A method and system for dynamically evaluating grid stability with new energy access

Abstract

The application relates to the technical field of power grid systems, and discloses a new energy access power grid stability dynamic evaluation method and system, which comprises the following steps: constructing a power grid dynamic model containing a new energy access unit to obtain a panoramic model; performing scale decomposition on the panoramic model to construct a slow system, a fast system and an ultra-slow disturbance system; analyzing and determining the stability of the fast system; determining the stability of the slow system; analyzing the change trend of a disturbance variable in a prediction period, and determining the stability of the ultra-slow disturbance system in combination with a disturbance influence sensitivity matrix; according to current system state, power flow distribution and topological structure information, assisting in determining whether there is a potential structural instability factor; evaluating whether the current new energy access system is in a stable state, and obtaining an evaluation conclusion. The application has high modeling granularity, clear determination mechanism and strong structure recognition capability, and improves the accuracy of power grid stability evaluation under a new energy complex access scenario.

Description

Technical Field

[0001] This invention relates to the field of power grid system technology, and more specifically, to a method and system for dynamic evaluation of power grid stability when new energy sources are integrated. Background Technology

[0002] With the global energy structure undergoing a low-carbon transformation, new energy sources such as photovoltaics, wind power, energy storage, and electric vehicles are being integrated into the power system in large quantities, prompting the power grid's operating characteristics to gradually evolve from being dominated by traditional synchronous generation to a distributed structure centered on power electronic devices. Against this backdrop, the dynamic behavior of the power grid has become more complex, and traditional stability assessment methods based on the steady-state characteristics of synchronous generators are gradually revealing their shortcomings. Especially in scenarios with a high proportion of new energy sources and frequent fluctuations in operating status, it is difficult to accurately assess the dynamic stability of the system, leading to problems such as unclear judgment of the system's response process after disturbances and inaccurate identification of structural instability points.

[0003] Existing technologies attempt to improve assessment accuracy by introducing new energy modeling and generalized stability analysis methods. For example, patent CN117937434A proposes a transient stability assessment method for new energy equipment connected to a power system. This method constructs a full model of the power system and performs transient stability analysis based on generalized motion state indices and stability assessment indices. However, this method still has certain shortcomings: it treats all new energy equipment as a disturbance source in model construction, failing to clearly distinguish the dynamic response differences between new energy power electronic interfaces and traditional units; simultaneously, the stability assessment mainly relies on overall index quantification, lacking decomposition and targeted assessment of multi-scale system dynamics, making it difficult to distinguish instability mechanisms at different time scales. Furthermore, this scheme does not introduce structural graph model analysis methods, failing to effectively identify potential structural vulnerabilities and risk transmission paths caused by system topology characteristics.

[0004] Therefore, it is necessary to design a dynamic evaluation method and system for grid stability when new energy sources are integrated to solve the problems existing in the current technology. Summary of the Invention

[0005] In view of this, the present invention proposes a dynamic evaluation method and system for grid stability of new energy access, aiming to solve the problems of insufficient model granularity, unclear time scale decomposition, and lack of identification of weak links in the current technology.

[0006] In one aspect, this invention proposes a method for dynamic evaluation of grid stability when new energy sources are integrated, comprising: A dynamic power grid model containing at least one new energy access unit is constructed and coupled with the main power grid model to obtain a panoramic model. The new energy access unit includes a photovoltaic power generation unit, a wind power generation unit, an energy storage device, or an electric vehicle charging and discharging unit. The panoramic model is decomposed into scales based on the singular perturbation method, and a slow system dominated by synchronous generators, a fast system dominated by power electronic equipment, and an ultra-slow perturbation system driven by perturbation are constructed. The fast system is linearized with small perturbations to construct a Jacobian matrix and determine its stability. The slow system is processed based on the equal area rule to calculate the fault acceleration area and the maximum deceleration area and determine its stability. The changing trend of the perturbation variable in the ultra-slow perturbation system within the prediction period is analyzed using a sliding window, and the perturbation influence sensitivity matrix is ​​combined to determine whether a critical transition of the system state is triggered and determine the stability of the ultra-slow perturbation system. Based on the current system state, power flow distribution, and topology information, a graph neural network-assisted model is used to help determine whether there are potential structural instabilities in the system. Based on the fast system stability, slow system stability, ultra-slow disturbance system stability, and auxiliary judgment results, assess whether the power grid of the current new energy access system is in a stable state and obtain the assessment conclusion.

[0007] Furthermore, when constructing a power grid dynamic model that includes at least one new energy access unit and coupling it with the main power grid model to obtain a panoramic model, the following steps are included: Based on the controller structure and electrical interface characteristics, a state space model is constructed for the photovoltaic power generation unit, wind power generation unit, energy storage device and electric vehicle charging and discharging unit. The state space model includes voltage, current, active power, frequency, controller output and battery state. Construct a main power grid model that includes synchronous generators, load models, node admittance matrices, and network topology; By coupling the modeling results of the new energy access unit with the main power grid model through node power injection, a unified time scale is established, and a consistent mapping relationship between electrical variables, control variables and structural variables is established to form a unified differential algebraic equation model that includes the new energy access unit and the main power grid model, thus obtaining the panoramic model.

[0008] Furthermore, when performing scale decomposition on the panoramic model based on the singular perturbation method, it includes: The system state variables in the panoramic model are divided into slow variables, fast variables, and disturbance variables according to their dynamic response characteristics. The slow variables include rotor angle, angular velocity, and frequency; the fast variables include inverter voltage, current control loop output, and phase-locked loop state; and the disturbance variables include wind speed, irradiance, and load power changes. The differential algebraic equation model is transformed into a singular perturbation form: , ; Where ε is the small parameter for scale separation, X represents the vector of slow variables, Y represents the vector of fast variables, and Z represents the vector of perturbation variables. This represents the derivative of the slow variable with respect to time. This represents the derivative of a fast variable with respect to time. Represents the dynamic function of a slow system. This represents the dynamic function of a fast system.

[0009] Furthermore, the fast system undergoes small-perturbation linearization to construct the Jacobian matrix. When determining the stability of the fast system, the following steps are taken: Select the equilibrium operating point of the fast system in the current operating state, and perform a first-order Taylor expansion of the nonlinear state equation of the fast system with the equilibrium operating point as the center. Constructing the incremental equations Δ of the fast system =A·ΔY, where A represents the Jacobian matrix of the partial derivative of the fast variable with respect to itself. Solve for all eigenvalues ​​of the Jacobian matrix A. If the real parts of all eigenvalues ​​of the matrix are negative, then the fast system is determined to be in a stable state under the current disturbance; otherwise, it is unstable.

[0010] Furthermore, based on the equal area rule, the slow system is processed to calculate the fault acceleration area and the maximum deceleration area. When determining the stability of the slow system, the following steps are taken: The stable equilibrium point δ0 of the slow system before the fault, the first stable equilibrium point δs after the fault is cleared, and the unstable equilibrium point δu after the fault are obtained, where the equilibrium point is a specific value of the rotor angle of the synchronous generator. Calculate the acceleration area during the fault, which is the area formed by integrating the mechanical power and the electromagnetic power after the fault within the interval [δ0, δu]. Calculate the maximum deceleration area, which is the area formed by integrating the mechanical power and the electromagnetic power after the fault within the interval [δs,δu]. The acceleration area is compared with the maximum deceleration area. If the acceleration area is smaller than the maximum deceleration area, the slow system is determined to be in a stable state; otherwise, it is considered unstable.

[0011] Furthermore, a sliding window analysis is used to analyze the changing trend of the disturbance variable in the ultra-slow disturbance system within the prediction period. Combined with the disturbance influence sensitivity matrix, it is determined whether a critical transition of the system state is triggered. When determining the stability of the ultra-slow disturbance system, the following steps are taken: A sliding window with a time length of 30 to 300 seconds is set, and the time series data of the disturbance variable is extracted within the sliding window; A time series prediction model is constructed based on the historical data of the disturbance variable to obtain the predicted change trend of the disturbance variable in future periods and extract the prediction slope, amplitude and volatility index. The sensitivity matrix to disturbance effects is obtained based on system identification; Multiply the disturbance prediction trend with the sensitivity matrix to obtain the predicted system state shift caused by future disturbances; If the predicted system state offset exceeds the stability safety boundary threshold, the ultra-slow disturbance system is determined to be unstable; otherwise, it is in a stable state.

[0012] Furthermore, when using graph neural network-assisted models to assist in determining whether there are potential structural instabilities in a system, the following steps are taken: Construct an input graph with the power grid topology as the graph, where the nodes of the graph represent generation units, load nodes or converter devices, and the edges represent bus connection relationships and line impedances; Construct a feature vector for each node in the graph. The feature vector includes the voltage amplitude, phase angle, active and reactive power injected into the node, frequency offset, connection type and control mode of the node. Based on the graph neural network model, the directed propagation and embedding update of features between nodes are performed through graph convolution operations, capturing the potential unstable structure in the topology, such as weak connections, single-point fault amplification paths, and low-redundancy distribution modes. Output node-level stability scores, identify nodes with scores below the score threshold as key nodes with potential structural instability factors, and output structurally weak subgraphs.

[0013] Furthermore, the graph neural network model, during training and application, includes: Training samples were constructed using historical disturbance scenarios and fault records. The sample labels were whether the system experienced voltage collapse, frequency oscillation, node islanding, or tie line overload after the disturbance. The graph attention network structure in graph neural networks is adopted, and a node attention mechanism is added during the feature propagation process. The graph neural network model is trained through supervised learning. In real-time evaluation, the current system state and topology information are input into a trained graph neural network to obtain auxiliary judgment results.

[0014] Furthermore, based on the fast system stability, slow system stability, ultra-slow disturbance system stability, and auxiliary judgment results, when assessing whether the power grid of the current new energy access system is in a stable state, including... The stability determination results output by the fast system, slow system and ultra-slow disturbance system are numerically encoded, wherein the stable state is set to +1 and the unstable state is set to -1. The stability scores of all key nodes output by the graph neural network are normalized. The normalized scores are then fused with the encoding results to construct a comprehensive stability index. The comprehensive stability index is as follows: S=α·Sf+β·Ss+γ·Sus+δ·(1-R); Where S represents the comprehensive stability index, Sf, Ss, and Sus represent the stability codes of the fast system, slow system, and ultra-slow perturbation system, respectively, R represents the normalized score, and the value range of R is (0, 1], and α, β, γ, and δ represent the weight coefficients. A stability threshold T is determined. If S≥T, the new energy access system is determined to be in a stable state; if S≤T, the new energy access system is determined to be in an unstable state; otherwise, it is determined to be in a critical state. Output stability assessment results and mark the locations of unstable links or structural risks in each subsystem.

[0015] Compared with existing technologies, the advantages of this invention are as follows: By introducing coupled modeling of the new energy access unit and the main power grid model, and combining the singular perturbation method to divide the complex dynamic behavior of the power grid into three time-scale subsystems of fast, slow, and ultra-slow disturbances, and employing quantitative methods such as small disturbance linearization, equal area rule, and disturbance trend prediction for different subsystems, hierarchical stability analysis is performed, effectively improving the accuracy of dynamic stability state discrimination under multiple time scales; the introduction of a graph neural network model to assist in the identification of structural instability risks based on the power grid topology enables intelligent perception and early warning of weak structural nodes, overcoming the deficiency of existing technologies that cannot identify structural risks based solely on the overall system state. This application features high modeling granularity, a clear judgment mechanism, and strong structural identification capabilities, improving the accuracy of power grid stability assessment in complex new energy access scenarios.

[0016] On the other hand, this application also provides a dynamic evaluation system for grid stability of renewable energy access, used to apply the above-mentioned dynamic evaluation method for grid stability of renewable energy access, including: The construction unit is configured to construct a dynamic power grid model containing at least one new energy access unit and couple it with the main power grid model to obtain a panoramic model. The new energy access unit includes a photovoltaic power generation unit, a wind power generation unit, an energy storage device, or an electric vehicle charging and discharging unit. The decomposition unit is configured to perform scale decomposition on the panoramic model based on the singular perturbation method, constructing a slow system dominated by synchronous generators, a fast system dominated by power electronic equipment, and a perturbation-driven ultra-slow perturbation system. The processing unit is configured to perform small perturbation linearization processing on the fast system, construct the Jacobian matrix, and determine the stability of the fast system; process the slow system based on the equal area rule, calculate the fault acceleration area and the maximum deceleration area, and determine the stability of the slow system; and use a sliding window to analyze the changing trend of the perturbation variables in the ultra-slow perturbation system within the prediction period, and combine the perturbation influence sensitivity matrix to determine whether a critical transition of the system state is triggered, and determine the stability of the ultra-slow perturbation system. The judgment unit is configured to make an auxiliary judgment on whether there are potential structural instability factors in the system based on the current system state, power flow distribution and topology information, using a graph neural network-assisted model. The evaluation unit is configured to evaluate whether the power grid of the current new energy access system is in a stable state based on the fast system stability, slow system stability, ultra-slow disturbance system stability and auxiliary judgment results, and obtain an evaluation conclusion.

[0017] It is understandable that the aforementioned methods and systems for dynamic evaluation of grid stability for new energy access have the same beneficial effects, and will not be elaborated further here. Attached Figure Description

[0018] Various other advantages and benefits will become apparent to those skilled in the art upon reading the following detailed description of preferred embodiments. The accompanying drawings are for illustrative purposes only and are not intended to limit the invention. Furthermore, the same reference numerals denote the same parts throughout the drawings. In the drawings: Figure 1 A flowchart of a method for dynamic evaluation of grid stability for new energy access provided in an embodiment of the present invention; Figure 2 This is a functional block diagram of a dynamic grid stability assessment system for new energy access provided in an embodiment of the present invention. Detailed Implementation

[0019] Exemplary embodiments of the present disclosure will now be described in more detail with reference to the accompanying drawings. While exemplary embodiments of the present disclosure are shown in the drawings, it should be understood that the present disclosure may be implemented in various forms and should not be limited to the embodiments set forth herein. Rather, these embodiments are provided to enable a more thorough understanding of the present disclosure and to fully convey the scope of the disclosure to those skilled in the art. It should be noted that, unless otherwise specified, embodiments and features in the embodiments of the present invention can be combined with each other. The present invention will now be described in detail with reference to the accompanying drawings and embodiments.

[0020] In existing power grid stability assessment methods, the differences in dynamic response between renewable energy power electronic interfaces and traditional synchronous generators are not clearly distinguished, leading to coupled disturbances in dynamic behavior across multiple time scales and making it impossible to accurately identify the instability mechanisms of different subsystems. Existing models treat renewable energy units uniformly as disturbance sources, ignoring the interaction characteristics between their internal control loops and the main power grid. This results in a mixture of dynamics from rapidly changing power electronic devices and dynamics from slowly changing mechanical inertia, making it difficult to accurately quantify the stability margin of each subsystem. Furthermore, the lack of trend prediction and structural vulnerability analysis for ultra-slow disturbance variables means that the assessment results cannot reflect the critical transition risk of the system under sustained disturbances.

[0021] For example, in regional power grids with both photovoltaic (PV) and wind farms connected, when sudden changes in irradiance cause fluctuations in PV output, there is a time scale difference of milliseconds to seconds between the rapid adjustment of the inverter control loop and the adjustment of the synchronous generator speed. Existing evaluation methods use unified time-domain simulation, which cannot effectively separate the dynamic processes of fast and slow variables, leading to voltage oscillations caused by phase-locked loop (PLL) loss being misjudged as power angle instability. In scenarios with gradually changing load power, traditional methods do not establish a sensitivity relationship between disturbance variables and system state deviations, and cannot predict the gradual impact of load growth trends on tie-line transmission limits, resulting in deviations in the calculation of static safety boundaries.

[0022] If the above problems are not addressed, confusion regarding dynamic response mechanisms will lead to misjudgments of stability, potentially triggering malfunctions or delayed actions of protection devices, resulting in cascading failures. When multi-timescale coupling effects are not decoupled, the dynamic interaction between power electronic equipment and synchronous generators may induce subsynchronous oscillations, causing overvoltage or overcurrent damage to equipment. The lack of identification of ultra-slow disturbance trends will fail to provide early warning of the potential risk of the system operating point approaching the stability boundary, increasing the probability of voltage collapse or frequency instability. Insufficient structural vulnerability analysis makes it difficult to locate low-redundancy nodes in the topology, which may cause local power outages and spread to the main grid during N-1 faults.

[0023] For this, please refer to Figure 1 As shown, this application proposes a dynamic evaluation method for grid stability when new energy sources are integrated, including: S100: Construct a dynamic power grid model that includes at least one new energy access unit and couple it with the main power grid model to obtain a panoramic model. The new energy access unit includes a photovoltaic power generation unit, a wind power generation unit, an energy storage device, or an electric vehicle charging and discharging unit. S200: Based on the singular perturbation method, the panoramic model is decomposed into scales to construct a slow system dominated by synchronous generators, a fast system dominated by power electronic equipment, and an ultra-slow perturbation system driven by perturbation. S300: The fast system is linearized with small perturbations, the Jacobian matrix is ​​constructed, and the stability of the fast system is determined; the slow system is processed based on the equal area rule, the fault acceleration area and the maximum deceleration area are calculated, and the stability of the slow system is determined; the changing trend of the perturbation variable in the ultra-slow perturbation system within the prediction period is analyzed by using a sliding window, and the perturbation influence sensitivity matrix is ​​combined to determine whether the critical transition of the system state is triggered, and the stability of the ultra-slow perturbation system is determined. S400: Based on the current system state, power flow distribution, and topology information, a graph neural network-assisted model is used to help determine whether there are potential structural instabilities in the system. S500: Based on the stability of the fast system, the stability of the slow system, the stability of the ultra-slow disturbance system, and the auxiliary judgment results, assess whether the power grid of the current new energy access system is in a stable state and obtain the assessment conclusion.

[0024] Specifically, new energy access units refer to photovoltaic power generation units, wind power generation units, energy storage devices, or electric vehicle charging and discharging units. A state-space model can be used to construct the dynamic characteristics of each unit, including voltage, current, active power, frequency, controller output, and battery status, and coupled with the main grid model through node power injection. This feature addresses the insufficient model accuracy caused by traditional methods that uniformly treat new energy as a disturbance source by accurately modeling the dynamic response characteristics of new energy equipment. The singular perturbation method decomposes the system into subsystems at different time scales. Specifically, by dividing the system into slow variables, fast variables, and disturbance variables, the differential algebraic equations are transformed into a singular perturbation form containing scale separation parameters. This feature achieves dynamic decoupling across multiple time scales, solving the problem that traditional methods cannot distinguish the differences in dynamic response between synchronous machines and power electronic equipment. Small perturbation linearization involves constructing the Jacobian matrix through a first-order Taylor expansion of the fast system, and determining stability by solving for the sign of the real part of the eigenvalues. This feature enables rapid identification of high-frequency oscillation risks for the fast dynamic characteristics of power electronic equipment. The equal-area rule refers to assessing the stability of slow systems by comparing the acceleration and deceleration areas, specifically calculating the area index by integrating the difference between mechanical and electromagnetic power. This feature addresses the insufficient applicability of traditional transient stability assessment methods in new energy scenarios, particularly for low-frequency dynamics dominated by synchronous generators. The sliding window analysis uses a time window to extract trends in disturbance variables, specifically calculating state shifts through time series prediction combined with a sensitivity matrix. This feature enables advanced prediction of the impact of ultra-slow disturbances on system stability, overcoming the lag in response to long-term disturbance trends in traditional methods. The graph neural network-assisted model refers to a graph convolutional network based on topological structures, specifically identifying weak connections and single-point fault paths through node feature propagation. This feature addresses the limitation of traditional methods in identifying structural instability factors by mining topological correlations.

[0025] This application constructs a hierarchical evaluation framework through multi-timescale decomposition, decomposing the panoramic model into slow system, fast system and ultra-slow perturbation system, and using the equal area rule, eigenvalue analysis and trend prediction to achieve targeted stability discrimination. At the same time, it combines graph neural network to mine topological risks, forming a comprehensive evaluation system covering dynamic response characteristics and structural vulnerability.

[0026] The working process and principle of this application are as follows: First, a dynamic power grid model including new energy access units is constructed and coupled with the main power grid model to obtain a panoramic model. New energy access units include photovoltaic power generation, wind power generation, energy storage devices, or electric vehicle charging and discharging units. Then, based on the singular perturbation method, the panoramic model is scaled to construct a slow system dominated by synchronous generators, a fast system dominated by power electronic equipment, and a perturbation-driven ultra-slow perturbation system.

[0027] For fast systems, small perturbation linearization is performed, the Jacobian matrix is ​​constructed, and stability is determined. For slow systems, the fault acceleration area and maximum deceleration area are calculated based on the equal area rule to determine stability. For ultra-slow perturbation systems, a sliding window analysis is used to analyze the changing trend of the perturbation variable within the prediction period, and the perturbation influence sensitivity matrix is ​​combined to determine whether a critical transition of the system state is triggered, thereby determining stability.

[0028] Furthermore, based on the current system state, power flow distribution, and topology information, a graph neural network-assisted model is used to assist in determining whether there are potential structural instabilities in the system. Finally, based on the fast system stability, slow system stability, ultra-slow disturbance system stability, and the auxiliary judgment results, the grid status of the current renewable energy integration system is assessed, and an assessment conclusion is obtained.

[0029] By decomposing and independently analyzing each subsystem at multiple scales, instability mechanisms at different time scales can be accurately identified. Introducing a graph neural network-assisted model can capture structural vulnerability patterns that are difficult to quantify using traditional methods. This comprehensive assessment method improves the accuracy and comprehensiveness of stability assessments for renewable energy grid integration.

[0030] As a preferred embodiment, the solution of this application is specifically implemented as follows: When constructing the power grid dynamic model, the photovoltaic power generation unit model is established first, including the photovoltaic array, DC / DC converter, inverter, and its control system. The wind power generation unit model includes the wind turbine, generator, converter, and its control system. The energy storage device model includes the battery pack, bidirectional converter, and its control system. The electric vehicle charging and discharging unit model includes the on-board battery, charging pile, and its control system.

[0031] The main power grid model includes synchronous generators, loads, transmission lines, and transformers. The renewable energy access units are coupled to the main power grid model through nodal power injection equations, forming a unified set of differential-algebraic equations.

[0032] During scaling, state variables are divided into slow variables (such as rotor angle and angular velocity), fast variables (such as inverter voltage and current), and disturbance variables (such as wind speed and irradiance). The singular perturbation method is then used to transform the original equations into their standard form.

[0033] The fast system is linearized with small perturbations, the Jacobian matrix is ​​constructed, and the eigenvalues ​​are solved. If the real parts of all eigenvalues ​​are negative, the fast system is considered stable.

[0034] For slow systems, the equal area criterion is applied to calculate the acceleration area and the maximum deceleration area during the fault period. If the acceleration area is less than the maximum deceleration area, the slow system is considered stable.

[0035] For ultra-slow perturbation systems, a sliding window of 30-300 seconds is set to extract historical data of the perturbation variables and predict future trends. Combined with the perturbation impact sensitivity matrix, it is determined whether a critical transition of the system state has been triggered.

[0036] A neural network model based on the power grid topology is constructed, with node features including voltage, power, and other information. Potential structural instability factors are captured through graph convolution operations.

[0037] Finally, by combining the stability assessment results of each subsystem with the output of the graph neural network, a comprehensive stability index is calculated, and a final evaluation conclusion is given.

[0038] Through the above-described scheme, this application achieves multi-scale dynamic assessment of the stability of renewable energy grid integration. By decoupling fast and slow dynamics, the problem of mixed responses at different time scales in traditional methods is overcome, improving the accuracy of identifying instability mechanisms in each subsystem. The introduction of ultra-slow disturbance system analysis and graph neural network-assisted judgment enhances the early warning capability for critical transition risks and structural vulnerabilities under continuous disturbances. This application can more comprehensively and accurately assess the dynamic stability of the power grid under conditions of high renewable energy integration.

[0039] In some of the solutions described above in this application, there is a problem of insufficient model coupling accuracy when constructing the power grid dynamic model of the renewable energy access unit. Because the renewable energy access unit and the main power grid have significant differences in dynamic response characteristics, if the controller structure, electrical interface characteristics, and time scale uniformity are not fully considered, it can easily lead to distortion in the interaction between models, affecting the accuracy of the panoramic model.

[0040] This application further proposes a method for constructing a dynamic power grid model that includes at least one renewable energy access unit and coupling it with the main power grid model to obtain a panoramic model. This method includes: constructing state-space models of photovoltaic power generation units, wind power generation units, energy storage devices, and electric vehicle charging and discharging units based on controller structure and electrical interface characteristics. The state-space models include voltage, current, active power, frequency, controller output, and battery state. A main power grid model is constructed that includes synchronous generators, load models, node admittance matrices, and network topology. The modeling results of the renewable energy access units are coupled with the main power grid model through node power injection to unify the time scale, establish a consistent mapping relationship between electrical variables, control variables, and structural variables, and form a unified differential-algebraic equation model that includes the renewable energy access units and the main power grid model to obtain a panoramic model.

[0041] The state-space model is constructed based on the controller hardware parameters and electrical interface impedance characteristics of the renewable energy unit. For example, the dual-loop control structure parameters of the photovoltaic inverter are mapped to the feedback coefficient matrix in the state equation. The main grid model transforms the transmission line parameters into a mathematical expression of the network topology through the node admittance matrix, for example, by using a π-type equivalent circuit to calculate the line admittance value. Time scale unification is achieved by setting a synchronous sampling period, for example, aligning the renewable energy unit control cycle with the grid power flow calculation cycle to a millisecond-level time resolution.

[0042] Specifically, in the modeling of photovoltaic power generation units, the maximum power point tracking algorithm is transformed into nonlinear constraints in the state equations, and the dynamic characteristics of the inverter's phase-locked loop are described by second-order differential equations. In the main grid model, the dynamic equations of the synchronous generator's excitation system are coupled with the state equations of the renewable energy units through node-injected power equations, for example, by superimposing the renewable energy output current vector onto the right side of the node current balance equations. The time scale is unified by discretizing the differential-algebraic equations using the implicit trapezoidal integral method, for example, by simultaneously solving the renewable energy unit control equations and the grid electromagnetic transient equations using a 0.1 millisecond step size. The resulting unified differential-algebraic equation model can accurately characterize the dynamic interaction between the renewable energy access unit and the main grid during fault ride-through; for example, the power support response of the energy storage device during grid frequency drops is fully mapped into the model's state variables.

[0043] As a preferred embodiment, the solution of this application is specifically implemented as follows: When constructing a power grid dynamic model that includes at least one new energy access unit and coupling it with the main power grid model to obtain a panoramic model, the following steps are included: First, state-space models for photovoltaic (PV) power generation units, wind power generation units, energy storage devices, and electric vehicle (EV) charging / discharging units are constructed based on the controller structure and electrical interface characteristics. These state-space models include voltage, current, active power, frequency, controller output, and battery state. Specifically, for the PV power generation unit, a mathematical model is established containing state variables such as DC-side voltage, AC-side current, and the maximum power point tracking controller output; for the wind power generation unit, dynamic equations are constructed containing state variables such as generator speed, pitch angle, and converter current; for the energy storage device, mathematical descriptions of state variables such as battery state of charge and power converter output current are established; and for the EV charging / discharging unit, a dynamic model is constructed containing variables such as charging power and battery state of charge.

[0044] A main power grid model is constructed, including synchronous generators, load models, node admittance matrices, and network topology. Specifically, the Park model is used to describe the dynamic characteristics of synchronous generators, and a system of differential equations is established that includes state variables such as rotor angle, angular velocity, and excitation voltage. For the load model, a static or dynamic model is selected to describe it based on the load characteristics. Based on the power grid topology, a node admittance matrix is ​​constructed to reflect the electrical connection relationships between nodes.

[0045] Finally, the modeling results of the renewable energy access units are coupled with the main grid model through node power injection, unifying the time scale and establishing a consistent mapping relationship between electrical variables, control variables, and structural variables. This forms a unified differential-algebraic equation model that includes both the renewable energy access units and the main grid model, resulting in a comprehensive system model. Specifically, the renewable energy access units are equivalent to power injection sources and connected to the main grid model through power balance equations; the time scale of each sub-model is unified to ensure model consistency; and mapping relationships are established between electrical variables such as voltage and current, control variables such as controller output, and structural variables such as network topology, forming a complete set of differential-algebraic equations as the system's comprehensive model.

[0046] Through the above technical solutions, this application achieves effective coupling between the new energy access unit and the main power grid model, constructing a panoramic dynamic power grid model that includes various forms of new energy. This model accurately reflects the impact of new energy access on the dynamic characteristics of the power grid. By using a unified differential-algebraic equation model, a unified expression of devices with different time scales and characteristics is achieved, which helps to capture the multi-scale dynamic behavior of the system. Furthermore, this method improves the completeness and accuracy of the model by establishing a consistent mapping between electrical, control, and structural variables.

[0047] In some of the schemes mentioned above in this application, after constructing a unified differential algebraic equation model that includes the new energy access unit and the main power grid model, due to the time scale difference between the dynamic response characteristics of the new energy power electronic interface and the traditional synchronous generator, directly performing overall model analysis will result in excessive coupling of the dynamic process, making it difficult to effectively separate the stability influencing factors under different time scales, thereby affecting the accuracy of subsequent subsystem stability determination.

[0048] This application further proposes dividing the system state variables in the panoramic model into slow variables, fast variables, and disturbance variables according to their dynamic response characteristics. Slow variables include rotor angle, angular velocity, and frequency; fast variables include inverter voltage, current control loop output, and phase-locked loop state; and disturbance variables include wind speed, irradiance, and load power changes. The differential-algebraic equation model is transformed into a singular perturbation form, where ε is a scale-separated small parameter, X represents the vector of slow variables, Y represents the vector of fast variables, and Z represents the vector of disturbance variables. This represents the derivative of the slow variable with respect to time. This represents the derivative of a fast variable with respect to time. Represents the dynamic function of a slow system. This represents the dynamic function of a fast system.

[0049] The dynamic response characteristics are quantified based on the synchronous generator's mechanical inertia time constant, the power electronic equipment control cycle, and the rate of change of external environmental disturbances. For example, the synchronous generator rotor dynamic time constant is in the second range, the inverter control loop response time is in the millisecond range, and the wind speed change cycle is in the minute range. The singular perturbation form introduces a scale-separation small parameter ε, decomposing the original differential-algebraic equation into three interrelated subsystem equations. The slow system equation corresponds to the mechanical dynamic process dominated by the synchronous generator, the fast system equation corresponds to the electromagnetic dynamic process dominated by the power electronic equipment, and the ultra-slow perturbation system equation corresponds to the quasi-steady-state process driven by external environmental disturbances. The value of parameter ε is determined according to the proportion of the actual equipment's dynamic response time. For example, when the synchronous generator time constant is in the 10-second range and the inverter control cycle is in the 10-millisecond range, ε can be in the order of 0.001.

[0050] Specifically, decoupled analysis of multi-timescale dynamic processes is achieved through dynamic variable classification and singular perturbation model reconstruction. The slow variable set focuses on the synchronous generator rotor motion equations, whose dynamic equations do not contain fast variable differential terms, facilitating mechanical energy balance analysis using the equal area rule. The fast variable set independently forms fast dynamic equations, whose state change rate is controlled by the ε coefficient, allowing Jacobian matrix eigenvalue analysis to focus on the stability of the power electronic control loop. Disturbance variables, as ultra-slow dynamic inputs, are quantitatively assessed for the impact of external disturbances on the system state through sliding window prediction combined with a sensitivity matrix. The scale separation method allows each subsystem to adopt the stability criterion best suited to its dynamic characteristics; for example, eigenvalue analysis is used for fast systems, energy function methods for slow systems, and trend prediction methods for ultra-slow disturbance systems, effectively improving the accuracy and efficiency of stability assessment.

[0051] As a preferred embodiment, the solution of this application is specifically implemented as follows: The system state variables in the panoramic model are divided into slow variables, fast variables, and disturbance variables according to their dynamic response characteristics. Specifically, slow variables include rotor angle, angular velocity, and frequency; fast variables include inverter voltage, current control loop output, and phase-locked loop state; and disturbance variables include wind speed, irradiance, and load power changes.

[0052] Furthermore, the differential-algebraic equation model is transformed into a singular perturbation form: , Where ε is the small parameter for scale separation, X represents the vector of slow variables, Y represents the vector of fast variables, and Z represents the vector of perturbation variables. This represents the derivative of the slow variable with respect to time. This represents the derivative of a fast variable with respect to time. Represents the dynamic function of a slow system. This represents the dynamic function of a fast system.

[0053] By classifying system state variables into different types and transforming the differential-algebraic equation model into a singular perturbation form, scale decomposition of the panoramic model can be achieved. For example, for a simplified system including a synchronous generator and a photovoltaic inverter, the rotor angle and angular velocity of the synchronous generator can be treated as slow variables, the voltage and current control loops of the photovoltaic inverter as fast variables, and the irradiance as a perturbation variable. Through this classification and transformation, the originally complex system model can be decomposed into a slow system, a fast system, and a perturbation system, laying the foundation for subsequent stability analysis.

[0054] Through the above technical solution, this application enables multi-timescale decomposition of complex power grid systems including renewable energy integration. Since renewable energy equipment and traditional synchronous generators possess different dynamic characteristics, dividing the system state variables into slow variables, fast variables, and disturbance variables can more accurately reflect the dynamic response characteristics of various types of equipment. By transforming the differential-algebraic equation model into a singular perturbation form, the fast and slow dynamic processes can be clearly distinguished, which is beneficial for subsequent stability analysis at different time scales. This decomposition method provides a theoretical basis for analyzing the multi-scale dynamic behavior of complex power grid systems, helping to improve the accuracy and relevance of stability assessments.

[0055] In some of the above-mentioned schemes in this application, a panoramic model containing new energy access units is constructed and scale decomposition is performed. However, when dealing with the stability of fast systems, due to the high dynamic characteristics of power electronic equipment, the complexity of its nonlinear state equations makes it difficult to quickly and accurately assess the stability state, and misjudgment is easily caused by linearization deviation or eigenvalue solution error.

[0056] This application further proposes a method for linearizing the fast system with small perturbations, constructing the Jacobian matrix, and determining the stability of the fast system, including: selecting the equilibrium operating point of the fast system in its current operating state, and performing a first-order Taylor expansion of the nonlinear state equation of the fast system centered on the equilibrium operating point; constructing the incremental equation Δ of the fast system. =A·ΔY, where A represents the Jacobian matrix of the partial derivative of the fast variable with respect to itself. Solve for all eigenvalues ​​of the Jacobian matrix A. If the real parts of all eigenvalues ​​of the matrix are negative, the fast system is determined to be in a stable state under the current disturbance; otherwise, it is unstable.

[0057] The selection of the equilibrium operating point is based on the steady-state conditions of the fast system under its current operating state, such as the steady-state output value of the inverter voltage control loop or the steady-state phase angle of the phase-locked loop. The first-order Taylor expansion achieves linear approximation by retaining the first-order partial derivatives of the nonlinear state equations, for example, by calculating the partial derivatives of the dynamic equations of the voltage control loop. The Jacobian matrix A is constructed by extracting the partial derivatives of the fast variables with respect to their own changes to form a square matrix, such as the partial derivatives of the voltage control loop output with respect to current changes. The eigenvalues ​​are solved using the QR algorithm or the Arnoldi iterative method, for example, by obtaining the matrix eigenvalues ​​through numerical calculation tools.

[0058] Specifically, when linearizing near the equilibrium operating point, the nonlinear state equation is approximated as a linear incremental equation, thus simplifying the dynamic response of the fast system to linear system behavior. The sign of the real parts of the eigenvalues ​​of the Jacobian matrix A directly reflects the stability of the system near the equilibrium point: if all eigenvalues ​​have negative real parts, the system can return to the equilibrium point after being disturbed; if there are eigenvalues ​​with positive real parts, the disturbance will cause the system to deviate from the equilibrium point. For example, after the dynamic equation of the phase-locked loop (PLL) of the inverter control loop is linearized, the real parts of the eigenvalues ​​related to the PLL damping coefficient in the Jacobian matrix are negative, indicating that the PLL dynamics are stable. Solving for the matrix eigenvalues ​​using numerical calculation tools avoids the time consumption of traditional trial-and-error methods and improves evaluation efficiency. The linearization method effectively reduces the complexity of dynamic analysis of fast systems.

[0059] As a preferred embodiment, the solution of this application is specifically implemented as follows: The equilibrium operating point of the fast system under its current operating state is selected, and a first-order Taylor expansion of the nonlinear state equations of the fast system is performed with the equilibrium operating point as the center. For example, for a fast system including a photovoltaic inverter, the steady-state values ​​of the inverter's output voltage and current are selected as the equilibrium operating point.

[0060] Constructing the incremental equations Δ of the fast system =A·ΔY. Where A represents the Jacobian matrix of the partial derivatives of the fast variables with respect to themselves. Specifically, for n fast variables, construct an n×n Jacobian matrix A, where the matrix element aij represents the partial derivative of the i-th fast variable with respect to the j-th fast variable.

[0061] Find all the eigenvalues ​​of the Jacobian matrix A. Further, use the QR algorithm to calculate the eigenvalues ​​of matrix A. This yields n complex eigenvalues ​​λ1, λ2, ..., λn.

[0062] Determine if the real parts of all eigenvalues ​​of the matrix are negative. Specifically, check if the real part Re(λi) of each eigenvalue λi is less than zero. If all Re(λi) are zero, the fast system is considered stable under the current perturbation; otherwise, it is considered unstable.

[0063] Through the above technical solution, this application achieves accurate assessment of the stability of fast systems. By constructing the Jacobian matrix and analyzing its eigenvalues, the dynamic characteristics of fast systems can be effectively captured. Compared with traditional methods, this scheme can better handle the rapid dynamic processes brought about by the integration of new energy sources, improving the accuracy and reliability of grid stability assessment. Furthermore, this method is computationally simple, suitable for real-time online assessment, and conducive to the timely detection of potential stability problems.

[0064] In some of the solutions described above in this application, after scale decomposition of the panoramic model based on the singular perturbation method, it is necessary to determine the stability of the slow system. However, traditional methods, when quantifying the dynamic behavior of synchronous generators during faults, have the drawback of failing to accurately capture the impact of the interaction between mechanical and electromagnetic power on system stability, making it difficult to predict whether the rotor angle will tend towards a stable equilibrium point after the fault is cleared.

[0065] This application further proposes a method for processing slow systems based on the equal area rule, calculating the acceleration area and maximum deceleration area during faults, and determining the stability of the slow system. This includes: obtaining the stable equilibrium point δ0 before the fault, the first stable equilibrium point δs after the fault is cleared, and the unstable equilibrium point δu after the fault, where the equilibrium point is a specific value of the synchronous generator rotor angle; calculating the acceleration area during the fault, which is the area formed by integrating the mechanical power and the electromagnetic power after the fault within the interval [δ0, δu]; calculating the maximum deceleration area, which is the area formed by integrating the mechanical power and the electromagnetic power after the fault within the interval [δs, δu]; comparing the acceleration area with the maximum deceleration area, and if the acceleration area is less than the maximum deceleration area, the slow system is determined to be in a stable state; otherwise, it is considered unstable.

[0066] The stable equilibrium point δ0 is obtained through steady-state power flow calculations of the system before the fault, while the unstable equilibrium point δu is determined by solving the extreme rotor angle point where mechanical power equals electromagnetic power in the system's dynamic equations after the fault. During the acceleration area calculation, the integration interval begins at the pre-fault equilibrium point δ0 and ends at the unstable equilibrium point δu; this area reflects the accumulated kinetic energy of the synchronous generator rotor during the fault. For the maximum deceleration area calculation, the integration interval begins at the stable equilibrium point δs reachable by the system after fault clearance and ends at the same unstable equilibrium point δu; this area characterizes the system's maximum kinetic energy absorption capacity after fault clearance.

[0067] Specifically, during the period from the occurrence of a fault to its clearance, the synchronous generator rotor accelerates due to excess mechanical power, causing the rotor angle to deviate from δ0. The accumulated kinetic energy during this phase can be quantified by integrating the area between δ0 and δu, representing the difference between the mechanical power and the post-fault electromagnetic power. After the fault is cleared, the system's operating point shifts to a new stable equilibrium point δs. At this point, the region where electromagnetic power exceeds mechanical power forms a deceleration area. When the acceleration area is less than the maximum deceleration area, the system has sufficient energy absorption capacity to bring the rotor angle to converge to δs, thus maintaining stability. For example, in a certain scenario, the calculated acceleration area is 1200 MJ, and the maximum deceleration area is 1500 MJ, indicating system stability. If the acceleration area reaches 1600 MJ while the maximum deceleration area is 1400 MJ, the system is considered unstable. This method, through physically quantified area comparison, effectively overcomes the limitation of traditional qualitative analysis in accurately assessing the energy balance state.

[0068] As a preferred embodiment, the solution of this application is specifically implemented as follows: The system obtains the stable equilibrium point δ0 before the fault, the first stable equilibrium point δs after the fault is cleared, and the unstable equilibrium point δu after the fault in the slow system. The equilibrium point is a specific value of the synchronous generator rotor angle. For example, δ0 = 30°, δs = 45°, and δu = 60° can be calculated using power system simulation software.

[0069] Calculate the acceleration area during the fault. Specifically, integrate the mechanical power and the electromagnetic power after the fault over the interval [δ0, δu] to obtain the acceleration area. Assuming the mechanical power is a constant of 1.0 and the electromagnetic power after the fault is 0.8sin(δ), the acceleration area can be expressed as the integral ∫[1.0-0.8sin(δ)]dδ, over the interval [30°, 60°].

[0070] Calculate the maximum deceleration area. Further, integrate the mechanical power and the post-fault electromagnetic power over the interval [δs, δu] to obtain the maximum deceleration area. Using the assumed power function described above, the maximum deceleration area can be expressed as the integral ∫[0.8sin(δ)-1.0]dδ, over the interval [45°, 60°].

[0071] The acceleration area is compared with the maximum deceleration area. Therefore, if the calculated acceleration area is less than the maximum deceleration area, the slow system is considered stable; otherwise, it is considered unstable. For example, if the calculated acceleration area is 0.3 and the calculated maximum deceleration area is 0.4, the slow system is considered stable.

[0072] Through the above technical solution, this application can perform quantitative analysis of slow systems based on the equal area criterion, accurately determining the stability of slow systems by comparing the acceleration area and the maximum deceleration area. This method avoids complex time-domain simulations, improving computational efficiency. Simultaneously, by considering the changes in system state before and after a fault, it can more comprehensively evaluate the dynamic characteristics of the system, improving the accuracy of stability assessment. Furthermore, this method is intuitive and easy to understand, facilitating rapid mastery and application by engineers, thus enhancing the practicality of power grid stability assessment.

[0073] In some of the schemes described above in this application, after decomposing the system into slow system, fast system and ultra-slow perturbation system by singular perturbation method, it is possible to evaluate the dynamic characteristics of different time scales respectively. However, when dealing with ultra-slow perturbation system, the existing methods are difficult to accurately predict the long-term changing trend of perturbation variables and their cumulative impact on the system state, resulting in the inability to identify the risk of critical state transition caused by slow perturbation in a timely manner.

[0074] This application further proposes to use a sliding window to analyze the changing trend of the disturbance variable in the ultra-slow disturbance system within the prediction period, and to combine the disturbance influence sensitivity matrix to determine whether the critical transition of the system state is triggered, thereby determining the stability of the ultra-slow disturbance system.

[0075] The sliding window duration is set to 30 to 300 seconds, covering typical cycles of new energy output fluctuations and load changes, effectively capturing the dynamic evolution of disturbance variables. Time-series data of disturbance variables are extracted in segments through the sliding window to ensure data continuity and timeliness. The time-series prediction model is built based on historical data, employing autoregressive integral moving average or long short-term memory network algorithms to output the predicted slope, amplitude, and fluctuation indicators of disturbance variables in future cycles, quantifying their changing trends. The system identification method trains the disturbance impact sensitivity matrix using input and output data; this matrix represents the system state shift caused by a unit change in the disturbance variable. Multiplying the predicted trend by the sensitivity matrix yields the predicted system state shift value. By comparing this value with a preset stability safety boundary threshold, the stability state of the ultra-slow disturbance system is determined.

[0076] Specifically, the sliding window setting enables the system to dynamically track the time-varying characteristics of disturbance variables, avoiding prediction bias due to insufficient data length. The time series prediction model generates the trajectory of future disturbance variables by analyzing the periodic, trend, and random components in historical data, such as predicting the increasing trend of wind speed or the sudden drop in irradiance. The disturbance impact sensitivity matrix is ​​updated through offline training or online recursive algorithms, reflecting the system's sensitivity to disturbances under current operating conditions. Combining the predicted trend with the sensitivity matrix allows for accurate calculation of the cumulative impact of future disturbances on the system state; for example, predicting that an increase in load power will cause a voltage shift of 0.05 pu at critical nodes. The stability safety boundary threshold is set based on system operating procedures or historical fault data; for example, the voltage shift threshold is set to ±0.1 pu. When the predicted shift value exceeds the threshold, an early warning mechanism is triggered, indicating a risk of system instability. The advantage of this method lies in combining disturbance trend prediction with sensitivity analysis, avoiding the limitations of a single prediction model while quantifying the actual impact of disturbances on the system state, thus improving the accuracy of stability assessment for ultra-slow disturbance systems.

[0077] As a preferred embodiment, the solution of this application is specifically implemented as follows: A sliding window with a duration of 60 seconds is set, and time-series data of disturbance variables are extracted within the sliding window. Disturbance variables include wind speed, irradiance, and load power changes.

[0078] Time series forecasting models are constructed based on historical data of disturbance variables. An autoregressive moving average (ARMA) model is used to predict wind speed and irradiance, while exponential smoothing is used to predict load power changes. These models are used to obtain the predicted trends of disturbance variables over the next 5 minutes, and the predicted slope, amplitude, and volatility indicators are extracted.

[0079] System identification is performed to obtain the sensitivity matrix of disturbance effects. The least squares method is used to fit historical data to establish the sensitivity relationship between the disturbance variable and the system state variable.

[0080] The predicted system state shift caused by future disturbances is obtained by multiplying the disturbance prediction trend by the sensitivity matrix. Specifically, the predicted change in the disturbance variable is multiplied by the corresponding sensitivity coefficient, and the results are summed to obtain the shift of each state variable.

[0081] Set stability safety boundary thresholds. For example, for voltage stability, set a voltage deviation of ±5% as the threshold; for frequency stability, set a frequency deviation of ±0.2Hz as the threshold. If the predicted system state offset exceeds these thresholds, the ultra-slow disturbance system is determined to be in an unstable state; otherwise, it is in a stable state.

[0082] Through the above technical solution, this application achieves dynamic assessment of the stability of ultra-slow disturbance systems in renewable energy grid integration. By employing a sliding window analysis to analyze the changing trends of disturbance variables and combining this with a disturbance impact sensitivity matrix, the system state shift over a future period can be predicted. By comparing with a preset threshold, it is possible to promptly determine whether the system will trigger a critical state transition due to the cumulative effect of disturbances. This method overcomes the shortcomings of traditional static analysis methods in capturing the impact of long-term disturbances, improving the accuracy of assessing the volatility impact of renewable energy. Simultaneously, the sliding window setting allows the assessment process to be dynamically updated, promptly reflecting changes in system state and providing operators with more real-time and effective stability early warning information.

[0083] In some of the schemes mentioned above in this application, the dynamic behavior of the new energy access system is evaluated through multi-scale decomposition and stability analysis. However, the impact of the grid topology characteristics on stability is not fully considered. It is difficult to effectively identify potential structural instability factors caused by weak connections, single-point fault amplification paths, or low redundancy distribution patterns, which may lead to the omission of instability risks caused by topological defects in the evaluation results.

[0084] This application further proposes a scheme to assist in judging the existence of potential structural instability factors in a system by using a graph neural network-assisted model. The scheme includes: constructing a graph with the power grid topology as input, where nodes represent generation units, load nodes, or converter devices, and edges represent bus connection relationships and line impedances; constructing a feature vector for each node in the graph, including the node's voltage amplitude, phase angle, injected active and reactive power, frequency offset, access type, and control mode; based on the graph neural network model, performing directed propagation and embedding updates of features between nodes through graph convolution operations to capture potential unstable structures in the topology, such as weak connections, single-point fault amplification paths, and low-redundancy distribution modes; outputting a node-level stability score, identifying nodes with scores below a threshold as key nodes with potential structural instability factors, and outputting a structurally weak subgraph.

[0085] The power grid topology is mapped as a graph structure containing nodes and edges. Node feature vectors integrate electrical parameters and control states; for example, when the voltage amplitude deviation exceeds 5% of the rated value or the frequency offset exceeds 0.2Hz, the corresponding dimension value in the node feature vector is marked as abnormal. Graph convolution operations employ a multi-layer message passing mechanism, aggregating feature information from adjacent nodes in each layer. For example, the first convolutional layer extracts local connectivity strength, and the second convolutional layer identifies cross-regional coupling relationships. The node attention mechanism dynamically adjusts the information contribution during feature propagation by calculating the association weights between nodes. For example, the frequency offset feature of a load node is given higher weight when propagating to adjacent generator nodes.

[0086] Specifically, during real-time assessment, grid topology information and current operating parameters are converted into graph data and input into the model. Node feature vectors are normalized, converting voltage amplitudes to per-unit values ​​and frequency offsets to percentages. Graph convolutional layers perform feature transformations, such as using linear transformation matrices to map the original features to a high-dimensional space, and then performing weighted summation based on the connection relationships defined by the adjacency matrix. After multiple rounds of graph convolution and attention calculations, the model outputs a stability score for each node; nodes with scores below 0.6 are identified as potentially unstable nodes. Weak subgraphs are generated by extracting low-scoring nodes and their connecting edges; for example, subgraphs containing multiple low-redundancy load nodes and connecting line impedances higher than 0.5 pu are marked as high-risk areas. This process enables the assessment system to locate structural vulnerabilities caused by topological characteristics, such as identifying risk paths where overload of a single tie line leads to islanding of adjacent areas, thus supplementing the deficiencies of multi-scale dynamic stability analysis.

[0087] As a preferred embodiment, the solution of this application is specifically implemented as follows: Construct an input graph representing the power grid topology. Nodes in the graph represent generation units, load nodes, or converter devices, while edges represent bus connections and line impedances.

[0088] Construct a feature vector for each node in the graph. The feature vector includes the node's voltage magnitude, phase angle, injected active and reactive power, frequency offset, connection type, and control mode.

[0089] Based on a graph neural network model, directed propagation and embedding updates of features between nodes are performed through graph convolution operations. This captures potentially unstable structures in the topology, including weak connections, single-point fault amplification paths, and low-redundancy distribution patterns.

[0090] Output node-level stability scores. Nodes with scores below a threshold are identified as critical nodes with potential structural instability factors, and structurally weak subgraphs are output.

[0091] Specifically, a Graph Attention Network (GAT) can be used as the graph neural network model. The GAT model contains multiple graph attention layers, each consisting of multiple attention heads. Each attention head independently calculates the attention coefficients between nodes and aggregates the features of neighboring nodes.

[0092] The forward propagation process of the attention layer is as follows: For node i, its feature vector hi is first mapped to a high-dimensional space through a linear transformation W: hi'=W*hi Calculate the attention coefficient between node i and its neighbor node j: eij=LeakyReLU(a^T*[hi'||hj']) Where a is a learnable attention vector, and || represents the concatenation operation.

[0093] Normalize the attention coefficient: αij=softmax(eij)=exp(eij) / Σk∈Niexp(eik) Characteristics of aggregated neighbor nodes: hi''=σ(Σj∈Niαij*hj') Where σ is the activation function.

[0094] Output splicing or averaging of multi-head attention: hi'''=||Kk=1hik'' or hi'''=(1 / K)*ΣKk=1hik'' By stacking multiple layers of GAT, higher-order topology information can be captured. The final layer outputs a stability score for each node.

[0095] Through the above technical solutions, this application can effectively identify potential instability factors caused by the topology of the power grid. The graph neural network model can automatically learn the mutual influence between nodes and capture complex topological features. Through an attention mechanism, the model can focus on key connections, improving the ability to identify weak connections and fault propagation paths. Node-level scoring outputs enable precise location of critical nodes at risk, providing targeted guidance for subsequent preventative control. The output of the structurally weak subgraph intuitively displays vulnerable areas in the system, helping operators quickly grasp the overall stability of the system. Compared with traditional topology analysis methods, this method has stronger nonlinear feature extraction capabilities and adaptability, and can better cope with the increased complexity of the power grid structure after the large-scale integration of new energy sources.

[0096] In some of the solutions described above in this application, when using graph neural networks to assist in judging potential structural instability factors in the power grid, there are problems such as low model training efficiency, insufficient attention to key nodes during feature propagation, and limited real-time evaluation accuracy.

[0097] This application further proposes to construct training samples using historical disturbance scenarios and fault records during training and application. The sample labels are whether the system experiences voltage collapse, frequency oscillation, node islanding, or tie-line overload after the disturbance. The graph attention network structure in graph neural networks is adopted, and a node attention mechanism is added during feature propagation. The graph neural network model is trained through supervised learning. In real-time evaluation, the current system state and topology information are input into the trained graph neural network to obtain auxiliary judgment results.

[0098] The historical disturbance scenarios and fault records include topology, node characteristics, and fault consequences data under various power grid disturbance events. Sample labels are generated from actual fault records or simulations. The graph attention network structure dynamically adjusts the influence strength of adjacent nodes during feature propagation by calculating attention weights between nodes. The node attention mechanism introduces learnable attention coefficients in each graph convolution operation to capture the differences in the impact of different nodes on stability. Supervised learning uses the cross-entropy loss function to optimize model parameters, and updates the attention weights and feature extraction layer parameters through backpropagation during training.

[0099] Specifically, during the training sample construction phase, snapshots of the power grid topology before and after disturbance events, node electrical measurement data, and fault type labels are extracted from the historical database to form a mapping relationship between the input graph structure and the target classification labels. During feature propagation, the graph attention network first calculates the attention coefficients between node i and its neighboring node j, converts these coefficients into weights using a normalized exponential function, and then aggregates the features of neighboring nodes using weighted aggregation. In the multi-layered stacked attention layers, the model progressively learns higher-order association patterns between nodes, such as how a single-point fault can trigger a cascading effect through low-redundancy paths. In the real-time evaluation phase, the current power grid topology connections and node operating states are encoded as a graph structure input to the pre-trained model. The model calculates the stability score of each node through forward propagation and outputs the identifiers of nodes below a threshold. By visualizing the attention weights, the key connection paths leading to the reduced scores can be further traced, assisting in locating structural weaknesses.

[0100] As a preferred embodiment, the specific implementation of this application is as follows: During the training of the graph neural network model of the power grid dynamic assessment system, the historical disturbance scenario data comes from the fault event records of a provincial power grid over the past five years, including seven typical disturbance types such as voltage sag, cascading trip, and frequency anomalies. The training samples are constructed by extracting a snapshot of the power grid topology ten minutes before the event. The node feature vectors are composed of voltage amplitude, phase angle, injected power, and control mode recorded by the SCADA system, while the edge features are generated from line impedance parameters and real-time power flow data. Sample labels are binary-classified based on whether voltage collapse or frequency exceedance occurs after the event. The graph attention network adopts a three-layer graph convolutional layer structure, with a multi-head attention mechanism in each layer. The node attention weights are calculated using trainable parameters, and the voltage offset of adjacent nodes is given higher attention during feature propagation. The model training uses the cross-entropy loss function, and the Adam optimizer is used for parameter updates. After training, the model is deployed to the real-time assessment platform. During the application phase, the system acquires the topology connection relationship and node operation data of the EMS system every five minutes. After feature vectorization, the data is input into the trained graph neural network, which outputs a node-level stability score. Nodes with a score below 0.6 are marked as potential weak points.

[0101] Through the above technical solutions, this application achieves accurate identification of hidden structural risks in power grid topology, effectively captures the dynamic correlation characteristics between nodes through graph attention mechanism, solves the problem of insufficient identification of low-redundancy network structures and fault propagation paths by traditional methods, and improves the early warning capability of structural instability factors in the scenario of new energy access.

[0102] In some of the solutions mentioned above in this application, a method for identifying potential instability factors in power grid topology based on graph neural network models was proposed. However, in the process of model training and application, there are problems such as low efficiency of node feature propagation and insufficient capture of dynamic influence of key nodes, which leads to limited accuracy of auxiliary judgment results.

[0103] This application further proposes a method for training and applying a graph neural network model, including constructing training samples using historical disturbance scenarios and fault records, with the sample labels indicating whether the system experiences voltage collapse, frequency oscillation, node islanding, or tie-line overload after the disturbance; employing the graph attention network structure in graph neural networks, incorporating a node attention mechanism during feature propagation, and training the graph neural network model through supervised learning; and inputting the current system state and topology information into the trained graph neural network in real-time evaluation to obtain auxiliary judgment results.

[0104] The training samples are constructed using multi-dimensional feature vectors extracted from historical data, including voltage amplitude, phase angle, node injected power, and frequency offset. Sample labels are assigned as binary classification results based on actual fault records. The graph attention network employs a multi-head attention layer to distribute feature weights among nodes. Each attention head calculates the correlation coefficient between nodes, which is then normalized using a softmax function to generate an attention weight matrix. The node attention mechanism introduces learnable weight parameters during the feature aggregation stage, dynamically adjusting the influence of neighboring nodes on the central node. The supervised learning process uses a cross-entropy loss function to optimize model parameters and updates network weights through backpropagation.

[0105] Specifically, during the training phase, grid operation data with varying proportions of renewable energy integration is extracted from historical databases. Each sample corresponds to a topology snapshot at a specific time point and subsequent fault status labels. During graph attention network initialization, each node's feature vector undergoes a linear transformation to generate a query vector, key vector, and value vector. An unnormalized attention score is obtained by calculating the dot product similarity between the query vector and the key vector. This score is then further masked based on the adjacency relationships of the grid topology to filter interference from non-connected nodes. The normalized attention weights are then weighted and summed with the value vector to generate an updated node embedding representation. In the real-time evaluation phase, the current grid node voltage and power injection data are input into the trained model. The graph attention network automatically identifies correlation patterns between key nodes and outputs node-level stability scores. Nodes below a threshold are marked as potential instability factors. For example, in a test scenario involving photovoltaic fluctuations and load abrupt changes, the model accurately identifies a decline in the energy storage node score in weakly connected areas, triggering an early warning signal.

[0106] As a preferred embodiment, the specific implementation of this application is as follows: When the fast system stability determination result is stable, the fast system stability code is set to +1; the slow system obtains an acceleration area value of 120MW·s and a maximum deceleration area of ​​150MW·s in the fault acceleration area calculation, and is determined to be in a stable state, and is coded as +1; the ultra-slow disturbance system finds that the state offset prediction value caused by the disturbance variable is 0.85 through sliding window prediction, which is lower than the safety boundary threshold of 1.2, and is coded as +1. After the graph neural network calculates the node-level stability score, the normalized score R is 0.92. The weight coefficients α=0.3, β=0.4, γ=0.2, and δ=0.1 are set, and the comprehensive index S=0.992 is calculated by the formula S=0.3×1+0.4×1+0.2×1+0.1×(1-0.92). When the stability threshold T is set to 0.8, since S≥T, the system is determined to be in a stable state, the evaluation conclusion output is stable, and the stable operating state of each subsystem is marked.

[0107] Through the above technical solution, this application can effectively integrate dynamic stability analysis results from multiple time scales with structural risk assessment indicators, and achieve comprehensive evaluation through quantitative coding and weight allocation, avoiding misjudgments of stability in a single subsystem but overall instability. Simultaneously, the dynamic combination of normalized scoring and coding results can accurately capture critical states, solving the problem of insufficient identification of multi-scale coupled instability mechanisms in traditional methods, and improving the comprehensiveness and accuracy of dynamic stability assessment for power grids with a high proportion of new energy sources.

[0108] The above embodiments introduce coupled modeling of the new energy access unit and the main power grid model, and combine the singular perturbation method to divide the complex dynamic behavior of the power grid into three time-scale subsystems: fast, slow, and ultra-slow disturbances. For different subsystems, quantitative methods such as small disturbance linearization, equal area rule, and disturbance trend prediction are used to conduct hierarchical stability analysis, effectively improving the accuracy of dynamic stability state discrimination under multiple time scales. A graph neural network model is introduced to assist in the identification of structural instability risks based on the power grid topology, realizing intelligent perception and early warning of weak structural nodes, overcoming the deficiency of existing technologies that cannot identify structural risks based solely on the overall system state. This application features high modeling granularity, a clear judgment mechanism, and strong structural identification capabilities, improving the accuracy of power grid stability assessment in complex new energy access scenarios.

[0109] In another preferred embodiment based on the above embodiments, see [reference] Figure 2 As shown, this embodiment provides a dynamic evaluation system for grid stability when new energy sources are integrated, used to apply the aforementioned dynamic evaluation method for grid stability when new energy sources are integrated, including: The building unit is configured to build a dynamic model of the power grid that includes at least one new energy access unit and couple it with the main power grid model to obtain a panoramic model. The new energy access unit includes a photovoltaic power generation unit, a wind power generation unit, an energy storage device, or an electric vehicle charging and discharging unit. The decomposition unit is configured to perform scale decomposition of the panoramic model based on the singular perturbation method, constructing a slow system dominated by synchronous generators, a fast system dominated by power electronic equipment, and a perturbation-driven ultra-slow perturbation system. The processing unit is configured to perform small perturbation linearization on the fast system, construct the Jacobian matrix, and determine the stability of the fast system; process the slow system based on the equal area rule, calculate the fault acceleration area and the maximum deceleration area, and determine the stability of the slow system; and use a sliding window to analyze the changing trend of the perturbation variable in the ultra-slow perturbation system within the prediction period, and combine the perturbation influence sensitivity matrix to determine whether the critical transition of the system state is triggered, and determine the stability of the ultra-slow perturbation system. The judgment unit is configured to make an auxiliary judgment on whether there are potential structural instability factors in the system based on the current system state, power flow distribution and topology information, using a graph neural network-assisted model. The evaluation unit is configured to assess whether the power grid of the current renewable energy access system is in a stable state based on fast system stability, slow system stability, ultra-slow disturbance system stability, and auxiliary judgment results, and to obtain an evaluation conclusion.

[0110] Specifically, the construction unit first establishes a state-space model of the new energy unit, including voltage, current, and battery state variables, and couples it with the node admittance matrix of the main grid to form a panoramic model. The decomposition unit classifies variables according to dynamic response characteristics, transforms differential-algebraic equations into singular perturbation forms, and achieves time-scale separation. In the processing unit, the fast system module linearizes the nonlinear equations, constructs the Jacobian matrix, and solves for eigenvalues; stability is determined if the real parts of all eigenvalues ​​are negative. The slow system module calculates the acceleration area and maximum deceleration area during a fault, and judges the stability of the synchronizing machine by comparing the areas. The ultra-slow disturbance module uses a time series prediction model to analyze the future trends of disturbance variables such as wind speed and irradiance, and predicts state shifts using a sensitivity matrix. The judgment unit captures weak connections and single-point fault paths in the topology through a graph attention network, outputting node-level scores to identify weak subgraphs. The evaluation unit normalizes and weights the stability codes and structural scores of each subsystem, and when the comprehensive index exceeds a preset threshold, the system is determined to be unstable, and the specific unstable link and risk location are marked.

[0111] As a preferred embodiment, the solution of this application is implemented as follows: The dynamic evaluation system for grid stability includes a construction unit, a decomposition unit, a processing unit, a judgment unit, and an evaluation unit. The construction unit establishes a state-space model of the photovoltaic power generation unit based on its dual-closed-loop control structure. A sixth-order detailed model is used for the synchronous generator. The node admittance matrix is ​​generated through power flow calculation, and the power balance equation dynamically couples the new energy unit and the main grid at a common coupling point. The decomposition unit divides state variables using a constant time-order difference, sets the scale separation parameter ε to 0.01, classifies the synchronous generator rotor dynamic equation as a slow system, classifies the inverter current loop equation as a fast system, and models the irradiance change rate as an ultra-slow disturbance system. In the fast system analysis, the processing unit selects the grid connection point voltage as the equilibrium point, constructs a Jacobian matrix for eigenvalue calculation, and calculates the acceleration area as 50.3 MW·s and the maximum deceleration area as 62.1 MW·s based on the power angle curve in the slow system analysis. The ultra-slow disturbance analysis uses a 120-second sliding window to predict the load power change trend and combines the sensitivity matrix to determine if the state offset exceeds the safety threshold. The judgment unit constructs a topology graph containing 328 nodes. The node feature vectors include voltage deviation rate and frequency fluctuation rate. A three-layer graph attention network identifies buses B23 and B45 as low-redundancy weak nodes. The evaluation unit sets weight coefficients α=0.4, β=0.3, γ=0.2, and δ=0.1, calculates a comprehensive index S=0.85, which is higher than the preset threshold of 0.7. Finally, it outputs a system stability conclusion and marks node B23 as needing reinforcement.

[0112] Through the above technical solutions, this application realizes the decoupled assessment of dynamic stability at multiple time scales and the joint analysis of topological vulnerability. It effectively distinguishes the interaction mechanism between the fast dynamic response of power electronic equipment and the slow dynamic of synchronous machines. By using a sliding window prediction, it can identify the cascading instability risk caused by the fluctuation of new energy output in advance. Combined with graph neural networks, it can mine the low-redundancy connection paths hidden in the topology, thereby improving the comprehensiveness and accuracy of the dynamic stability assessment of high-proportion new energy power grids.

[0113] Those skilled in the art will understand that embodiments of this application can be provided as methods, systems, or computer program goods. Therefore, this application can take the form of a completely hardware embodiment, a completely software embodiment, or an embodiment combining software and hardware aspects. Furthermore, this application can take the form of a computer program goods embodied on one or more computer-usable storage media (including, but not limited to, disk storage, CD-ROM, optical storage, etc.) containing computer-usable program code.

[0114] This application is described with reference to flowchart illustrations and / or block diagrams of methods, apparatus (systems), and computer program goods according to embodiments of this application. It will be understood that each block of the flowchart illustrations and / or block diagrams, and combinations of blocks in the flowchart illustrations and / or block diagrams, can be implemented by computer program instructions. These computer program instructions can be provided to a processor of a general-purpose computer, special-purpose computer, embedded processor, or other programmable data processing apparatus to produce a machine, such that the instructions, which execute via the processor of the computer or other programmable data processing apparatus, generate instructions for implementing the flowchart... Figure 1 One or more processes and / or boxes Figure 1 A device that provides the functions specified in one or more boxes.

[0115] These computer program instructions may also be stored in a computer-readable storage medium that can direct a computer or other programmable data processing device to function in a particular manner, such that the instructions stored in the computer-readable storage medium produce an article of manufacture including instruction means, which are implemented in a process Figure 1 One or more processes and / or boxes Figure 1 The function specified in one or more boxes.

[0116] These computer program instructions may also be loaded onto a computer or other programmable data processing equipment to cause a series of operational steps to be performed on the computer or other programmable equipment to produce a computer-implemented process, thereby providing instructions that execute on the computer or other programmable equipment for implementing the process. Figure 1 One or more processes and / or boxes Figure 1 The steps of the function specified in one or more boxes.

[0117] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention and not to limit it. Although the present invention has been described in detail with reference to the above embodiments, those skilled in the art should understand that modifications or equivalent substitutions can still be made to the specific implementation of the present invention. Any modifications or equivalent substitutions that do not depart from the spirit and scope of the present invention should be covered within the scope of protection of the claims of the present invention.

Claims

1. A method for dynamic evaluation of grid stability when new energy sources are integrated, characterized in that, include: A dynamic power grid model containing at least one new energy access unit is constructed and coupled with the main power grid model to obtain a panoramic model. The new energy access unit includes a photovoltaic power generation unit, a wind power generation unit, an energy storage device, or an electric vehicle charging and discharging unit. The panoramic model is decomposed into scales based on the singular perturbation method, and a slow system dominated by synchronous generators, a fast system dominated by power electronic equipment, and an ultra-slow perturbation system driven by perturbation are constructed. The fast system is linearized with small perturbations to construct the Jacobian matrix and determine the stability of the fast system. The slow system is processed based on the equal area rule, the fault acceleration area and the maximum deceleration area are calculated, and the stability of the slow system is determined. The changing trend of the disturbance variable in the ultra-slow disturbance system within the prediction period is analyzed by using a sliding window, and the system stability is determined by combining the disturbance influence sensitivity matrix to determine whether a critical transition of the system state is triggered. Based on the current system state, power flow distribution, and topology information, a graph neural network-assisted model is used to help determine whether there are potential structural instabilities in the system. Based on the fast system stability, slow system stability, ultra-slow disturbance system stability, and auxiliary judgment results, assess whether the power grid of the current new energy access system is in a stable state and obtain the assessment conclusion.

2. The method for dynamic evaluation of grid stability for new energy access according to claim 1, characterized in that, When constructing a power grid dynamic model that includes at least one new energy access unit and coupling it with the main power grid model to obtain a panoramic model, the following steps are included: Based on the controller structure and electrical interface characteristics, a state space model is constructed for the photovoltaic power generation unit, wind power generation unit, energy storage device and electric vehicle charging and discharging unit. The state space model includes voltage, current, active power, frequency, controller output and battery state. Construct a main power grid model that includes synchronous generators, load models, node admittance matrices, and network topology; By coupling the modeling results of the new energy access unit with the main power grid model through node power injection, a unified time scale is established, and a consistent mapping relationship between electrical variables, control variables and structural variables is established to form a unified differential algebraic equation model that includes the new energy access unit and the main power grid model, thus obtaining the panoramic model.

3. The method for dynamic evaluation of grid stability for new energy access according to claim 2, characterized in that, When performing scale decomposition on the panoramic model based on the singular perturbation method, the following is included: The system state variables in the panoramic model are divided into slow variables, fast variables, and disturbance variables according to their dynamic response characteristics. The slow variables include rotor angle, angular velocity, and frequency; the fast variables include inverter voltage, current control loop output, and phase-locked loop state; and the disturbance variables include wind speed, irradiance, and load power changes. The differential algebraic equation model is transformed into a singular perturbation form: ， ； Where ε is the small parameter for scale separation, X represents the vector of slow variables, Y represents the vector of fast variables, and Z represents the vector of perturbation variables. This represents the derivative of the slow variable with respect to time. This represents the derivative of a fast variable with respect to time. Represents the dynamic function of a slow system. This represents the dynamic function of a fast system.

4. The method for dynamic evaluation of grid stability for new energy access according to claim 3, characterized in that, When performing small-perturbation linearization on the fast system, constructing the Jacobian matrix, and determining the stability of the fast system, the following steps are included: Select the equilibrium operating point of the fast system in the current operating state, and perform a first-order Taylor expansion of the nonlinear state equation of the fast system with the equilibrium operating point as the center. Constructing the incremental equations Δ of the fast system =A·ΔY, where A represents the Jacobian matrix of the partial derivative of the fast variable with respect to itself. Solve for all eigenvalues ​​of the Jacobian matrix A. If the real parts of all eigenvalues ​​of the matrix are negative, then the fast system is determined to be in a stable state under the current disturbance; otherwise, it is unstable.

5. The method for dynamic evaluation of grid stability for new energy access according to claim 4, characterized in that, When processing the slow system based on the equal area rule, calculating the fault acceleration area and the maximum deceleration area, and determining the stability of the slow system, the following steps are included: The stable equilibrium point δ0 of the slow system before the fault, the first stable equilibrium point δs after the fault is cleared, and the unstable equilibrium point δu after the fault are obtained, where the equilibrium point is a specific value of the rotor angle of the synchronous generator. Calculate the acceleration area during the fault, which is the area formed by integrating the mechanical power and the electromagnetic power after the fault within the interval [δ0, δu]. Calculate the maximum deceleration area, which is the area formed by integrating the mechanical power and the electromagnetic power after the fault within the interval [δs,δu]. The acceleration area is compared with the maximum deceleration area. If the acceleration area is smaller than the maximum deceleration area, the slow system is determined to be in a stable state; otherwise, it is considered unstable.

6. The method for dynamic evaluation of grid stability for new energy access according to claim 5, characterized in that, The stability of the ultra-slow perturbation system is determined by analyzing the changing trend of the perturbation variable within the prediction period using a sliding window method, and by combining this with the perturbation influence sensitivity matrix to determine whether a critical transition of the system state has been triggered. A sliding window with a time length of 30 to 300 seconds is set, and the time series data of the disturbance variable is extracted within the sliding window; A time series prediction model is constructed based on the historical data of the disturbance variable to obtain the predicted change trend of the disturbance variable in future periods and extract the prediction slope, amplitude and volatility index. The sensitivity matrix to disturbance effects is obtained based on system identification; Multiply the disturbance prediction trend with the sensitivity matrix to obtain the predicted system state shift caused by future disturbances; If the predicted system state offset exceeds the stability safety boundary threshold, the ultra-slow disturbance system is determined to be unstable; otherwise, it is in a stable state.

7. The method for dynamic evaluation of grid stability for new energy access according to claim 6, characterized in that, When using graph neural network-assisted models to assist in determining whether there are potential structural instabilities in a system, the following are included: Construct an input graph with the power grid topology as the graph, where the nodes of the graph represent generation units, load nodes or converter devices, and the edges represent bus connection relationships and line impedances; Construct a feature vector for each node in the graph. The feature vector includes the voltage amplitude, phase angle, active and reactive power injected into the node, frequency offset, connection type and control mode of the node. Based on the graph neural network model, the directed propagation and embedding update of features between nodes are performed through graph convolution operations, capturing the potential unstable structure in the topology, such as weak connections, single-point fault amplification paths, and low-redundancy distribution modes. Output node-level stability scores, identify nodes with scores below the score threshold as key nodes with potential structural instability factors, and output structurally weak subgraphs.

8. The method for dynamic evaluation of grid stability for new energy access according to claim 7, characterized in that, The graph neural network model, during its training and application, includes: Training samples were constructed using historical disturbance scenarios and fault records. The sample labels were whether the system experienced voltage collapse, frequency oscillation, node islanding, or tie line overload after the disturbance. The graph attention network structure in graph neural networks is adopted, and a node attention mechanism is added during the feature propagation process. The graph neural network model is trained through supervised learning. In real-time evaluation, the current system state and topology information are input into a trained graph neural network to obtain auxiliary judgment results.

9. The method for dynamic evaluation of grid stability for new energy access according to claim 8, characterized in that, When assessing whether the power grid of the current renewable energy access system is in a stable state based on the fast system stability, slow system stability, ultra-slow disturbance system stability, and auxiliary judgment results, including... The stability determination results output by the fast system, slow system and ultra-slow disturbance system are numerically encoded, wherein the stable state is set to +1 and the unstable state is set to -1. The stability scores of all key nodes output by the graph neural network are normalized. The normalized scores are then fused with the encoding results to construct a comprehensive stability index. The comprehensive stability index is as follows: S=α·Sf+β·Ss+γ·Sus+δ·(1-R); Where S represents the comprehensive stability index, Sf, Ss, and Sus represent the stability codes of the fast system, slow system, and ultra-slow perturbation system, respectively, R represents the normalized score, and the value range of R is (0, 1], and α, β, γ, and δ represent the weight coefficients. Determine the stability threshold T. If S≥T, then the new energy access system is determined to be in a stable state. If S≤T, then the new energy access system is determined to be in an unstable state; Otherwise, it is determined to be a critical state; Output stability assessment results and mark the locations of unstable links or structural risks in each subsystem.

10. A dynamic power grid stability assessment system for new energy access, used to apply the dynamic power grid stability assessment method for new energy access as described in any one of claims 1-9, characterized in that, include: The construction unit is configured to construct a dynamic power grid model containing at least one new energy access unit and couple it with the main power grid model to obtain a panoramic model. The new energy access unit includes a photovoltaic power generation unit, a wind power generation unit, an energy storage device, or an electric vehicle charging and discharging unit. The decomposition unit is configured to perform scale decomposition on the panoramic model based on the singular perturbation method, constructing a slow system dominated by synchronous generators, a fast system dominated by power electronic equipment, and a perturbation-driven ultra-slow perturbation system. The processing unit is configured to perform small-perturbation linearization on the fast system, construct the Jacobian matrix, and determine the stability of the fast system. The slow system is processed based on the equal area rule, the fault acceleration area and the maximum deceleration area are calculated, and the stability of the slow system is determined. The changing trend of the disturbance variable in the ultra-slow disturbance system within the prediction period is analyzed by using a sliding window, and the system stability is determined by combining the disturbance influence sensitivity matrix to determine whether a critical transition of the system state is triggered. The judgment unit is configured to make an auxiliary judgment on whether there are potential structural instability factors in the system based on the current system state, power flow distribution and topology information, using a graph neural network-assisted model. The evaluation unit is configured to evaluate whether the power grid of the current new energy access system is in a stable state based on the fast system stability, slow system stability, ultra-slow disturbance system stability and auxiliary judgment results, and obtain an evaluation conclusion.

Images