A system and method for suppressing geomagnetic disturbances and high-altitude electromagnetic pulses

By combining time-coupled sub-blob bar optimization and active safety model prediction reinforcement learning with solid-state transformers and physical information heterogeneous graph neural networks, the problem of low-frequency common-mode current caused by geomagnetic disturbances and high-altitude electromagnetic pulses was solved, achieving fast response and stable grid flux control.

CN123121548APending Publication Date: 2026-11-17SHAANXI UNIV OF SCI & TECH
View PDF 1 Cites 0 Cited by

Patent Information

Application Number
CN202611430433.X
Authority / Receiving Office
CN · China
Patent Type
Applications(China)
Current Assignee / Owner
Filing Date
2026-09-15
Publication Date
2026-11-17

AI Technical Summary

Technical Problem

Existing technologies cannot effectively suppress low-frequency common-mode currents caused by geomagnetic disturbances and high-altitude electromagnetic pulses. Furthermore, traditional deployment optimization methods ignore the dynamic process of geomagnetic disturbances, leading to rapid saturation of the transformer core and an inability to maintain stability in scenarios with high flux rise rates.

Method used

By combining time-coupled sub-Blule bar optimization and active safety model prediction reinforcement learning, the deployment location and parameter configuration of solid-state transformers are predicted and optimized through solid-state transformers, physical information heterogeneous graph neural networks, and two-layer Wasserstein fuzzy sets, and common-mode current is detected and suppressed in real time.

Benefits of technology

It achieves millisecond-level flux response prediction, shortens controller response time by 65%, reduces total harmonic distortion to 1.9%, maintains asymptotic stability under extreme GIC pulses, and has high robustness and economy.

✦ Generated by Eureka AI based on patent content.

Smart Images

  • Figure CN123121548A_ABST
    Figure CN123121548A_ABST
Patent Text Reader

Abstract

This invention belongs to the field of electromagnetic protection technology for power systems, specifically relating to a system and method for suppressing geomagnetic disturbances and high-altitude electromagnetic pulses. It includes: a physical information heterogeneous graph neural network deployment optimization module, configured to model the power grid topology as a heterogeneous graph, embedding the physical laws of the power system into the message passing process using a physical information neural network to predict the peak magnetic flux response of the core at each node; and a time-coupled split-Brow bar optimization layer, employing a two-layer Wasserstein fuzzy set to characterize the spatial distribution uncertainty and temporal evolution path uncertainty of geomagnetic disturbances or high-altitude electromagnetic pulses. The two modules collaboratively output the minimum number of solid-state transformer deployment locations and parameter configurations. This invention can output the minimum number of solid-state transformer deployment locations and parameter configurations while simultaneously considering the spatial distribution uncertainty and temporal evolution path uncertainty of geomagnetic disturbances and high-altitude electromagnetic pulses.
Need to check novelty before this filing date? Find Prior Art

Description

Technical Field

[0001] This invention belongs to the field of electromagnetic protection technology for power systems, specifically relating to a system and method for suppressing geomagnetic disturbances and high-altitude electromagnetic pulses. Background Technology

[0002] Existing suppression technologies are mainly divided into three categories: passive blocking devices (such as neutral point series capacitors or resistors), line switching, and traditional neutral point injection / cancellation devices. These technologies each have drawbacks, including the potential to induce ferroresonance, computational complexity and inability to eliminate geomagnetically induced current (GIC) sources, and slow response speed. Furthermore, existing deployment optimization methods all assume that the spatial distribution of the geoelectric field during geomagnetic disturbances (GMD) events is static, neglecting the dynamic process of geomagnetic disturbance evolution over time. In actual disasters, the geoelectric field can surge several times within tens of seconds, causing the transformer core to rapidly enter deep saturation from the linear region. Traditional sub-Browser optimization only considers the uncertainty of spatial distribution and cannot constrain the rate of change in the time dimension; therefore, the deployment scheme may fail in scenarios with excessively high flux rise rates. Simultaneously, security reinforcement learning based on Lyapunov filters is a post-event correction mechanism, and a security gap exists when the rate of change of disturbance exceeds the controller's response capability.

[0003] Chinese patent publication number CN120810605A, entitled "A Peak-Shaving Method for Electricity Management Based on Multi-Objective Programming," discloses: constructing a heterogeneous graph network structure of the power system, defining an uncertainty representation model on the graph structure, constructing an uncertainty propagation dynamics model, implementing a graph structure-aware uncertainty propagation algorithm to generate the spatiotemporal evolution trajectory of uncertainty, constructing a graph structure-aware robust optimization framework, implementing a multi-level risk zoning management strategy, and discovering the utilization of the uncertainty complementarity effect; however, it cannot achieve: physical suppression of low-frequency common-mode current induced in the power grid by high-altitude electromagnetic pulses or geomagnetic disturbances. Summary of the Invention

[0004] In order to overcome the problems existing in the prior art, the present invention aims to provide a geomagnetic disturbance and high-altitude electromagnetic pulse suppression system and method, which simultaneously makes up for the lack of dynamic adaptability from the two levels of deployment planning and control execution by using time-coupled sub-Bruker optimization and active safety model prediction reinforcement learning.

[0005] To achieve the above objectives, the technical solution adopted by the present invention is as follows: In a first aspect, the present invention provides a geomagnetic disturbance and high-altitude electromagnetic pulse suppression system, comprising: A solid-state transformer includes an input stage, an isolation stage, and an output stage connected in sequence. The input stage is a four-arm AC-DC converter, the isolation stage is a high-frequency isolation DC-DC converter, and the output stage is a DC-AC inverter. The physical information heterogeneous graph neural network deployment optimization module is configured to take a power grid topology graph as input, the power grid topology graph having multiple node types and edge types, and the physical information heterogeneous graph neural network combines a heterogeneous graph neural network with a physical information neural network to embed the physical laws of the power system into the message transmission process of the graph neural network, and is used to predict the peak magnetic flux response of the iron core of each node under any deployment scheme. A time-coupled sub-Browser bar optimization layer is coupled to the physical information heterogeneous graph neural network deployment optimization module. It is configured to: employ a two-layer Wasserstein fuzzy set to characterize the spatial distribution uncertainty and temporal evolution path uncertainty of geomagnetic disturbances and upper-air electromagnetic pulses in the ground electric field, respectively. The inner layer is an autoregressive generative model jointly driven by the three parameters of the ground electric field peak value, duration, and rise rate to construct a time path set. The physical information heterogeneous graph neural network deployment optimization module is used in conjunction with the time-coupled sub-Browser bar optimization layer to output the minimum number of solid-state transformer deployment locations and parameter configurations. An active safety model predictive reinforcement learning adaptive controller is configured to: after the solid-state transformer is deployed, learn and execute the optimal control strategy online through a dual-agent architecture, and output the switching sequence control command of the four-arm AC-DC converter; The deployment-control closed-loop online update framework system is configured to use a mirror descent algorithm to feed back the control performance during the operation phase to the deployment optimization module in real time.

[0006] Optionally, the optimization objective of the time-coupled sub-bar optimization layer is:

[0007] Where x is the binary deployment decision vector, and ξ(t) is the time-dependent geomagnetic disturbance intensity. Let λ be the peak magnetic flux of the core of the i-th conventional transformer at time t, and λ be the flux change rate penalty coefficient; output a robust deployment scheme that minimizes the worst-case expected magnetic flux across the entire time domain.

[0008] Optionally, the four-arm AC-DC converter includes: a first switch S1, a second switch S2, a third switch S3, a fourth switch S4, a fifth switch S5, a sixth switch S6, a seventh switch S7, and an eighth switch S8; the drain of the first switch S1 is connected to the positive terminal of the DC bus, the source of the first switch S1 is connected to the drain of the second switch S2, and the source of the second switch S2 is connected to the negative terminal of the DC bus; the drain of the third switch S3 is connected to the positive terminal of the DC bus. The source of the third switch S3 is connected to the drain of the fourth switch S4, and the source of the fourth switch S4 is connected to the negative terminal of the DC bus; the drain of the fifth switch S5 is connected to the positive terminal of the DC bus, and the source of the fifth switch S5 is connected to the drain of the sixth switch S6, and the source of the sixth switch S6 is connected to the negative terminal of the DC bus; the drain of the seventh switch S7 is connected to the positive terminal of the DC bus, and the source of the seventh switch S7 is connected to the drain of the eighth switch S8, and the source of the eighth switch S8 is connected to the negative terminal of the DC bus.

[0009] Optionally, the solid-state transformer suppression system further includes an energy storage system electrically coupled to the DC bus of the solid-state transformer.

[0010] Optionally, the physical information heterogeneous graph neural network deployment optimization module uses a proxy power flow model for offline training. The training parameters of the proxy power flow model include: transmission line length, resistance, grounding resistance, network topology, and transformer nonlinear magnetization characteristic curve. The physical information heterogeneous graph neural network is used to learn the end-to-end mapping relationship from the deployment scheme to the magnetic flux response of the entire power grid core during training.

[0011] Optionally, in the two-layer Wasserstein fuzzy set of the time-coupled sub-Bruker bar optimization layer, the outer Wasserstein sphere is constructed based on the spatial distribution statistics of historical geomagnetic storm events, and the inner time evolution path is generated by a conditional variational autoencoder and is configured to sample the set of possible evolution paths within the future time window based on the current geoelectric field state; the magnetic flux change rate penalty coefficient is selected through cross-validation.

[0012] Optionally, the proactive safety model predictive reinforcement learning adaptive controller includes: The main control unit employs a deep reinforcement learning policy network to balance the power of the solid-state transformer and reduce common-mode current stress with the goal of maximizing long-term cumulative rewards. The safety prediction processing unit is used for gradient training based on the diffusion strategy. It takes the current system state and the predicted trajectory for the next k steps as input and outputs the early correction action. The value of k is dynamically determined by the safety horizon function, which calculates the shortest safety warning time online based on the real-time detected rate of rise of the geomagnetic induced current and the system inertial parameters. The safety constraint internalization processing unit is used to embed Lyapunov constraints into the objective function of the main control unit in the form of Karl-Kun-Tucker conditions during policy gradient updates, so that the main control unit itself tends to generate safe actions.

[0013] Optionally, the mirror descent algorithm defines Bregman divergence as a geometric metric for the deployment scheme. In each iteration of mirror descent, the gradient estimate is updated based on the cumulative performance residual of the current deployment scheme, and the updated deployment decision vector is output.

[0014] Optionally, the safety horizon function for:

[0015] Where α and β are weighting coefficients. The saturation flux threshold, To estimate the magnetic flux, V is the rate of change of magnetic flux. dc ε is the DC bus voltage, and ε is a small constant to prevent division by zero.

[0016] Secondly, the present invention provides a method for suppressing geomagnetic disturbances and upper-level electromagnetic pulses, employing the aforementioned geomagnetic disturbance and upper-level electromagnetic pulse suppression system, comprising the following steps: Acquire power grid topology data and construct a power grid heterogeneity diagram; Using the aforementioned heterogeneous power grid graph as the input to the physical information heterogeneous graph neural network deployment optimization module, the physical laws of the power system are embedded into the message transmission process of the graph neural network through the physical information heterogeneous graph neural network, thereby predicting the peak magnetic flux response of the iron core of each node under any deployment scheme. By using a time-coupled sub-Bruker optimization layer, a two-layer Wasserstein fuzzy set is used to characterize the spatial distribution uncertainty and temporal evolution path uncertainty of geomagnetic disturbances and upper-air electromagnetic pulses in the ground electric field. The inner layer is constructed by an autoregressive generative model driven by three parameters: peak value, duration, and rise rate of the ground electric field to build a time path set. By coordinating the physical information heterogeneous graph neural network deployment optimization module with the time-coupled sub-Bruker optimization layer, the minimum number of solid-state transformer deployment locations and parameter configurations are output under the premise of simultaneously considering spatial distribution uncertainty and temporal evolution path uncertainty. Install the solid-state transformer according to the deployment location and parameter configuration described; During operation, the zero-sequence common-mode current component in the power grid is detected in real time by the active safety model prediction reinforcement learning adaptive controller; when the common-mode current exceeds the preset threshold, the four-arm AC-DC converter is controlled to perform common-mode current redirection operation.

[0017] Compared with the prior art, the present invention has the following beneficial effects: This invention models the power grid topology as a heterogeneous graph and embeds the physical laws of the power system into a physical information neural network to achieve millisecond-level magnetic flux response prediction. A time-coupled split-Bruker optimization layer uses a two-layer Wasserstein fuzzy set to simultaneously characterize the static uncertainty of the spatial distribution of geomagnetic disturbances and the uncertainty of the temporal evolution path. A robust deployment scheme that minimizes the worst-case expected magnetic flux is solved by penalizing the gradient term of the magnetic flux change rate. An active safety model predictive reinforcement learning controller replaces the passive correction mode of the traditional Lyapunov filter with a dual-agent architecture of a main control unit and a safety prediction processing unit. It utilizes a diffusion strategy gradient to generate safety correction actions k steps in advance, achieving an active safety barrier. This invention injects real-time control performance feedback into deployment optimization based on mirror descent, enabling the deployment scheme to continuously evolve with the power grid.

[0018] Simulations show that the present invention achieves full-grid safety threshold constraints with less than 50% deployment quantity, reduces controller response time by 65%, reduces total harmonic distortion to 1.9%, and can maintain asymptotic stability under extreme GIC pulse rise rates, demonstrating high robustness, high economy, and strong adaptability. Attached Figure Description

[0019] The accompanying drawings described herein are for illustrative purposes only and are not intended to limit the scope of the invention in any way. Furthermore, the shapes and proportions of the components in the drawings are merely schematic to aid in understanding the invention and do not specifically limit the shapes and proportions of the components. In the drawings: Figure 1 This is a main circuit topology diagram of a solid-state transformer four-arm AC-DC converter provided in an embodiment of the present invention; Figure 2 This is the overall system architecture for HEMP and GMD suppression in the SST power system according to an embodiment of the present invention; Figure 3 This is a schematic diagram of the physical information heterogeneous graph neural network algorithm according to an embodiment of the present invention; Figure 4 This is a schematic diagram of the principle of TC-DRO time-coupled split-Brow bar optimization of two-layer Wasserstein fuzzy sets according to an embodiment of the present invention; Figure 5 This is a diagram of the dual-agent architecture for predictive reinforcement learning in the Safe-MPRL active security model according to an embodiment of the present invention. Figure 6 This is a flowchart of the control closed-loop online update image descent iteration process according to an embodiment of the present invention. Detailed Implementation

[0020] To enable those skilled in the art to better understand the technical solutions of this invention, the technical solutions of the embodiments of this invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of this invention, and not all embodiments. Based on the embodiments of this invention, all other embodiments obtained by those skilled in the art without creative effort should fall within the scope of protection of this invention.

[0021] Therefore, the following detailed description of the embodiments of the invention provided in the accompanying drawings is not intended to limit the scope of the claimed invention, but merely to illustrate selected embodiments of the invention. All other embodiments obtained by those skilled in the art based on the embodiments of the invention without inventive effort are within the scope of protection of the invention.

[0022] In the description of the embodiments of the present invention, it should be noted that if terms such as "upper", "lower", "horizontal", "inner" indicate the orientation or positional relationship based on the orientation or positional relationship shown in the accompanying drawings, or the orientation or positional relationship in which the product of the invention is usually placed during use, they are only for the convenience of describing the present invention and simplifying the description, and do not indicate or imply that the device or element referred to must have a specific orientation, or be constructed and operated in a specific orientation, and therefore should not be construed as a limitation of the present invention.

[0023] When an element is referred to as being "set on" another element, it can be directly on the other element or there may be an intervening element. When an element is referred to as being "connected to" another element, it can be directly connected to the other element or there may be an intervening element. The terms "vertical," "horizontal," "left," "right," and similar expressions used herein are for illustrative purposes only and do not represent the only embodiments. The use of the term "horizontal" does not imply that the component is required to be absolutely horizontal, but rather that it may be slightly tilted. "Horizontal" simply means that its direction is more horizontal than "vertical," and does not mean that the structure must be completely horizontal, but rather that it may be slightly tilted.

[0024] It should be noted that similar reference numerals and letters in the following figures denote similar items; therefore, once an item is defined in one figure, it does not need to be further defined and explained in subsequent figures. In the description of this invention, it should be understood that the terms "comprising" and "including" indicate the presence of the described features, integrals, steps, operations, elements, and / or components, but do not exclude the presence or addition of one or more other features, integrals, steps, operations, elements, components, and / or collections thereof.

[0025] Unless otherwise defined, all technical and scientific terms used herein have the same meaning as commonly understood by one of ordinary skill in the art to which this invention pertains. The terminology used herein in the specification of the invention is for the purpose of describing particular embodiments only and is not intended to be limiting of the invention. As used in the specification and appended claims, the singular forms “a,” “an,” and “the” are intended to include the plural forms unless the context clearly indicates otherwise.

[0026] The present invention will now be described in detail with reference to the accompanying drawings.

[0027] Power systems, especially long-distance transmission lines, are highly susceptible to high-altitude electromagnetic pulses (HEMPs) and geomagnetic disturbances (GMDs). These events induce low-frequency common-mode currents similar to DC in the power grid, known as geomagnetically induced currents (GICs). When GICs flow into the grounded neutral point of traditional power transformers, they bias the magnetic flux operating point of the core, causing half-cycle saturation in the transformer. This leads to excitation current distortion, increased harmonics, a sharp rise in reactive power losses, localized overheating, and may even trigger malfunctions in protection devices, resulting in a chain reaction of power outages.

[0028] Solid-state transformers (SSTs) have flexible functions such as voltage regulation and power flow control, and can be used to suppress GMD and HEMP.

[0029] This invention provides a geomagnetic disturbance and high-altitude electromagnetic pulse suppression system, comprising: A solid-state transformer includes an input stage, an isolation stage, and an output stage connected in sequence. The input stage is a four-arm AC-DC converter, the isolation stage is a high-frequency isolation DC-DC converter, and the output stage is a DC-AC inverter. The physical information heterogeneous graph neural network deployment optimization module is configured to take a power grid topology graph as input, the power grid topology graph having multiple node types and edge types, and the physical information heterogeneous graph neural network combines a heterogeneous graph neural network with a physical information neural network to embed the physical laws of the power system into the message transmission process of the graph neural network, and is used to predict the peak magnetic flux response of the iron core of each node under any deployment scheme. A time-coupled sub-Browser bar optimization layer is coupled to the physical information heterogeneous graph neural network deployment optimization module. It is configured to: employ a two-layer Wasserstein fuzzy set to characterize the spatial distribution uncertainty and temporal evolution path uncertainty of geomagnetic disturbances and upper-air electromagnetic pulses in the ground electric field, respectively. The inner layer is an autoregressive generative model jointly driven by the three parameters of the ground electric field peak value, duration, and rise rate to construct a time path set. The physical information heterogeneous graph neural network deployment optimization module is used in conjunction with the time-coupled sub-Browser bar optimization layer to output the minimum number of solid-state transformer deployment locations and parameter configurations. An active safety model predictive reinforcement learning adaptive controller is configured to: after the solid-state transformer is deployed, learn and execute the optimal control strategy online through a dual-agent architecture, and output the switching sequence control command of the four-arm AC-DC converter; The deployment-control closed-loop online update framework system is used to enable the deployment scheme to continuously evolve with the power grid. It adopts the mirror descent algorithm to feed back the control performance of the operation phase to the deployment optimization module in real time: the Bregman divergence is defined as the geometric metric of the deployment scheme. Each iteration of mirror descent updates the gradient estimate based on the cumulative performance residual of the current deployment scheme and outputs the updated deployment decision vector. The update is implemented at the edge side through block coordinate mirror descent. Local re-optimization is triggered only when the cumulative performance residual exceeds a preset threshold to avoid frequent global calculations.

[0030] This invention provides a method for suppressing geomagnetic disturbances and upper-level electromagnetic pulses, employing the aforementioned geomagnetic disturbance and upper-level electromagnetic pulse suppression system, comprising the following steps: Acquire power grid topology data and construct a power grid heterogeneity diagram; Using the aforementioned heterogeneous power grid graph as the input to the physical information heterogeneous graph neural network deployment optimization module, the physical laws of the power system are embedded into the message transmission process of the graph neural network through the physical information heterogeneous graph neural network, thereby predicting the peak magnetic flux response of the iron core of each node under any deployment scheme. By using a time-coupled sub-Bruker optimization layer, a two-layer Wasserstein fuzzy set is used to characterize the spatial distribution uncertainty and temporal evolution path uncertainty of geomagnetic disturbances and upper-air electromagnetic pulses in the ground electric field. The inner layer is constructed by an autoregressive generative model driven by three parameters: peak value, duration, and rise rate of the ground electric field to build a time path set. The physical information heterogeneous graph neural network deployment optimization module and the time-coupled sub-Blu-ray bar optimization layer work together to output the minimum number of solid-state transformer deployment locations and parameter configurations. Install the solid-state transformer according to the deployment location and parameter configuration described; The zero-sequence common-mode current component in the power grid is detected in real time. When the common-mode current exceeds a preset threshold, the four-arm AC-DC converter is controlled to perform a common-mode current redirection operation.

[0031] This invention addresses the lack of dynamic adaptability by employing time-coupled bibliometric optimization and proactive safety model predictive reinforcement learning, simultaneously compensating for deficiencies in dynamic adaptation at both the deployment planning and control execution levels. This overcomes the shortcomings of existing solid-state transformer suppression technologies, such as deployment optimization ignoring temporal evolution uncertainties, safety reinforcement learning being passive correction rather than proactive prediction, and deployment and control being isolated. It achieves the effect of outputting the minimum number of solid-state transformer deployment locations and parameter configurations while simultaneously considering the spatial distribution uncertainties and temporal evolution path uncertainties of geomagnetic disturbances and high-altitude electromagnetic pulses, thereby ensuring that the core magnetic flux of the transformers in the entire power grid does not exceed limits at the lowest cost.

[0032] Example 1 This embodiment of a geomagnetic disturbance and high-altitude electromagnetic pulse suppression system includes: The deployment optimization subsystem is used to output the minimum number of solid-state transformer deployment schemes during the power grid planning phase.

[0033] An operation control subsystem is used to perform adaptive common-mode current suppression in deployed solid-state transformers. The deployment optimization subsystem includes a PIHGNN (Physics-Informed Heterogeneous Graph Neural Network) deployment optimization module and a Temporal-Coupled Distributionally Robust Optimization (TC-DRO) layer; the operation control subsystem also includes a Safe-MPRL (Safe Model Predictive Reinforcement Learning) adaptive controller.

[0034] The deployment-control closed-loop online update framework system is used to enable the deployment scheme to continuously evolve with the power grid.

[0035] The input to the PIHGNN deployment optimization module is a power grid topology graph G=(V,E), where V has multiple node types (substation nodes, traditional transformer nodes, and solid-state transformer candidate nodes), and E has multiple edge types (transmission lines and grounding connections). PIHGNN combines a heterogeneous graph neural network (HGNN) with a physical information neural network (PINN). By embedding the physical laws of the power system into the message passing process of the graph neural network, it learns the path from the deployment decision vector x to the peak magnetic flux Φ of the core at each node. i The end-to-end mapping relationship.

[0036] Specifically, each layer of message passing in PIHGNN has a loss regularization term that embeds the physical laws of the power system. This ensures that the output of the neural network conforms to the physical equations of transformer magnetization characteristics and GIC propagation. The PIHGNN deployment optimization module uses a surrogate power flow model for offline training, learning the end-to-end mapping relationship from deployment schemes to the magnetic flux response of the entire power grid core during training, thereby compressing the single fitness evaluation from minutes to milliseconds.

[0037] TC-DRO layer: Introduces temporal uncertainty into the deployment of GIC (Geomagnetically Induced Current) suppression devices. Traditional Wasserstein DRO (Distributionally Robust Optimization) only considers spatially distributed perturbations.

[0038] This invention constructs a two-layer Wasserstein fuzzy set: Outer Wasserstein sphere: Based on spatial interpolation data of historical geomagnetic storm events (such as the USGS, United States Geological Survey, 50 events recorded by the U.S. Geological Survey), a joint empirical distribution of geoelectric field intensity at each node is constructed, and the radius W is selected through cross-validation.

[0039] Inner time path generation: A Conditional Variational Autoencoder (CVAE) is used to generate a set of possible geoelectric field evolution paths within a future time window [0,T], conditioned on the current geoelectric field intensity E(t). Each path has three key parameters: peak value... Duration Maximum rate of increase The training data for this CVAE comes from the time series of measured geomagnetic storms (sampling rate 1Hz), with latent variables... Uncertainty between capture paths.

[0040] Optimization goal:

[0041] The second term is a penalty term for the rate of change of magnetic flux, with a coefficient λ > 0. The physical meaning of this objective function is not only to minimize the maximum magnetic flux at any node at any time throughout the entire event cycle, but also to suppress the abrupt change rate of magnetic flux, avoiding deployment schemes that rely on dangerous GMD (Geomagnetic Disturbance) waveforms. By solving this minimax problem, the deployment scheme output by the TC-DRO layer guarantees that for all possible spatial distributions and temporal path combinations within the Wasserstein sphere, the maximum magnetic flux response and rate of increase of the entire power grid satisfy the safety constraints.

[0042] Safe-MPRL (Safe Model Predictive Reinforcement Learning) Controller: This invention replaces the traditional passive correction architecture with active safe model predictive reinforcement learning.

[0043] The active safety model prediction reinforcement learning controller adopts a dual-agent architecture: Main control unit (Main Agent): Employs the Soft Actor-Critic (SAC) algorithm, adding the first derivative of magnetic flux to the state space. The reward function includes a security cost term (output by the security agent). The goal of the main control unit is to maximize the cumulative reward with security constraints.

[0044] Safety Predictive Processing Unit This agent does not participate in the main task reward optimization; its input is the current state. The predicted trajectory is obtained by using a short-time trajectory predictor (a lightweight GRU, gated recurrent unit) network to predict the state in the next k steps. The output is the corrective action a. safe The security agent employs a diffusion strategy for gradient training: the strategy is viewed as a denoising diffusion process, gradually recovering the optimal security action from random noise. The time step of the diffusion process is decoupled from the time scale of GMD events, enabling the security agent to generate smooth and robust correction sequences in highly dynamic scenarios.

[0045] Safety horizon function S(r): Dynamically determines the number of prediction steps k. Based on the real-time detected GIC rise rate.

[0046] Calculate the shortest safe warning time based on the current magnetic flux margin of the iron core from the saturation boundary:

[0047] Where T s To control the sampling period, S(r) decreases when the magnetic flux rapidly approaches saturation or the DC bus voltage drops rapidly, forcing the safety agent to intervene more steps in advance.

[0048] Safety constraint internalization unit: Used to add Lyapunov constraints to the policy gradient update formula using the Karush-Kuhn-Tucker (KKT) conditions. Specifically, each time the main control unit is updated, the following constrained optimization problem is solved:

[0049] Where τ is the trajectory. For strategy The expectation of the sampled trajectory, r Let be the reward function at time t. Let be the system state at time t, and at be the control action at time t. for The derivative of the Lyapunov function with respect to time under the given state, where η is the decay rate coefficient. The Lyapunov function value in the state.

[0050] By using the Lagrange multiplier method to transform it into an unconstrained objective, the trained main control unit naturally tends to avoid unsafe regions without the need for post-processing correction by external filters.

[0051] The deployment-control closed-loop online update framework system is used to further break down information silos between deployment and control.

[0052] During operation, the Safe-MPRL adaptive controller of each solid-state transformer records control performance indicators (actual peak flux, THD (Total Harmonic Distortion), energy consumption, etc.) and aggregates them into a global performance residual ΔJ via a graph message passing network. When the accumulated ΔJ exceeds the threshold ΔJ... th When the deployment plan is updated online, the update uses a mirror descent algorithm: the deployment decision vector x is treated as a point on the probability simplex (through continuous relaxation), and the Bregman divergence is defined as... (i.e., entropy regularization), the update rule for each iteration is:

[0053] in The gradient is estimated from real-time performance data (and can be obtained through finite difference or simultaneous perturbation random approximation). xt is the decision vector at the t-th iteration. The gradient is mirrored at xt. Let be the step size for the t-th iteration.

[0054] Due to the large scale of the entire power grid, a block-based coordinate mirroring descent approach is adopted: the power grid is divided into multiple electrically coupled blocks, each of which independently updates its local deployment decisions, while weak coupling is achieved between blocks by passing boundary conditions. The updated deployment plan is then sent to the corresponding solid-state transformers after safety verification, adjusting their switching status or control parameters.

[0055] This closed-loop framework enables deployment schemes to adapt to grid topology changes, equipment aging, and geomagnetic storm statistical characteristic drift caused by climate change, without the need for periodic and costly offline global re-optimization.

[0056] In this embodiment, the solid-state transformer is a cascaded modular multilevel architecture, and its power conversion link includes, in sequence: an input stage (a four-arm AC-DC converter, used to rectify high-voltage AC into DC), an isolation stage, and an output stage.

[0057] The input end of the solid-state transformer is connected to the primary side of the power grid, and the output end is connected to the secondary side of the power grid or the load.

[0058] The four-arm AC-DC converter includes: a first switch S1, a second switch S2, a third switch S3, a fourth switch S4, a fifth switch S5, a sixth switch S6, a seventh switch S7, and an eighth switch S8.

[0059] The drain of the first switch S1 is connected to the positive terminal of the DC bus, the source of the first switch S1 is connected to the drain of the second switch S2, and the source of the second switch S2 is connected to the negative terminal of the DC bus; the drain of the third switch S3 is connected to the positive terminal of the DC bus, the source of the third switch S3 is connected to the drain of the fourth switch S4, and the source of the fourth switch S4 is connected to the negative terminal of the DC bus; the drain of the fifth switch S5 is connected to the positive terminal of the DC bus, the source of the fifth switch S5 is connected to the drain of the sixth switch S6, and the source of the sixth switch S6 is connected to the negative terminal of the DC bus; the drain of the seventh switch S7 is connected to the positive terminal of the DC bus, the source of the seventh switch S7 is connected to the drain of the eighth switch S8, and the source of the eighth switch S8 is connected to the negative terminal of the DC bus.

[0060] Specifically, the three live wires on the primary side of the power grid are respectively connected to the bridge arm between the source of the third switch S3 and the drain of the fourth switch S4, and between the source of the fifth switch S5 and the drain of the sixth switch S6. The neutral wire on the primary side of the power grid is connected to the bridge arm between the source of the first switch S1 and the drain of the second switch S2.

[0061] The isolation stage is a high-frequency isolated DC-DC converter that achieves voltage conversion and electrical isolation. It operates at frequencies from kHz to MHz and is used to significantly reduce the size of magnetic components.

[0062] The output stage is a DC-AC inverter, used to convert the DC bus voltage into an AC voltage with the required amplitude and frequency. The DC bus of the solid-state transformer also provides an external interface for electrically coupled energy storage systems and new energy power generation equipment; the above three-level architecture achieves unified power scheduling and flexible allocation through a shared DC bus.

[0063] The high-frequency isolated DC-DC converter includes a DC bus capacitor Cdc, a DC filter inductor Ldc, a first damping resistor Rc1, a second damping resistor Rc2, an absorption capacitor Cp, an output filter capacitor Cs, a ninth switch S9, a tenth switch S10, an eleventh switch S11, a twelfth switch S12, a high-frequency transformer, a thirteenth switch S13, a fourteenth switch S14, a fifteenth switch S15, and a sixteenth switch S16.

[0064] The first terminal of the DC bus capacitor Cdc is connected to the positive terminal of the DC bus and the first terminal of the DC filter inductor Ldc. The second terminal of the DC bus capacitor Cdc is connected to the negative terminal of the DC bus, the second terminal of the first damping resistor Rc1, the source of the tenth switch S10, and the source of the twelfth switch S12. The second terminal of the DC filter inductor Ldc is connected to the first terminal of the absorption capacitor Cp, the drain of the ninth switch S9, and the drain of the eleventh switch S11; the second terminal of the absorption capacitor Cp is connected to the first terminal of the first damping resistor Rc1; the drain of the ninth switch S9 is connected to the positive bus, the source of the ninth switch S9 is connected to the drain of the tenth switch S10, and the source of the tenth switch S10 is connected to the negative terminal of the DC bus; the drain of the eleventh switch S11 is connected to the positive terminal of the DC bus, the source of the eleventh switch S11 is connected to the drain of the twelfth switch S12, and the source of the twelfth switch S12 is connected to the negative terminal of the DC bus. The connection point between the source of the ninth switch S9 and the drain of the tenth switch S10 forms the first AC terminal of the primary side, and the connection point between the source of the eleventh switch S11 and the drain of the twelfth switch S12 forms the second AC terminal of the primary side. The first AC terminal and the second AC terminal of the primary side are respectively connected to the two ends of the primary side of the high-frequency transformer. The drain of the thirteenth switch S13 is connected to the drain of the fifteenth switch S15, the first terminal of the second capacitor Cs, and the positive DC output bus; the source of the thirteenth switch S13 is connected to the drain of the fourteenth switch S14 and the first terminal of the secondary side of the high-frequency transformer; the source of the fourteenth switch S14 is connected to the source of the sixteenth switch S16, the negative DC output bus, and the second terminal of the second resistor Rc2; the drain of the fifteenth switch S15 is connected to the positive DC output bus, and the source of the fifteenth switch S15 is connected to the second terminal of the secondary side of the high-frequency transformer and the drain of the sixteenth switch S16; the second terminal of the second capacitor Cs is connected to the first terminal of the second resistor Rc2.

[0065] The output stage is a four-arm DC-AC inverter, with its positive DC input terminal connected to the positive DC output bus and its negative DC input terminal connected to the negative DC output bus. Its AC output terminal is used to connect to the secondary side of the power grid or the load.

[0066] Optionally, the first switch S1, the second switch S2, the third switch S3, the fourth switch S4, the fifth switch S5, the sixth switch S6, the seventh switch S7, and the eighth switch S8 are all connected in anti-parallel with fast recovery diodes.

[0067] The method of using the geomagnetic disturbance and high-altitude electromagnetic pulse suppression system includes the following steps: Obtain power grid topology data, including node types, transmission line parameters, and transformer parameters, and construct a heterogeneous power grid diagram.

[0068] By employing the PHIGNN deployment optimization module in conjunction with the TC-DRO layer, and considering both spatial distribution uncertainty and temporal evolution path uncertainty, the minimum number of solid-state transformer deployment locations and parameter configurations are output.

[0069] Solid-state transformers are installed in the actual power grid according to the deployment plan.

[0070] During operation, the zero-sequence common-mode current component in the power grid is detected in real time by the Safe-MPRL adaptive controller.

[0071] When the common-mode current exceeds the preset threshold, the safety horizon function dynamically calculates the warning time, and the safety prediction processing unit generates an early correction action; the main control unit simultaneously outputs the optimal control sequence, and the two are weighted and fused before being output to the four-arm converter.

[0072] In series mode, the Safe-MPRL adaptive controller generates a command that is opposite to the direction of the source-side common-mode current, producing a canceling flux.

[0073] After a geomagnetic disturbance or high-altitude electromagnetic pulse event ends, the absorbed energy is fed back to the power grid through an energy storage system.

[0074] Closed-loop update steps: Collect control performance metrics in this event, calculate performance residuals, and if the cumulative residuals exceed the threshold, trigger mirror descent iteration to update the deployment scheme or controller hyperparameters.

[0075] Example 2 This embodiment uses a 20-node transmission network from the U.S. Department of Energy's GMD research standard use case as a test system to verify the effectiveness of the TC-DRO collaborative deployment optimization method proposed in this invention.

[0076] Uncertainty set configuration: Minute-by-minute geoelectric field sequences for each historical geomagnetic storm event are extracted to construct a CVAE training set. The latent space dimension is set to 8, and the generation time window length is T=300 seconds. The outer Wasserstein radius W=0.35 (consistent with the original patent). The magnetic flux change rate penalty coefficient λ is selected through grid search: when λ=0.5, the maximum magnetic flux in the worst scenario is 1.12 pu, and the maximum magnetic flux change rate decreases from 0.18 pu / s to 0.09 pu / s, significantly improving the safety margin. In contrast, the original DRO solution has a maximum magnetic flux change rate as high as 0.15 pu / s under the same test set, which is close to the action boundary of the protection device.

[0077] Optimization results: The TC-DRO output deployment nodes are {2,4,7,9,12,15} (the same as the original DRO, but with an additional node 5 as an auxiliary compensation point, for a total of 7 nodes). Under the measured waveform of the 1989 Quebec event, the maximum magnetic flux predicted by PIHGNN is 1.07 pu (1.12 pu for the original DRO), and the maximum magnetic flux change rate is 0.07 pu / s. In particular, under the artificially constructed "delay spike" waveform (flat for the first 200 seconds, then steeply rising for the last 100 seconds), the original DRO deployment scheme showed a magnetic flux overshoot of 1.23 pu, while the TC-DRO scheme consistently remained below 1.15 pu, verifying the effectiveness of the time coupling constraint.

[0078] Example 3 This embodiment uses a 220kV / 35kV main transformer as the protection object. The secondary winding of the solid-state transformer of the present invention is connected in series in the neutral point grounding circuit of the main transformer. The control mode is set to series flux cancellation to verify the effectiveness of the traditional PIDTRL (Physics-Informed Deep Transfer Reinforcement Learning) adaptive controller.

[0079] Security agent training: Employs a gradient diffusion strategy with a diffusion step count T. diff=100, noise scheduling is cosine beta scheduling. The training data has 500 safe trajectories and 200 artificially injected unsafe trajectories (obtained by forcibly increasing the GIC rise rate). The safety cost function recovered by inverse reinforcement learning shows that... Its high sensitivity aligns with physical intuition.

[0080] Dynamic safety horizon function: set α=0.6, β=0.4, ε=1e -5 In the 1989 Quebec incident, the maximum rate of increase in GIC occurred at 45 seconds, dI GIC / dt=18A / s, at this time the calculated value of S(r) is 0.32 seconds, and the rounding k=4 steps (step size 0.1 seconds). The safety agent outputs the correction action 0.4 seconds in advance, reduces the duty cycle and temporarily increases the energy storage absorption power, so that the measured maximum magnetic flux drops from 1.14pu when there is no safety agent to 1.06pu, and no protection alarm is triggered.

[0081] Comparative Experiment: The Safe-MPRL of this invention is compared with a traditional PID (Physics-Informed Deep) reinforcement learning controller plus a Lyapunov filter (the perturbation is an extreme pulse with a rise rate of 25 A / s): The comparison is shown in the table below: Table 1

[0082] Example 4 During the continuous operation of the seven solid-state transformers deployed in Example 1 for six months, a new long-distance transmission branch (120km long, connected to node 11) was added to the power grid. Traditional methods require a global re-deployment optimization, taking several hours. The closed-loop framework of this invention, upon detecting that the cumulative performance residual ΔJ exceeded the threshold (0.15), automatically triggered a mirror descent update, performing only local re-optimization on the blocks containing nodes 9, 11, and 12, taking 3 minutes, and outputting a new deployment scheme: adding a solid-state transformer at node 11 and removing the auxiliary compensation point at node 5. The updated system was verified under actual disturbances, with the maximum magnetic flux recovering from 1.13 pu to 1.08 pu.

[0083] Example 5 TC-DRO's CVAE module can pre-compute and store the results as a lookup table. Online deployment optimization requires only table lookups and simple numerical integration, making it suitable for real-time deployment and adjustments. Safe-MPRL's dual-agent network is implemented, with inference latency below 5ms. The closed-loop mirror descent update module runs on a central server and interacts with edge devices via MQTT (Message Queuing Telemetry Transport), with communication overhead below 10kbps.

[0084] In summary, the time-coupled sub-bar optimization, active safety model prediction reinforcement learning, and deployment-control closed-loop evolutionary framework proposed in this invention further improve the system's adaptability to dynamic extreme events, safety, and long-term economic efficiency based on existing patents, demonstrating significant technological advancements and industrial application value.

[0085] Unless otherwise specified, the equipment components involved in the above embodiments are all conventional equipment components, and the structural settings, working methods or control methods involved are all conventional settings, working methods or control methods in the art unless otherwise specified.

[0086] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention and are not intended to limit it. Any other modifications or equivalent substitutions made by those skilled in the art to the technical solutions of the present invention, as long as they do not depart from the spirit and scope of the technical solutions of the present invention, should be covered within the scope of the claims of the present invention.

Claims

1. A geomagnetic disturbance and high-altitude electromagnetic pulse suppression system, characterized in that, include: A solid-state transformer includes an input stage, an isolation stage, and an output stage connected in sequence. The input stage is a four-arm AC-DC converter, the isolation stage is a high-frequency isolation DC-DC converter, and the output stage is a DC-AC inverter. The physical information heterogeneous graph neural network deployment optimization module is configured to take a power grid topology graph as input, the power grid topology graph having multiple node types and edge types, and the physical information heterogeneous graph neural network combines a heterogeneous graph neural network with a physical information neural network to embed the physical laws of the power system into the message transmission process of the graph neural network, and is used to predict the peak magnetic flux response of the iron core of each node under any deployment scheme. A time-coupled sub-Browser bar optimization layer is coupled to the physical information heterogeneous graph neural network deployment optimization module. It is configured to: employ a two-layer Wasserstein fuzzy set to characterize the spatial distribution uncertainty and temporal evolution path uncertainty of geomagnetic disturbances and upper-air electromagnetic pulses in the ground electric field, respectively. The inner layer is an autoregressive generative model jointly driven by the three parameters of the ground electric field peak value, duration, and rise rate to construct a time path set. The physical information heterogeneous graph neural network deployment optimization module is used in conjunction with the time-coupled sub-Browser bar optimization layer to output the minimum number of solid-state transformer deployment locations and parameter configurations. An active safety model predictive reinforcement learning adaptive controller is configured to: after the solid-state transformer is deployed, learn and execute the optimal control strategy online through a dual-agent architecture, and output the switching sequence control command of the four-arm AC-DC converter; The deployment-control closed-loop online update framework system is configured to use a mirror descent algorithm to feed back the control performance during the operation phase to the deployment optimization module in real time.

2. The system of claim 1, wherein, The optimization objective of the time-coupled sub-bar optimization layer is: where x is the binary deployment decision vector, ξ(t) is the time-dependent geomagnetic disturbance strength, ) is the core peak flux of the ith conventional transformer at time t, λ is the flux rate of change penalty coefficient; the output is a robust deployment scheme that minimizes the full-time worst-case expected flux.

3. The system of claim 1, wherein the system is configured to: The four-arm AC-DC converter includes: a first switch S1, a second switch S2, a third switch S3, a fourth switch S4, a fifth switch S5, a sixth switch S6, a seventh switch S7, and an eighth switch S8; the drain of the first switch S1 is connected to the positive terminal of the DC bus, the source of the first switch S1 is connected to the drain of the second switch S2, and the source of the second switch S2 is connected to the negative terminal of the DC bus; the drain of the third switch S3 is connected to the positive terminal of the DC bus, and the... The source of the third switch S3 is connected to the drain of the fourth switch S4, and the source of the fourth switch S4 is connected to the negative terminal of the DC bus; the drain of the fifth switch S5 is connected to the positive terminal of the DC bus, and the source of the fifth switch S5 is connected to the drain of the sixth switch S6, and the source of the sixth switch S6 is connected to the negative terminal of the DC bus; the drain of the seventh switch S7 is connected to the positive terminal of the DC bus, and the source of the seventh switch S7 is connected to the drain of the eighth switch S8, and the source of the eighth switch S8 is connected to the negative terminal of the DC bus.

4. The system of claim 3, wherein the system is configured to: The solid-state transformer suppression system also includes an energy storage system electrically coupled to the DC bus of the solid-state transformer.

5. The system of claim 1, wherein, The physical information heterogeneous graph neural network deployment optimization module uses a proxy power flow model for offline training. The training parameters of the proxy power flow model include: transmission line length, resistance, grounding resistance, network topology, and transformer nonlinear magnetization characteristic curve. The physical information heterogeneous graph neural network is used to learn the end-to-end mapping relationship from the deployment scheme to the core magnetic flux response of the entire power grid during training.

6. The system of claim 1, wherein, The time-coupled sub-Bruker bar optimization layer consists of a two-layer Wasserstein fuzzy set, where the outer Wasserstein sphere is constructed based on the spatial distribution statistics of historical geomagnetic storm events, and the inner time evolution path is generated by a conditional variational autoencoder and configured to sample the set of possible evolution paths within the future time window based on the current geoelectric field state; the magnetic flux change rate penalty coefficient is selected through cross-validation.

7. The geomagnetic disturbance and high-altitude electromagnetic pulse mitigation system of claim 1, wherein, The proactive safety model predicts a reinforcement learning adaptive controller, which includes: The main control unit employs a deep reinforcement learning policy network to balance the power of the solid-state transformer and reduce common-mode current stress with the goal of maximizing long-term cumulative rewards. The safety prediction processing unit is used for gradient training based on the diffusion strategy. It takes the current system state and the predicted trajectory for the next k steps as input and outputs the early correction action. The value of k is dynamically determined by the safety horizon function, which calculates the shortest safety warning time online based on the real-time detected rate of rise of the geomagnetic induced current and the system inertial parameters. The safety constraint internalization processing unit is used to embed Lyapunov constraints into the objective function of the main control unit in the form of Karl-Kun-Tucker conditions during policy gradient updates, so that the main control unit itself tends to generate safe actions.

8. The geomagnetic disturbance and high-altitude electromagnetic pulse mitigation system of claim 1, wherein, The mirror descent algorithm defines Bregman divergence as the geometric metric of the deployment scheme. In each iteration of mirror descent, the gradient estimate is updated based on the cumulative performance residual of the current deployment scheme, and the updated deployment decision vector is output.

9. The system of claim 7, wherein the system further comprises a magnetic field sensor configured to detect a magnetic field of the Earth and to generate a magnetic field signal indicative of the magnetic field of the Earth. The safety view function is: where a, β are weight coefficients, is a saturation flux threshold, is an estimated flux, is a flux rate of change, V dc is a DC bus voltage, ε is a small constant to prevent division by zero.

10. A method of suppressing geomagnetic disturbances and high-altitude electromagnetic pulses, characterized by, The geomagnetic disturbance and upper-level electromagnetic pulse suppression system according to any one of claims 1 to 9 includes the following steps: Acquire power grid topology data and construct a heterogeneous power grid diagram; Using the aforementioned heterogeneous power grid graph as the input to the physical information heterogeneous graph neural network deployment optimization module, the physical laws of the power system are embedded into the message transmission process of the graph neural network through the physical information heterogeneous graph neural network, thereby predicting the peak magnetic flux response of the iron core of each node under any deployment scheme. By using a time-coupled sub-Bruker optimization layer, a two-layer Wasserstein fuzzy set is used to characterize the spatial distribution uncertainty and temporal evolution path uncertainty of geomagnetic disturbances and upper-air electromagnetic pulses in the ground electric field. The inner layer is constructed by an autoregressive generative model driven by three parameters: peak value, duration, and rise rate of the ground electric field to build a time path set. By coordinating the physical information heterogeneous graph neural network deployment optimization module with the time-coupled sub-Bruker optimization layer, the minimum number of solid-state transformer deployment locations and parameter configurations are output under the premise of simultaneously considering spatial distribution uncertainty and temporal evolution path uncertainty. Install the solid-state transformer according to the deployment location and parameter configuration described; During operation, the zero-sequence common-mode current component in the power grid is detected in real time by the active safety model prediction reinforcement learning adaptive controller; when the common-mode current exceeds the preset threshold, the four-arm AC-DC converter is controlled to perform common-mode current redirection operation.

Citation Information

Patent Citations

  • Power consumption management peak avoidance method based on multi-objective planning

    CN120810605A