Virtual power plant-power distribution network voltage dynamic cooperative control method and device

Abstract

The invention provides a virtual power plant-power distribution network voltage dynamic cooperative control method and device, and the method comprises the steps: obtaining multi-source heterogeneous data to construct a mixed data set, and mapping the mixed data set to a quantum state space for feature extraction; fusing the quantum features, the satellite cloud picture and the voltage waveform data, and modeling by adopting quantum probability; generating a DERs output scene based on a quantum state superposition principle, searching and quantifying the voltage out-of-limit probability amplitude of each node, and generating a power distribution network risk thermodynamic diagram; constructing a hierarchical control decision; the hierarchical control decision is executed, then a causal logic chain of a voltage out-of-limit event is generated through an invariant graph reasoning network, the contribution degree of features to the control decision is quantified, and iterative optimization of model parameters is achieved in combination with a meta-learning mechanism. Based on the method, the invention further provides a voltage dynamic cooperative control device, self-adaptive response of distributed energy uncertainty is achieved, and a solution is provided for voltage stabilization of the power distribution network in a high-proportion new energy access scene.

Description

Technical Field

[0001] This invention belongs to the field of power grid collaborative control technology, and specifically relates to a method and device for dynamic collaborative control of voltage between a virtual power plant and a distribution network. Background Technology

[0002] With the current acceleration of energy transition, virtual power plants, by aggregating distributed energy resources (DERs) and connecting them to urban distribution networks, have become an important development direction for new power systems. However, the randomness and intermittent nature of high-proportion DERs presents several technical bottlenecks for distribution network voltage control: On the one hand, DERs and resources such as energy storage and controllable loads have significant differences in response speed over time scales, and traditional hierarchical control frameworks rely on fixed priority allocation, making it difficult to dynamically coordinate the regulation capabilities of different resources; on the other hand, the fusion analysis of multi-source heterogeneous data such as meteorological data, grid topology status, and DER output still relies on manual feature engineering, and existing models still struggle to capture complex spatiotemporal coupling relationships; furthermore, while advanced machine learning algorithms can improve control real-time performance, their black-box decision-making model remains fundamentally contradictory to the power system's requirements for compliance and interpretability of control strategies.

[0003] Currently, traditional technical approaches have significant shortcomings in addressing the aforementioned challenges: hierarchical control strategies based on fixed rules cannot adapt to the dynamic changes in DERs output, often leading to voltage regulation lag; data analysis models relying on manual feature engineering have limited generalization capabilities for complex scenarios, affecting voltage prediction accuracy; and algorithms such as deep reinforcement learning, due to a lack of interpretability support, pose a risk of violating operating procedures in practical engineering applications. Summary of the Invention

[0004] To address the aforementioned technical challenges, this invention proposes a method and device for dynamic coordinated voltage control of a virtual power plant and distribution network. Breaking through the limitations of fixed logic in traditional voltage control, it achieves adaptive response to uncertainties in distributed energy resources, providing an intelligent solution for voltage stability in distribution networks with high-proportion renewable energy integration scenarios.

[0005] To achieve the above objectives, the present invention adopts the following technical solution: A method for dynamic coordinated control of voltage in a virtual power plant and distribution network includes the following steps: A hybrid dataset is constructed by acquiring multi-source heterogeneous data for collaborative control, and the hybrid dataset is mapped to quantum state space for feature extraction; the multi-source heterogeneous data includes distributed energy output data, distribution network topology status, and meteorological monitoring data. An improved Transformer-XL model is used to fuse quantum features, satellite cloud images, and voltage waveform data, and quantum probabilistic modeling is employed. Based on the principle of quantum superposition, DERs output scenarios are generated, and the voltage over-limit probability amplitude of each node is quantified through quantum Monte Carlo tree search to generate a heat map of distribution network risks. A hierarchical control decision-making mechanism is constructed. Specifically, the reactive power compensation problem of energy storage is modeled as a quantum optimization problem to generate real-time adjustment commands for reactive power of energy storage. Based on the risk heat map and topology characteristics of the distribution network, the topology characteristics of the distribution network are extracted through an interpretable graph reasoning network. The constraints of the distribution network operation procedure are represented by linear time-series logic to generate a controllable load dispatching strategy. Based on the day-ahead DERs output prediction, the energy aggregation strategy is optimized to generate interpretable dispatching rules. The hierarchical control decision is executed, and then a causal logic chain of voltage over-limit events is generated through an invariant graph inference network. The contribution of features to the control decision is quantified using SHAP values, and the model parameters are rapidly iteratively optimized by combining a meta-learning mechanism.

[0006] The present invention also proposes a virtual power plant-distribution network voltage dynamic collaborative control device, comprising: a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the computer program implements the steps of the method when executed by the processor.

[0007] The effects described in the invention are merely those of the embodiments, and not all the effects of the invention. One of the above technical solutions has the following advantages or beneficial effects: This invention proposes a method and device for dynamic coordinated control of voltage in a virtual power plant and distribution network. The method includes the following steps: acquiring multi-source heterogeneous data for coordinated control to construct a hybrid dataset; mapping the hybrid dataset to a quantum state space for feature extraction; the multi-source heterogeneous data includes distributed energy output data, distribution network topology status, and meteorological monitoring data; fusing quantum features, satellite cloud images, and voltage waveform data using an improved Transformer-XL model, and employing quantum probability modeling; generating DERs output scenarios based on the quantum state superposition principle; quantizing the voltage exceedance probability amplitude of each node through quantum Monte Carlo tree search; and generating a distribution network risk heatmap; and constructing... The hierarchical control decision-making process involves: modeling the reactive power compensation problem of energy storage as a quantum optimization problem to generate real-time adjustment commands for reactive power of energy storage; extracting topological features of the distribution network through an interpretable graph inference network based on the distribution network risk heat map and topological features of the distribution network, representing the constraints of the distribution network operation procedure using linear time-series logic, and generating a controllable load scheduling strategy; optimizing the energy aggregation strategy based on day-ahead DERs output prediction to generate interpretable scheduling rules; executing the hierarchical control decision-making process, and then generating a causal logic chain of voltage over-limit events through an invariant graph inference network, quantifying the contribution of features to the control decision-making using SHAP values, and combining a meta-learning mechanism to achieve rapid iterative optimization of model parameters. Based on this method, the present invention also proposes a virtual power plant-distribution network voltage dynamic collaborative control device. This invention systematically solves the core challenges faced by traditional voltage control in scenarios with a high proportion of new energy access by deeply integrating cutting-edge technologies such as quantum machine learning, neural symbolic integration, and dynamic architecture search, achieving a dual breakthrough in technical performance and engineering value.

[0008] At the technical performance level, this invention constructs a full-chain intelligent system of "prediction-evaluation-control-interpretation": the multimodal quantum feature learning model significantly improves the prediction accuracy of distributed energy output, and combined with quantum Monte Carlo risk assessment technology, it achieves probabilistic and accurate prediction of voltage limit risks; the hierarchical neural symbolic control architecture improves the voltage transient response speed to the millisecond level through the synergy of quantum annealing and deep reinforcement learning, and at the same time, it uses the joint reasoning of graph neural networks and power rules to ensure that the control strategy fully complies with the operating specifications, thus resolving the contradiction between the black box nature of artificial intelligence decision-making and the safety of power systems.

[0009] At the engineering application level, this invention significantly reduces edge computing costs and distribution network expansion investment through quantum-classical hybrid hardware deployment and dynamic architecture optimization, achieving efficient utilization of hardware resources. The interpretability enhancement technology shortens the root cause location time for voltage over-limit events, significantly improving operation and maintenance efficiency. In uncertain scenarios such as extreme weather, the robust control strategy based on quantum probability significantly improves the renewable energy absorption rate, providing key technical support for the construction of a highly resilient power grid under the "dual carbon" target.

[0010] From an industry development perspective, this invention breaks through the traditional power system's reliance on centralized power sources, providing an intelligent paradigm for the collaborative control of virtual power plants and distribution networks, and driving the power system's upgrade from "rigid control" to "flexible adaptation." Its engineering implementation will significantly enhance the grid's capacity to accommodate new energy sources, strengthen the resilience and security of urban distribution networks, and lay the technological foundation for emerging application scenarios such as smart cities and vehicle-to-grid (V2G) interaction, demonstrating significant economic value and social benefits. Attached Figure Description

[0011] Figure 1 This is a flowchart of a virtual power plant-distribution network voltage dynamic collaborative control method proposed in Embodiment 1 of the present invention; Figure 2 This is a diagram of the topology of a city power distribution network proposed in Embodiment 1 of the present invention; Figure 3 This is a graph showing the voltage distribution of each node in an urban distribution network after different algorithm control methods proposed in Embodiment 1 of the present invention, taking into account the VPP. Figure 4 The thermal distribution of the SHAP value of the feature proposed in Embodiment 1 of the present invention on the contribution of voltage control decision; Figure 5 This is the photovoltaic power output prediction comparison curve proposed in Embodiment 1 of the present invention; Figure 6 This is a bar chart comparing the benefits of the hardware-algorithm collaborative scheme proposed in Embodiment 1 of the present invention; Figure 7 This is a bar chart comparing the traditional scheme and the quantum-dynamic architecture scheme in terms of three quantitative indicators: voltage limit exceedance probability, new energy absorption rate, and average control error under the uncertainty scenario proposed in Embodiment 1 of the present invention. Figure 8 This is a schematic diagram of a virtual power plant-distribution network voltage dynamic collaborative control device proposed in Embodiment 2 of the present invention. Detailed Implementation

[0012] To clearly illustrate the technical features of this solution, the invention will be described in detail below through specific embodiments and in conjunction with the accompanying drawings. The following disclosure provides many different embodiments or examples for implementing different structures of the invention. To simplify the disclosure of the invention, components and arrangements of specific examples are described below. Furthermore, reference numerals and / or letters may be repeated in different examples. This repetition is for simplification and clarity and does not in itself indicate a relationship between the various embodiments and / or arrangements discussed. It should be noted that the components illustrated in the drawings are not necessarily drawn to scale. Descriptions of well-known components, processing techniques, and processes are omitted in this invention to avoid unnecessarily limiting the invention.

[0013] Example 1 Embodiment 1 of this invention proposes a virtual power plant-distribution network voltage dynamic collaborative control method. By constructing a multi-time-scale adaptive control architecture, a multi-source data deep fusion model, and an interpretable decision-making mechanism, it systematically solves the core pain points of traditional technologies in voltage dynamic control.

[0014] Figure 1 This is a flowchart of a virtual power plant-distribution network voltage dynamic collaborative control method proposed in Embodiment 1 of the present invention; In step S100, the process begins.

[0015] In step S110, multi-source heterogeneous data for collaborative control is acquired to construct a hybrid dataset. The system uses edge computing nodes to access distributed energy output data, distribution network topology status, meteorological monitoring data and user load curves in real time to form multi-source heterogeneous data.

[0016] In step S120, the hybrid dataset is mapped to the quantum state space for feature extraction. Quantum variational feature learning technology is employed to map heterogeneous data to the quantum state space for feature extraction. The nonlinear correlation between DERs output and meteorological factors is captured using quantum entanglement properties. Simultaneously, classical signal processing methods are used to filter out measurement noise, providing high-dimensional feature vectors for subsequent artificial intelligence analysis.

[0017] Considering meteorological data DERs' historical contributions and distribution network topology The hybrid dataset Parameterized quantum circuits (PQCs) are designed to map classical data to quantum Hilbert space: (1) In the formula, This indicates that the hybrid dataset, after being encoded by quantum states, The corresponding quantum state is the state vector in quantum Hilbert space; This represents a parameterized quantum encoding operation used to map heterogeneous data to a quantum Hilbert space, where It consists of a rotating gate R (ф1) and an entanglement gate C (ф2); Θ={ф1,ф2}; ф1 represents the rotating gate parameter; ф2 represents the entanglement gate parameter. The encoding effect can be optimized by adjusting these parameters. Represents the initial quantum state; Extracting quantum features through quantum measurement ; (2) In the formula, It is a quantum state The dual state of is used in quantum mechanics for operations such as calculating inner products, and . Together they constitute the Dirac notation system for describing quantum states and related operations; It is an observable operator that can be decomposed into the tensor product of Pauli matrices.

[0018] In step S130, uncertainty quantization is performed; to achieve co-training of quantum features and classical prediction models, a joint loss function is defined. : (3) In the formula, This is the set of parameters to be optimized for the classic prediction model; A loss function that measures the ability to represent quantum features, used to evaluate the quality of features extracted from the quantum part; The loss function evaluates the prediction error of the classical prediction model, reflecting the magnitude of the error when the classical model makes predictions using quantum features, etc.; λ is a balance coefficient used to adjust the proportion of quantum part loss and classical part loss in the joint loss.

[0019] Update parameters using an alternating optimization algorithm: (4) (5) In the formula, This represents the set of parameters in the quantum part after a parameter update, which is the updated parameter state. To update the parameter set of the pre-quantum part; The learning rate is used to update the parameters of the quantum part, controlling the step size of parameter updates in each iteration; The loss function LQVFL is a function of quantum parameters. exist The gradient at a point reflects the quantum parameters. In the current state, in which direction should the adjustment be made to achieve... Decrease, that is, guide how quantum parameters are updated to optimize the expression of quantum features. This is the set of parameters of the classical prediction model after one parameter update, representing the updated parameter state of the classical model. This is the set of parameters for the classic prediction model before the update. The learning rate used to update the parameters of the classic part; It is a loss function For classical model parameters exist The gradient at that point is used to guide the parameters of the classical model. How to update the model to reduce the prediction error of the classic model.

[0020] In step S140, the improved Transformer-XL model is used to fuse quantum features, satellite cloud images and voltage waveform data, and quantum probability modeling is adopted; the DERs output scenario is generated based on the quantum superposition principle, and the voltage over-limit probability amplitude of each node is quantified by quantum Monte Carlo tree search to generate a distribution network risk heat map; To handle temporal dependencies, an improved Transformer-XL architecture is introduced, which enhances the ability to model long sequences through relative positional encoding: (6) In the formula, Attention(Q,K,V) represents the attention calculation function; Q is the query matrix; K is the key matrix; V is the value matrix; R is the relative position encoding matrix, which captures the relative distance information between sequence elements; softmax is the normalization function; and d is usually a parameter related to the dimension of matrices such as Q and K.

[0021] The model output is mapped to a probability distribution via a hybrid density network: (7) In the formula, The target variable for prediction in the model; These are the input variables for the model; To represent the given input Under the condition of output The probability distribution; The number of Gaussian distributions in the mixed density network; For the first The weights of a Gaussian distribution in the mixture; For the first A Gaussian distribution, It is the mean of the Gaussian distribution. These are the standard deviations; they are both determined by the model based on the input. Calculated.

[0022] The improved Transformer-XL model is used to fuse quantum features, satellite cloud image features, and voltage waveform data, and quantum probabilistic modeling is employed as follows: (8) in, This represents the features after being fused by the attention fusion module; To represent the various features input to the fusion module; Features of satellite cloud images; Voltage waveform data; The attention weight is given by the following formula: (9) Quantum probabilistic modeling is used to address the uncertainty in DERs output, specifically: (10) In the formula, Historical contribution data for DERs; To indicate that the output of DERs is The quantum probability at time; Tr[] represents the trace operation; This is the quantum state density matrix corresponding to the data. Let be the density matrix of the reference state. Probability prediction intervals are generated through Monte Carlo sampling to improve model robustness.

[0023] This invention breaks through the limitations of traditional machine learning frameworks by achieving deep fusion and uncertainty quantification of multi-source heterogeneous data through a quantum-classical hybrid architecture. The quantum feature extraction module utilizes the properties of quantum entanglement to capture complex correlations between data, transforming traditional feature engineering into a trainable quantum circuit optimization problem; the improved Transformer-XL architecture effectively handles the long-term time-series dependencies and fluctuation characteristics of power systems through relative position encoding and a probability output layer.

[0024] The probability distribution of DER output is predicted by adaptively adjusting the network structure through dynamic neural architecture search.

[0025] The process of generating DERs output scenarios based on the quantum superposition principle, and generating a distribution network risk heatmap by quantizing the voltage over-limit probability amplitude of each node through quantum Monte Carlo tree search, includes: Define the quantum state corresponding to the DERs output set as: (29) In the formula, This represents the quantum state corresponding to the output set of DERs; For the first The ground state of a power output scenario. For complex amplitude coefficients, satisfying ; Define search tree nodes Indicates the state of the distribution network, edge The node value function represents the change in DER output: (30) In the formula, For the edge Number of visits; For the first The reward value of this simulation; The selection strategy employs quantum upper confidence bounds. Specifically: (31) In the formula, For nodes Total number of visits; When expanding nodes, the ground state of the expanded output scenario is generated through quantum gate operations. : (32) In the formula, For parameterized quantum gates; By converting superposition states into classical probability distributions through quantum measurement, a voltage limit operator is defined. Its quantum state The expected value is: (33) The quantum probability amplitude of voltage exceeding the limit is expressed as: : (34) In the formula, This represents the voltage limit exceedance factor for a single constraint. The mathematical expression for mapping the probability amplitude to the distribution network topology and generating a risk heatmap is as follows: (35) In the formula, For nodes Risk value; path( ) for scene The affected node paths; Based on quantum probability distribution, a robust optimization model for energy storage charging and discharging plans is constructed; the energy storage state vector is defined. The optimization objective is: (36) In the formula: To optimize variables; It is a quantum state; For expectation value operators; To control the cost function, Risk aversion coefficient; For variance operators; This is the loss function.

[0026] In step S150, a hierarchical control decision is constructed; specifically: the energy storage reactive power compensation problem is modeled as a quantum optimization problem to generate real-time adjustment instructions for energy storage reactive power; based on the distribution network risk heat map and distribution network topology characteristics, the distribution network topology characteristics are extracted through an interpretable graph reasoning network, and the constraints of the distribution network operation procedure are represented by linear time-series logic to generate a controllable load dispatching strategy; based on the day-ahead DERs output prediction, the energy aggregation strategy is optimized to generate interpretable dispatching rules; To address the collaborative control requirements of multi-timescale resources in virtual power plants, this invention constructs a three-layer neural symbolic control architecture. Precise control is achieved through the cross-fusion of quantum computing, graph reasoning, and dynamic optimization. The mathematical relationships of the three-layer architecture can be expressed as follows: (11) In the formula, This represents the distribution network state vector, including voltage, power flow, and DERS states; S() is the strategic layer decision function, C() is the coordination layer constraint handling function, and F() is the fast layer execution function. This is the final control command.

[0027] The process of modeling the energy storage reactive power compensation problem as a quantum optimization problem to generate real-time energy storage reactive power adjustment commands includes: Define a quadratic unconstrained binary optimization model : (12) In the formula, Let ∈{-1,+1}n be the state vector of the qubit. Let be the coupling matrix. It is the bias vector; The optimal state q* is solved using a quantum annealer and serves as the initial policy for the gradient of the deep deterministic policy: (13) In the formula, A set of information describing the environment in which an intelligent agent exists; For the agent in state The following actions were taken; For deterministic policy functions, The standard deviation of noise. To explore noise, the q* generated by quantum annealing is used for initialization. This accelerates DDPG convergence.

[0028] Based on the distribution network risk heat map and distribution network topology characteristics, the process of extracting distribution network topology characteristics through an interpretable graph reasoning network, representing the constraints of distribution network operation procedures using linear time-series logic, and generating a controllable load dispatching strategy includes: Message passing mechanism for building graph neural networks: (14) In the formula, For nodes of Layer node characteristics; It is the set of adjacent nodes; For nodes of Layer node characteristics; For the nodes in the graph The set of directly connected neighboring nodes; For nodes A neighbor node In the Layer node representation; For the first A trainable weight matrix in the layer used to process edge information or neighbor node information; For the first The trainable weight matrix in the layer used to process information about its own nodes; For activation functions; The constraints of the distribution network operation procedures are represented using linear sequential logic and then transformed into penalty terms using a weighted automaton. (15) In the formula, For time range; Indicates time The degree of violation of the rules; wt is the time weight; For a moment The corresponding linear sequential logic formula; The joint loss function is: (16) In the formula, For the joint loss function; To reinforce learning loss; This is the first weighting coefficient.

[0029] The process of generating interpretable scheduling rules based on optimizing energy aggregation strategies according to day-ahead DERs output predictions includes: Define the neural architecture search space Optimize architecture parameters using genetic algorithms : (17) In the formula, β is the set of parameters of the neural architecture; Assess the loss for the architecture; Construct interpretable rules for generating DERs aggregation strategies using symbolic systems: (18) In the formula, Interpretable rules representing the aggregation strategy of DERs generated by the symbol system; This is a state judgment based on logical predicates. For the corresponding aggregation action; Genetic algorithms simultaneously optimize the parameters of neural architecture and symbolic rules: (19) In the formula, Optimal sign rule; Lsym( ) is the loss function for the symbolic rule, used to measure the symbolic rule. The advantages and disadvantages; Kirchhoff's voltage law can be expressed as a loop constraint in graph theory: (20) In the formula, For the line Voltage drop; It is a closed loop in the circuit; Voltage over-limit threshold is transformed into state-space constraint: ;(twenty one) In the formula, For nodes At any moment The voltage value is a state quantity that describes the dynamic changes in the voltage of a distribution network node; Minimum voltage threshold; Maximum voltage threshold; The above rules are embedded as prior knowledge into the action space of reinforcement learning: ;(twenty two) In the formula, For safe actions within the search space; To enhance the learning agent's ability to learn at any time The actions taken.

[0030] This invention achieves a three-level synergy of rapid response, rule constraints, and global optimization through deep coupling of neural symbols: the rapid layer uses quantum annealing to accelerate the policy search of DDPG, reducing the exponential complexity of traditional optimization problems to polynomial level; the coordination layer extracts grid topology features through invariant graph reasoning networks and transforms LTL rules into differentiable constraints, ensuring that the control strategy is 100% compliant with engineering specifications; the strategy layer achieves adaptive generation and interpretable expression of DERs aggregation strategies through joint optimization of dynamic neural architecture search and genetic symbol system.

[0031] This invention also includes dynamic neural architecture optimization based on conditional computation graphs and meta-learning, wherein the mathematical expression of the dynamic architecture is: ;(twenty three) In the formula, For a dynamic gated vector, satisfying ; For each sub-module in the dynamic neural architecture The set of parameters; This is a set of parameters related to dynamic gating. For feature vectors; For the first Parameters within each sub-model; for Each submodule; Design a gating function based on input features. To achieve dynamic switching of network architecture, a fluctuation feature extractor is defined. ; ;(twenty four) In the formula, The sequence of changes in the output of DERs; MaxPooling() is the maximum pooling operation; For convolutional neural networks The output of the processing; In scenarios with stable loads, the gating mechanism switches to a lightweight sub-network. Its structure is defined as follows: (25) In the formula, This is for fully connected layer operations; The mask vector is used to implement parameter pruning through sparse regularization; (26) In the formula, For sparsity regularization loss, The regularization coefficient is used. A model-independent meta-learning framework is adopted to define the meta-learner. To optimize initial parameters ; (27) In the formula, The optimal parameters; For a set of tasks; Task-specific data; Design a two-level optimization problem to achieve coordinated optimization of dynamic architecture and model parameters: (28) In the formula, The expected loss function is used to solve the problem through an alternating optimization strategy: first, fix the architecture parameters. Optimize model parameters Based on the updated Optimize architecture parameters.

[0032] This invention achieves adaptive matching of neural network architecture to the output characteristics of DERs through the deep integration of conditional computation graphs and meta-learning. The conditional computation graph uses a gating mechanism to dynamically activate attention modules or lightweight networks, reducing the computational overhead of edge nodes while ensuring prediction accuracy; the meta-learning framework enables the model to quickly adapt to the access of new energy storage devices, avoiding the retraining costs of traditional methods.

[0033] In step S160, hierarchical control decisions are executed, and control effect feedback is provided.

[0034] In step S170, hierarchical control decisions are executed, and then a causal logic chain of voltage over-limit events is generated through an invariant graph inference network. The contribution of features to the control decisions is quantified using SHAP values, and rapid iterative optimization of model parameters is achieved by combining a meta-learning mechanism. Specifically, this includes: Invariant graph inference networks extract topological association features through multi-layer message passing: (37) In the formula, For nodes In the Layer feature representation; For activation functions; Characteristics of line impedance; Adjacent nodes To the current node Passing by In the Layer weight matrix; Adjacent nodes In the Layer feature representation; With line impedance characteristics The relevant weight matrix; For nodes In the Layer feature representation; For nodes The set of all neighboring nodes; The rule engine maps graph features into causal logic chains: (38) In the formula, The output is the weight matrix; RuleEngine transforms the probability distribution into a natural language interpretation. Combine SHAP values ​​with graph features to generate a feature importance heatmap: (39) In the formula, Features In the importance distribution across the entire network, Normalize() is the normalization function; For selection functions; node eigenvectors; Design the joint loss function Strengthen the consistency between interpretation and control: (40) In the formula, This is the second weighting coefficient; Attribution function; For interpretation functions; defined as: (41) In the formula, The preset feature weights; The explanatory or attribution function generated for the model is for the first... The output value of each feature.

[0035] This invention achieves end-to-end interpretability for topological reasoning, feature attribution, and logic generation through the deep integration of invariant graph reasoning networks and SHAP values. The invariant graph reasoning network utilizes graph neural networks to capture the physical correlations of the power distribution network, and combines a rule engine to transform abstract features into natural language causal explanations; SHAP values ​​quantify feature contribution from a data-driven perspective, compensating for the subjective deficiencies of traditional rule-based explanations.

[0036] In step S180, the process ends.

[0037] To resolve the conflict between real-time performance and computational complexity in virtual power plant voltage control, this invention employs a quantum-classical hybrid hardware architecture, dynamically allocating computational tasks to classical processors and quantum chips. The task partitioning model for edge nodes can be represented as: (42) In the formula, For the overall task set, It includes tasks such as real-time data preprocessing and model inference, which are handled by high-performance classical chips; It includes tasks such as high-dimensional optimization and probabilistic modeling, and delivers a quantum annealer solution.

[0038] Design a bidirectional data interaction interface to achieve collaborative computing between two types of hardware. The classical processor sends the optimization problem parameters Θ to the quantum chip, and the quantum chip returns the probabilistic solution P(Θ). The mathematical expression of the interface protocol is as follows: (43) To achieve millisecond-level control response, a mathematical model of communication latency is constructed: (44) In the formula, For quantum key distribution time, For 5G slice transmission latency, Calculate the time for edge nodes.

[0039] (45) In the formula, This is the delay percentage coefficient. This represents the total response time required by the system.

[0040] Design a load-aware hardware resource scheduling algorithm to monitor the CPU utilization ι of classical chips and the idle rate μ of quantum chips in real time. (46) The core design philosophy of this architecture is embodied in the principle of "dedicated hardware division of labor and collaboration". This invention achieves a significant increase in the processing power of complex optimization tasks by modularly upgrading the quantum-classical heterogeneous computing interface while retaining the original classical computing hardware, thus avoiding the high costs associated with replacing the entire system hardware. This invention constructs a hierarchical response mechanism of classical computing and quantum computing—the classical computing module is responsible for handling observable real-time control tasks, while the quantum computing module predicts the potential fluctuation trend of distributed energy output through probabilistic modeling, forming a dual closed-loop control system of "real-time response-look-forward prediction". This invention employs quantum key distribution (QKD) technology to construct a physical layer encryption mechanism for control commands, and combines it with 5G network slicing technology to achieve communication channel isolation for control data streams, thus comprehensively meeting the stringent requirements of power systems for control reliability.

[0041] To fully illustrate the technical effect of the virtual power plant-distribution network voltage dynamic collaborative control method proposed in Embodiment 1 of the present invention, a city was selected with an actual improved three-phase balanced 10.5kV radial urban distribution network containing high-penetration distributed photovoltaic units to verify the effectiveness of the proposed method. Figure 2 This is a topology diagram of a city power distribution network proposed in Embodiment 1 of the present invention.

[0042] This simulation example includes two distributed photovoltaic (PV) power sources, two energy storage systems, one 5G base station, one data center (IDC), and two electric vehicle charging stations. The distributed PV system considers the correlation and temporal characteristics of solar irradiance between the two power stations. The load model primarily uses commercial and residential electricity loads, which are weighted at 55.8% and 44.2% respectively to represent the load curve for the entire community. Furthermore, the model uses typical days from the four seasons to approximate the annual load curve, specifying 91, 90, 91, and 93 days for spring, summer, autumn, and winter, respectively.

[0043] Figure 3 This is a graph showing the voltage distribution curves of various nodes in an urban distribution network considering VPP after different algorithm control methods proposed in Embodiment 1 of the present invention; from Figure 3 It is evident that, compared to no control or other single-module control algorithms, the voltage fluctuation amplitude of each node in the urban distribution network considering VPP is smaller after the strategic-coordination-fast three-layer collaborative control algorithm is controlled, and there are no voltage over-limit situations, thus ensuring the stability of the voltage of the urban distribution network considering VPP.

[0044] Figure 4 This figure shows the heat map of the SHAP values ​​of the features proposed in Embodiment 1 of this invention on voltage control decisions. It clearly shows that the contributions of various features to voltage control decisions differ significantly under different functional node groups. For example, in centralized photovoltaic node groups, new energy-related features such as DERs output fluctuations and photovoltaic power prediction errors have high SHAP values, significantly contributing to decision-making. In backbone network interconnection node groups, distribution network topology changes have a prominent contribution. This figure can intuitively quantify the influence of each feature on voltage control decisions at different nodes, providing a basis for optimizing voltage control strategies and identifying key influencing factors. It helps improve the accuracy and effectiveness of distribution network voltage regulation, echoing the logic of the strategy-coordination-fast three-layer collaborative control algorithm in improving voltage stability, and supporting the optimization direction of voltage control strategies from the perspective of feature contribution. Figure 5 This is the photovoltaic power output prediction comparison curve proposed in Embodiment 1 of the present invention; it includes three curves: actual photovoltaic power output, traditional LSTM prediction, and quantum feature + Transformer-XL prediction, supplemented by markings for extreme fluctuation periods and nighttime periods. From Figure 5As can be seen, compared with traditional LSTM prediction, the Quantum Feature + Transformer-XL prediction curve has a higher degree of fit with the actual photovoltaic power output curve, especially during periods of extreme fluctuations, and can more accurately capture power output changes; moreover, during the nighttime period (8 pm to 5 am), the photovoltaic power output of all three curves is 0. Therefore, the Quantum Feature + Transformer-XL prediction method, compared with traditional LSTM, is more accurate in capturing power output fluctuations in photovoltaic power output prediction, which can improve the accuracy of photovoltaic power output prediction, provide a more reliable basis for the control of distribution networks containing photovoltaics, and ensure the power balance and stable operation of the distribution network.

[0045] Figure 6 This is a bar chart comparing the benefits of the hardware-algorithm collaborative scheme proposed in Embodiment 1 of the present invention; it includes three dimensions: edge node hardware cost, edge node energy consumption, and distribution network expansion investment, comparing the traditional scheme with the hardware-algorithm collaborative scheme. Figure 6 As can be seen, compared with traditional solutions, the hardware-algorithm collaborative solution significantly reduces edge node hardware costs, edge node energy consumption, and distribution network expansion investment. This demonstrates that the hardware-algorithm collaborative solution, through dynamic neural architecture search and a quantum-classical hybrid computing model, can effectively reduce hardware costs, lower energy consumption, and decrease investment in distribution network expansion and upgrades. It also optimizes resource allocation through intelligent means, showcasing the cost control advantages of this collaborative solution in distribution network construction and operation.

[0046] Figure 7 This is a bar chart comparing the traditional scheme and the quantum-dynamic architecture scheme in terms of three quantitative indicators: voltage over-limit probability, new energy absorption rate, and mean control error under the uncertainty scenario proposed in Embodiment 1 of this invention; from Figure 7 As can be seen, under different scenarios such as typhoons, severe convection, normal weather, and continuous rain, the quantum-dynamic architecture scheme significantly reduces the voltage exceedance probability (e.g., from 0.72 to 0.12 in typhoon scenarios), greatly improves the renewable energy absorption rate (e.g., from 0.62 to 0.88 in normal weather scenarios), and significantly reduces the mean control error (e.g., from 15.60 to 2.10 in typhoon scenarios) compared to the traditional scheme. This demonstrates that the risk assessment technology based on quantum probability modeling and the adaptive optimization mechanism of the dynamic neural architecture work synergistically to accurately quantify the uncertainty of distributed energy output. This enhances the distribution network's resistance to renewable energy fluctuations in both extreme and normal scenarios, effectively improves renewable energy absorption efficiency, and significantly enhances the system's robustness under uncertain scenarios.

[0047] The present invention, in Embodiment 1, proposes a method for dynamic coordinated control of voltage in a virtual power plant and distribution network. By deeply integrating cutting-edge technologies such as quantum machine learning, neural symbol integration, and dynamic architecture search, it systematically solves the core challenges faced by traditional voltage control in scenarios with a high proportion of new energy access, and achieves a dual breakthrough in technical performance and engineering value.

[0048] Example 2 This invention also proposes a virtual power plant-distribution network voltage dynamic collaborative control device, comprising: Memory, used to store computer programs; When a processor executes the computer program, the method steps are as follows: In step S100, the process begins.

[0049] In step S110, multi-source heterogeneous data for collaborative control is acquired to construct a hybrid dataset. The system uses edge computing nodes to access distributed energy output data, distribution network topology status, meteorological monitoring data and user load curves in real time to form multi-source heterogeneous data.

[0050] In step S120, the hybrid dataset is mapped to the quantum state space for feature extraction. Quantum variational feature learning technology is employed to map heterogeneous data to the quantum state space for feature extraction. The nonlinear correlation between DERs output and meteorological factors is captured using quantum entanglement properties. Simultaneously, classical signal processing methods are used to filter out measurement noise, providing high-dimensional feature vectors for subsequent artificial intelligence analysis.

[0051] In step S130, uncertainty quantification is performed; In step S140, the improved Transformer-XL model is used to fuse quantum features, satellite cloud images and voltage waveform data, and quantum probability modeling is adopted; the DERs output scenario is generated based on the quantum superposition principle, and the voltage over-limit probability amplitude of each node is quantified by quantum Monte Carlo tree search to generate a distribution network risk heat map; In step S150, a hierarchical control decision is constructed; specifically: the energy storage reactive power compensation problem is modeled as a quantum optimization problem to generate real-time adjustment instructions for energy storage reactive power; based on the distribution network risk heat map and distribution network topology characteristics, the distribution network topology characteristics are extracted through an interpretable graph reasoning network, and the constraints of the distribution network operation procedure are represented by linear time-series logic to generate a controllable load dispatching strategy; based on the day-ahead DERs output prediction, the energy aggregation strategy is optimized to generate interpretable dispatching rules; In step S160, hierarchical control decisions are executed, and control effect feedback is provided; In step S170, a causal logic chain for voltage limit exceedance events is generated using an invariant graph inference network. The contribution of features to control decisions is quantified using SHAP values, and a meta-learning mechanism is combined to achieve rapid iterative optimization of model parameters. In step S180, the process ends.

[0052] The present invention, in embodiment 2, proposes a virtual power plant-distribution network voltage dynamic collaborative control device. By deeply integrating cutting-edge technologies such as quantum machine learning, neural symbol integration, and dynamic architecture search, it systematically solves the core challenges faced by traditional voltage control in scenarios with a high proportion of new energy access, and achieves a dual breakthrough in technical performance and engineering value.

[0053] The present invention, in embodiment 2, proposes a virtual power plant-distribution network voltage dynamic collaborative control device. At the technical performance level, it constructs a full-chain intelligent system of "prediction-evaluation-control-interpretation": a multimodal quantum feature learning model significantly improves the prediction accuracy of distributed energy output; combined with quantum Monte Carlo risk assessment technology, it achieves probabilistic and accurate prediction of voltage limit exceedance risks; a hierarchical neural symbolic control architecture, through the synergy of quantum annealing and deep reinforcement learning, improves the voltage transient response speed to the millisecond level; simultaneously, by utilizing the joint reasoning of graph neural networks and power rules, it ensures that the control strategy fully complies with operational specifications, resolving the contradiction between the black-box nature of artificial intelligence decision-making and power system security.

[0054] The virtual power plant-distribution network voltage dynamic collaborative control device proposed in Embodiment 2 of this invention significantly reduces edge computing costs and distribution network expansion investment through quantum-classical hybrid hardware deployment and dynamic architecture optimization at the engineering application level, achieving efficient utilization of hardware resources. The interpretability enhancement technology shortens the root cause location time of voltage over-limit events, significantly improving operation and maintenance efficiency. In uncertain scenarios such as extreme weather, the robust control strategy based on quantum probability significantly improves the renewable energy consumption rate, providing key technical support for the construction of a highly resilient power grid under the "dual carbon" target.

[0055] The virtual power plant-distribution network voltage dynamic collaborative control device proposed in Embodiment 2 of this invention, from an industry development perspective, breaks through the traditional power system's dependence on centralized power sources, providing an intelligent paradigm for the collaborative control of virtual power plants and distribution networks, and promoting the upgrade of the power system from "rigid control" to "flexible adaptation." Its engineering implementation will significantly improve the grid's ability to accept new energy sources, enhance the resilience and security of urban distribution networks, and lay a technological foundation for emerging application scenarios such as smart cities and vehicle-to-grid interaction, possessing significant economic value and social benefits.

[0056] It should be noted that the present invention also provides an electronic device, including: a communication interface capable of interacting with other devices such as network devices; and a processor connected to the communication interface to enable information interaction with other devices, used to execute a virtual power plant-distribution network voltage dynamic collaborative control method provided by one or more of the above technical solutions when running a computer program, wherein the computer program is stored in a memory. In practical applications, the various components of the electronic device are coupled together through a bus system. It is understood that the bus system is used to realize the connection and communication between these components. In addition to a data bus, the bus system also includes a power bus, a control bus, and a status signal bus. The memory in the embodiments of this application is used to store various types of data to support the operation of the electronic device. Examples of this data include any computer program used to operate on the electronic device. It is understood that the memory can be volatile memory or non-volatile memory, or both. Non-volatile memory can be read-only memory (ROM), programmable read-only memory (PROM), erasable programmable read-only memory (EPROM), electrically erasable programmable read-only memory (EEPROM), ferromagnetic random access memory (FRAM), flash memory, magnetic surface memory, optical disc, or compact disc read-only memory (CD-ROM); magnetic surface memory can be disk storage or magnetic tape storage. Volatile memory can be random access memory (RAM), which is used as an external cache.By way of example, but not limitation, many forms of RAM are available, such as Static Random Access Memory (SRAM), Synchronous Static Random Access Memory (SSRAM), Dynamic Random Access Memory (DRAM), Synchronous Dynamic Random Access Memory (SDRAM), Double Data Rate Synchronous Dynamic Random Access Memory (DDRSDRAM), Enhanced Synchronous Dynamic Random Access Memory (ESDRAM), SyncLink Dynamic Random Access Memory (SLDRAM), and Direct Rambus Random Access Memory (DRRAM). The memories described in the embodiments of this application are intended to include, but are not limited to, these and any other suitable types of memory. The methods disclosed in the embodiments of this application can be applied to or implemented by a processor. The processor may be an integrated circuit chip with signal processing capabilities. In implementation, each step of the above method can be completed by integrated logic circuits in the processor hardware or by instructions in software form. The processor can be a general-purpose processor, a DSP (Digital Signal Processing, i.e., a chip capable of implementing digital signal processing technology), or other programmable logic devices, discrete gate or transistor logic devices, discrete hardware components, etc. The processor can implement or execute the methods, steps, and logic block diagrams disclosed in the embodiments of this application. A general-purpose processor can be a microprocessor or any conventional processor, etc. The steps of the methods disclosed in the embodiments of this application can be directly manifested as execution by a hardware decoding processor, or execution by a combination of hardware and software modules in the decoding processor. The software modules can be located in a storage medium, which is located in memory. The processor reads the program from the memory and, in conjunction with its hardware, completes the steps of the aforementioned method. When the processor executes the program, it implements the corresponding processes in the various methods of the embodiments of this application; for simplicity, these will not be elaborated further here.

[0057] The description of the relevant parts of the virtual power plant-distribution network voltage dynamic collaborative control device provided in Embodiment 2 of this application can be found in the detailed description of the corresponding parts of the virtual power plant-distribution network voltage dynamic collaborative control method provided in Embodiment 1 of this application, and will not be repeated here.

[0058] It should be noted that, in this document, relational terms such as "first" and "second" are used merely to distinguish one entity or operation from another, and do not necessarily require or imply any such actual relationship or order between these entities or operations. Furthermore, the terms "comprising," "including," or any other variations thereof are intended to cover non-exclusive inclusion, such that the elements inherent in a process, method, article, or apparatus that includes a list of elements are included. Without further limitations, an element defined by the phrase "comprising one..." does not exclude the presence of other identical elements in the process, method, article, or apparatus that includes said element. Additionally, portions of the technical solutions provided in the embodiments of this application that are consistent with the implementation principles of corresponding technical solutions in the prior art have not been described in detail to avoid excessive elaboration.

[0059] While specific embodiments of the present invention have been described above in conjunction with the accompanying drawings, this is not intended to limit the scope of protection of the present invention. Those skilled in the art can make other modifications or variations based on the above description. It is neither necessary nor possible to exhaustively describe all embodiments here. Various modifications or variations that can be made by those skilled in the art without creative effort based on the technical solutions of the present invention are still within the scope of protection of the present invention.

Claims

1. A method for dynamic coordinated control of voltage in a virtual power plant and distribution network, characterized in that, Includes the following steps: A hybrid dataset is constructed by acquiring multi-source heterogeneous data for collaborative control, and the hybrid dataset is mapped to quantum state space for feature extraction; the multi-source heterogeneous data includes distributed energy output data, distribution network topology status, and meteorological monitoring data. The improved Transformer-XL model is used to fuse quantum features, satellite cloud images, and voltage waveform data, and quantum probabilistic modeling is employed. The power output scenario of DERs is generated based on the principle of quantum superposition. The voltage over-limit probability amplitude of each node is quantified by quantum Monte Carlo tree search, and a heat map of distribution network risk is generated. Constructing a hierarchical control decision-making mechanism; specifically: modeling the energy storage reactive power compensation problem as a quantum optimization problem to generate real-time adjustment instructions for energy storage reactive power; based on the distribution network risk heat map and distribution network topology characteristics, extracting distribution network topology characteristics through an interpretable graph reasoning network, and using linear time-series logic to represent the constraints of the distribution network operation procedure to generate a controllable load scheduling strategy; Based on the day-ahead DERs output prediction, optimize the energy aggregation strategy and generate interpretable scheduling rules; The hierarchical control decision is executed, and then a causal logic chain of voltage over-limit events is generated through an invariant graph inference network. The contribution of features to the control decision is quantified using SHAP values, and the model parameters are rapidly iteratively optimized by combining a meta-learning mechanism.

2. The method for dynamic coordinated control of virtual power plant-distribution network voltage according to claim 1, characterized in that, The process of mapping multi-source heterogeneous data to quantum state space for feature extraction includes: ；（1） In the formula, This indicates that the hybrid dataset, after being encoded by quantum states, The corresponding quantum state; The parameterized quantum coding operation is represented by Θ = {ф1, ф2}; ф1 represents the rotation gate parameter; ф2 represents the entanglement gate parameter; Represents the initial quantum state; Extracting quantum features through quantum measurement ; ； In the formula, It is a quantum state The dual state.

3. The method for dynamic coordinated control of virtual power plant-distribution network voltage according to claim 2, characterized in that, The improved Transformer-XL model is used to fuse quantum features, satellite cloud image features, and voltage waveform data, and quantum probabilistic modeling is employed as follows: ； in, This represents the features after being fused by the attention fusion module; To represent the various features input to the fusion module; Features of satellite cloud images; Voltage waveform data; The attention weight is calculated using the following formula: ； Quantum probabilistic modeling is used to address the uncertainty in DERs output, specifically: ； In the formula, Historical contribution data for DERs; To indicate that the output of DERs is The quantum probability at time; Tr[] represents the trace operation; This is the quantum state density matrix corresponding to the data. Let be the density matrix of the reference state.

4. The method for dynamic coordinated control of virtual power plant-distribution network voltage according to claim 3, characterized in that, The process of generating DERs output scenarios based on the quantum superposition principle, and generating a distribution network risk heatmap by quantizing the voltage over-limit probability amplitude of each node through quantum Monte Carlo tree search, includes: Define the quantum state corresponding to the DERs output set as: ； In the formula, This represents the quantum state corresponding to the output set of DERs; For the first The ground state of a power output scenario. For complex amplitude coefficients, satisfying ; Define search tree nodes Indicates the state of the distribution network, edge The node value function represents the change in DER output: ； In the formula, For the edge Number of visits; For the first The reward value of this simulation; The selection strategy employs quantum upper confidence bounds. Specifically: ； In the formula, For nodes Total number of visits; When expanding nodes, the ground state of the expanded output scenario is generated through quantum gate operations. : ； In the formula, For parameterized quantum gates; By converting superposition states into classical probability distributions through quantum measurement, a voltage limit operator is defined. Its quantum state The expected value is: ； The quantum probability amplitude of voltage exceeding the limit is expressed as: : ； In the formula, This represents the voltage limit exceedance factor for a single constraint. The mathematical expression for mapping the probability amplitude to the distribution network topology and generating a risk heatmap is as follows: ； In the formula, For nodes Risk value; path( ) for scene The affected node paths; Based on quantum probability distribution, a robust optimization model for energy storage charging and discharging plans is constructed; the energy storage state vector is defined. The optimization objective is: ； In the formula: To optimize variables; It is a quantum state; For expectation value operators; To control the cost function, Risk aversion coefficient; For variance operators; This is the loss function.

5. The method for dynamic coordinated control of virtual power plant-distribution network voltage according to claim 4, characterized in that, The process of modeling the energy storage reactive power compensation problem as a quantum optimization problem to generate real-time energy storage reactive power adjustment commands includes: Define a quadratic unconstrained binary optimization model : ； In the formula, Let ∈{-1,+1}n be the state vector of the qubit. Let be the coupling matrix. It is the bias vector; The optimal state q* is solved using a quantum annealer and serves as the initial policy for the gradient of the deep deterministic policy: ； In the formula, A set of information describing the environment in which an intelligent agent exists; For the agent in state The following actions were taken; For deterministic policy functions, The standard deviation of noise. To explore noise.

6. The method for dynamic coordinated control of virtual power plant-distribution network voltage according to claim 5, characterized in that, Based on the distribution network risk heat map and distribution network topology characteristics, the process of extracting distribution network topology characteristics through an interpretable graph reasoning network, representing the constraints of distribution network operation procedures using linear time-series logic, and generating a controllable load dispatching strategy includes: Message passing mechanism for building graph neural networks: ； In the formula, For nodes of Layer node characteristics; It is the set of adjacent nodes; For nodes of Layer node characteristics; For the nodes in the graph The set of directly connected neighboring nodes; For nodes A neighbor node In the Layer node representation; For the first A trainable weight matrix in the layer used to process edge information or neighbor node information; For the first The trainable weight matrix in the layer used to process information about its own nodes; For activation functions; The constraints of the distribution network operation procedures are represented using linear sequential logic and then transformed into penalty terms using a weighted automaton. ； In the formula, For time range; Indicates time The degree of violation of the rules; wt is the time weight; For a moment The corresponding linear sequential logic formula; The joint loss function is: ; In the formula, For the joint loss function; To reinforce learning loss; This is the first weighting coefficient.

7. The method for dynamic coordinated control of virtual power plant-distribution network voltage according to claim 6, characterized in that, The process of generating interpretable scheduling rules based on optimizing energy aggregation strategies according to day-ahead DERs output predictions includes: Define the neural architecture search space Optimize architecture parameters using genetic algorithms : ； In the formula, β is the set of parameters of the neural architecture; Assess the loss for the architecture; Construct interpretable rules for generating DERs aggregation strategies using symbolic systems: ； In the formula, Interpretable rules representing the aggregation strategy of DERs generated by the symbol system; For state judgment based on logical predicates, For the corresponding aggregation action; Genetic algorithms simultaneously optimize the parameters of neural architecture and symbolic rules: ； In the formula, Optimal sign rule; Lsym( ) is the loss function for the sign rule; Kirchhoff's voltage law can be expressed as a loop constraint in graph theory: ； In the formula, For the line Voltage drop; It is a closed loop in the circuit; Voltage over-limit threshold is transformed into state-space constraint: ； In the formula, For nodes At any moment The voltage value; Minimum voltage threshold; Maximum voltage threshold; The above rules are embedded as prior knowledge into the action space of reinforcement learning: ； In the formula, For safe actions within the search space; To enhance the learning agent's ability to learn at any time The actions taken.

8. The method for dynamic coordinated control of virtual power plant-distribution network voltage according to claim 7, characterized in that, The method further includes dynamic neural architecture optimization based on conditional computation graphs and meta-learning, wherein the mathematical expression of the dynamic architecture is: ； In the formula, For a dynamic gated vector, satisfying ; For each sub-module in the dynamic neural architecture The set of parameters; This is a set of parameters related to dynamic gating. For feature vectors; For the first Parameters within each sub-model; for Each submodule; Design a gating function based on input features. To achieve dynamic switching of network architecture, a fluctuation feature extractor is defined. ; ； In the formula, The sequence of changes in the output of DERs; MaxPooling() is the maximum pooling operation; For convolutional neural networks The output of the processing; In scenarios with stable loads, the gating mechanism switches to a lightweight sub-network. Its structure is defined as follows: ； In the formula, This is for fully connected layer operations; The mask vector is used to implement parameter pruning through sparse regularization; ； In the formula, For sparsity regularization loss, The regularization coefficient is used. A model-independent meta-learning framework is adopted to define the meta-learner. To optimize initial parameters ; In the formula, The optimal parameters; For a set of tasks; Task-specific data; Design a two-level optimization problem to achieve coordinated optimization of dynamic architecture and model parameters: ； In the formula, Let be the expected loss function.

9. The method for dynamic coordinated control of virtual power plant-distribution network voltage according to claim 8, characterized in that, The process of generating causal logic chains for voltage limit exceedance events using invariant graph inference networks, quantifying the contribution of features to control decisions using SHAP values, and combining this with a meta-learning mechanism to achieve rapid iterative optimization of model parameters includes: Invariant graph inference networks extract topological association features through multi-layer message passing: ； In the formula, For nodes In the Layer feature representation; For activation functions; Characteristics of line impedance; Adjacent nodes To the current node Passing by In the Layer weight matrix; Adjacent nodes In the Layer feature representation; With line impedance characteristics The relevant weight matrix; For nodes In the Layer feature representation; For nodes The set of all neighboring nodes; The rule engine maps graph features into causal logic chains: ； In the formula, The output is the weight matrix; RuleEngine transforms the probability distribution into a natural language interpretation. Combine SHAP values ​​with graph features to generate a feature importance heatmap: ； In the formula, Features In the importance distribution across the entire network, Normalize() is the normalization function; For selection functions; node eigenvectors; Design the joint loss function Strengthen the consistency between interpretation and control: ； In the formula, This is the second weighting coefficient; Attribution function; For interpretation functions; defined as: ； In the formula, These are the preset feature weights; The explanatory or attribution function generated for the model is for the first... The output value of each feature.

10. A virtual power plant-distribution network voltage dynamic collaborative control device, characterized in that, include: A memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the computer program, when executed by the processor, implements the steps of the method as described in any one of claims 1 to 9.

Images