A method for early identification and adaptive suppression of post-contingency oscillation resonance based on physical constraints and hierarchical reinforcement

Abstract

The application provides a method for early identification and adaptive suppression of post-connection-variation oscillation resonance based on physical constraints and layered reinforcement, relates to the technical field of power system control, and obtains transient electrical quantity data to extract a high-dimensional feature matrix; inputs a precursor identification model embedded with a physical constraint loss term, and outputs a risk probability and an evidence chain; splices a state space vector to input a layered strategy model, reduces dimension, maps to macro discrete action elements and micro continuous adjustment parameters to generate an action sequence; constructs a safety projection constraint based on a power grid physical topology to generate a final safety action sequence and execute it; extracts an index set within an evaluation window, and triggers a state machine jump when it does not converge. The application overcomes the divergence defect of reinforcement learning control in a long-time-delay power grid, realizes early adaptive suppression of oscillation, and guarantees control convergence and equipment limit safety.

Description

Technical Field

[0001] This invention relates to the field of power system control technology, and in particular to a method for early identification and adaptive suppression of post-connection oscillation resonance based on physical constraints and hierarchical reinforcement. Background Technology

[0002] In modern power grids with a high proportion of renewable energy integration, interconnecting transformers are frequently used for power transfer between different busbars, feeder sections, or power sources. In actual operation, low-frequency power oscillations or harmonic amplification often occur in the system within seconds to tens of seconds after the interconnecting switch is closed. Existing relay protection and interconnecting control largely rely on fixed electrical quantity thresholds, which are post-event triggering mechanisms. This static threshold-based approach struggles to accurately distinguish between normal, brief power fluctuations and true signs of instability in the early stages, easily leading to control lag or equipment malfunction.

[0003] To overcome the limitations of traditional control methods, existing technologies, such as "Mitigating sub-synchronous oscillation using intelligent damping control of DFIG based on improved TD3 algorithm with knowledge fusion," attempt to introduce deep reinforcement learning models to generate continuous damping control signals online. However, in modern power grids with massive and high-proportion renewable energy integration, the oscillations caused by interconnection are essentially instability processes of high-dimensional nonlinear time-varying dynamic systems with severe time delays. The aforementioned existing technologies directly force reinforcement learning models to perform trial and error optimization in a vast continuous action space. Due to the physical lag of the power grid's dynamic response, an inappropriate control command (such as prematurely cutting off reactive power compensation) often only evolves into a voltage drop or system disconnection several seconds later. This misalignment of control feedback in the time dimension leads to the extreme problem of sparsity of reward signals when reinforcement learning evaluates the value of actions. The lack of dense and effective feedback in the real physical environment easily leads to training failure to converge or the output of dangerously divergent control commands.

[0004] In transient power grid environments with long time delays and high-dimensional nonlinear coupling, reinforcement learning control based on continuous action space trial and error suffers from system divergence due to the severe lag and sparsity of reward signals. Since this optimization mechanism, which directly maps continuous control quantities, cannot accurately reflect the long-term physical boundary of the control action's effect on the grid's damping characteristics, it introduces evaluation failures due to feedback gaps during strategy evolution, causing the control system to deviate from its millisecond-level adaptive suppression real-time stable state. Summary of the Invention

[0005] To address the shortcomings of existing technologies, this invention aims to provide a method for early identification and adaptive suppression of post-connection oscillation resonance based on physical constraints and hierarchical reinforcement. This method utilizes a hybrid action space constraint mechanism combining action elementization and a multi-level template library to reduce the dimensionality of complex continuous adjustment processes and map them as discrete action sequence combinations derived from expert knowledge. Simultaneously, it introduces dense auxiliary reward signals based on physical indicators into the model evaluation system. This limits the exploration space of artificial intelligence algorithms, completes the physical feedback link, and effectively overcomes the convergence failure and strategy oscillation problems caused by reward sparsity in transient power grid control models.

[0006] To solve the above-mentioned technical problems, the technical solution adopted by the present invention is as follows: The present invention provides a method for early identification and adaptive suppression of post-connection oscillation resonance based on physical constraints and hierarchical reinforcement, comprising the following steps: S1. Obtain transient electrical quantity data after connecting the transformer and construct the original time series matrix. Extracting the original time series matrix Generate a high-dimensional feature matrix ; to high-dimensional feature matrix Input to a loss term with embedded energy dissipation physical constraints The precursor recognition model outputs precursor categories. Risk Probability And generate an interpretable chain of evidence. ; S2, Concatenate high-dimensional feature matrices Precursor categories Risk probability With an explainable chain of evidence Generate global state space vector ;Transfer the global state space vector The input hierarchical strategy model is dimensionality-reduced and mapped to macroscopic discrete action elements. With micro-continuous adjustment parameters To generate the original mixed action sequence Based on the power grid physical topology, a safe projection constraint is constructed to transform the original hybrid action sequence. Projecting the sequence of actions into the safe and feasible domain generates the final safe action sequence. ; S3. During the calibrated control evaluation window Internal extraction of multidimensional evaluation index set In the multidimensional evaluation index set When the power grid fails to converge, the hierarchical strategy model is triggered to execute state machine transitions to switch macroscopic discrete action elements. and return to generate the final security action sequence. Steps; S4, at the edge intelligent unit With station controller Physical isolation is maintained between them; station-side controller Final safety action sequence Includes macroscopic discrete motion elements The mapping is converted into physical execution instructions and then issued.

[0007] In the preferred embodiment, step S1 includes the following sub-steps: S11, Upon receiving the successful closing signal After starting the stable observation time window Through the substation's underlying communication protocol Acquire digital electrical messages and construct a sequence containing instantaneous three-phase voltage values. Three-phase current instantaneous value sequence and frequency sampling sequence The original time series matrix For the original time series matrix Perform phase calculation to generate a three-phase voltage waveform sequence With three-phase current waveform sequence Active power sequence is generated using instantaneous reactive power theory. With reactive power sequence Combined with frequency sampling sequence The difference operation generates a frequency change rate sequence. Aligning three-phase voltage waveform sequences Three-phase current waveform sequence Active power sequence Reactive power sequence With frequency change rate sequence Generate the basic time series matrix ; in low-frequency characteristic sliding window Internal active power sequence Calculate the active power growth rate by performing linear regression and Hilbert transform Damping ratio fitted with envelope ; in high-frequency feature sliding window Internal three-phase voltage waveform sequence Perform Fast Fourier Transform Calculate the energy change rate of the target frequency band With the increasing trend of total harmonic distortion splicing active power growth rate Envelope fitting damping ratio Target frequency band energy change rate With the increasing trend of total harmonic distortion Generate high-dimensional feature matrix ; High-dimensional feature matrix Including active power growth rate Envelope fitting damping ratio Target frequency band energy change rate With the increasing trend of total harmonic distortion ; S12, Transform the high-dimensional feature matrix The input is a physical information neural network consisting of a one-dimensional time-dilated convolution operator and a fully connected mapping layer. Through physical information neural networks Output layer output precursor category Risk Probability Extracting the basic time series matrix Combined with the reference value of the inherent damping coefficient Utilizing the system's transient total energy Derivatives and phase difference sequences Algebraic summation of the squares of derivatives within a stable observation window The energy dissipation physical constraint loss term is calculated by summing the squared residuals at discrete time steps. ; S13. Traverse feature subsets that do not contain the target feature dimension. Computational physical information neural network In the input feature subset The difference between the biased risk probability and the value after incorporating the target feature dimension is used to generate an attribution contribution score matrix by accumulating the factorial weights. Extract the attribution contribution score matrix. The original physical observation values ​​corresponding to the core features are compared with preset safety thresholds using Boolean logic, and the comparison results are encapsulated to generate an interpretable chain of evidence. .

[0008] In the preferred embodiment, step S2 includes the following sub-steps: S21. Use one-hot coding algorithm to classify precursors. Convert to numerical category vector Using word embedding operators to interpretable chains of evidence Mapped to feature vectors of the evidence chain ; Concatenate high-dimensional feature matrices Numerical category vectors Risk probability With the feature vector of the chain of evidence Generate global state space vector ;Transfer the global state space vector Input hierarchical reinforcement learning architecture Utilizing a low-level discrete action decision network Feedforward fully connected layer on global state space vector Perform linear transformations and nonlinear activations to output a state-action value vector. Extract the state-action value vector The index corresponding to the maximum value determines the macroscopic discrete action element. ;Transfer the global state space vector Macroscopic discrete motion elements splicing input top-level continuous parameter fine-tuning network The policy network submodule outputs a mean vector and a standard deviation vector. Based on these mean and standard deviation vectors, reparameterized sampling is performed, and the output is controlled by the time window length. Active power limit upper limit With reactive power support compensation coefficient Microscopic continuous adjustment parameters ; Combine macroscopic discrete motion elements With micro-continuous adjustment parameters Generate the original mixed action sequence ; S22. Calculate the envelope fitting damping ratio observed at the current time. The difference between the historical damping ratio and the previous action is used to establish a direct proportional mapping relationship to generate a positive reward for damping improvement. Compare the upward trend of total harmonic distortion (THD). Actual voltage amplitude of the bus Compared with the system safety operation benchmark matrix, in the upward trend of total harmonic distortion (THD) Or the actual voltage amplitude of the busbar When the limit is exceeded, a preset maximum negative value is assigned to generate a safety boundary violation penalty item. Extract the original mixed action sequence The data includes the types of actuators and the adjustment range, quantifying the physical costs of equipment wear and tear, and the economic losses from active power limiting, generating control action cost items. ; Improve weighting coefficients using damping Margin penalty weighting coefficient With cost weight coefficient Positive incentives for improved damping Penalties for exceeding safety boundaries Cost of control actions Perform weighted fusion and calculate the comprehensive reward function. To drive the underlying discrete action decision network With top-level continuous parameter fine-tuning network Update the weight parameters.

[0009] In the preferred embodiment, step S2 further includes the following sub-steps: S23. Obtain the static topology parameters of the distribution network and construct the distribution network connection matrix. Node association matrix Resistance matrix With reactance matrix Using the node association matrix Derivation and calculation of the power transfer matrix Extract the original mixed action sequence Included active power regulation components With reactive power regulation component Using a quadratic programming solver Construct an optimization function with the objective of minimizing the distance. Extract the lower limit critical value of the busbar safe voltage. upper limit critical value of busbar safe voltage Relaxation compensation variables Reference voltage of the reference node The inverse of the node association matrix Initial net active power of nodes Initial net reactive power of the node By combining the power flow linearization model of the distribution network, a set of linear equations with upper and lower voltage limits is constructed to provide constraints; a quadratic programming solver is then used. The interior-point method is used for iterative solution in the original mixed action sequence. When the location is outside the safe feasible region defined by the linear equations of the upper and lower voltage limits, search for the distance from the original hybrid action sequence on the boundary of the safe feasible region. The nearest projection point, truncating the dangerous component to generate a component containing the final active power regulation component. With the final reactive power regulation component Final safety action sequence ; S24. Calculate the attribution contribution score matrix. Information entropy quantification generates overall decision confidence. Compare the overall decision confidence level With a preset high security threshold in non-volatile memory With low security threshold ; In overall decision confidence Greater than or equal to the high security threshold At that time, extract the micro-continuous adjustment parameters. Included control time window length Generate control evaluation window And issue the final safety action sequence ; In overall decision confidence Given the low security threshold With high security judgment threshold In between, the confidence degradation protocol is activated to eliminate the final security action sequence. The inclusion of rigid motion elements generates limited inhibition of motion and extends the control evaluation window. ; In overall decision confidence Strictly below the low security threshold At that time, the circuit breaker is triggered to transfer control authority to the local stability control device and to limit the ramp rate, without generating a control evaluation window. .

[0010] In the preferred embodiment, in step S3, within the calibrated control evaluation time window... Internal extraction of multidimensional evaluation index set This includes the following sub-steps: The three-phase voltage waveform sequence is continuously subscribed to and read through the substation process layer network. With three-phase current waveform sequence The instantaneous power calculation operator is used based on the three-phase voltage waveform sequence. With three-phase current waveform sequence Real-time solution control evaluation window Active power time series within ; Active power time series location using peak extraction algorithm At the start of the time window and the finish line Nearby envelope peak values, extract control evaluation window Peak active power envelope at the end time With control evaluation window Peak active power envelope at the start time Using mathematical calculation equations Calculate the amplitude decay rate used to characterize the degree of oscillation energy decay. ; Calculate the real-time envelope fitting damping ratio within the current time window using the damping identification operator. The real-time envelope fitting damping ratio Damping ratio fitted to the initial envelope before the action is executed The damping evolution difference is obtained by performing a subtraction operation. ; The real-time operating margin is obtained by calculating the physical distance between the current state of the system and the preset voltage collapse boundary based on the real-time collected bus voltage amplitude. The amplitude attenuation rate Damping evolution difference With real-time runtime margin Perform feature encapsulation to generate a multi-dimensional evaluation index set .

[0011] In the preferred scheme, in step S3, the multidimensional evaluation index set... When the power grid fails to converge, the hierarchical strategy model is triggered to execute state machine transitions to switch macroscopic discrete action elements. and return to generate the final security action sequence. The steps also include the following sub-steps: Read the global state space vector , global state space vector With multidimensional evaluation index set Perform tensor concatenation along the time dimension to generate an evolutionary state vector. ; Evolutionary state vector Input hierarchical reinforcement learning architecture Utilizing a low-level discrete action decision network Based on the comprehensive reward function Assess the return on investment; Determine the difference in damping evolution Is it greater than 0 and has an amplitude decay rate? Is it less than 1? In the difference of damping evolution Greater than 0 and amplitude decay rate When the condition of less than 1 is met, the comprehensive reward function Output positive gain, utilizing the underlying discrete action decision network Maintain current macroscopic discrete action elements The goal is to fine-tune the network using continuous parameters at the top level. Under the current template, fine-tune the micro-continuous adjustment parameters based on positive feedback. ; In the difference of damping evolution When the condition of being less than or equal to 0 is not met, the comprehensive reward function... This generates a huge negative penalty gradient, forcibly triggering the underlying discrete action decision network. The strategy index jump involves leaping from the expert template library to select novel macroscopic discrete action elements with stronger intervention power based on the updated state value estimate. Based on novel macroscopic discrete motion elements Re-trigger the quadratic programming projection to generate the final safe action sequence for the next iteration. .

[0012] In the preferred embodiment, step S3 further includes the following sub-steps: Extracting a multi-dimensional evaluation index set using a low-level hardware comparator Included real-time runtime margin Real-time runtime margin With the limit of the crash margin threshold in the register Perform continuous comparisons; in real-time runtime margin Always less than the ultimate collapse margin threshold or hierarchical reinforcement learning architecture Composite reward function for three consecutive evaluations When the behavior is monotonically decreasing, issuing a non-maskable circuit breaker command triggers the ultimate circuit breaker operation. The underlying circuit breaker is tripped to exit the interconnection state; Clock alignment is performed after the transient process ends, and the high-dimensional feature matrix is... Physical information neural network Predicted precursor categories Explanable chain of evidence Hierarchical reinforcement learning architecture Final safety action sequence issued in each iteration With the corresponding multidimensional evaluation index set Encapsulate and generate full event data tuples ; All event data tuples Uploaded to centralized retraining platform The offline experience replay pool utilizes the offline gradient descent algorithm to analyze physical information neural networks. With hierarchical reinforcement learning architecture Perform parameter tuning.

[0013] In the preferred embodiment, step S4 includes the following sub-steps: Obtain the final safety action sequence Includes macroscopic discrete motion elements Using the station-side controller Extracting the physical resource mapping matrix inside a non-volatile flash memory chip. According to macroscopic discrete motion elements In the physical resource mapping matrix Perform row address lookup and column attribute matching to output the device operation vector. And issue instructions; Utilizing the underlying photovoltaic converter control unit Obtain the apparent power capacity of the converter hard-coded in read-only memory AC voltage transformers collect actual bus voltage amplitude. ; Extracting micro-continuous adjustment parameters Included reactive power support compensation coefficient With active power limit upper limit Combined with reactive power support compensation coefficient Actual voltage amplitude of the bus The required reactive power for deviation calculation ; Extract the upper limit of active power. Required reactive power With converter apparent power capacity Active and reactive coupled boundary constraint equations When the active and reactive power coupling boundary constraint equations hold, the upper limit of active power is used. With required reactive power Generate quadrature-axis and direct-axis reference currents and synthesize a space vector pulse width modulation signal; When the active and reactive power coupling boundary constraint equations do not hold, the dynamic reactive power priority compensation strategy is activated. Maintain the required reactive power With the numerical value unchanged, a dynamic adaptive limiting equation is used. Truncated, dimension-reduced, and amplitude-limited calculations are used to output the actual active power output command. Output commands using actual active power With required reactive power The quadrature-axis and direct-axis reference currents are generated and synthesized into a space vector pulse width modulation signal.

[0014] In the preferred embodiment, step S4 further includes the following sub-steps: Utilizing edge intelligent units During the control evaluation window Intra-cycle to station controller Send heartbeat message sequence from the register address segment ; Using station-side controller The direct memory access controller receives heartbeat message sequences The watchdog circuit is accessed via low-level assembly-level hardware addressing instructions. The reset pin sends a reset level, activating the full-link watchdog circuit. Internal countdown decrement register Restored to full value; Extract countdown decrement register The value and the timeout overflow threshold In the countdown decrement register The value decreases to zero and meets the timeout overflow threshold. At that time, a full-link watchdog circuit is used. Send a non-maskable interrupt signal to the non-maskable pin of the main central processing unit. ; Utilizing the main central processing unit to receive non-maskable interrupt signals Control the program counter pointer to jump to the memory degradation interlocking strategy The physical address range; Extracting downgrade interlocking strategies Disable specific isolated buses The receiving physical port sends a conservative zero-power ramp command to the converter. Disconnect the incremental active reference current loop of the converter.

[0015] In a preferred embodiment, the present invention also provides a computer device, the computer device including at least one processor coupled to at least one memory, the memory storing at least one computer program or instruction, wherein the computer program or instruction is loaded and executed by the processor to implement the steps of the method for early identification and adaptive suppression of post-connection oscillation resonance based on physical constraints and hierarchical reinforcement as described in any of the preceding embodiments.

[0016] In a preferred embodiment, the present invention further provides a computer-readable storage medium, characterized in that the computer-readable storage medium stores a computer program or instructions, which, when executed by a processor, implement the steps of the method for early identification and adaptive suppression of post-connection oscillation resonance based on physical constraints and hierarchical reinforcement as described in any of the preceding claims.

[0017] In a preferred embodiment, the present invention further provides a computer program product, including a computer program or instructions, characterized in that, when the computer program or instructions are executed by a processor, they implement the steps of the method for early identification and adaptive suppression of post-connection oscillation resonance based on physical constraints and hierarchical reinforcement as described in any of the preceding claims.

[0018] This invention provides a method for early identification and adaptive suppression of post-connection oscillation resonance based on physical constraints and hierarchical reinforcement. Through the coordination between the aforementioned structures, it offers the following advantages compared to existing methods: First, based on the extraction of physical features across multiple time scales, the energy dissipation differential equation of the power system is embedded as a constraint into the underlying mapping logic of the neural network, and an attribution explanation engine based on Shapley values ​​is introduced simultaneously. This enables accurate identification of true instability precursors in the complex transient fluctuations at the initial closing stage, outputting prediction results and structured evidence chains supported by physical dynamics, significantly improving the early agility of precursor identification and the reliability of engineering decisions.

[0019] Secondly, it addresses the technical problem of conventional reinforcement learning algorithms easily falling into policy divergence when facing transient power grids with long time delays and high-dimensional mixed action spaces, due to the severe lag and sparsity of reward signals. By adopting a hierarchical reinforcement learning architecture, the complex inhibition control is decoupled into a series optimization process of macroscopic discrete mode decision-making and microscopic continuous parameter fine-tuning, and a comprehensive evaluation system integrating real-time physical boundaries and damped improvement gradients is constructed. This mechanism of hierarchical dimensionality reduction and dense physical feedback effectively reduces the disordered exploration space of intelligent algorithms and significantly improves the convergence speed and environmental adaptability of action policy models in transient environments.

[0020] Third, a quadratic programming expert safety layer based on the power flow linearization equation of the distribution network is connected in series between the intelligent strategy output end and the physical execution end. When the tentative actions generated by the intelligent model deviate from the safe feasible region, this safety layer can reproject them into the required physical boundary with minimal space cost. In this way, highly uncertain action probes are effectively transformed into safety instructions with deterministic electrical bottom lines, ensuring the stability of bus voltage during transient evolution.

[0021] Fourth, by relying on multi-dimensional physical feedback indicators extracted in real time within the control evaluation window, the underlying state machine of the reinforcement learning architecture is driven to perform dynamic elastic transitions. This mechanism enables the system to nonlinearly reconstruct the intervention level and suppression sequence based on the strengths and weaknesses of the power grid's physical resistance characteristics, thereby giving the control system resilience in extreme environments and expanding the boundaries of technical protection from the algorithmic mechanism level.

[0022] Fifth, by introducing adaptive limiting of electromagnetic capabilities controlled by the apparent power circle of the converter, and constructing a hard real-time takeover defense line based on physical isolation of computing power and a full-link hardware watchdog, the system can instantly strip the intelligent algorithm of its control and fall back to a conservative stability protection state within a limited response time through a pure hardware clock overflow interrupt when encountering abnormal communication interruption or full computing power load. This meets the power grid's physical security access standards for grid-connected intelligent devices. Attached Figure Description

[0023] The present invention will be further described below with reference to the accompanying drawings and embodiments: Figure 1 This is a main view diagram of the process structure of this invention; Figure 2 This is a schematic diagram of the structure of the computer device of the present invention. Detailed Implementation

[0024] To better understand the purpose, system architecture, and functional implementation of this embodiment, the embodiments and features described herein can be combined with each other without conflict. The exemplary embodiments disclosed herein will be described below with reference to the accompanying drawings, including specific technical details disclosed to aid understanding; however, these details should be considered exemplary rather than restrictive. Therefore, those skilled in the art should understand that various improvements and adjustments can be made to the embodiments described herein without departing from the scope and core ideas of the invention. Similarly, for clarity, detailed descriptions of well-known technologies, functions, and structures are omitted in the following description.

[0025] Example 1: like Figure 1 As shown, this embodiment provides a method for early identification and adaptive suppression of post-connection oscillation resonance based on physical constraints and hierarchical reinforcement. The specific steps are as follows: S1: Precursor identification stage of high-speed sampling and physical information fusion after communication This step involves utilizing the critical time window after the connection transformer is closed to acquire transient electrical data through high-frequency sampling, and combining the physical mechanisms of the power system with artificial intelligence algorithms to achieve accurate identification and attribution explanation of oscillation and resonance precursors.

[0026] S11: Closing State Triggering and Multi-Timescale Feature Deep Processing After the connecting switch of the transformer is closed, the station-end main control unit monitors and receives the closing success signal in real time. The closing success signal Serving as the logical starting point of the entire suppression and control system, it is used to synchronously trigger and initiate the stable observation time window. This stable observation time window The initial duration is preferably set to 10 seconds to fully cover the entire transient evolution process during the initial closing phase.

[0027] During the stable observation window During operation, the edge computing unit communicates via the substation's underlying communication protocol. The edge computing unit subscribes to and acquires digital electrical messages from the underlying merging unit at a sampling frequency of no less than 10kHz. Specifically, the edge computing unit continuously reads and caches electrical quantity data from both sides of the transformer using this protocol to construct the original timing matrix. The original time series matrix Includes a sequence of instantaneous three-phase voltage values Three-phase current instantaneous value sequence and frequency sampling sequence .

[0028] Subsequently, the edge computing unit processed the original time series matrix. Preprocessing and physical quantity transformation are performed to generate the basic time series matrix. .

[0029] Specifically, the steps for performing basic time series data processing are as follows: First, the edge computing unit calls the phase calculation operator to process the instantaneous three-phase voltage value sequence. With the instantaneous value sequence of three-phase current Converted into the corresponding three-phase voltage waveform sequence With three-phase current waveform sequence And simultaneously calculate the active power sequence. With reactive power sequence .

[0030] Next, the frequency sampling sequence Perform first-order finite difference operations to obtain the frequency change rate sequence. It is used to reflect the intensity of transient fluctuations in the system frequency.

[0031] Subsequently, the above sequences are aligned according to timestamps and encapsulated into a basic time series matrix. .

[0032] Edge computing units in the basic timing matrix Based on this, a secondary feature extraction is performed using a multi-timescale sliding window mechanism to generate a high-dimensional feature matrix. This high-dimensional feature matrix This is used to provide physically distinctive feature vectors for subsequent AI inference.

[0033] Specifically, the steps for performing multi-dimensional physical feature extraction are as follows: First, the edge computing unit opens a sliding window for low-frequency features. This low-frequency characteristic sliding window The step size is preferably 0.1 seconds, and the window length is limited to cover at least 0.5 cycles of the inherent low-frequency oscillation of the power grid. This is achieved through a sliding window for low-frequency characteristics. Within, edge computing units process active power sequences. Perform linear regression slope calculation to determine the active power growth rate. Simultaneously, the active power sequence is extracted using the Hilbert transform. The oscillation envelope is used to calculate the envelope fitting damping ratio, which reflects the stability margin of the system, through an exponential fitting method. .

[0034] Next, the edge computing units open high-frequency feature sliding windows in parallel. This high-frequency characteristic sliding window The preferred step size is 0.02 seconds to capture millisecond-level harmonic distortion. This is applied to each high-frequency feature sliding window. Within, the edge computing unit processes the three-phase voltage waveform sequence. Perform Fast Fourier Transform The amplitude distribution of each harmonic is obtained.

[0035] Subsequently, the rate of change of the total energy within the target frequency band over time is calculated to obtain the energy change rate of the target frequency band. Simultaneously, the total harmonic distortion (THD) rate of the current window is calculated and differs from that of the previous window, outputting the upward trend of the THD rate. .

[0036] Finally, the active power growth rate Envelope fitting damping ratio Target frequency band energy change rate With the increasing trend of total harmonic distortion The data is then normalized and concatenated to generate a high-dimensional feature matrix. Its mathematical expression is as follows: ; in, It is a high-dimensional feature matrix that aggregates multi-dimensional physical operation characteristics; The active power growth rate is used to reflect the degree of power oscillation divergence. The envelope fitting damping ratio is used to reflect the system's recovery capability after being disturbed; The target frequency band energy change rate is used to reflect the rate of energy rise in a specific resonant frequency band. The total harmonic distortion rate (THD) is used to reflect the increasing trend of voltage distortion.

[0037] S12: Precursor Classification and Risk Assessment Based on Physical Information Neural Networks Physical information neural networks are deployed and run inside edge computing units. Used to determine the characteristics of a high-dimensional feature matrix Achieve accurate mapping of precursor categories. This physical information neural network. By embedding the physical evolution equations of the power system into the loss function, the "logical illusion" defect that may occur in the transient analysis of conventional deep learning models, which violates the physical mechanism, is overcome.

[0038] Specifically, the steps for performing precursor classification and physical constraint calculation are as follows: First, the high-dimensional feature matrix Input to physical information neural network The input layer is composed of multiple layers of temporal convolutional operators stacked with fully connected mapping layers.

[0039] Next, the network generates intermediate state feature tensors through nonlinear mapping. The predicted precursor category is ultimately output by the Softmax output layer. Corresponding risk probability This category of precursors It includes discrete tags such as normal brief oscillations, precursors to low-frequency oscillations, and precursors to resonant amplification.

[0040] Subsequently, to ensure the output conforms to physical principles, the edge computing unit invokes the built-in physical constraint verification module in real time. This module verifies the physical constraints based on the high-dimensional feature matrix. The evolution trend of electrical quantities in the system is analyzed to calculate the physical constraint residuals corresponding to the current system state. Its core is based on the system energy dissipation differential equation, summing the squared residuals step-by-step over the entire stable observation window to construct the energy dissipation physical constraint loss term. The formula is as follows: ; in, This is the energy dissipation physical constraint loss term, used to measure the overall deviation between the predicted state and the system's energy conservation law; A discrete-time step summation operator covering the entire stable observation time window; For discrete time step variables; The initial time step triggered by the successful closing signal; To stabilize the observation time window; To be based on the high-dimensional feature matrix Real-time estimation of the system's transient total energy; The reference value for the inherent damping coefficient is determined based on the known parameters of the system. This represents the instantaneous value of the phase difference sequence at the current time step.

[0041] Finally, physical information neural networks The parameter updates are controlled by the total loss function. .

[0042] ; in, The total loss function used for network training; Precursor category The corresponding classification cross-entropy loss; This is the physical penalty weight coefficient, used to adjust the strength of the correction of physical laws on the model output.

[0043] Through the aforementioned physical information fusion mechanism, the physical information neural network Output risk probability It not only has statistical significance, but also has rigorous power grid dynamics support, thus providing a reliable input for subsequent suppression decisions.

[0044] S13: Generation of Interpretable Evidence Chains Based on Shapley Values To provide operators with transparent decision-making support and to support targeted adjustments to subsequent action strategies, edge computing units in the precursor category... Once determined, immediately initiate the attribution explanation process based on Shapley values ​​to generate an explainable chain of evidence. .

[0045] Specifically, the steps for performing an explainable attribution analysis are as follows: First, the edge computing unit will use the high-dimensional feature matrix Each feature component in the model is treated as a player in the game, and the probability of each feature pair against risk is calculated using a sampling estimation method. Increased marginal contribution.

[0046] Next, an attribution contribution score matrix is ​​generated. Its core calculation is shown in the following formula: ; in, For the first Attribution contribution score for each feature dimension; A summation operator with strict bounds, used to iterate over all summation operators that do not contain the first... A subset of features in each feature dimension; High-dimensional feature matrix The set of all feature dimensions in; For set difference operators; For only containing the first A single-element set of each feature dimension; For any characteristic subset belonging to the difference set of the above sets; This represents the number of dimensions contained in the feature subset. It is a factorial operator used to calculate the permutation weights under specific feature combinations. This represents the total number of feature dimensions. For set union operators; To incorporate the first feature subset After each feature dimension, the physical information neural network Output update risk probability; For physical information neural networks that only input this subset of features The output bias risk probability.

[0047] Subsequently, the edge computing unit scored the attribution contribution matrix. The scores are sorted by absolute value. The top-K core features with the highest scores are extracted, and their corresponding original observations are extracted from the high-dimensional feature matrix. It was retrieved from [the database].

[0048] Finally, the observed values ​​are logically compared with the system's preset safety thresholds and encapsulated into a structured, interpretable chain of evidence. .

[0049] For example, if a pre-resonance precursor is detected, it can explain the chain of evidence. Automatically generated text description: Due to the energy change rate of the target frequency band The observed value exceeds the threshold and its Shapley score The percentage reached 60%, indicating a risk of resonance amplification.

[0050] The risk probability generated by the aforementioned steps Precursor categories Explanable chain of evidence and the high-dimensional feature matrix as input Together, they form a diagnostic information package. This package will be stored in real time in the shared memory of the edge computing unit.

[0051] S2: Adaptive Action Policy Generation Stage Based on Hierarchical Reinforcement Learning and Power Grid Flow Quadratic Programming This stage involves receiving the multidimensional physical features and diagnostic information extracted from the preceding stage, constructing a cross-domain hybrid continuous and discrete action layered architecture, and connecting it with a quadratic programming safety layer based on power flow linearization of the distribution network to achieve safe, rapid, and adaptive generation of suppression action strategies under transient conditions after grid interconnection.

[0052] S21: Hierarchical network architecture mapping for suppressing action strategies in mixed action spaces In the transient control environment following grid interconnection, station-end actuators need to simultaneously decide on discrete control modes and continuous physical quantity adjustments. Conventional single-layer reinforcement learning algorithms suffer from a fundamental technical flaw when facing such high-dimensional hybrid action spaces: they are prone to getting trapped in local optima and failing to achieve stable convergence. To overcome this technical problem, the inhibition action strategy model within the edge computing unit is reconstructed into a hierarchical reinforcement learning architecture. .

[0053] Specifically, the steps for constructing the global state space are as follows: The edge computing unit reads the diagnostic information packet generated in the preceding stage through the internal high-speed bus and extracts the high-dimensional feature matrix from it. Precursor categories Risk probability and an explanatory chain of evidence .

[0054] Subsequently, the edge computing unit uses a one-hot coding algorithm to classify the discrete precursors. Convert to numerical category vector .

[0055] Next, the edge computing unit uses word embedding operators from natural language processing to create interpretable chains of evidence in text form. Mapped to a fixed-dimensional chain of evidence feature vector .

[0056] Finally, the edge computing unit will use the high-dimensional feature matrix Numerical category vectors Risk probability and evidence chain feature vectors Multi-channel concatenation along the feature dimension generates a global state space vector that comprehensively represents the current transient operating state of the power grid. Its mathematical state concatenation equation is as follows: ; in, The global state space vector characterizing the current transient physical operating environment and diagnostic conclusions of the power grid; This is a multi-channel feature vector concatenation operator; This is the high-dimensional feature matrix extracted in the preceding stage; It is a numerical category vector; Predicting risk probability based on early warning signs; This is the feature vector of the evidence chain.

[0057] Hierarchical reinforcement learning architecture Includes underlying discrete action decision network With top-level continuous parameter fine-tuning network Discrete Action Decision Network at the Bottom A deep Q-network structure is employed to handle discrete control variables at the macroscopic level. A top-level continuously parameter fine-tuning network is also used. A flexible actor-critic network structure is employed to handle continuous control variables at the micro level.

[0058] Specifically, the steps for performing discrete action decisions and mapping macroscopic elements are as follows: Low-level discrete action decision network The input terminal receives the global state space vector. The feedforward fully connected layers inside the network sequentially process the global state space vector. Perform linear transformations and nonlinear activations to output a state-action value vector. The state action value vector Each dimension corresponds to a macroscopic discrete action element in the system's preset expert template library. .

[0059] During the model inference phase, the edge computing unit employs a greedy strategy to extract the state-action value vector. The index corresponding to the maximum value is used to determine the current optimal macroscopic discrete action element. This macroscopic discrete actuator Preferably, it includes discrete decision instructions such as control mode switching combination number and filter parameter reconstruction level.

[0060] During the model training phase, the underlying discrete action decision network Utilizing the experience replay pool Historical transition tuple data are extracted, and the time-series difference target value is calculated based on the Bellman equation. Furthermore, by minimizing the difference between the evaluation network output and the temporal difference objective value... Mean squared error loss function between The underlying discrete action decision network is updated using gradient descent. Weight parameters. Mean squared error loss function. The formula is as follows: ; in, For the underlying discrete action decision network The mean squared error loss function; This represents the number of samples in a randomly sampled batch. This is a summation operator that iterates through all random samples within a batch. For the sample sequence index variable; In order to target the The temporal difference target value obtained by calculating the target network for each sample; For the first The global state space vector of each sample; For the first Macroscopic discrete moving elements of a sample; To evaluate the network for the first The input features of a sample are used to predict the value of a specific state action.

[0061] Specifically, the steps for performing continuous parameter fine-tuning and micro-instruction generation are as follows: Top-level continuous parameter fine-tuning network Receive global state space vector Simultaneously, receive the underlying discrete action decision network. Macroscopic discrete actuators that have been identified As a prerequisite constraint.

[0062] Top-level continuous parameter fine-tuning network It contains one policy network submodule and two parallel critic network submodules. The policy network submodule stores the global state space vector. Macroscopic discrete motion elements The concatenated vectors are used as input, and the output vectors are the mean and standard deviation vectors of a specific Gaussian distribution.

[0063] Subsequently, the edge computing unit performs reparameterized sampling based on the mean vector and standard deviation vector, and outputs micro-continuous adjustment parameters. This microscopic continuously adjustable parameter Used to define specific continuous physical quantities within a macroscopic template framework, preferably including a specific control time window length. Active power limit upper limit and reactive power support compensation coefficient .

[0064] To prevent the strategy from getting trapped in local optima, a top-level continuously parameter fine-tuning network is used. A policy entropy regularization term is introduced into the objective function. By maximizing the reward and policy entropy regularization term. The weighted sum allows the policy network to achieve a dynamic balance between exploiting known high-reward actions and exploring unknown actions. The parameter updates of the policy network submodules are influenced by the loss function. Strict control over the loss function. The formula is as follows: ; in, The loss function for the policy network submodule; This represents the number of samples in a randomly sampled batch. This is a summation operator that iterates through all random samples within a batch. For the sample sequence index variable; An adaptive temperature coefficient for controlling the intensity of exploration; It is the natural logarithm operator; For a given first The global state space vector of each sample Macroscopic discrete motion elements Under these conditions, the policy network output corresponds to the micro-continuous adjustment parameter. The probability density value; For finding the minimum value operator; and These are two parallel commentator network submodules for the first... The value assessment of the current mixed action of each sample is obtained by taking the minimum value to alleviate the overestimation problem.

[0065] Ultimately, the underlying discrete action decision network Macroscopic discrete motion element determined With top-level continuous parameter fine-tuning network Sampled micro-continuous adjustment parameters Merge the sequences to generate the complete original mixed action sequence. The original mixed action sequence This constitutes a preliminary draft of control decisions proposed by the artificial intelligence model in response to the current precursors of transient instability.

[0066] S22: Construction of a Multidimensional Constraint Reward Function to Overcome the Sparsity of Long-Latency Rewards In the transient dynamic evolution of modern power distribution networks, due to the response hysteresis of electrical components and the physical delay of topology transmission, an erroneous action command often only induces a sudden drop in bus voltage or system disconnection several seconds later. This misalignment of feedback over time scales causes conventional reinforcement learning to face the extremely severe problem of reward signal sparsity, leading to the inability of the model to converge during trial-and-error training. Therefore, a hierarchical reinforcement learning architecture is proposed. The evolutionary fine-tuning process abandons the traditional delayed trial-and-error feedback and instead adopts a comprehensive reward function that integrates real-time physical boundary-assisted evaluation. .

[0067] Specifically, the steps for extracting the multidimensional physical components of the reward function are as follows: First, the edge computing unit fits the damping ratio based on the envelope extracted in the previous stage. A damping improvement assessment is performed. The edge computing unit calculates the envelope-fitted damping ratio observed at the current moment. The difference between the damping ratio and the historical damping ratio from the previous action. If this difference is positive, the system determines that the oscillation divergence trend has been physically suppressed, and at this point, a positive reward for damping improvement is calculated and generated. Positive incentives for improved damping. The value of is strictly proportional to the difference.

[0068] Next, the edge computing unit extracts the upward trend of the total harmonic distortion rate calculated in the preceding stage. Compared with the actual voltage amplitude of the bus at the current moment Used to calculate the penalty for exceeding the safety boundary. The edge computing unit has a pre-installed system security benchmark matrix in memory, which is activated when the total harmonic distortion rate increases. The distortion threshold specified in the safe operation reference matrix or the actual voltage amplitude of the bus exceeds the value specified in the reference matrix. When the voltage exceeds the upper and lower limits of the steady-state voltage specified in the safe operation reference matrix, a penalty term for exceeding the safety boundary limit will be applied. It is assigned a preset, extremely negative value. This negative value is used to severely punish out-of-bounds behavior, thereby forcibly cutting off the exploration gradient leading to that dangerous action space during model training.

[0069] Finally, the edge computing unit extracts the original hybrid action sequence. The types and adjustment ranges of the motion elements involved are used to calculate the cost of control actions. Control action cost This is used to quantify the physical costs of equipment wear and tear and the economic losses caused by active power limiting resulting from the execution of corresponding control strategies. The greater the intensity of the intervention and the more underlying devices involved, the greater the cost of the control action. The higher the value, the better.

[0070] Subsequently, the various sub-dimensional indicators mentioned above are weighted and fused to calculate the final comprehensive reward function. The formula is as follows: ; in, For hierarchical reinforcement learning architecture The output scalar of the comprehensive reward function upon which backpropagation and value assessment depend; The damping improvement weighting coefficient is used to amplify the positive returns of damping; Positive incentives for damping improvements that reflect enhanced system stability; This is a margin penalty weighting coefficient used to ensure that the physical bottom line cannot be crossed; This refers to the penalty for exceeding the safety boundary due to triggering the rigid operating boundary of the power grid; This refers to the cost weighting coefficient used to optimize adjustment costs; Costs incurred for controlling actions.

[0071] The comprehensive reward function constructed through the above mechanism Capable of providing underlying discrete action decision networks With top-level continuous parameter fine-tuning network It provides dense auxiliary gradient feedback with clear physical mechanism guidance, which completely eliminates the policy oscillation and divergence problem caused by long time delay from the bottom layer of the algorithm.

[0072] S23: Quadratic Programming Expert Safety Layer Projection Based on Distribution Network Power Flow Linearization Specifically, the steps for performing the physical projection of the expert knowledge security layer are as follows: First, the edge computing unit pre-acquires and loads the static topology parameters of the distribution network to construct the distribution network connection matrix. The edge computing unit extracts a node association matrix reflecting the network topology based on the line connection relationships of the power distribution network. And diagonalized resistance matrices are constructed based on the actual impedance parameters of the lines. With diagonalized reactance matrix .

[0073] Next, the edge computing unit utilizes the node association matrix Derivation and calculation of the power transfer matrix used to reflect power flow characteristics Because there is a complex nonlinear implicit relationship between the square of the node voltage amplitude and the injected power in the distribution network, in order to achieve millisecond-level real-time solution, the edge computing unit adopts the distribution network power flow linearization model to ignore the quadratic term component that generates high-order network losses and constructs a linear mapping partial derivative matrix to characterize the voltage response state before and after the action is issued.

[0074] Subsequently, expert knowledge security layer Reading hierarchical reinforcement learning architecture The original mixed action sequence output And extract the corresponding active power adjustment component from the action sequence. With reactive power regulation component .

[0075] To find the safe action closest to the original action while satisfying the rigid voltage boundary of the power grid, the edge computing unit activates its built-in quadratic programming solver. Quadratic Programming Solver Construct an optimization function with the objective of minimizing the distance. The formula is as follows: ; in, For finding the minimum value of a quadratic programming problem; For the quadratic programming solver The final safety action sequence output after recalibration; For hierarchical reinforcement learning architecture The output is the original mixed action sequence; This is a 2-norm operator for calculating the squared Euclidean distance between two high-dimensional action vectors.

[0076] Quadratic Programming Solver Solving the above optimization function At that time, the following linear equation system of voltage upper and lower limits, which characterizes the physical hard constraints of the power grid, must be strictly obeyed: ; in, The lower limit critical value of the busbar safety voltage specified by the system; This is the critical value of the upper limit of the busbar safety voltage specified by the system. For relaxation compensation variables; The reference voltage for the reference node; It is the inverse of the node association matrix; It is a resistance matrix; This is the reactance matrix; This is the power transfer matrix; This represents the initial net active power of the node; This represents the initial net reactive power of the node. For the final safety action sequence The final active power regulation component included; For the final safety action sequence The final reactive power adjustment component included.

[0077] The relaxation compensation variable Specifically designed to compensate for the physical error of secondary network losses that was neglected during the linearization of the power flow in the aforementioned distribution network, its value is preferably calibrated to 0.002 per unit. Based on historical error statistics of large-scale distribution networks, setting this specific value for the relaxation compensation variable ensures that, under different network topology sizes, the relaxation boundary of the linearization equation can perfectly cover the extreme values ​​of the actual voltage offset error generated by the model approximation, thereby guaranteeing that the safety action obtained by projection will never cause the voltage of the actual system to exceed the limit.

[0078] Quadratic Programming Solver The interior-point method is used to solve this convex optimization problem at high speed through iteration. (The original mixed action sequence is then described.) The quadratic programming solver is located within the safe feasible region polyhedron defined by the linear constraint equations. Directly use the original mixed action sequence Assigned to the final safety action sequence And output; if the original mixed action sequence Located outside the safe feasible region, the quadratic programming solver The projected point closest to the original action vector will be found on the feasible region boundary. Dangerous components that may lead to instability will be truncated, and the vector corresponding to the projected point will be used as the final safe action sequence. Output.

[0079] S24: Control evaluation time window calibration and confidence-driven degradation for the final safety action sequence Specifically, the steps for performing confidence-driven security auditing and time window calibration are as follows: Edge computing units extract attribution contribution score matrices generated by physical information neural networks in the preceding stages. Due to this attribution contribution score matrix This reflects the strength of each physical characteristic dimension in predicting the final precursor risk. The edge computing unit calculates this attribution contribution score matrix. Information entropy is used to quantify the overall reliability of current AI decision-making, thereby obtaining the overall decision confidence level. The lower the information entropy, the more concentrated the support of a few core physical features for decision-making, the clearer the physical logic of precursor identification, and consequently, the higher the overall decision confidence. The higher.

[0080] The edge computing unit will calculate the overall decision confidence level. With a high security threshold preset in non-volatile memory With low security threshold Perform conditional branch comparison and judgment.

[0081] If the overall decision confidence level Greater than or equal to the high security threshold The edge computing unit determines that the current decision is supported by sufficient physical evidence. The edge computing unit extracts the final security action sequence. The specific parameters included internally, especially the micro-continuous adjustment parameters. Obtain the length of the control time window determined by the model's dynamic optimization. The edge computing unit controls the time window length. Based on the system's real-time clock, a corresponding control evaluation window is generated. Control evaluation window Used to define the sequence of subsequent safety actions. The closed-loop monitoring and verification cycle of physical execution effects. Subsequently, the edge computing unit will finalize the security action sequence. The complete command is sent to the station-side controller for electrical execution.

[0082] If the overall decision confidence level Given the low security threshold With high security judgment threshold Between these events, the edge computing unit activates a built-in confidence degradation protocol. This confidence degradation protocol forces a scan of the final security action sequence. The system shields and eliminates rigid operating elements that may cause irreversible changes to the power grid topology (preferably, blocking circuit breaker tripping commands that directly disconnect the connection), allowing only limited suppression actions with flexible adjustment capabilities (preferably, reactive power support or slight filter parameter adjustment actions), while appropriately extending the control evaluation window corresponding to the action. In order to conduct safety observations over a longer period of time.

[0083] If the overall decision confidence level Strictly below the low security threshold The edge computing unit determines that the system is untrustworthy due to abnormal acquisition signals or severe logical divergence in the model. In this case, the edge computing unit proactively executes a circuit breaker, without generating any control evaluation window. The system then transfers control back to the traditional local stability control device via a system hardware interrupt mechanism, allowing the underlying equipment to perform the most conservative action to limit the ramp rate, thus preventing a chain reaction of power grid failures.

[0084] S3: Hybrid action execution, time window closed-loop evaluation, and model co-evolution stage S31: Physical closed-loop extraction of multidimensional indicators based on control evaluation time window The final safety action sequence is executed at the station controller. The system then enters closed-loop monitoring mode. The edge computing unit continuously adjusts parameters at the microscopic level. Determined control evaluation window Internally, the system obtains feedback electrical and physical quantities in real time through the underlying high-speed bus in order to determine the actual effect of the current suppression action based on the feedback results.

[0085] Specifically, implement a multi-dimensional evaluation index set. The specific steps for extraction are as follows: First, the edge computing unit continuously subscribes to and reads the three-phase voltage waveform sequence through the substation process layer network. With three-phase current waveform sequence .

[0086] Next, the edge computing unit uses the instantaneous power calculation operator based on the three-phase voltage waveform sequence. With three-phase current waveform sequence Real-time solution control evaluation window Active power time series within .

[0087] Subsequently, the edge computing unit locates the target location using a peak extraction algorithm. At the start of the time window and the finish line Nearby envelope peak values ​​are used to calculate the amplitude attenuation rate. Its mathematical equation is as follows: ; in, The amplitude decay rate is used to characterize the degree of oscillation energy decay. To control the evaluation window Peak value of active power envelope at the end point; To control the evaluation window Peak value of active power envelope at the start time.

[0088] Simultaneously, the edge computing unit invokes the damping identification operator to calculate the real-time envelope fitting damping ratio within the current time window. And fit the damping ratio to the initial envelope before the action is executed. Perform a subtraction operation to obtain the damping evolution difference value. .

[0089] Finally, the edge computing unit calculates the physical distance between the current system state and the preset voltage collapse boundary based on the real-time collected bus voltage amplitude, thus obtaining the real-time operating margin. Amplitude attenuation rate Damping evolution difference and real-time operational margin Perform feature encapsulation to generate a multi-dimensional evaluation index set. .

[0090] S32: Anti-avoidance dynamic elastic transitions and template sequence reconstruction This embodiment is based on a multidimensional evaluation index set. Drive the hierarchical reinforcement learning architecture to perform state transitions and action jumps.

[0091] Specifically, the steps for executing the dynamic elastic transition of the action sequence are as follows: First, the edge computing unit reads the global state space vector generated in the previous stage. And combine it with the multidimensional evaluation index set extracted at the current stage. Perform tensor concatenation along the time dimension to generate an evolutionary state vector that describes the system's evolutionary trend. .

[0092] Next, the evolution state vector Input to hierarchical reinforcement learning architecture Discrete Action Decision Network at the Bottom Based on the comprehensive reward function defined in the preceding step S22 Evaluate the immediate rewards of the current action.

[0093] Subsequently, the edge computing unit determines the damping evolution difference. Is it greater than 0 and has an amplitude decay rate? Is it less than 1?

[0094] If the above criteria are met, it indicates that the current action has a positive inhibitory effect, and the overall reward function... Output positive gain. Underlying discrete action decision network. Maintain the current macroscopic discrete actuators The direction is to fine-tune the network by the top-level continuous parameters. Under the current template, fine-tune subsequent micro-continuous adjustment parameters based on positive feedback. .

[0095] If the above judgment conditions are not met (i.e.) This indicates that the system's oscillation trend has not been effectively curbed, and the comprehensive reward function... This generates a huge negative penalty gradient. At this point, the underlying discrete action decision network... Forced policy index jumps are triggered, and based on the updated state value estimate, novel macroscopic discrete action elements with stronger intervention power are selected from the expert template library in a leapfrog manner. .

[0096] Finally, based on the novel macroscopic discrete motion element Re-trigger the quadratic planning projection of the preceding step S23 to generate the final safety action sequence for the next iteration.

[0097] This dynamic jump mechanism allows the suppression sequence to no longer be limited by static execution logic, but to nonlinearly reconstruct the strategy based on the real-time quality of physical feedback, thereby ensuring extremely rapid convergence to complex oscillation precursors.

[0098] S33: Ultimate circuit breaker isolation based on physical margin limits When system risks continue to rise and intelligent adjustment methods are unable to restore stability within physical limits, this system implements a rigid safety net through physical margin monitoring.

[0099] Specifically, the steps for performing the ultimate circuit breaker operation are as follows: First, the underlying hardware comparator built into the edge computing unit extracts a multi-dimensional evaluation index set in real time. Real-time runtime margin .

[0100] Next, the real-time runtime margin will be determined. Compared to the preset limit of the crash margin threshold in the register Perform continuous comparisons.

[0101] Subsequently, if real-time runtime margin It remains below the critical collapse margin threshold throughout the preset reduction period. Or a hierarchical reinforcement learning architecture Composite reward function for three consecutive evaluations If all values ​​show a monotonically decreasing trend, then the system is determined to have entered an irreversible unstable state.

[0102] Finally, the edge computing unit sends a non-maskable circuit breaker command to the station controller via hardwired logic, triggering the ultimate circuit breaker operation. This drives the underlying circuit breaker to trip and exit the interconnection state, thereby ensuring the safety of the main network equipment.

[0103] S34: Co-evolution of a Central Cloud Model Based on Full Event Tuples Specifically, the steps for performing full event tuple encapsulation and co-evolution are as follows: First, the edge computing unit clocks the multidimensional data throughout the entire post-connection transient process.

[0104] Next, the edge computing unit will use the initial high-dimensional feature matrix Physical information neural network Predicted precursor categories Explanable chain of evidence Hierarchical reinforcement learning architecture Final safety action sequence issued in each iteration and the corresponding multidimensional evaluation index set Structured encapsulation is performed according to the execution order to generate full event data tuples. .

[0105] Subsequently, the edge computing unit transmits the full event data tuple via an encrypted communication link. Uploaded to centralized retraining platform .

[0106] Finally, centralized retraining platform All event data tuples Stored in the offline experience replay pool, and the physical information neural network is then used with the offline gradient descent algorithm. With hierarchical reinforcement learning architecture Parameters are fine-tuned to generate an updated version of the intelligent model, which is then deployed at an opportune time. The training objective is to minimize the prediction and classification error and maximize the cumulative physical reward on historical failure samples.

[0107] S4: Mapping stage between computing power isolation topology and underlying physical control elements S41: Two-layer physical isolation and communication bus topology construction between edge intelligent units and station controllers The final safety action sequence generated in the preliminary stage The actual physical response must be generated through the underlying electrical control loop. To isolate the high-performance computing inference process from the hard real-time control loop, this system adopts a heterogeneous computing architecture, physically dividing the core hardware layer into edge intelligent units. With station controller .

[0108] Specifically, the steps for performing cross-layer data interaction and physical isolation are as follows: Edge intelligent unit It employs a computing architecture equipped with a high-performance graphics processor, specifically designed to handle high-dimensional feature matrices. Secondary calculation, physical information neural network Precursor reasoning and hierarchical reinforcement learning architecture Strategy evolution. Station-side controller. It adopts an industrial-grade control architecture based on programmable gate arrays and a hard real-time operating system, which is specifically used for the direct acquisition of underlying power grid analog quantities, rigid discrimination of hardware interlocking logic, and the issuance of action messages.

[0109] Edge intelligent unit With station controller They do not share any memory address space; they are connected only by a specific isolation bus. It enables bidirectional interaction of commands and status. This specific isolated bus Preferably, a high-speed industrial Ethernet physical layer with opto-isolation characteristics is used to completely block common-mode interference and the spread of electrical faults at the physical medium level.

[0110] During the uplink flow of sampled data, the station-side controller process layer network for accessing substations Subscribing to the underlying merging unit via the International Electrotechnical Commission 61850 standard The sampled value message. Station controller. Extracting high-speed sampled three-phase voltage waveform sequences With three-phase current waveform sequence To ensure data alignment across devices, the station-side controller... A local high-precision clock source was used to add a synchronization timestamp accurate to the microsecond level to the above waveform sequence. Subsequently via a specific isolation bus Transmit it unidirectionally to the edge intelligent unit Shared memory space.

[0111] During the flow of control commands downlink, the edge intelligent unit Based on the above, which has a synchronized timestamp The time-series data is used to calculate and output the final safety action sequence according to the preceding steps. Edge intelligent unit The final safety action sequence Encapsulated as standard control command data packets, transmitted via a specific isolation bus Reverse transmission to station controller .

[0112] Station controller Upon receiving the control command data packet, no floating-point matrix operations are performed. Instead, the control command is directly extracted and verified, and then transmitted through the process layer network. Towards underlying intelligent terminals It issues substation event control messages for general objects, thereby driving circuit breakers or converters to perform precise electrical actions with a physical delay of microseconds.

[0113] S42: Macroscopic Discrete Action Element Analysis and Interlocking Execution Based on Physical Resource Mapping Matrix Specifically, the steps for performing macroscopic discrete action analysis are as follows: Station controller The physical resource mapping matrix is ​​pre-installed inside the non-volatile memory. The physical resource mapping matrix It records in detail the communication addresses and action attributes of all controllable resources within the connected area.

[0114] When the station controller From the final safety action sequence Extract macroscopic discrete motion elements Then, based on this macroscopic discrete action element The logical number in the physical resource mapping matrix This involves performing row lookups and column matching to obtain the precise device operation vector. .

[0115] Preferably, if the macroscopic discrete action element The indicator is a control mode switching action. Physical resource mapping matrix Map this command to a mode reconfiguration message for a specific grid-connected converter. (Station controller) Send device operation vectors to the converter The digital signal processor at the bottom layer of the forced converter cuts off the original grid-type phase-locked loop control circuit.

[0116] Preferably, if the macroscopic discrete action element Indicates filter parameter reconstruction action Physical resource mapping matrix Map this command to a weight modification command for the active damped virtual impedance. (Station controller) Based on device operation vectors The instruction to the underlying controller modifies the virtual inductance parameters of the current inner loop, thereby changing the equivalent impedance of the system in a specific harmonic frequency band, and thus cutting off the resonant amplification path of high-frequency energy.

[0117] Preferably, if the macroscopic discrete action element Indicates a freeze action for the tap strategy. Physical resource mapping matrix This instruction is interpreted as a highest-priority hardware latching signal. (Station controller) The electrical hard-contact lockout command is sent directly to the on-load tap changer of the main transformer, thereby forcibly shielding its automatic tap change logic and avoiding secondary oscillation caused by tap change action.

[0118] Station controller After issuing the above-mentioned device operation vectors At the same time, strictly follow the electrical interlocking timing sequence standard of putting on before cutting off and flexible before rigid to ensure that the generated macro decision is adapted to the basic physical operation procedures of the power grid.

[0119] S43: Adaptive constraints on microscopic continuous adjustment parameters controlled by the electromagnetic physical boundaries of the converter. Microscopic continuous adjustment parameters of the top-level continuous parameter fine-tuning network output in the preliminary stage This involves specific physical adjustment quantities, requiring continuous modulation at the millisecond level by the underlying converter equipment. Specifically, the steps for performing secondary dimensionality reduction and limiting of micro-parameters are as follows: Converter control unit After receiving the parameters from the station controller, the system first reads the apparent power capacity of the converter, which is embedded in the equipment nameplate. And collect the actual voltage amplitude of the bus at the grid connection point in real time. .

[0120] Next, the converter control unit Parameters are continuously adjusted at the micro level. Extracting the reactive power support compensation coefficient With active power limit upper limit .

[0121] Subsequently, the converter control unit According to the reactive power support compensation coefficient Compared with the current actual bus voltage amplitude The theoretical reactive current command used to support grid voltage recovery is calculated. And further convert it into the required reactive power. .

[0122] To ensure equipment safety, the converter control unit... The required reactive power With active power limit upper limit Substituting the active and reactive power coupled boundary constraint equations, thermodynamic and electromagnetic capabilities are verified. The active and reactive power coupled boundary constraint equations are as follows: ; in, For micro-continuous adjustment parameters The planned upper limit of expected active power; Based on the reactive power support compensation coefficient The calculated expected system supports reactive power; The maximum apparent power capacity is determined by the physical limits of hardware insulation and heat dissipation of the underlying converter equipment.

[0123] If the above equation holds, it indicates that the microscopic continuous adjustment parameters generated by the artificial intelligence operator do not exceed the physical execution capability of the underlying hardware, and the converter control unit... Direct basis and The final pulse width modulation signal is synthesized.

[0124] If the above equation does not hold (i.e., the sum of squares exceeds the apparent power capacity) The square of this value indicates a risk of overload in the command. At this point, the converter control unit... Trigger the built-in dynamic reactive power priority compensation strategy .

[0125] This dynamic reactive power priority compensation strategy Prioritize transient voltage stabilization. Converter control unit. Forced to maintain reactive power command The active power limit remains unchanged, and the upper limit is restricted by a dynamic adaptive limiting equation. Perform physical truncation calculations and output the actual active power output command. The dynamic adaptive limiting equation is as follows: ; in, For converter control unit The actual active power tracking threshold used when finally synthesizing the pulse width modulation waveform; This is the minimum value operator.

[0126] Through this constraint mechanism, the continuous exploration variables issued by the artificial intelligence model are firmly confined within the absolute safety domain of the converter's inherent capabilities, thereby maximizing the effective control to suppress oscillations while ensuring that the hardware does not suffer thermal overload damage.

[0127] S44: End-to-end watchdog physical takeover mechanism for scenarios of computing power exhaustion and communication failure Edge intelligent units under extremely harsh operating conditions There may be issues such as computing power being fully loaded, leading to inference timeouts, or specific isolated buses. The probability of a physical wire breakage is extremely low. This system has a station-side controller. Internally, a pure hardware takeover defense line was built, independent of the main program.

[0128] Specifically, the steps for implementing end-to-end watchdog monitoring and physical takeover are as follows: Station controller It is equipped with a full-link watchdog circuit. This end-to-end watchdog circuit This is an independent timer / counter module driven by an external passive crystal oscillator, which internally contains a countdown decrement register. .

[0129] Under normal operating conditions, the edge intelligent unit Control evaluation windows must be generated in each preceding stage. During the cycle, actively send data to the station controller. Send a sequence of heartbeat messages carrying a specific identifier. .

[0130] Station controller Each successful reception and verification of a heartbeat message sequence That is, through the underlying hardware addressing instructions to the end-to-end watchdog circuit. Send a reset level to decrement the countdown register. Restore to the initial full value state.

[0131] Full-link watchdog circuit An unchangeable timeout overflow threshold is hard-coded internally. The timeout overflow threshold Strictly greater than two consecutive control evaluation windows The maximum possible duration.

[0132] If edge intelligent unit Output stopped due to a fault, station-side controller Will be unable to receive heartbeat message sequences Consequently, it becomes impossible to supply the end-to-end watchdog circuit. Send reset level. Countdown decrement register. It will continue to decrease with the crystal oscillator pulse.

[0133] When the countdown decrement register Decrease to zero and exceed the timeout overflow threshold At that time, the full-link watchdog circuit The output pin level of [something] changes instantaneously, thereby sending a signal to the station controller. The main central processing unit triggers the highest priority non-maskable interrupt signal. .

[0134] Station controller The main central processing unit captures the non-maskable interrupt signal. Then, all currently executing Ethernet communication and floating-point operation tasks are forcibly suspended, and the interrupt vector table directly jumps to the degradation interlocking strategy embedded in the underlying read-only memory. .

[0135] This downgraded interlocking strategy Immediately disable the specific isolated bus via register configuration The receiving port completely deprives the edge intelligent unit of its receiving port. The system control was then transferred. Subsequently, the station-side controller... The state machine instantly reverts to the physical takeover state and broadcasts a conservative zero-power ramp command to all controlled underlying converters. .

[0136] The converter received a conservative zero-power ramp command. Subsequently, the active power growth rate is forcibly limited to zero, while the current output steady-state reactive current parameters are locked, thereby preventing a chain reaction of grid collapses and buying valuable physical time for dispatchers to intervene manually.

[0137] Example 2 Further explanation in conjunction with Example 1, such as Figure 2 The structure shown is... Figure 2 A schematic diagram of the structure of a computer device provided in an embodiment of this application. The computer device includes: Processor, memory, communication bus, and computer programs stored in memory that can run on the processor.

[0138] The processor can call the computer program in memory, and when executing the program, implement the method for early identification and adaptive suppression of oscillation resonance after connection change based on physical constraints and hierarchical reinforcement provided in the above embodiments. The method includes: S1, acquiring transient electrical quantity data after connection change and constructing the original time series matrix. Extracting the original time series matrix Generate a high-dimensional feature matrix ; to high-dimensional feature matrix Input to a loss term with embedded energy dissipation physical constraints The precursor recognition model outputs precursor categories. Risk Probability And generate an interpretable chain of evidence. ; S2, Concatenate high-dimensional feature matrices Precursor categories Risk probability With an explainable chain of evidence Generate global state space vector ;Transfer the global state space vector The input hierarchical strategy model is dimensionality-reduced and mapped to macroscopic discrete action elements. With micro-continuous adjustment parameters To generate the original mixed action sequence Based on the power grid physical topology, a safe projection constraint is constructed to transform the original hybrid action sequence. Projecting the sequence of actions into the safe and feasible domain generates the final safe action sequence. ; S3. During the calibrated control evaluation window Internal extraction of multidimensional evaluation index set In the multidimensional evaluation index set When the power grid fails to converge, the hierarchical strategy model is triggered to execute state machine transitions to switch macroscopic discrete action elements. and return to generate the final security action sequence. Steps; S4, at the edge intelligent unit With station controller Physical isolation is maintained between them; station-side controller Final safety action sequence Includes macroscopic discrete motion elements The mapping is converted into physical execution instructions and then issued.

[0139] Furthermore, computer equipment also includes: The Communications Interface (CI) is used for communication between the memory and the processor.

[0140] The memory may include high-speed RAM, and may also include non-volatile memory, such as at least one disk drive.

[0141] If the memory, processor, and communication interface are implemented independently, they can be interconnected via a bus to communicate with each other. The bus can be an Industry Standard Architecture (ISA) bus, a Peripheral Component Interconnect (PCI) bus, or an Extended Industry Standard Architecture (EISA) bus, etc. Buses can be categorized as address buses, data buses, control buses, etc. For ease of representation, Figure 2 The bus is represented by a single thick line, but this does not mean that there is only one bus or one type of bus.

[0142] To provide interaction with a user, the systems and techniques described herein can be implemented on a computer having: a display device for displaying information to the user (e.g., a CRT (cathode ray tube) or LCD (liquid crystal display) monitor); and a keyboard and pointing device (e.g., a mouse or trackball) through which the user provides input to the computer. Other types of devices can also be used to provide interaction with the user; for example, feedback provided to the user can be any form of sensory feedback (e.g., visual feedback, auditory feedback, or tactile feedback); and input from the user can be received in any form (including sound input, voice input, or tactile input).

[0143] Display devices are used to display images, videos, etc. Display devices may include display panels, which may employ liquid crystal displays (LCDs), organic light-emitting diodes (OLEDs), active-matrix organic light-emitting diodes (AMOLEDs), flexible light-emitting diodes (FLEDs), MiniLEDs, MicroLEDs, Micro-OLEDs, quantum dot light-emitting diodes (QLEDs), etc.

[0144] Alternatively, in a specific implementation, if the memory, processor, and communication interface are integrated on a single chip, then the memory, processor, and communication interface can communicate with each other through an internal interface.

[0145] On the other hand, embodiments of this application also provide a computer-readable storage medium storing a computer program thereon. When executed by a processor, the program implements the above-mentioned method for early identification and adaptive suppression of oscillation resonance after connection change based on physical constraints and hierarchical reinforcement. The method includes: S1, acquiring transient electrical quantity data after connection change and constructing an original time series matrix. Extracting the original time series matrix Generate a high-dimensional feature matrix ; to high-dimensional feature matrix Input to a loss term with embedded energy dissipation physical constraints The precursor recognition model outputs precursor categories. Risk Probability And generate an interpretable chain of evidence. ; S2, Concatenate high-dimensional feature matrices Precursor categories Risk probability With an explainable chain of evidence Generate global state space vector ;Transfer the global state space vector The input hierarchical strategy model is dimensionality-reduced and mapped to macroscopic discrete action elements. With micro-continuous adjustment parameters To generate the original mixed action sequence Based on the power grid physical topology, a safe projection constraint is constructed to transform the original hybrid action sequence. Projecting the sequence of actions into the safe and feasible domain generates the final safe action sequence. ; S3. During the calibrated control evaluation window Internal extraction of multidimensional evaluation index set In the multidimensional evaluation index set When the power grid fails to converge, the hierarchical strategy model is triggered to execute state machine transitions to switch macroscopic discrete action elements. and return to generate the final security action sequence. Steps; S4, at the edge intelligent unit With station controller Physical isolation is maintained between them; station-side controller Final safety action sequence Includes macroscopic discrete motion elements The mapping is converted into physical execution instructions and then issued.

[0146] In another aspect, embodiments of this application also provide a computer program product, which includes a computer program that can be stored on a computer-readable storage medium. The computer program can execute computer instructions. When the computer program is executed by a processor, the computer can perform the aforementioned methods for early identification and adaptive suppression of post-connection oscillation resonance based on physical constraints and hierarchical reinforcement. This method includes: S1, acquiring transient electrical quantity data after connection change and constructing an original time-series matrix. Extracting the original time series matrix Generate a high-dimensional feature matrix ; to high-dimensional feature matrix Input to a loss term with embedded energy dissipation physical constraints The precursor recognition model outputs precursor categories. Risk Probability And generate an interpretable chain of evidence. ; S2, Concatenate high-dimensional feature matrices Precursor categories Risk probability With an explainable chain of evidence Generate global state space vector ;Transfer the global state space vector The input hierarchical strategy model is dimensionality-reduced and mapped to macroscopic discrete action elements. With micro-continuous adjustment parameters To generate the original mixed action sequence Based on the power grid physical topology, a safe projection constraint is constructed to transform the original hybrid action sequence. Projecting the sequence of actions into the safe and feasible domain generates the final safe action sequence. ; S3. During the calibrated control evaluation window Internal extraction of multidimensional evaluation index set In the multidimensional evaluation index set When the power grid fails to converge, the hierarchical strategy model is triggered to execute state machine transitions to switch macroscopic discrete action elements. and return to generate the final security action sequence. Steps; S4, at the edge intelligent unit With station controller Physical isolation is maintained between them; station-side controller Final safety action sequence Includes macroscopic discrete motion elements The mapping is converted into physical execution instructions and then issued.

[0147] The logic and / or steps represented in the flowchart or otherwise described herein, for example, can be considered as a sequenced list of executable instructions for implementing logical functions, and can be embodied in any computer-readable medium for use by, or in conjunction with, an instruction execution system, apparatus or device (such as a computer-based system, a processor-included system or other system that can fetch and execute instructions from, an instruction execution system, apparatus or device).

[0148] The processing module is used to perform calculations and processing on the data. In another implementation, the processing module is one or more processors, such as application processors (APs), application-specific integrated circuits (ASICs), modem processors, graphics processing units (GPUs), image signal processors (ISPs), controllers, memory, video codecs, digital signal processors (DSPs), baseband processors, and / or neural network processing units (NPUs). Different processing units can be independent devices or integrated into one or more processors. The controller can generate operation control signals based on instruction opcodes and timing signals to control instruction fetching and execution. The processor's cache memory can store recently used or repeatedly used instructions or data, reducing processor latency.

[0149] The controller can serve as a central nervous system and command center. Based on the instruction opcode and timing signals, the controller generates operation control signals to control instruction fetching and execution. The processor may also include memory for storing instructions and data. In some embodiments, the processor's memory is a cache memory. This memory can store instructions or data that the processor has just used or that is used repeatedly. If the processor needs to reuse the instruction or data, it can directly retrieve it from the memory. This avoids repeated accesses, reduces processor waiting time, and thus improves system efficiency.

[0150] In another implementation, the processing module can achieve certain functions through the logical relationships of hardware circuits. These logical relationships can be fixed or reconfigurable. For example, the processor can be a hardware circuit implemented using an Application-Specific Integrated Circuit (ASIC) or a Programmable Logic Device (PLD), such as a Field-Programmable Gate Array (FPGA). In reconfigurable hardware circuits, the processor loads a configuration document to configure the hardware circuit. Furthermore, for the complex quadratic programming and semi-positive definite matrix decomposition involved in the trajectory optimization in this application, hardware circuits designed for artificial intelligence or high-performance matrix operations can also be used, such as Neural Processing Units (NPUs), Tensor Processing Units (TPUs), and Deep Learning Processing Units (DPUs). The aforementioned processors can call and execute instructions from memory.

[0151] For the purposes of this specification, "computer-readable medium" can be any means that can contain, store, communicate, propagate, or transmit a program for use in or in conjunction with an instruction execution system, apparatus, or device. More specific examples (a non-exhaustive list) of computer-readable media include: an electrical connection having one or more wires (electronic device), a portable computer disk drive (magnetic device), random access memory (RAM), read-only memory (ROM), erasable and editable read-only memory (EPROM or flash memory), fiber optic devices, and portable optical disc read-only memory (CDROM), optical storage devices, magnetic storage devices, or any suitable combination of the foregoing. Furthermore, a computer-readable medium can even be paper or other suitable media on which the program can be printed, since the program can be obtained electronically by optically scanning the paper or other medium, followed by editing, interpreting, or otherwise processing as necessary, and then stored in a computer memory.

[0152] Various embodiments of the systems and techniques described above herein can be implemented in digital electronic circuit systems, integrated circuit systems, field-programmable gate arrays (FPGAs), application-specific integrated circuits (ASICs), application-specific standard products (ASSPs), systems-on-a-chip (SoCs), complex programmable logic devices (CPLDs), computer hardware, firmware, software, and / or combinations thereof. These various embodiments may include implementations in one or more computer programs that can be executed and / or interpreted on a programmable system including at least one programmable processor, which may be a dedicated or general-purpose programmable processor, capable of receiving data and instructions from a storage system, at least one input device, and at least one output device, and transmitting data and instructions to the storage system, the at least one input device, and the at least one output device.

[0153] The program code used to implement the methods of this disclosure may be written in any combination of one or more programming languages. This program code may be provided to a processor or controller of a general-purpose computer, special-purpose computer, or other programmable data processing apparatus, such that when executed by the processor or controller, the program code causes the functions / operations specified in the flowcharts and / or block diagrams to be implemented. The program code may be executed entirely on a machine, partially on a machine, as a standalone software package partially on a machine and partially on a remote machine, or entirely on a remote machine or server.

[0154] The systems and technologies described herein can be implemented in computing systems that include backend components (e.g., as a data server), or computing systems that include middleware components (e.g., an application server), or computing systems that include frontend components (e.g., a user computer with a graphical user interface or web browser through which a user can interact with implementations of the systems and technologies described herein), or any combination of such backend, middleware, or frontend components. The components of the system can be interconnected via digital data communication of any form or medium (e.g., a communication network). Examples of communication networks include local area networks (LANs), wide area networks (WANs), and the Internet.

[0155] Computer systems can include clients and servers. Clients and servers are generally located far apart and typically interact through communication networks. Client-server relationships are created by computer programs running on the respective computers and having a client-server relationship with each other.

[0156] It should be understood that the various forms of processes shown above can be used to reorder, add, or delete steps. For example, the steps described in this disclosure can be executed in parallel, sequentially, or in different orders, as long as the desired result of the technical solution disclosed in this disclosure can be achieved, and this is not limited herein.

[0157] The specific embodiments described above do not constitute a limitation on the scope of protection of this disclosure. Those skilled in the art should understand that various modifications, combinations, sub-combinations, and substitutions can be made according to design requirements and other factors. Although embodiments of this application have been shown and described above, it is understood that the above embodiments are exemplary and should not be construed as limiting this application. Those skilled in the art can make changes, modifications, substitutions, and variations to the above embodiments within the scope of this application.

Claims

1. A method for early identification and adaptive suppression of post-connection oscillation resonance based on physical constraints and hierarchical reinforcement, characterized in that, Includes the following steps: S1. Obtain transient electrical quantity data after connecting the transformer and construct the original time series matrix. Extracting the original time series matrix Generate a high-dimensional feature matrix ; to high-dimensional feature matrix Input to a loss term with embedded energy dissipation physical constraints The precursor recognition model outputs precursor categories. Risk Probability And generate an interpretable chain of evidence. ; S2, Concatenate high-dimensional feature matrices Precursor categories Risk probability With an explainable chain of evidence Generate global state space vector ; global state space vector The input hierarchical strategy model is dimensionality-reduced and mapped to macroscopic discrete action elements. With micro-continuous adjustment parameters To generate the original mixed action sequence ; Based on the power grid physical topology, a safe projection constraint is constructed, which transforms the original hybrid action sequence. Projecting the sequence of actions into the safe and feasible domain generates the final safe action sequence. ; S3. During the calibrated control evaluation window Internal extraction of multidimensional evaluation index set In the multidimensional evaluation index set When the power grid fails to converge, the hierarchical strategy model is triggered to execute state machine transitions to switch macroscopic discrete action elements. and return to generate the final security action sequence. Steps; S4, at the edge intelligent unit With station controller Physical isolation is maintained between them; Station controller Final safety action sequence Includes macroscopic discrete motion elements The mapping is converted into physical execution instructions and then issued.

2. The method for early identification and adaptive suppression of post-connection oscillation resonance based on physical constraints and hierarchical reinforcement according to claim 1, characterized in that, Step S1 includes the following sub-steps: S11, Upon receiving the successful closing signal After starting the stable observation time window Through the substation's underlying communication protocol Acquire digital electrical messages and construct a sequence containing instantaneous three-phase voltage values. Three-phase current instantaneous value sequence and frequency sampling sequence The original time series matrix For the original time series matrix Perform phase calculation to generate a three-phase voltage waveform sequence With three-phase current waveform sequence Active power sequence is generated using instantaneous reactive power theory. With reactive power sequence Combined with frequency sampling sequence The difference operation generates a frequency change rate sequence. Aligning three-phase voltage waveform sequences Three-phase current waveform sequence Active power sequence Reactive power sequence With frequency change rate sequence Generate the basic time series matrix ; in low-frequency characteristic sliding window Internal active power sequence Calculate the active power growth rate by performing linear regression and Hilbert transform Damping ratio fitted with envelope ; in high-frequency feature sliding window Internal three-phase voltage waveform sequence Perform Fast Fourier Transform Calculate the energy change rate of the target frequency band With the increasing trend of total harmonic distortion splicing active power growth rate Envelope fitting damping ratio Target frequency band energy change rate With the increasing trend of total harmonic distortion Generate high-dimensional feature matrix ; High-dimensional feature matrix Including active power growth rate Envelope fitting damping ratio Target frequency band energy change rate With the increasing trend of total harmonic distortion ; S12, Transform the high-dimensional feature matrix The input is a physical information neural network consisting of a one-dimensional time-dilated convolution operator and a fully connected mapping layer. Through physical information neural networks Output layer output precursor category Risk Probability Extracting the basic time series matrix Combined with the reference value of the inherent damping coefficient Utilizing the system's transient total energy Derivatives and phase difference sequences Algebraic summation of the squares of derivatives within a stable observation window The energy dissipation physical constraint loss term is calculated by summing the squared residuals at discrete time steps. ; S13. Traverse feature subsets that do not contain the target feature dimension. Computational physical information neural network In the input feature subset The difference between the biased risk probability and the value after incorporating the target feature dimension is used to generate an attribution contribution score matrix by accumulating the factorial weights. Extract the attribution contribution score matrix. The original physical observation values ​​corresponding to the core features are compared with preset safety thresholds using Boolean logic, and the comparison results are encapsulated to generate an interpretable chain of evidence. .

3. The method for early identification and adaptive suppression of post-connection oscillation resonance based on physical constraints and hierarchical reinforcement according to claim 1 or 2, characterized in that, Step S2 includes the following sub-steps: S21. Use one-hot coding algorithm to classify precursors. Convert to numerical category vector Using word embedding operators to interpretable chains of evidence Mapped to feature vectors of the evidence chain ; Concatenate high-dimensional feature matrices Numerical category vectors Risk probability With the feature vector of the chain of evidence Generate global state space vector ;Transfer the global state space vector Input hierarchical reinforcement learning architecture Utilizing a low-level discrete action decision network Feedforward fully connected layer on global state space vector Perform linear transformations and nonlinear activations to output a state-action value vector. Extract the state-action value vector The index corresponding to the maximum value determines the macroscopic discrete action element. ;Transfer the global state space vector Macroscopic discrete motion elements splicing input top-level continuous parameter fine-tuning network The policy network submodule outputs a mean vector and a standard deviation vector. Based on these mean and standard deviation vectors, reparameterized sampling is performed, and the output is controlled by the time window length. Active power limit upper limit With reactive power support compensation coefficient Microscopic continuous adjustment parameters ; Combine macroscopic discrete motion elements With micro-continuous adjustment parameters Generate the original mixed action sequence ; S22. Calculate the envelope fitting damping ratio observed at the current time. The difference between the historical damping ratio and the previous action is used to establish a direct proportional mapping relationship to generate a positive reward for damping improvement. ; Compare the upward trend of total harmonic distortion rate Actual voltage amplitude of the bus Compared with the system safety operation benchmark matrix, in the upward trend of total harmonic distortion (THD) Or the actual voltage amplitude of the busbar When the limit is exceeded, a preset maximum negative value is assigned to generate a safety boundary violation penalty item. ; Extracting the original mixed action sequence The data includes the types of actuators and the adjustment range, quantifying the physical costs of equipment wear and tear, and the economic losses from active power limiting, generating control action cost items. ; Improve weighting coefficients using damping Margin penalty weighting coefficient With cost weight coefficient Positive incentives for improved damping Penalties for exceeding safety boundaries Cost of control actions Perform weighted fusion and calculate the comprehensive reward function. To drive the underlying discrete action decision network With top-level continuous parameter fine-tuning network Update the weight parameters.

4. The method for early identification and adaptive suppression of post-connection oscillation resonance based on physical constraints and hierarchical reinforcement according to claim 3, characterized in that, Step S2 also includes the following sub-steps: S23. Obtain the static topology parameters of the distribution network and construct the distribution network connection matrix. Node association matrix Resistance matrix With reactance matrix Using the node association matrix Derivation and calculation of the power transfer matrix Extract the original mixed action sequence Included active power regulation components With reactive power regulation component Using a quadratic programming solver Construct an optimization function with the objective of minimizing the distance. Extract the lower limit critical value of the busbar safe voltage. upper limit critical value of busbar safe voltage Relaxation compensation variables Reference voltage of the reference node The inverse of the node association matrix Initial net active power of nodes Initial net reactive power of the node By combining the power flow linearization model of the distribution network, a set of linear equations with upper and lower voltage limits is constructed to provide constraints. Using a quadratic programming solver The interior-point method is used for iterative solution in the original mixed action sequence. When the location is outside the safe feasible region defined by the linear equations of the upper and lower voltage limits, search for the distance from the original hybrid action sequence on the boundary of the safe feasible region. The nearest projection point, truncating the dangerous component to generate a component containing the final active power regulation component. With the final reactive power regulation component Final safety action sequence ; S24. Calculate the attribution contribution score matrix. Information entropy quantification generates overall decision confidence. Compare the overall decision confidence level With a preset high security threshold in non-volatile memory With low security threshold ; In overall decision confidence Greater than or equal to the high security threshold At that time, extract the micro-continuous adjustment parameters. Included control time window length Generate control evaluation window And issue the final safety action sequence ; In overall decision confidence Given the low security threshold With high security judgment threshold In between, the confidence degradation protocol is activated to eliminate the final security action sequence. The inclusion of rigid motion elements generates limited inhibition of motion and extends the control evaluation window. ; In overall decision confidence Strictly below the low security threshold At that time, the circuit breaker is triggered to transfer control authority to the local stability control device and to limit the ramp rate, without generating a control evaluation window. .

5. The method for early identification and adaptive suppression of post-connection oscillation resonance based on physical constraints and hierarchical reinforcement according to claim 1, characterized in that, In step S3, within the calibrated control evaluation time window Internal extraction of multidimensional evaluation index set This includes the following sub-steps: The three-phase voltage waveform sequence is continuously subscribed to and read through the substation process layer network. With three-phase current waveform sequence The instantaneous power calculation operator is used based on the three-phase voltage waveform sequence. With three-phase current waveform sequence Real-time solution control evaluation window Active power time series within ; Active power time series location using peak extraction algorithm At the start of the time window and the finish line Nearby envelope peak values, extract control evaluation window Peak active power envelope at the end time With control evaluation window Peak active power envelope at the start time Using mathematical calculation equations Calculate the amplitude decay rate used to characterize the degree of oscillation energy decay. ; Calculate the real-time envelope fitting damping ratio within the current time window using the damping identification operator. The real-time envelope fitting damping ratio Damping ratio fitted to the initial envelope before the action is executed The damping evolution difference is obtained by performing a subtraction operation. ; The real-time operating margin is obtained by calculating the physical distance between the current state of the system and the preset voltage collapse boundary based on the real-time collected bus voltage amplitude. The amplitude attenuation rate Damping evolution difference With real-time runtime margin Perform feature encapsulation to generate a multi-dimensional evaluation index set .

6. The method for early identification and adaptive suppression of post-connection oscillation resonance based on physical constraints and hierarchical reinforcement according to claim 1 or 5, characterized in that, In step S3, in the multidimensional evaluation index set When the power grid fails to converge, the hierarchical strategy model is triggered to execute state machine transitions to switch macroscopic discrete action elements. and return to generate the final security action sequence. The steps also include the following sub-steps: Read the global state space vector , global state space vector With multidimensional evaluation index set Perform tensor concatenation along the time dimension to generate an evolutionary state vector. ; Evolutionary state vector Input hierarchical reinforcement learning architecture Utilizing a low-level discrete action decision network Based on the comprehensive reward function Assess the return on investment; Determine the difference in damping evolution Is it greater than 0 and has an amplitude decay rate? Is it less than 1? In the difference of damping evolution Greater than 0 and amplitude decay rate When the condition of less than 1 is met, the comprehensive reward function Output positive gain, utilizing the underlying discrete action decision network Maintain current macroscopic discrete action elements The goal is to fine-tune the network using continuous parameters at the top level. Under the current template, fine-tune the micro-continuous adjustment parameters based on positive feedback. ; In the difference of damping evolution When the condition of being less than or equal to 0 is not met, the comprehensive reward function... This generates a huge negative penalty gradient, forcibly triggering the underlying discrete action decision network. The strategy index jump involves leaping from the expert template library to select novel macroscopic discrete action elements with stronger intervention power based on the updated state value estimate. Based on novel macroscopic discrete motion elements Re-trigger the quadratic programming projection to generate the final safe action sequence for the next iteration. .

7. The method for early identification and adaptive suppression of post-connection oscillation resonance based on physical constraints and hierarchical reinforcement according to claim 6, characterized in that, Step S3 also includes the following sub-steps: Extracting a multi-dimensional evaluation index set using a low-level hardware comparator Included real-time runtime margin Real-time runtime margin With the limit of the crash margin threshold in the register Perform continuous comparisons; in real-time runtime margin Always less than the ultimate collapse margin threshold or hierarchical reinforcement learning architecture Composite reward function for three consecutive evaluations When the behavior is monotonically decreasing, issuing a non-maskable circuit breaker command triggers the ultimate circuit breaker operation. The underlying circuit breaker is tripped to exit the interconnection state; Clock alignment is performed after the transient process ends, and the high-dimensional feature matrix is... Physical information neural network Predicted precursor categories Explanable chain of evidence Hierarchical reinforcement learning architecture Final safety action sequence issued in each iteration With the corresponding multidimensional evaluation index set Encapsulate and generate full event data tuples ; All event data tuples Uploaded to centralized retraining platform The offline experience replay pool utilizes the offline gradient descent algorithm to analyze physical information neural networks. With hierarchical reinforcement learning architecture Perform parameter tuning.

8. The method for early identification and adaptive suppression of post-oscillatory resonance in connection-dependent communication based on physical constraints and hierarchical reinforcement as described in claim 1, characterized in that, Step S4 includes the following sub-steps: Obtain the final safety action sequence Includes macroscopic discrete motion elements Using the station-side controller Extracting the physical resource mapping matrix inside a non-volatile flash memory chip. According to macroscopic discrete motion elements In the physical resource mapping matrix Perform row address lookup and column attribute matching to output the device operation vector. And issue instructions; Utilizing the underlying photovoltaic converter control unit Obtain the apparent power capacity of the converter hard-coded in read-only memory AC voltage transformers collect actual bus voltage amplitude. ; Extracting micro-continuous adjustment parameters Included reactive power support compensation coefficient With active power limit upper limit Combined with reactive power support compensation coefficient Actual voltage amplitude of the bus The required reactive power for deviation calculation ; Extract the upper limit of active power. Required reactive power With converter apparent power capacity Active and reactive coupled boundary constraint equations When the active and reactive power coupling boundary constraint equations hold, the upper limit of active power is used. With required reactive power Generate quadrature-axis and direct-axis reference currents and synthesize a space vector pulse width modulation signal; When the active and reactive power coupling boundary constraint equations do not hold, the dynamic reactive power priority compensation strategy is activated. Maintain the required reactive power With the numerical value unchanged, a dynamic adaptive limiting equation is used. Truncated, dimension-reduced, and amplitude-limited calculations are used to output the actual active power output command. Output commands using actual active power With required reactive power The quadrature-axis and direct-axis reference currents are generated and synthesized into a space vector pulse width modulation signal.

9. A computer device comprising at least one processor coupled to at least one memory storing at least one computer program or instruction, characterized in that, The computer program or instructions are loaded and executed by the processor to implement the steps of the method for early identification and adaptive suppression of post-connection oscillation resonance based on physical constraints and hierarchical reinforcement as described in any one of claims 1 to 8.

10. A computer-readable storage medium, characterized in that, The computer-readable storage medium stores a computer program or instructions, which, when executed by a processor, implement the steps of the method for early identification and adaptive suppression of post-connection oscillation resonance based on physical constraints and hierarchical reinforcement as described in any one of claims 1 to 8.

Images