semantic relevance score, representing the coupling strength between type p and type q, is defined as follows:

[0179]

[0180] Where g(·) is a scoring function based on node feature similarity, and γ is a balancing parameter, normalized to semantic attention weights:

[0181]

[0182] This mechanism achieves functional coupling adaptation, takes into account both type prior and feature dynamics, and improves the efficiency of information fusion between heterogeneous nodes.

[0183] S24. To integrate the three-dimensional attention mechanism, design a cross-dimensional joint attention fusion strategy:

[0184]

[0185] Where, λ s ,λ t ,λ sem w is the global adjustment coefficient. l These are spatial hierarchy weights, all of which are trainable parameters.

[0186] Based on the learning results of the multi-view high-order heterogeneous graph constructed by S1 and the multi-dimensional joint attention mechanism of S2, an optimization scheduling model driven by the goals of power gap identification, compensation allocation and optimization is constructed to achieve real-time power balance between source, load and storage and improve system resilience.

[0187] The specific implementation steps are as follows:

[0188] S31. To address the time-varying uncertainty of distributed renewable energy power output, a power gap detection and dynamic assessment mechanism is designed. The real-time power gap ΔP of each load node is estimated using the node characteristics output by a graph neural network. i (t):

[0189]

[0190] Among them, Pd,i (t) represents the actual demand at time t of load node i. For the predicted total output of new energy sources and energy storage, the power gap vector ΔP(t) serves as the input to the dynamic compensation strategy, reflecting the power balance state of the system in real time.

[0191] S32. Based on the node functional attributes and system constraints, design a priority-based power compensation allocation strategy and define the compensation priority weight w. i The weighting comprehensively considers energy storage capacity, response speed, and regulation cost:

[0192] Among them, C i For energy storage capacity, T resp,i For response time, Cost i To adjust costs, the weighting coefficients ω1, ω2, and ω3 can be determined through training and optimization. The power gap is shared by energy storage and flexible load nodes according to the weighting priority, achieving efficient dynamic compensation.

[0193] S33. Based on the power gap and compensation resources, construct a joint scheduling optimization model. The objective function is to minimize the total system operating cost while satisfying power balance and operational constraints.

[0194]

[0195] Among them, P i (t) represents the adjustment power of node i at time t. The model constraints include:

[0196] 1) Power balance constraints:

[0197]

[0198] Among them, P i (t) represents the adjustment power of node i at time t, ΔP net (t)=∑ i ΔP i (t) represents the total power deficit of the system;

[0199] 2) Node power upper and lower limits:

[0200] P i min ≤P i (t)≤P i max

[0201] 3) Energy storage state constraints:

[0202]

[0203] Among them, SOC i(t) represents the energy storage state, η i The charging / discharging efficiency is given by Δt, which is the time step.

[0204] S34. To address the complex nonlinear characteristics of the joint optimization model, a heuristic solution strategy based on improved particle swarm optimization (PSO) is designed. This strategy includes:

[0205] S341, Particle Encoding: The adjustment power of each node is used as the particle position vector;

[0206] S342. Fitness function: defined as the objective function J plus a penalty function for violating constraints;

[0207] S343, Dynamic Weight Adjustment: Combining real-time load and power gap, the algorithm parameters are dynamically adjusted to improve search efficiency and convergence speed;

[0208] S344 Distributed computing architecture: Design a parallel computing framework for large-scale networks to achieve fast scheduling and response.

[0209] Through real-time iterative optimization, efficient compensation and economical scheduling of system power deficit are achieved, enhancing the resilience and stability of distributed new energy systems. This provides an initial model structure and action space boundary for subsequent S4 deep reinforcement learning, improving the convergence speed and feasibility of policy training.

[0210] S4. Based on the joint power scheduling modeling problem of multi-view high-order heterogeneous graphs in S3, a graph embedding-driven deep reinforcement learning solution framework is designed to construct a state, action, and reward triplet to obtain the distributed source-load control policy, so as to realize the adaptive learning and global optimization of the distributed source-load control policy in complex environments.

[0211] like Figure 3 As shown, the specific implementation steps are as follows:

[0212] S41. Embed the node features generated by encoding the multi-view high-order heterogeneous graph in S1 through the joint attention mechanism in S2. As the system state s t The main components, while also incorporating auxiliary information such as environmental power deficit and energy storage status:

[0213] s t =[H t ,ΔP net (t),SOC t V t ,L t ]

[0214] in, Represents the embedding representation of heterogeneous graph nodes; ΔP net (t) represents the current total power deficit of the system; V represents the state of charge of each energy storage node; t ,L t Represents the node voltage and load demand vector;

[0215] S42, Action Space A is defined as the set of adjustment operations for all controllable nodes, including photovoltaic output regulation, energy storage charging and discharging power adjustment, and the response amplitude of controllable loads:

[0216]

[0217] Among them, P pv,i (t) represents the output adjustment value of photovoltaic node i; This represents the charging and discharging power of energy storage node i; The adjustable partial response of load node i must satisfy the following physical constraints in its action space A:

[0218] Photovoltaic regulation range:

[0219] Energy storage capacity and power limitations:

[0220] Load regulation rate limit:

[0221] S43. Design a multi-objective reward function that integrates volatility suppression and control cost minimization:

[0222]

[0223] Where Var[ΔP(t)] represents the system power fluctuation rate, which is used to measure the stability of regulation; The control cost of node i can be defined as: λ1, λ2, and λ3 represent voltage deviation and node over-limit penalty items; λ1, λ2, and λ3 represent multi-objective weights, which are dynamically adjusted according to the operational safety level.

[0224] S44. Employ a graph embedding-aware deep reinforcement learning framework, with state s t Input Policy Network π θ (a t |s t Output the optimal action a) t A heterogeneous graph feature residual channel + state attention fusion structure is introduced to improve the policy's sensitivity to structural changes. The training objective is to maximize the cumulative discount reward.

[0225]

[0226] The graph structure is updated every τ steps, and training oscillations are mitigated through experience replay. In addition, a policy stability term J is introduced.stable Ensure that the strategy does not change drastically:

[0227]

[0228] The final output scheduling strategy π * (a|s) is used to guide subsequent control execution (S5) and feedback correction (S6) processes.

[0229] S5. Based on the distributed source-load control strategy obtained in S4, a coordination control mechanism for heterogeneous node types is constructed to ensure that each control unit can independently execute strategy instructions, and system-level consistency and responsiveness are achieved through graph structure.

[0230] The specific implementation steps are as follows:

[0231] S51, Control instruction mapping and node behavior driving, outputting action a from the reinforcement learning policy. t Mapped to physical control commands that can be executed by the node, including:

[0232] S511. Generate output adjustment command for the source node: u pv,i (t)=P pv,i (t);

[0233] S512, for energy storage nodes, decoding is performed to update charging / discharging power and SOC:

[0234] S513. Generate load adjustment commands for load nodes:

[0235] The unified instruction is driven by the embedding results of the node's local state and global strategy, and supports the concurrent execution of heterogeneous control instructions;

[0236] S52. To address the regional boundary coordination problem in distributed control structures, a coordination factor constraint mechanism based on graph structure propagation is designed, defining the set B of boundary nodes at the intersection between sub-regions B and C. bd Their power coordination relationship is expressed as:

[0237]

[0238] Where κ is the allowable error tolerance, propagated through the edge weight propagation function φ(e) in the graph. ij Dynamic adjustments are made, and if the error exceeds the tolerance, a local boundary control mechanism is triggered to recalculate the instruction.

[0239] S53, By constructing a "graph propagation coordination graph" ctrl =(V,δ ctrlTo achieve synchronous propagation of control states between different nodes and avoid over-adjustment or mismatch, a residual consistency propagation function is used.

[0240]

[0241] Where η ij This is the propagation coefficient between adjacent nodes, ensuring that control commands are quickly and consistently transmitted within the neighborhood, thus improving system stability.

[0242] S6. Based on the real-time monitoring results of the system's operating status, construct an operation feedback and adaptive adjustment module to dynamically correct the control strategy and graph structure representation, thereby improving the system's adaptive capability to disturbances and topology changes.

[0243] The specific steps are as follows:

[0244] S61. Real-time collection of operational data from each node X t ={V i (t),I i (t),P i (t),SOC i (t)}, through the desired state By comparison, construct the error function:

[0245]

[0246] If the error exceeds the set threshold κ th Then, we will move on to the next stage of intensive training.

[0247] S62. Based on the error source classification results, design an adaptive adjustment strategy:

[0248] S621. If the error is concentrated in the node power control, fine-tune the adjustment cost term λ2 in the reward function.

[0249] S622. If voltage exceeds the limit frequently, increase the penalty weight in λ3.

[0250] S623. If the strategy has poor stability, increase the regularization coefficient γ of the strategy smoothing term.

[0251] S63. When the system topology changes, such as the addition of a node or a break in a line, the graph structure migration mechanism is triggered. The migration mechanism includes:

[0252] S631, Graph Node Update: Dynamically adding or removing nodes and their connecting edges;

[0253] S632, Node Embedding Initialization: Initialize new nodes using the average embedding vector of nearest neighbor nodes, etc.

[0254] Further explanation is needed:

[0255] Those skilled in the art will understand that, in order to clearly and systematically illustrate the technical solution of the present invention, the functional modules in the above embodiments are logically divided and described, including but not limited to a high-order graph construction module, a structure embedding module, a joint scheduling modeling module, a two-layer optimization module, and a control feedback module. It should be understood that this modular division is only used to express the technical ideas and illustrate the implementation path, and is not a limitation on the functional deployment method. In practical applications, related modules can be flexibly adjusted according to system resource conditions, control strategy constraints, or computing architecture requirements, such as merging, splitting, reconstructing, or replacing modules, without affecting the substantial technical effect of the method of the present invention. Therefore, the organizational form of the functional units should not be regarded as a limitation on the technical protection scope of the present invention.

[0256] It should also be noted that the embodiments listed above are only used to illustrate the technical principles and application methods of the present invention, and do not constitute a limitation on the scope of protection of the present invention. Although this specification has described the present invention in detail with reference to several preferred embodiments, those skilled in the art should understand that various modifications, variations, or equivalent substitutions can be made to the relevant technical details without departing from the core technical ideas of the present invention. For example, simplification of graph construction methods, replacement of graph neural network structures, switching of reinforcement learning algorithms, and adjustment of inter-module interface protocols should all be considered reasonable technical extensions of the present invention. Therefore, all equivalent changes or substitutions made within the spirit and technical principles of the present invention should be covered within the scope of protection of the present invention.

Claims

1. A distributed source-load collaborative optimization method based on high-order topology and multi-scale attention, characterized in that, Includes the following steps: S1: Based on the physical connection, control dependency and interaction behavior information in the distributed source-load-storage system, construct a multi-view high-order heterogeneous graph that integrates structural topology and functional relationship; S2: Based on the multi-view high-order heterogeneous graph constructed by S1, a multi-dimensional joint attention mechanism is designed to integrate multi-scale spatial, temporal dynamics and semantic heterogeneous information. S3: Based on the learning results of the multi-view high-order heterogeneous graph constructed by S1 and the multi-dimensional joint attention mechanism of S2, an optimization scheduling model driven by power gap identification, compensation allocation and optimization objectives is constructed. S4: Based on the joint power scheduling modeling problem of multi-view high-order heterogeneous graphs in S3, a graph embedding-driven deep reinforcement learning solution framework is designed to construct state, action, and reward triples to obtain distributed source-load control policies. S5: Based on the distributed source-load control strategy obtained in S4, a coordination control mechanism for heterogeneous node types is constructed to ensure that each control unit can independently execute strategy instructions, and to achieve system-level consistency and responsiveness through a graph structure; S6: Based on the real-time monitoring results of the system's operating status, construct an operation feedback and adaptive adjustment module to dynamically correct the control strategy and graph structure representation.

2. The distributed source-load collaborative optimization method based on high-order topology and multi-scale attention according to claim 1, characterized in that, Step S1 includes the following steps: S11, Construction of Hyperedge-Line Graph Dual Mapping Mechanism Define the system node set as V = {v1, v2, ..., v...} N }, and introduce a set of superedges. And define the hyperedge mapping function: Mapping hyperedges as virtual nodes in a line graph Constructing a line graph ζ l = (V l , δ l ), where: Edge set: S12, Structural Function Attribute Coding Assign an action attribute vector to each node v i , edge (v i ,v j ) and take the attributes of nodes and edges as input features of the graph neural network to realize unified coding of structure and function; S13, Multi-subgraph Modeling Based on functional attributes, the high-order heterogeneous graph is split into subgraphs, as follows: S131, source-load energy supply sub-diagram ζ (1) = (V s ∪V l ,δ (1) ), where the edge set δ (1) = {(v s , v l ) | v s ∈ V s , v l ∈ V l} represents the power supply path; S132, charge-adjustment sub-diagram ζ (2) = (V e ∪V l ,δ (2) ), where the edge set δ (2) = {(v e ,v l ) | v l , v e ∈ V e} e l l represents the relationship between energy storage charging and discharging and load regulation;​​ S133, source-source collaboration subgraph ζ (3) = (V s , δ (3) ), δ (3) represents the synergistic relationship between each source; S134. Combine the above subgraphs into a multi-view high-order heterogeneous graph: ζ = {ζ (1) , ζ (2) , ζ (3)}.

3. The distributed source-load collaborative optimization method based on high-order topology and multi-scale attention according to claim 1, characterized in that, Step S2 includes the following steps: S21. Design a multi-scale spatial attention mechanism, as follows: The formula for calculating the weight is: Through hierarchical weight normalization: S22. Design a multi-scale temporal attention mechanism, as follows: By constructing temporal feature descriptions across multiple time windows, the temporal correlation weights are calculated respectively: The normalized weights are: Ultimately, multi-scale temporal attention is achieved through weighted fusion: where β τ is a trainable parameter that reflects the contribution of different time scales to the node dynamics. S23. Design a semantic heterogeneous attention mechanism, as follows: Normalized to semantic attention weights: S24. Design a cross-dimensional joint attention fusion strategy, as follows: where λ s ,λ t ,λ sem are global adjustment coefficients, w l are spatial level weights, all are trainable parameters.

4. The distributed source-load collaborative optimization method based on high-order topology and multi-scale attention according to claim 1, characterized in that, Step S3 includes the following steps: S31. Design power gap detection and dynamic evaluation mechanism: The node feature output by the graph neural network estimates the real-time power gap ΔP of each load node i (t): ΔP(t) is the input of the dynamic compensation strategy, which reflects the power balance state of the system in real time; S32. Based on the node functional attributes and system constraints, design a priority-based power compensation allocation strategy: S33. Based on real-time power gap and power compensation allocation strategies, construct an optimized scheduling model with the objective function of minimizing the total system operating cost while satisfying power balance and operational constraints. where P i (t) is the regulated power of node i at time t, the model constraints include: Power balance constraints: where P i (t) is the regulated power of node i at time t, ΔP net (t) =∑ i ΔP i (t) represents the total power gap of the system; Node power upper and lower limits: P i min ≤P i (t)≤P i max ; Energy storage state constraints: SOC i (t+1) = SOC i (t) - η i P i (t)Δt, where SOC i (t) is the state of charge, η i is the charge-discharge efficiency, and Δt is the time step. S34. Design a heuristic solution strategy based on improved particle swarm optimization, which includes: S341, Particle Encoding: The adjustment power of each node is used as the particle position vector; S342. Fitness function: defined as the objective function J plus a penalty function for violating constraints; S343, Dynamic Weight Adjustment: Combines real-time load and power gap to dynamically adjust algorithm parameters; S344, Distributed Computing Architecture: Designing a parallel computing framework for large-scale networks.

5. The distributed source-load collaborative optimization method based on high-order topology and multi-scale attention as described in claim 1, characterized in that, Step S4 includes the following steps: S41. Encode the node embedding features generated by the multi-view high-order heterogeneous graph in S1 through the multi-dimensional joint attention mechanism in S2. As the system state s t The components, while introducing auxiliary information: s t = [H t , ΔP net (t), SOC t , V t , L t ] ; S42, Action Space A is defined as the set of adjustment operations for all controllable nodes, including photovoltaic output regulation, energy storage charging and discharging power adjustment, and the response amplitude of controllable loads: Action space A must satisfy the following physical constraints: Photovoltaic regulation range: Energy storage capacity and power limitations: Load regulation rate limit: S43. Design a multi-objective reward function that integrates volatility suppression and control cost minimization: S44. Employ a graph embedding-aware deep reinforcement learning framework, with state s t Input Policy Network π θ (a t |s t Output the optimal action a) t A heterogeneous graph feature residual channel plus state attention fusion structure is introduced, and the training objective is to maximize the cumulative discount reward: Furthermore, a policy stability term J stable is introduced Final output scheduling policy π * (a | s).

6. The distributed source-load collaborative optimization method based on high-order topology and multi-scale attention as described in claim 1, characterized in that, Step S5 includes the following steps: S51, control instruction mapping and node behavior driving, output action a of reinforcement learning strategy t mapping physical control instructions executable by the node, comprising: S511. Generate output adjustment command for the source node: u pv,i (t)=P pv,i (t); S512, for energy storage nodes, decoding is performed to update charging / discharging power and SOC: S513. Generate load adjustment commands for load nodes: S514, the unified instruction is driven by the embedding results of the node's local state and global strategy, and supports the concurrent execution of heterogeneous control instructions; S52, a coordination factor constraint mechanism based on graph structure propagation is designed, and a boundary node set B between the sub-regions B and C is defined bd The power coordination relationship is represented as: where k is the allowable error tolerance, dynamically adjusted by the edge weight propagation function φ(e ij ) in the graph, and if the error exceeds the tolerance, a local boundary regulation mechanism is triggered to recalculate the instructions. S53, by constructing a "graph propagation coordination graph" ζ ctrl = (V, δ ctrl ), to achieve synchronization propagation of control states between different nodes, using a residual consistency propagation function: where η ij is the propagation coefficient between adjacent nodes.

7. The distributed source-load collaborative optimization method based on high-order topology and multi-scale attention as described in claim 1, characterized in that, Step S6 includes the following steps: S61. Real-time collection of operational data from each node X t ={V i (t),I i (t),P i (t),SOC i (t)}, through the desired state By comparison, construct the error function: If the error exceeds a set threshold κ th then go to S61 for a retraining phase; S62. Based on the error source classification results, design an adaptive adjustment strategy: S621. If the error is concentrated in the node power control, fine-tune the adjustment cost term λ2 in the reward function. S622. If voltage exceeds the limit frequently, increase the penalty weight in λ3. S623. If the strategy has poor stability, increase the regularization coefficient γ of the strategy smoothing term. S63. When the system topology changes, the graph structure migration mechanism is triggered. The migration mechanism includes: S631, Graph Node Update: Dynamically adding or removing nodes and their connecting edges; S632, Node Embedding Initialization: Initialize new nodes using the average embedding vector of nearest neighbor nodes.

Images