Dynamic self-adaptive low-frequency load shedding method and system for high-proportion distributed new energy power grid

Abstract

The invention provides a dynamic self-adaptive low-frequency load shedding method and system for a high-proportion distributed new energy power grid, and the method comprises the following steps: processing the data of a multi-source power system through self-adaptive time sequence sampling and multi-dimensional anomaly detection, and extracting voltage class features; analyzing voltage grade correlation characteristics and dynamic influence characteristics, and evaluating a voltage grade coupling degree; inertia estimation and trend analysis are carried out based on frequency response characteristics; respectively extracting distributed and centralized new energy features and predicting output of the distributed and centralized new energy features; performing multi-scale decomposition and feature fusion on the system inertia and the new energy features; determining load removal priority and power distribution based on multi-dimensional risk assessment; and a hierarchical control method is adopted to generate a control instruction and dynamically optimize a control strategy. According to the invention, the dynamic characteristics of the system can be accurately identified, accurate load removal control is realized, and the stability and reliability of the system are improved.

Description

Technical Field

[0001] The present invention relates to a low-frequency load shedding method, in particular to a dynamic adaptive low-frequency load shedding method and system for a high-proportion distributed new energy power grid. Background Art

[0002] With the deepening of the energy transformation, the penetration rate of distributed new energy in the power grid is continuously increasing, bringing unprecedented challenges to traditional low-frequency load shedding schemes. In a power grid with a high proportion of distributed new energy access, the system inertia is significantly reduced, and the frequency dynamic characteristics change fundamentally, resulting in a significant decline in the effectiveness of traditional low-frequency load shedding methods. At the same time, the randomness and volatility of distributed new energy make the system operation state more complex, increasing the risk of system instability. Therefore, studying a low-frequency load shedding method that adapts to the characteristics of a high-proportion distributed new energy is of great significance for ensuring the safe and stable operation of the power grid.

[0003] At present, a large number of studies on low-frequency load shedding technologies for distributed new energy power grids have been carried out at home and abroad. Traditional low-frequency load shedding schemes are mainly based on fixed frequency thresholds and cut-off capacities, and use pre-set cut-off strategies for control. Some studies have tried to introduce an adaptive adjustment mechanism to dynamically adjust the cut-off capacity according to the frequency change rate. There are also studies that have proposed a low-frequency load shedding method based on inertia identification, attempting to optimize the cut-off strategy by estimating the real-time inertia of the system. In terms of distributed new energy access, some studies have begun to consider the impact of new energy output characteristics on the system frequency and proposed a simple compensation strategy considering new energy power fluctuations.

[0004] However, the existing studies still have the following technical problems: First, in terms of system feature extraction, it is difficult to accurately describe the dynamic coupling relationship between different voltage levels, especially the inaccurate quantification of the propagation characteristics and influence degree under system disturbance conditions; second, in terms of inertia assessment, the existing methods do not fully consider the impact of the spatial distribution characteristics of distributed new energy on the system inertia, resulting in insufficient inertia estimation accuracy; third, in terms of new energy characteristic analysis, it is difficult to take into account the characteristic differences between distributed and centralized new energy at the same time, lacking a unified feature extraction and prediction framework; fourth, in terms of multi-time scale feature fusion, the existing methods are difficult to effectively handle the coupling relationship between different time scale features, affecting the accuracy of system state assessment; fifth, in terms of load shedding optimization, there is a lack of a hierarchical control strategy considering the voltage level coupling degree, making it difficult to achieve coordinated control between different voltage levels. These technical problems seriously restrict the application effect of low-frequency load shedding in a high-proportion distributed new energy power grid. Summary of the Invention

[0005] Objective of the invention: To provide a dynamic adaptive low-frequency load shedding method and system for a high-proportion distributed new energy power grid. By accurately characterizing system characteristics and inertia assessment, it adapts to the stochastic characteristics and multi-time scale characteristics of new energy, and meets the requirements of frequency stability with fewer rounds of load shedding and less load cut-off.

[0006] Technical solution: The present invention provides a dynamic adaptive low-frequency load shedding method for a high-proportion distributed new energy power grid, including the following steps:

[0007] Step S1: Obtain the original frequency data, original power data, original voltage data, original current data, original new energy output data, and original energy storage state data from power system equipment, perform real-time sampling on them to obtain a time-series sampling data set, conduct data quality assessment on the time-series sampling data set to obtain data quality indicators and outlier marker sets, use the outlier marker sets to preprocess the time-series sampling data set to obtain a multi-source preprocessed data set, and extract voltage level characteristics based on the multi-source preprocessed data set to obtain a voltage level characteristic matrix;

[0008] Step S2: Obtain the multi-source preprocessed data set and the voltage level characteristic matrix, analyze the correlation characteristics between different voltage levels to obtain a voltage level correlation matrix, evaluate the dynamic influence characteristics based on the voltage level correlation matrix and the multi-source preprocessed data set to obtain a dynamic influence characteristic matrix, and conduct coupling characteristic analysis based on the dynamic influence characteristic matrix to obtain a voltage level influence characteristic matrix and a voltage level coupling degree index;

[0009] Step S3: Obtain the multi-source preprocessed data set and the voltage level influence characteristic matrix, perform inertia estimation based on the frequency response characteristic to obtain an instantaneous inertia estimation value, analyze the inertia change trend based on the instantaneous inertia estimation value and the voltage level influence characteristic matrix to obtain an inertia change trend index, and conduct feature extraction based on the instantaneous inertia estimation value and the inertia change trend index to obtain a system inertia characteristic matrix;

[0010] Step S4: Obtain the multi-source preprocessed data set, use a distributed feature extraction method to obtain a distributed new energy characteristic matrix, use a centralized feature extraction method to obtain a centralized new energy characteristic matrix, and conduct new energy output prediction based on the distributed new energy characteristic matrix and the centralized new energy characteristic matrix to obtain a new energy prediction matrix;

[0011] Step S5: Obtain the system inertia characteristic matrix, the distributed new energy characteristic matrix, and the centralized new energy characteristic matrix, use a multi-scale decomposition method to obtain a multi-scale component matrix, analyze the time scale correlation relationship based on the multi-scale component matrix to obtain a time scale correlation degree matrix, and conduct feature fusion based on the multi-scale component matrix and the time scale correlation degree matrix to obtain a multi-scale feature tensor;

[0012] Step S6: Obtain the multi-scale feature tensor, time-scale correlation matrix, inertia change trend index, and new energy prediction matrix. Use a multi-dimensional risk assessment method to obtain the system risk assessment matrix. Based on the system risk assessment matrix and the new energy prediction matrix, perform load priority assessment to obtain the load shedding priority sequence. Optimize the shedding strategy based on the load shedding priority sequence and the system risk assessment matrix to obtain the shedding power quantity matrix;

[0013] Step S7: Obtain the load shedding priority sequence, the shedding power quantity matrix, and the voltage level coupling degree index. Use a hierarchical control method to generate hierarchical control instructions. Based on the hierarchical control instructions and the multi-source preprocessing data set, evaluate the control effect to obtain the control effect evaluation index. Optimize the strategy based on the control effect evaluation index and the system risk assessment matrix to obtain the control strategy adjustment parameters and the updated control instructions.

[0014] According to another aspect of the present invention, there is also provided a high-proportion distributed new energy grid dynamic adaptive under-frequency load shedding system, including: at least one processor; and a memory communicatively connected to at least one of the processors; wherein, the memory stores instructions executable by the processor, and the instructions are used to be executed by the processor to perform the high-proportion distributed new energy grid dynamic adaptive under-frequency load shedding method described in any one of the above technical solutions.

[0015] Compared with the prior art, the beneficial effects of the present invention are as follows: A complete set of high-proportion distributed new energy grid dynamic adaptive under-frequency load shedding method systems is constructed, realizing the whole-process intelligent processing from data acquisition, feature extraction, state evaluation to control strategy optimization. Through the organic combination of innovative technologies such as multi-source data real-time processing, voltage level correlation analysis, system inertia assessment, new energy feature extraction, multi-scale feature fusion, risk assessment, and hierarchical control, the technical defects of traditional under-frequency load shedding methods under the condition of high-proportion new energy access are solved. This solution can accurately perceive the changes in the system state, accurately evaluate the system risk, optimize the load shedding strategy, realize precise load shedding control, and significantly improve the stability and reliability of the system. Especially under complex conditions such as new energy output fluctuations and system disturbances, it shows excellent adaptability and control effects. Description of the Drawings

[0016] Figure 1 is the flowchart of the present invention.

[0017] Figure 2 is the flowchart of step S1 of the present invention.

[0018] Figure 3 is the flowchart of step S2 of the present invention.

[0019] Figure 4It is the flowchart of step S3 of the present invention.

[0020] Figure 5 It is the flowchart of step S4 of the present invention.

[0021] Figure 6 It is the flowchart of step S5 of the present invention.

[0022] Figure 7 It is the flowchart of step S6 of the present invention.

[0023] Figure 8 It is the flowchart of step S7 of the present invention. Detailed implementation manner

[0024] As Figure 1 shown, according to the high - proportion distributed new - energy power - grid dynamic adaptive low - frequency load - shedding method of the present invention, the following steps are included:

[0025] Step S1: Obtain the original frequency data, original power data, original voltage data, original current data, original new - energy output data, and original energy - storage state data from the power - system equipment, perform real - time sampling on them to obtain a time - series sampling data set, perform data - quality evaluation on the time - series sampling data set to obtain a data - quality index and an outlier - marked set, use the outlier - marked set to pre - process the time - series sampling data set to obtain a multi - source pre - processed data set, and extract voltage - level features based on the multi - source pre - processed data set to obtain a voltage - level feature matrix;

[0026] Step S2: Obtain the multi - source pre - processed data set and the voltage - level feature matrix, analyze the correlation characteristics between different voltage levels to obtain a voltage - level correlation matrix, evaluate the dynamic influence characteristics according to the voltage - level correlation matrix and the multi - source pre - processed data set to obtain a dynamic influence feature matrix, and perform coupling - characteristic analysis based on the dynamic influence feature matrix to obtain a voltage - level influence feature matrix and a voltage - level coupling - degree index;

[0027] Step S3: Obtain the multi - source pre - processed data set and the voltage - level influence feature matrix, perform inertia estimation based on the frequency - response characteristic to obtain an instantaneous inertia - estimation value, analyze the inertia - change trend according to the instantaneous inertia - estimation value and the voltage - level influence feature matrix to obtain an inertia - change trend index, and perform feature extraction based on the instantaneous inertia - estimation value and the inertia - change trend index to obtain a system - inertia feature matrix;

[0028] Step S4: Obtain the multi - source pre - processed data set, use a distributed feature - extraction method to obtain a distributed new - energy feature matrix, use a centralized feature - extraction method to obtain a centralized new - energy feature matrix, and perform new - energy output prediction according to the distributed new - energy feature matrix and the centralized new - energy feature matrix to obtain a new - energy prediction matrix;

[0029] Step S5: Obtain the system inertia characteristic matrix, distributed new energy characteristic matrix, and centralized new energy characteristic matrix. Use the multi-scale decomposition method to obtain the multi-scale component matrix. Analyze the time-scale correlation relationship based on the multi-scale component matrix to obtain the time-scale correlation degree matrix. Perform feature fusion based on the multi-scale component matrix and the time-scale correlation degree matrix to obtain the multi-scale feature tensor;

[0030] Step S6: Obtain the multi-scale feature tensor, time-scale correlation degree matrix, inertia change trend index, and new energy prediction matrix. Use the multi-dimensional risk assessment method to obtain the system risk assessment matrix. Perform load priority assessment based on the system risk assessment matrix and the new energy prediction matrix to obtain the load shedding priority sequence. Optimize the shedding strategy based on the load shedding priority sequence and the system risk assessment matrix to obtain the shedding power quantity matrix;

[0031] Step S7: Obtain the load shedding priority sequence, shedding power quantity matrix, and voltage level coupling degree index. Use the hierarchical control method to generate hierarchical control instructions. Evaluate the control effect based on the hierarchical control instructions and the multi-source preprocessing data set to obtain the control effect evaluation index. Perform strategy optimization based on the control effect evaluation index and the system risk assessment matrix to obtain the control strategy adjustment parameters and the updated control instructions.

[0032] The present invention constructs a complete set of dynamic adaptive low-frequency load shedding method systems for high-proportion distributed new energy power grids, realizing the whole-process intelligent processing from data acquisition, feature extraction, state evaluation to control strategy optimization. Through the organic combination of innovative technologies such as multi-source data real-time processing, voltage level correlation analysis, system inertia assessment, new energy feature extraction, multi-scale feature fusion, risk assessment, and hierarchical control, it solves many challenges faced by traditional low-frequency load shedding methods under the condition of high-proportion new energy access. This solution can accurately perceive the system state change, precisely evaluate the system risk, optimize the load shedding strategy, achieve precise control, and significantly improve the stability and reliability of the system. Especially under complex conditions such as new energy output fluctuations and system disturbances, it shows excellent adaptability and control effect.

[0033] According to one aspect of the present invention, as Figure 2 shown, step S1 is specifically

[0034] Step S11: Real-time collect the original frequency data, original power data, original voltage data, original current data, original new energy output data, and original energy storage state data from the power system equipment. Use the adaptive time-series sampling method to process the data to obtain the time-series sampling data set;

[0035] Among them, the dynamic sampling interval T(t) = T0·exp(-α·V(t))·B(t); the rate-of-change feature V(t) = ||dx(t) / dt||2 / max(||x(t)||2, ε); the sampling quality factor B(t) = 1 / (1 + exp(-κ·(SNR(t) - SNR0))); where: x(t) is the original data sequence; T0 is the reference sampling interval; SNR(t) is the signal-to-noise ratio; SNR0 is the reference signal-to-noise ratio; α, κ are adjustment parameters; ε is a small constant to prevent division by zero; t is the time index; ||·||2 is the L2 norm.

[0036] Step S12: Obtain a time-series sampling data set, and analyze the time-series sampling data set by using an improved multi-dimensional anomaly detection method to generate a data quality index and an outlier label set.

[0037] Among them, the multi-dimensional anomaly detection method is specifically: the anomaly degree A(i) = ρ(i)·δ(i)·W(i); the local density ρ(i) = ∑exp(-(d(i,j) / d k ) 2 ); the relative distance δ(i) = min(d(i,j)), where j satisfies ρ(j) > ρ(i); the time-varying weight W(i) = exp(-β·|dX(i) / dt|); where: d(i,j) is the distance between samples; d k is the truncation distance; X(i) is the sample feature vector; β is the time decay coefficient; i, j are sample indices.

[0038] Step S13: Obtain a time-series sampling data set and an outlier label set, process the labeled outliers by using an adaptive filtering algorithm, and then perform noise reduction by using a multi-level signal decomposition method to obtain a multi-source preprocessed data set.

[0039] The adaptive filtering algorithm is specifically: the filtering output y(t) = ∑(h(k,t)·x(t - k)); the adaptive filter h(k,t) = h0(k)·exp(-α·e(t))·G(k,t); the error signal e(t) = y(t) - y*(t); the gain adjustment where: x(t) is the input signal; y*(t) is the desired output; h0(k) is the initial filter coefficient; α, β are adaptive coefficients; k is the filter order; t is the time index; is the gradient operator.

[0040] Step S14: Obtain a multi-source preprocessed data set, and analyze the data of different voltage levels in the multi-source preprocessed data set by using an adaptive feature extraction method to obtain a voltage level feature matrix.

[0041] Adaptive feature extraction method, specifically: feature vector F(i,t) = W(t)·X(i,t)·D(i,t); feature weight W(t) = softmax(Q·K^T / √d); dynamic feature X(i,t) = σ(U·x(i,t)+b); scale factor where: x(i,t) is the original data; Q, K are attention matrices; U is a transformation matrix; b is a bias vector; σ(·) is an activation function; λ is an attenuation coefficient; i is the voltage level index; t is the time index; d is the feature dimension.

[0042] By establishing a real-time acquisition and processing mechanism for multi-source heterogeneous data, high-quality acquisition of key parameters of the power system is achieved. Specifically, the adaptive time-series sampling method can dynamically adjust the sampling frequency according to the system state, avoiding data redundancy or information loss problems caused by traditional fixed sampling rates; through an improved multi-dimensional anomaly detection method, abnormal data can be accurately identified and marked, improving data quality; using an adaptive filtering algorithm and a multi-level signal decomposition method to process and denoise abnormal data, ensuring data reliability; finally, through the adaptive feature extraction method, accurate extraction of data features at different voltage levels is achieved. The combined application of this series of innovative technologies solves problems such as unreasonable sampling, inaccurate anomaly identification, and severe noise interference in the traditional data acquisition process, providing a high-quality data basis for subsequent analysis and decision-making. Especially in power grids with a high proportion of distributed new energy access, it can effectively cope with the challenges of rapid fluctuations in system parameters, improving the real-time performance and accuracy of data processing.

[0043] According to one aspect of the present invention, as Figure 3 shown, step S2 is specifically

[0044] Step S21, obtain a multi-source preprocessed data set and a voltage level feature matrix, and use a multi-level network analysis method to construct a voltage level association model to obtain a voltage level association matrix;

[0045] The multi-level network analysis method is specifically: network strength N(i,j) = C(i,j)·T(i,j)·P(i,j); connectivity C(i,j) = exp(-d(i,j) / r0); time series correlation T(i,j) = ∑(x(i,t)·x(j,t+τ))·exp(-α·|τ|); power flow P(i,j) = |S(i,j)| / S_max·exp(-β·θ(i,j)); where: d(i,j) is the electrical distance; r0 is the feature distance; x(i,t) is the node state; S(i,j) is the apparent power; θ(i,j) is the power angle; α, β are adjustment parameters; i, j are level indices; t is the time index; τ is the time delay.

[0046] Step S22: Obtain the voltage level correlation matrix and the multi-source preprocessed dataset, and use an improved time-varying causal analysis method to evaluate the dynamic influence relationship between voltage levels, so as to obtain the dynamic influence feature matrix;

[0047] The time-varying causal analysis method is specifically as follows: Causal strength C(i,j,t)=I(i,j,t)·D(i,j,t)·M(i,j,t); Mutual information I(i,j,t)=∑p(x,y,t)·log(p(x,y,t) / (p(x,t)·p(y,t))); Dynamic dependence Time-varying modulation where: p(x,y,t) is the joint probability distribution; p(x,t), p(y,t) are the marginal distributions; x(t), y(t) are time series; α, β are adjustment parameters; i, j are level indices; t is the time index; τ is the time delay.

[0048] Step S23: Obtain the dynamic influence feature matrix, and use an adaptive coupling evaluation method to analyze the coupling relationship between voltage levels, so as to obtain the voltage level influence feature matrix and the voltage level coupling degree index.

[0049] The adaptive coupling evaluation method is specifically as follows: Coupling degree K(i,j,t)=E(i,j,t)·S(i,j,t)·V(i,j,t); Energy coupling E(i,j,t)=|P(i,j,t)| / P_max·exp(-α·|θ(i,j,t)|); State coupling S(i,j,t)=exp(-d(i,j) / r0)·ρ(i,j,t); Voltage coupling V(i,j,t)=1 / (1+β·||ΔV(i,j,t)||2); where: P(i,j,t) is the power flow; θ(i,j,t) is the phase angle difference; d(i,j) is the electrical distance; ρ(i,j,t) is the correlation coefficient; ΔV(i,j,t) is the voltage difference; α, β are adjustment parameters; r0 is the characteristic distance; i, j are voltage level indices; t is the time index.

[0050] A complete set of voltage level correlation analysis and impact assessment systems has been established, enabling the accurate characterization of the complex interaction relationships between different voltage levels. By constructing a voltage level correlation model through a multi-level network analysis method, the topological connection relationships and electrical coupling characteristics between different voltage levels can be accurately captured; an improved time-varying causality analysis method is adopted to achieve a quantitative assessment of the dynamic impact relationships between voltage levels, especially the propagation paths and impact degrees under system disturbance conditions; through an adaptive coupling assessment method, a coupling strength calculation model between voltage levels is established, effectively identifying key influencing factors and weak links. The comprehensive application of these innovative technologies solves the problem that traditional analysis methods are difficult to accurately describe the dynamic coupling relationships between voltage levels. Especially in the case of high-proportion distributed new energy access, it can effectively evaluate the mutual influence degrees between different voltage levels, providing an important basis for system stability analysis and control strategy formulation.

[0051] According to one aspect of the present invention, as Figure 4 shown, step S3 is specifically as follows

[0052] Step S31: Obtain a multi-source preprocessed data set, and analyze the system frequency response characteristics using an improved state estimation algorithm to obtain an instantaneous inertia estimation value;

[0053] The state estimation algorithm is specifically: state vector Gain matrix K(t) = P(t)·H^T·R -1 ; Prediction error P(t) = (I - K(t)·H)·P0(t); Dynamic correction G(t) = exp(-α·||r(t)||2); Residual where: z(t) is the measurement vector; h(·) is the measurement equation; H is the measurement matrix; R is the measurement noise covariance; α is the adjustment parameter; t is the time index; T is the transpose; I is the identity matrix.

[0054] Step S32: Obtain the instantaneous inertia estimation value and the voltage level influence characteristic matrix, and use a deep learning optimization algorithm to predict the system inertia change trend to obtain an inertia change trend index;

[0055] The deep learning optimization algorithm is specifically: prediction output Deep network f(X) = σ(W2·σ(W1·X + b1) + b2); Dynamic weight Compensation term Φ(t) = β·∑(ε(t - k)·exp(-γ·k)); Loss function where: X(t) is the input feature; σ(·) is the activation function; W1, W2 are weight matrices; b1, b b2 are bias vectors; ε(t) is the prediction error; α, β, γ, λ are adjustment parameters; t is the time index; k is the historical time index.

[0056] Step S33: Obtain the instantaneous inertia estimation value and the inertia change trend index, and use the multi-dimensional feature extraction method to construct the inertia feature space to obtain the system inertia feature matrix.

[0057] The multi-dimensional feature extraction method is specifically: the feature matrix F(t) = T(t)·S(t)·D(t); the time-series feature The spatial feature S(t) = GCN(A(t))·V(t); the dynamic feature D(t) = Attention(Q(t), K(t), V(t)); where: x(t) is the input data; A(t) is the adjacency matrix; V(t) is the node feature; Q(t), K(t) are the attention parameters; LSTM(·) is the long short-term memory network; GCN(·) is the graph convolutional network; α is the attenuation coefficient; t is the time index.

[0058] A real-time evaluation and prediction framework for the system inertia characteristics is constructed, realizing an accurate grasp of the dynamic characteristics of the power system. By analyzing the system frequency response characteristics through an improved state estimation algorithm, the system instantaneous inertia can be estimated in real time and accurately; a deep learning optimization algorithm is used to predict the system inertia change trend, effectively capturing the dynamic change law of the system inertia; the multi-dimensional feature extraction method is used to construct the inertia feature space, realizing a comprehensive characterization of the system inertia characteristics. The application of this set of innovative technologies solves the problem of insufficient accuracy of traditional inertia estimation methods under the condition of high proportion of new energy access. Especially in the case of system disturbances and new energy output fluctuations, it can accurately evaluate the dynamic response ability of the system, providing a reliable decision-making basis for under-frequency load shedding control and significantly improving the safety and stability of the system.

[0059] According to one aspect of the present invention, as Figure 5 shown, step S4 is specifically

[0060] Step S41: Obtain the multi-source preprocessed data set, and use the improved deep learning feature extraction method to analyze the characteristics of distributed new energy to obtain the distributed new energy feature matrix;

[0061] The extraction process of the distributed feature is specifically: the node feature F(i,t) = P(i,t)·C(i,t)·R(i,t); the power feature P(i,t) = |S(i,t)| / S_max·exp(-α·|dS(i,t) / dt|); the spatial correlation C(i,t) = ∑(w(i,j)·exp(-d(i,j) / r0)·ρ(i,j,t)); the reliability R(i,t) = 1 / (1 + β·σ 2 (i,t)); where: S(i,t) is the apparent power; d(i,j) is the node distance; ρ(i,j,t) is the correlation coefficient; σ2 (i, t) is the fluctuation variance; w(i, j) is the spatial weight; α and β are adjustment parameters; r0 is the characteristic distance; i and j are node indices; t is the time index.

[0062] Step S42: Obtain a multi-source preprocessed dataset, and use an enhanced neural network method to extract centralized new energy features to obtain a centralized new energy feature matrix.

[0063] The enhanced neural network method is specifically: the predicted output Local network The global network f2(x) = Transformer(x) · G(t); the residual network f3(x) = ResNet(x) · R(t); the dynamic weight W(t) = softmax(Q · K^T / √d); where: x(t) is the input feature; CNN(·), Transformer(·), ResNet(·) are neural network models; G(t), R(t) are gain factors; α is the attenuation coefficient; Q, K are attention parameters; t is the time index; d is the feature dimension

[0064] Step S43: Obtain a distributed new energy feature matrix and a centralized new energy feature matrix, and use an ensemble learning prediction method to perform multi-step prediction to obtain a new energy prediction matrix.

[0065] The ensemble learning prediction method is specifically: the prediction result Y(t) = ∑(w(i, t) · y(i, t)) · Q(t); the model weight w(i, t) = exp(-α · e(i, t)) / ∑exp(-α · e(i, t)); the prediction error e(i, t) = ||y(i, t) - y*(t)||2 · (1 + β · |dy(i, t) / dt|); the quality factor Q(t) = 1 / (1 + γ · σ 2 (t)); where: y(i, t) is the single model prediction value; y*(t) is the actual value; σ 2 (t) is the prediction variance; α, β, γ are adjustment parameters; i is the model index; t is the time index.

[0066] A unified framework for feature extraction and prediction of distributed and centralized new energy is established, achieving accurate characterization and prediction of the output characteristics of new energy. Through an improved deep learning feature extraction method, the spatial distribution characteristics and temporal variation laws of distributed new energy can be accurately captured; an enhanced neural network method is used to extract the characteristics of centralized new energy, effectively identifying the operation modes and output characteristics of large-scale new energy power stations; an ensemble learning prediction method is used for multi-step prediction, realizing accurate prediction of new energy output. The application of this technological innovation solves the problem that traditional new energy prediction methods are difficult to simultaneously consider the characteristics of distributed and centralized new energy. Especially under complex meteorological conditions, it can accurately predict the output changes of new energy, providing important support for system dispatching and control, and significantly improving the new energy consumption capacity and system operation efficiency.

[0067] According to one aspect of the present invention, as Figure 6 shown, step S5 is specifically as follows,

[0068] Step S51: Obtain the system inertia feature matrix, the distributed new energy feature matrix, and the centralized new energy feature matrix, and perform multi-time scale decomposition using an improved modal decomposition method to obtain a multi-scale component matrix;

[0069] The modal decomposition method is specifically: signal decomposition S(t) = ∑c(k,t) + r(t); modal component Instantaneous amplitude A(k,t) = |H(c(k,t))|·exp(-α·|dc(k,t) / dt|); instantaneous phase where: H(·) is the Hilbert transform; r(t) is the residual term; α is the attenuation coefficient; k is the modal index; t is the time index.

[0070] Step S52: Obtain the multi-scale component matrix, and use an improved information entropy analysis method to calculate the dynamic correlation relationship between different time scales to obtain a time scale correlation degree matrix;

[0071] The information entropy analysis method is specifically: scale correlation degree R(i,j) = I(i,j)·T(i,j)·E(i,j); mutual information I(i,j) = ∑p(x,y)·log(p(x,y) / (p(x)·p(y))); transfer entropy T(i,j) = ∑p(x_t+1,x_t,y_t)·log(p(x_t+1|x_t,y_t) / p(x_t+1|x_t)); energy correlation E(i,j) = |∑(X(i,t)·X(j,t))| / √(∑X 2 (i,t)·∑X 2(j, t)); where: p(·) is a probability distribution; X(i, t) is a time series; i, j are scale indices; t is a time index.

[0072] Step S53: Obtain the multi-scale component matrix and the time-scale correlation matrix, and perform feature fusion using an adaptive tensor decomposition method to obtain a multi-scale feature tensor.

[0073] The adaptive tensor decomposition method is specifically as follows: Tensor reconstruction X = G ×1 U ×2 V ×3 W + R(t); Core tensor G(t) = X ×1 U^T ×2 V^T ×3 W^T; Dynamic correction R(t) = α·||X - X'||_F·exp(-β·t); Feature matrix U(t) = argmin||X - G ×1 U ×2 V ×3 W||_F + λ·||U||1; where: × n is the n-mode tensor product; X' is the reconstructed tensor; α, β, λ are adjustment parameters; ||·||_F is the Frobenius norm; ||·||1 is the L1 norm; t is a time index.

[0074] It realizes the adaptive decomposition and fusion of multi-time-scale features and establishes a multi-dimensional characterization model of system characteristics. Through the improved mode decomposition method for multi-time-scale decomposition, it can effectively separate the system dynamic characteristics of different time scales; by using the improved information entropy analysis method to calculate the dynamic correlation relationship between time scales, it accurately identifies the mutual influence of different time-scale features; and by using the adaptive tensor decomposition method for feature fusion, it realizes the unified expression of multi-dimensional features. The application of this innovative technology solves the limitation of traditional analysis methods in dealing with multi-time-scale coupling problems. Especially in the case of rapid changes in system dynamic characteristics, it can accurately capture the feature changes of different time scales, provides a comprehensive feature expression for system stability analysis, and significantly improves the accuracy and reliability of the analysis.

[0075] According to one aspect of the present invention, as Figure 7 shown, step S6 is specifically as follows,

[0076] Step S61: Obtain the multi-scale feature tensor, the time-scale correlation matrix, and the inertia change trend index, and construct a system vulnerability model using an improved multi-dimensional evaluation method to obtain a system risk assessment matrix;

[0077] The multi-dimensional evaluation method is specifically as follows: risk index R(i,t) = V(i,t)·S(i,t)·I(i,t); vulnerability V(i,t) = 1 / (1 + exp(-α·|ΔF(t)| / F0)); severity S(i,t) = |ΔP(t)| / P_max·exp(-β·t); correlation degree I(i,t) = ∑(w(i,j)·ρ(i,j,t)·exp(-d(i,j) / r0)); where: ΔF(t) is the frequency deviation; ΔP(t) is the power change; ρ(i,j,t) is the correlation coefficient; d(i,j) is the electrical distance; w(i,j) is the weight coefficient; α and β are adjustment parameters; i and j are node indices; t is the time index.

[0078] Step S62: Obtain the system risk assessment matrix and the new energy prediction matrix, and use the comprehensive evaluation optimization method to determine the load shedding priority order to obtain the load shedding priority sequence.

[0079] The comprehensive evaluation optimization method is specifically as follows: priority P(i,t) = W(t)·L(i,t)·C(i,t); load weight W(t) = exp(-α·R(i,t)) / (1 + β·|dP(i,t) / dt|); location influence L(i,t) = ∑(w(i,j)·exp(-d(i,j) / r0)); constraint condition C(i,t) = 1 / (1 + exp(γ·(V(i,t) - V0))); where: R(i,t) is the risk degree; P(i,t) is the power; V(i,t) is the voltage; d(i,j) is the electrical distance; w(i,j) is the coupling weight; α, β, and γ are adjustment parameters; i and j are node indices; t is the time index.

[0080] Step S63: Obtain the load shedding priority sequence and the system risk assessment matrix, and use the improved swarm intelligence algorithm to optimize the cut-off power distribution plan to obtain the cut-off power quantity matrix.

[0081] The swarm intelligence algorithm is specifically as follows: fitness function F(x) = w1·f1(x) + w2·f2(x) + w3·f3(x); frequency target f1(x) = ∑|ΔF(t)| 2 ·exp(-α·t); power target f2(x) = ∑|ΔP(t)| 2 ·exp(-β·t); voltage target f3(x) = ∑|ΔV(t)| 2·exp(-γ·t); The position is updated as x(t + 1) = x(t) + v(t); The velocity is updated as v(t + 1) = ω·v(t) + c1·r1·(p_best - x(t)) + c2·r2·(g_best - x(t)); where: ΔF(t), ΔP(t), and ΔV(t) are system deviations; w1, w2, and w3 are weight coefficients; α, β, and γ are time decay coefficients; ω is the inertia weight; c1 and c2 are learning factors; r1 and r2 are random numbers; and t is the number of iterations.

[0082] An optimization framework for load shedding strategies based on multi-dimensional risk assessment is established to achieve precise load shedding control. By using an improved multi-dimensional assessment method to construct a system vulnerability model, the weak links of the system can be accurately identified; a comprehensive evaluation optimization method is adopted to determine the priority order of load shedding, effectively balancing system security and user importance; an improved swarm intelligence algorithm is used to optimize the load shedding power distribution scheme to achieve the optimal allocation of the load shedding amount. The application of this innovative technology solves the problem that traditional load shedding schemes are difficult to adapt to the changes in system characteristics under the condition of high proportion of new energy access. Especially under system disturbances and new energy fluctuations, it can achieve precise control of load shedding, significantly improving the stability and reliability of the system.

[0083] According to one aspect of the present invention, as Figure 8 shown, step S7 is specifically as follows

[0084] Step S71: Obtain the load shedding priority sequence, the load shedding power amount matrix, and the voltage level coupling degree index, and use an improved intelligent learning method to construct a hierarchical control strategy to obtain hierarchical control instructions;

[0085] The intelligent learning method is specifically: The control strategy π(s,a) = N(μ(s),σ(s)); The mean function The variance function σ(s) = f_σ(s) / (1 + β·|dF(s) / dt|); The value function V(s) = E[∑γ t r(s_t,a_t)]; The reward function r(s,a) = -(w1·ΔF 2 + w2·ΔP 2 + w3·ΔV 2 ); where: s is the system state; a is the control action; f_μ(·) and f_σ(·) are neural networks; ΔF, ΔP, and ΔV are system deviations; α, β, and γ are adjustment parameters; w1, w2, and w3 are weight coefficients; and t is the time step.

[0086] Step S72: Obtain the hierarchical control instructions and the multi-source preprocessed data set, and use an adaptive assessment method to analyze the control execution effect to obtain control effect evaluation indicators;

[0087] The adaptive evaluation method is specifically as follows: The evaluation index E(t) = P(t)·Q(t)·R(t); The execution accuracy P(t) = exp(-α·||e(t)||2) / (1 + β·|de(t) / dt|); The response quality Q(t) = 1 / (1 + γ·τ(t))·exp(-μ·σ 2 (t)); The reliability R(t) = (1 - exp(-λ·t))·∏(1 - p_f(i,t)); Where: e(t) is the control error; τ(t) is the response time; σ 2 (t) is the fluctuation variance; p_f(i,t) is the failure probability; α, β, γ, μ, λ are adjustment parameters; i is the node index; t is the time index.

[0088] Step S73: Obtain the control effect evaluation index and the system risk assessment matrix, and use the online learning optimization method to adjust the control strategy to obtain the control strategy adjustment parameter and the updated control instruction.

[0089] The online learning optimization method is specifically as follows: Policy update The objective function J(θ) = E[R(t) + α·H(π) - β·D_KL(π_new||π_old)]; The advantage function A(s,a) = Q(s,a) - V(s); The policy gradient The entropy regularization H(π) = -∑π(a|s)·logπ(a|s); Where: π_θ is the policy function; Q(s,a) is the action value function; V(s) is the state value function; D_KL is the KL divergence; η is the learning rate; α, β are adjustment parameters; s is the state; a is the action; t is the time step.

[0090] It realizes the adaptive optimization and dynamic adjustment of the hierarchical control strategy, and establishes a complete control effect evaluation and optimization mechanism. By constructing the hierarchical control strategy through the improved intelligent learning method, the coordinated control between different voltage levels can be realized; The adaptive evaluation method is used to analyze the control execution effect, and the actual effect of the control strategy is accurately evaluated; The online learning optimization method is used to adjust the control strategy, and the real-time optimization of the control strategy is realized. The application of this innovative technology solves the problem that traditional control methods are difficult to adapt to the changes in the dynamic characteristics of the system. Especially in the case of rapid changes in the system state, the dynamic optimization of the control strategy can be realized, significantly improving the accuracy and reliability of the control.

[0091] According to another aspect of the present invention, there is also provided a high-proportion distributed new energy power grid dynamic adaptive low-frequency load shedding system, characterized in that it includes:

[0092] At least one processor; And,

[0093] A memory communicatively connected to at least one of the processors; wherein,

[0094] The memory stores instructions executable by the processor, and the instructions are used to be executed by the processor to implement the high-proportion distributed new energy grid dynamic adaptive low-frequency load shedding method described in any one of the above technical solutions.

[0095] According to one aspect of the present invention, step S11 is specifically as follows:

[0096] Step S111: Perform time-frequency analysis on the original frequency data, original power data, original voltage data, original current data, original new energy output data, and original energy storage state data to obtain a time-frequency feature matrix; extract the main feature components of the signal from the time-frequency feature matrix to obtain a feature component sequence.

[0097] Wherein the time-frequency analysis is specifically: the time-frequency feature matrix TF(t,f) = ∑(x(τ)·h*(τ - t)·e^(-j2πfτ))dτ + λ·R(t,f); where: x(τ) is the input signal; h(t) is the improved adaptive window function, h(t) = g(t)·exp(-αt 2 / σ(t)); g(t) is the Gaussian function; σ(t) is the time-varying window width function, σ(t) = σ0 + β·|dx(t) / dt|; R(t,f) is the regularization term, λ is the regularization coefficient; t is the time variable; f is the frequency variable; α, β are adaptive parameters; is the Laplace operator; ||·||2 is the L2 norm.

[0098] Step S112: According to the data items in the feature component sequence, calculate the change gradient of adjacent data points in the data sequence, perform probability statistics on the change gradient according to a preset interval to generate a probability distribution vector, construct an information entropy matrix based on the probability distribution vector; group the data features in the information entropy matrix according to similarity, establish a feature category index table, and generate a data characteristic classification matrix; segment the data characteristic classification matrix according to feature categories, calculate the minimum information loss sampling interval for each segmented data, and construct a basic sampling parameter set.

[0099] Wherein, the information entropy matrix H(t,k) = -∑(p(i,k)·log2(p(i,k))) + γ·D(t,k); where: p(i,k) is the probability distribution of the k-th type of data within the time window t, p(i,k) = n(i,k) / N(k)·exp(-μ·v(i,k)); n(i,k) is the number of samples of the k-th type of data in the interval i; N(k) is the total number of samples of the k-th type of data; v(i,k) is the data change rate; D(t,k) is the time-varying correction term, γ, μ, and ω are adaptive weight coefficients; is the gradient operator; ||·||1 is the L1 norm; t is the time window index; k is the data category index; i is the probability interval index.

[0100] Step S113: Using a preset length as the window interval, continuously slide on the basic sampling parameter set to calculate the average value and standard deviation of the data change rate within the window, generating a change rate sequence; divide the change rate sequence into multiple time subsequences, establish a local trend model for each subsequence, predict the data change range at the next moment, obtaining a trend prediction matrix; according to the predicted change range in the trend prediction matrix, dynamically calculate the optimal sampling interval at the next sampling moment, forming a dynamic sampling parameter set.

[0101] Step S114: Compare the dynamic sampling parameter set with the frequency components of the time-frequency feature matrix, calculate the ratio of the sampling frequency to the signal frequency, generating a sampling effectiveness index; based on the sampling effectiveness index, establish a sampling parameter optimization objective function, obtaining an optimized sampling parameter set.

[0102] Step S115: According to the sampling moments determined by the optimized sampling parameter set, extract the data at the corresponding moments from the original data, forming an initial sampling data set; for the data missing intervals in the initial sampling data set, calculate the change characteristics of the data at both ends of the interval, and generate supplementary data points according to the change characteristics, obtaining a supplementary sampling data set.

[0103] Step S116: Smooth the abnormal fluctuations in the supplementary sampling data set to obtain a smoothed sampling sequence; perform frequency domain conversion on the smoothed sampling sequence, extract frequency domain characteristic parameters, generating a spectral characteristic matrix; align and combine the smoothed sampling sequence and the spectral characteristic matrix according to the time stamp, obtaining a time series sampling data set.

[0104] By constructing an adaptive time series sampling scheme, the efficient acquisition and preprocessing of key parameters in the power system are realized. Specifically, first, the time-frequency analysis method is used to extract the characteristic components of the original data to obtain the main characteristics of the signal; then, based on the calculation of the data change gradient and information entropy, a sampling parameter optimization model is established to realize the dynamic adjustment of the sampling parameters; through sliding window analysis and local trend prediction, the data change law is accurately grasped; finally, the frequency domain characteristic verification is used to ensure the sampling quality. This innovative sampling method solves the problem that it is difficult for the traditional fixed sampling method to balance the sampling efficiency and data quality. Especially in the case of high proportion of new energy access leading to rapid fluctuations in system parameters, it can adaptively adjust the sampling strategy, avoiding data redundancy and preventing the loss of key information. At the same time, through data supplementation and smoothing processing, the continuity and reliability of the sampling data are ensured, providing a high-quality data basis for subsequent analysis.

[0105] According to one aspect of the present invention, step S12 specifically includes:

[0106] Step S121: Map the data points in the time series sampling data set to the feature space, calculate the local density value and distance value of each data point, and generate a density feature vector; mark the positions of the abnormal points according to the product of the local density value and the distance value in the density feature vector to obtain a local abnormal point set; calculate the abnormal degree values of the points in the local abnormal point set to form an abnormal degree sequence.

[0107] Among them, the local density ρ(i)=∑exp(-(d(i,j) / d k ) 2 )+λ·C(i); the distance calculation d(i,j)=||x(i)-x(j)||2+α·|f(i)-f(j)|; where: x(i) is the feature vector of data point i; f(i) is the frequency feature of data point i; d k is the local truncation distance, is the average distance; σ(d) is the distance standard deviation; C(i) is the connectivity index, C(i)=∑(exp(-d(i,j) / rc)); rc is the connectivity radius; α, β, λ are adaptive parameters; i, j are data point indices.

[0108] Step S122: Select the density distribution parameters of the normal data points from the density feature vector, establish a reference density model, calculate the dynamic density threshold, and obtain a density threshold sequence; perform a weighted combination of the density threshold sequence and the time correlation degree of the data points to generate a comprehensive determination threshold set.

[0109] Step S123: Compare the data points in the abnormal degree sequence with the comprehensive determination threshold set, mark the data positions exceeding the threshold, and generate an initial abnormal marking set; calculate the aggregation degree of adjacent abnormal points in the initial abnormal marking set to obtain an abnormal aggregation degree matrix.

[0110] Step S124: Combine adjacent abnormal points into abnormal intervals according to the aggregation degree values in the abnormal aggregation degree matrix to generate an abnormal interval sequence; calculate the duration and intensity features of each interval in the abnormal interval sequence to obtain an interval feature matrix.

[0111] Step S125: Calculate the credibility scores of each abnormal interval based on the duration and intensity values of the interval feature matrix to generate an abnormal credibility sequence; screen the abnormal credibility sequence according to the score from high to low to obtain an abnormal value marking set.

[0112] Step S126: Calculate the spatial distribution density of the outlier points in the outlier label set to obtain a spatial distribution index; statistically analyze the time distribution characteristics of the outlier points in the outlier label set to generate a time distribution index; combine the spatial distribution index and the time distribution index into a data quality index.

[0113] A data quality evaluation system based on multi-dimensional outlier detection is established, realizing comprehensive quality monitoring of sampled data. By mapping data points to the feature space to calculate local density and distance values, outlier points can be effectively identified; based on a comprehensive decision-making strategy of dynamic density threshold and time correlation, the accuracy of outlier detection is improved; through the analysis of the aggregation degree and credibility scoring of outlier intervals, accurate positioning and hierarchical processing of outlier data are achieved. This innovative outlier detection method solves the problem of high false alarm and miss rate of traditional methods in complex power systems. Especially under system disturbances and new energy fluctuations, it can accurately distinguish real outliers and normal fluctuations. At the same time, through the comprehensive analysis of spatial distribution and time distribution indicators, reliable quantitative indicators are provided for data quality evaluation, significantly improving the efficiency and accuracy of data quality management.

[0114] According to one aspect of the present invention, step S13 is specifically as follows:

[0115] Step S131: Extract the data points marked by the outlier label set from the time-series sampled data set, calculate the neighborhood window size of these data points, and generate an outlier processing window sequence; calculate the change trend parameters of the data for each window in the outlier processing window sequence to obtain a trend parameter matrix.

[0116] Among them, the window size W(t) = W0·exp(-η·V(t)) + ξ·S(t); the change trend parameter The smoothing coefficient S(t) = 1 / (1 + exp(-κ·(SNR(t) - SNR0))); where: W0 is the reference window size; x(t) is the time-series data; SNR(t) is the signal-to-noise ratio; SNR0 is the reference signal-to-noise ratio; η, ξ, κ are adjustment parameters; ε is a small constant to prevent division by zero; t is the time index; is the gradient operator.

[0117] Step S132: Based on the parameter values of each window in the trend parameter matrix, calculate the local smoothing coefficient of the data within the window to generate a smoothing coefficient sequence; perform a weighted average operation on the smoothing coefficient sequence and the data within the corresponding window to obtain a local smoothing sequence.

[0118] Step S133: Replace the local smoothing sequence back to the position of the abnormal point in the time series sampling dataset to generate an initial filtering sequence; calculate the local change rate of the data in the initial filtering sequence to obtain a change rate matrix; set the decomposition scale parameter according to the change rate matrix to generate a decomposition parameter set.

[0119] Step S134: Multilevel decompose the initial filtering sequence according to the scale parameter determined by the decomposition parameter set to obtain a multilevel component sequence; calculate the energy distribution of each component in the multilevel component sequence to generate an energy distribution matrix.

[0120] Step S135: Sort the multilevel component sequence according to the energy values in the energy distribution matrix to screen the main components and obtain a main component sequence; calculate the reconstruction weights of each component in the main component sequence to generate a reconstruction weight vector.

[0121] Step S136: Weight and combine the components in the main component sequence according to the reconstruction weight vector to generate a reconstructed data sequence; calculate the signal-to-noise ratio index of the reconstructed data sequence to obtain a signal-to-noise ratio matrix; verify and fine-tune the reconstruction result according to the signal-to-noise ratio matrix to finally obtain a multi-source preprocessing dataset.

[0122] It realizes the unified optimization of the adaptive processing of abnormal data and signal denoising, and establishes a high-quality data preprocessing system. Through the adaptive window design and local smoothing processing, the influence of abnormal points is effectively removed; based on the multilevel signal decomposition and energy distribution analysis, the accurate identification and separation of noise are realized; through the reconstruction weight optimization and signal-to-noise ratio verification, the reliability of the processing result is guaranteed. This innovative data processing method solves the limitations of traditional methods in processing high-noise and strong-interference data. Especially in a complex power system environment, it can effectively extract useful signals, not only retain the true change characteristics of the data, but also significantly improve the signal quality, providing reliable data support for subsequent analysis and decision-making.

[0123] According to one aspect of the present invention, step S14 is specifically as follows:

[0124] Step S141: Group the data in the multi-source preprocessing dataset according to the voltage level to generate a voltage grouping sequence; calculate the timestamp alignment error of each group of data in the voltage grouping sequence to obtain a time alignment error set; interpolate and supplement the data according to the time alignment error set to generate an aligned data sequence.

[0125] Step S142: Divide the aligned data sequence into multiple time windows, calculate the fluctuation parameters of the data within the window to obtain a fluctuation feature sequence; calculate the stability index of the data of each voltage level according to the fluctuation feature sequence to generate a stability index matrix.

[0126] Step S143: Based on the index values in the stability index matrix, calculate the data correlation between voltage levels to obtain a correlation matrix; extract the main association patterns in the correlation matrix to generate an association pattern set; construct the topological structure of the voltage level data according to the association pattern set to obtain a topological feature sequence.

[0127] Step S144: Project the data in the aligned data sequence into the feature space, calculate the feature vectors of each voltage level to obtain a feature vector group; perform orthogonalization processing on the feature vector group to generate an orthogonal feature basis; calculate the projection coefficients of the data based on the orthogonal feature basis to obtain a projection coefficient matrix.

[0128] Step S145: Calculate the feature contribution degrees of each voltage level according to the coefficient values in the projection coefficient matrix to generate a feature weight sequence; combine the feature weight sequence with the topological feature sequence to obtain a comprehensive feature sequence.

[0129] Step S146: Construct the feature descriptors of each voltage level based on the comprehensive feature sequence to obtain a feature description sequence; correspond the feature description sequence with the voltage level identifiers to generate an identifier mapping table; combine the feature description sequence and the identifier mapping table to finally obtain the voltage level feature matrix.

[0130] Among them, the voltage feature vector The stability index E(i,t) = exp(-λ·σ 2 (v(i,t)) / μ 2 (v(i,t))); the fluctuation feature F(i,t) = ∑|Δv(i,t)| / Δt·exp(-γ·|t - τ|); the correlation matrix M(i,t) = softmax(K(i)·Q(i) T / √d); where: v(i,t) is the voltage time series data; σ 2 (·) is the variance; μ(·) is the mean; K(i), Q(i) are the key-value matrices of the attention mechanism; θ(i,t), is the adaptive weight; λ, γ are the attenuation coefficients; d is the feature dimension; i is the voltage level index; t is the time index.

[0131] A feature extraction and expression system for data of different voltage levels is established, realizing the accurate characterization of voltage level features. Through data alignment and fluctuation characteristic analysis, the dynamic characteristics of each voltage level are accurately grasped; based on correlation analysis and topological feature extraction, the internal connections between voltage levels are revealed; through feature orthogonalization and weight optimization, a complete feature description system is constructed. This innovative feature extraction method solves the problem that traditional methods are difficult to comprehensively characterize voltage level features. Especially in the collaborative analysis of multi-voltage levels, it can accurately identify the feature differences and correlation relationships of each level, providing reliable feature support for subsequent coordinated control, and significantly improving the accuracy of analysis and the pertinence of control.

[0132] According to one aspect of the present invention, step S21 is specifically as follows:

[0133] Step S211: Stratify the feature data in the voltage level feature matrix according to the voltage level to generate a voltage stratification sequence; calculate the time distribution characteristics of the data in each layer of the voltage stratification sequence to obtain a time feature set; construct an inter-layer time sequence correlation graph according to the time feature set to generate a time sequence correlation table;

[0134] Step S212: Extract the topological connection relationships of each voltage level node from the multi-source preprocessed data set to obtain a topological connection sequence; quantify the connection relationships in the topological connection sequence into connection strength values to generate a connection strength matrix; calculate the electrical distance between nodes based on the connection strength matrix to obtain an electrical distance sequence;

[0135] Step S213: Construct an inter-layer coupling strength calculation model according to the time sequence correlation table and the electrical distance sequence to obtain a coupling strength matrix; perform normalization processing on the coupling strength matrix to generate a normalized strength sequence; calculate the inter-layer influence weight based on the normalized strength sequence to obtain an inter-layer weight matrix;

[0136] Step S214: Convert the inter-layer weight matrix into a network connection structure to generate a network structure sequence; calculate the centrality index of each node in the network structure sequence to obtain a centrality index set; determine the key connection path according to the centrality index set to generate a key path sequence;

[0137] Step S215: Establish a hierarchical propagation model based on the key path sequence, calculate the propagation coefficient to obtain a propagation coefficient matrix; combine the propagation coefficient matrix with the coupling strength matrix to generate a comprehensive correlation sequence; perform feature extraction on the comprehensive correlation sequence to obtain a voltage level correlation matrix.

[0138] Among them, the association strength R(i, j, t) = ψ(i, j)·exp(-d(i, j) / r0)·G(i, j, t); the correlation G(i, j, t) of the time-series association = ∑(x(i, τ)·x(j, τ + Δt))·exp(-β·|Δt|); the electrical distance d(i, j) = ∑(Z(k)·I(k)); where: ψ(i, j) is the topological correlation coefficient; r0 is the characteristic distance threshold; x(i, τ) is the state of level i at time τ; Z(k) is the impedance on path k; I(k) is the current on path k; β is the time decay coefficient; i, j are level indices; t is the time index; k is the path index.

[0139] A multi-level association analysis framework between voltage levels is constructed to achieve an accurate expression of the system hierarchical structure. Through time-series association and topological connection analysis, the mutual influences between levels are comprehensively characterized; based on the calculation of the coupling strength and the analysis of the network structure, the action mechanism between levels is quantified; through the identification of key paths and the optimization of the propagation coefficient, the action chain between levels is revealed. This innovative association analysis method solves the limitations of traditional methods in dealing with complex hierarchical structures. Especially in large-scale power systems, it can accurately grasp the interaction relationship between different voltage levels, providing an important basis for system stability analysis and coordinated control, and significantly improving the comprehensiveness and accuracy of the analysis.

[0140] According to one aspect of the present invention, step S22 is specifically as follows:

[0141] Step S221: Divide the association data in the voltage level association matrix according to time windows to obtain a time window sequence; calculate the change rate within the window for the data in the time window sequence to generate a change rate matrix; identify key change points based on the change rate matrix to obtain a change point sequence.

[0142] Step S222: Extract the moment data corresponding to the change point sequence from the multi-source preprocessed data set to generate a critical moment data set; calculate the state parameters of each voltage level in the critical moment data set to obtain a state parameter sequence; construct a state transition probability model according to the state parameter sequence to generate a transition probability matrix.

[0143] Step S223: Calculate the influence delay time between each voltage level based on the transition probability matrix to obtain a delay time sequence; combine the delay time sequence with the change rate matrix to generate a time-varying feature set; perform a causality test on the time-varying feature set to obtain a causality test matrix.

[0144] Step S224: Extract significant causal relationships according to the causality test matrix to generate a causal relationship sequence; calculate the strength values of each relationship in the causal relationship sequence to obtain a causal strength matrix; convert the causal strength matrix into an influence propagation network to generate a propagation network sequence.

[0145] Step S225: Dynamically extract features from the propagation network sequence to obtain a dynamic influence feature matrix.

[0146] Among them, the rate of change matrix time-varying weight Coupling term C(i,t) = ∑(w(i,j)·|V(i,t) - V(j,t)|·exp(-β·d(i,j))); where: V(i,t) is the state quantity of voltage level i at time t; w(i,j) is the coupling weight; d(i,j) is the electrical distance; α, ζ, k, β are adjustment parameters; V0 is the reference threshold; i, j are voltage level indices; t is the time index; is the gradient operator.

[0147] A time-varying analysis framework for the dynamic influence relationship between voltage levels is established, realizing the accurate evaluation of the system's dynamic characteristics. Through time-window analysis and change-point identification, the critical transition moments of the system state are accurately captured; based on state transition probability and influence delay analysis, the temporal characteristics of dynamic influence are quantified; through causality test and propagation network construction, the internal mechanism of influence transmission is revealed. This innovative dynamic analysis method solves the problem that traditional static analysis methods are difficult to reflect the dynamic characteristics of the system. Especially in the case of rapid system state changes, it can accurately identify and quantify the dynamic influence relationship between voltage levels, providing reliable dynamic feature support for system stability assessment, and significantly improving the timeliness and accuracy of analysis.

[0148] According to one aspect of the present invention, step S23 is specifically as follows:

[0149] Step S231: Decompose the dynamic influence feature matrix into different frequency components to generate a frequency component sequence; calculate the energy distribution of each component in the frequency component sequence to obtain an energy distribution matrix; identify the main coupling modes based on the energy distribution matrix to generate a coupling mode set.

[0150] Step S232: Calculate the characteristic parameters for each mode in the coupling mode set to obtain a mode characteristic sequence; sort the mode characteristic sequence according to the energy contribution to generate a characteristic sorting table; construct a coupling degree calculation model based on the characteristic sorting table to obtain a coupling calculation parameter set.

[0151] Step S233: Calculate the coupling strength between each voltage level according to the coupling calculation parameter set to generate a coupling strength sequence; perform normalization processing on the coupling strength sequence to obtain a normalized coupling degree matrix; convert the normalized coupling degree matrix into a feature vector form to generate a voltage level influence feature matrix.

[0152] Step S234: Calculate the comprehensive coupling degree index based on the eigenvalues in the voltage level influence feature matrix to obtain the voltage level coupling degree index.

[0153] The accurate evaluation and quantitative expression of the coupling characteristics between voltage levels are realized, and a complete coupling degree evaluation system is established. Through frequency component decomposition and energy distribution analysis, the main coupling modes are identified; based on mode feature ranking and coupling degree calculation, the coupling intensity is quantified; through feature matrix construction and comprehensive evaluation, the systematic expression of coupling characteristics is realized. This innovative coupling evaluation method solves the limitations of traditional methods in evaluating complex coupling relationships. Especially in multi-voltage level systems, it can accurately identify and quantify the coupling degree between each level, providing an important basis for system coordinated control, and significantly improving the pertinence and effectiveness of control.

[0154] According to one aspect of the present invention, step S31 is specifically as follows:

[0155] Step S311: Pair the frequency and power data in the multi-source preprocessed dataset to generate a frequency-power sequence; perform segmentation processing on the frequency-power sequence to obtain a segmented dataset; calculate the frequency change rate of each segment in the segmented dataset to generate a frequency change rate sequence.

[0156] Step S312: Establish a power perturbation model based on the frequency change rate sequence to obtain a power perturbation matrix; project the data in the power perturbation matrix into the state space to generate a state space sequence; calculate the system response characteristics according to the state space sequence to obtain a response characteristic set.

[0157] Step S313: Extract frequency response parameters from the response characteristic set to generate a response parameter matrix; construct an inertia calculation model according to the response parameter matrix to obtain a calculation model sequence; perform parameter optimization on the calculation model sequence to generate an optimized parameter set.

[0158] Step S314: Calculate the instantaneous inertia of the system based on the optimized parameter set to obtain an instantaneous inertia estimation value.

[0159] Among them, the characteristic of power response P(t) = H(s)·ΔF(s) + Φ(t); the transfer function H(s) = K·(1 + sT1) / (1 + sT2)·exp(-τs)·[1 + ψ(t)·G(s)]; the correction term Φ(t) = μ·∑(P(t - k)·exp(-λ·k))·η(t); where: ΔF(s) is the Laplace transform of the frequency deviation; ψ(t) is the time-varying gain; G(s) is the compensation link; η(t) is the adaptive coefficient; K is the proportional coefficient; T1, T2 are time constants; τ is the time delay; μ, λ are weight coefficients; t is the time index; s is the Laplace operator; k is the historical time index.

[0160] An inertia real-time estimation method based on frequency response characteristics is constructed, realizing the accurate evaluation of the system's dynamic characteristics. Through frequency-power pairing analysis and power disturbance modeling, the dynamic response characteristics of the system are accurately described; based on state space mapping and response characteristic extraction, the inertia characteristics of the system are quantified; through calculation model optimization and parameter adaptive adjustment, the accuracy of estimation is improved. This innovative inertia estimation method solves the problem of insufficient estimation accuracy of traditional methods under the condition of high proportion of new energy access. Especially in the case of system disturbance, it can accurately evaluate the instantaneous inertia characteristics of the system, providing a reliable decision-making basis for under-frequency load shedding control and significantly improving the accuracy and reliability of control.

[0161] According to one aspect of the present invention, step S32 is specifically as follows:

[0162] Step S321: Recombine the instantaneous inertia estimation values according to the time series to generate an inertia time series matrix; divide the inertia time series matrix into time windows to obtain a window sequence set; calculate the statistical characteristics of each window in the window sequence set to generate a window feature matrix.

[0163] Step S322: Extract influence factors from the voltage level influence feature matrix to obtain an influence factor sequence; fuse the influence factor sequence with the window feature matrix to generate a fused feature set; perform dimensionality reduction processing on the fused feature set to obtain a dimensionality-reduced feature sequence.

[0164] Step S323: Construct a prediction model based on the dimensionality-reduced feature sequence to generate a prediction model parameter set; use the prediction model parameter set to calculate the trend prediction values to obtain a trend prediction sequence; perform error analysis on the trend prediction sequence to generate an error index matrix.

[0165] Step S324: Optimize the prediction result according to the error index matrix to obtain an optimized prediction sequence; convert the optimized prediction sequence into a trend index to generate an inertia change trend index.

[0166] Among them, the predicted value of inertia J(t + Δt) = f(X(t))·B(t) + ω·M(t); the feature mapping f(X(t)) = σ(W1·X(t) + b1)·σ(W2·X(t) + b2); the confidence Correction term M(t) = ∑(α(k)·ε(t - k))·exp(-γ·k); where: X(t) is the input feature vector; W1, W2 are weight matrices; b1, b2 are bias vectors; σ(·) is the activation function; ε(t) is the historical error; α(k) is the error weight; ω, γ are adjustment parameters; J0 is the reference inertia; t is the time index; k is the historical time index.

[0167] A prediction and evaluation system for the trend of system inertia change is established to achieve dynamic prediction of inertia characteristics. By reorganizing the inertia time series and extracting window features, the change law of inertia is captured; based on the fusion of influencing factors and feature dimensionality reduction processing, key prediction features are extracted; through the optimization of deep learning models and error compensation, the prediction accuracy is improved. This innovative trend prediction method solves the problem that traditional methods are difficult to accurately predict the change of system inertia. Especially under the condition of new energy output fluctuation, it can accurately predict the change trend of system inertia, providing forward-looking support for system stability control and significantly improving the predictability and initiative of control.

[0168] According to one aspect of the present invention, step S33 is specifically as follows:

[0169] Step S331: Combine the instantaneous inertia estimation value and the inertia change trend index to generate a combined feature sequence; perform normalization processing on the combined feature sequence to obtain a normalized feature set; calculate the correlation between features in the normalized feature set to generate a feature correlation matrix.

[0170] Step S332: Construct a feature space based on the feature correlation matrix to obtain a feature space parameter set; convert the feature space parameter set into orthogonal basis vectors to generate an orthogonal basis sequence; calculate feature projection coefficients according to the orthogonal basis sequence to obtain a projection coefficient matrix.

[0171] Step S333: Perform feature reconstruction on the projection coefficient matrix to generate a reconstructed feature sequence; correspond the reconstructed feature sequence to the physical characteristics of inertia to obtain a physical feature matrix; construct a feature description model based on the physical feature matrix to generate a system inertia feature matrix.

[0172] The multi-dimensional expression and feature fusion of system inertia characteristics are realized, and a complete inertia feature description system is established. Through feature combination and normalization processing, the unified expression of features is realized; based on the construction of the feature space and orthogonal basis transformation, the feature expression form is optimized; through feature reconstruction and physical property mapping, the physical meaning of features is guaranteed. This innovative feature expression method solves the problem of one-sidedness of traditional methods in inertia feature description. Especially under complex system conditions, it can comprehensively describe the inertia characteristics of the system, providing reliable feature support for system stability analysis and significantly improving the comprehensiveness and accuracy of analysis.

[0173] According to one aspect of the present invention, step S41 is specifically as follows:

[0174] Step S411: Extract the output data of distributed new energy nodes from the multi-source preprocessed dataset to generate a distributed output sequence; divide the distributed output sequence according to the time scale to obtain a time scale matrix; calculate the statistical parameters of each scale in the time scale matrix to generate a statistical feature set.

[0175] Step S412: Identify the output fluctuation patterns based on the statistical feature set to obtain a fluctuation pattern sequence; quantify the pattern parameters in the fluctuation pattern sequence to generate a pattern parameter matrix; perform clustering analysis on the pattern parameter matrix to obtain a clustering feature set.

[0176] Step S413: Construct a spatial association network of distributed nodes based on the clustering feature set to generate a spatial association sequence; calculate the influence intensity between nodes in the spatial association sequence to obtain an influence intensity matrix; extract key node features according to the influence intensity matrix to generate a node feature set.

[0177] Step S414: Combine the node feature set and the pattern parameter matrix to obtain a combined feature sequence; perform feature dimensionality reduction on the combined feature sequence to generate a dimensionality-reduced feature matrix; construct a feature description model based on the dimensionality-reduced feature matrix to obtain a distributed new energy feature matrix.

[0178] Among them, the extraction algorithm for distributed features: Feature vector F(i,t) = T(i,t)·S(i,t)·D(i,t); temporal features Spatial feature S(i,t) = ∑(w(i,j)·GCN(A(t))·x(j,t)); Dynamic association D(i,t) = softmax(Q(i,t)·K(i,t) T / √d)·V(i,t); where: x(i,t) is the state of node i at time t; LSTM(·) is a long short-term memory network; GCN(·) is a graph convolutional network; A(t) is the adjacency matrix; w(i,j) is the spatial weight; Q(i,t), K(i,t), V(i,t) are attention mechanism parameters; λ is the temporal decay coefficient; d is the feature dimension; i, j are node indices; t is the time index.

[0179] A deep feature extraction framework for distributed new energy nodes is established, realizing the accurate characterization of distributed characteristics. Through multi-scale analysis of output data and fluctuation pattern recognition, the output characteristics are accurately grasped; based on the construction of a spatial association network and node influence evaluation, the interaction between distributed nodes is revealed; through feature dimensionality reduction and description model construction, efficient expression of features is achieved. This innovative feature extraction method solves the problem that traditional methods are difficult to effectively process large-scale distributed new energy features. Especially in the case of a large number of nodes and wide distribution, it can accurately identify and express the operating characteristics of each node and their mutual influence, providing reliable feature support for new energy consumption and system regulation, and significantly improving the efficiency and accuracy of new energy management.

[0180] According to one aspect of the present invention, step S42 is specifically as follows:

[0181] Step S421: Extract the data of centralized new energy power stations from the multi-source preprocessed dataset to generate a power station data sequence; classify the power station data sequence according to the operating status to obtain a status classification matrix; calculate the characteristic parameters of each category in the status classification matrix to generate a characteristic parameter set.

[0182] Step S422: Perform time series decomposition on the characteristic parameter set to obtain a decomposition sequence set; calculate the spectral characteristics of each sequence in the decomposition sequence set to generate a spectral characteristic matrix; extract the main frequency components according to the spectral characteristic matrix to obtain a main frequency component sequence.

[0183] Step S423: Construct a power station characteristic model based on the main frequency component sequence to generate a characteristic model sequence; optimize the parameters of the characteristic model sequence to obtain an optimized parameter set; calculate the operating characteristics of the power station according to the optimized parameter set to generate an operating characteristic matrix.

[0184] Step S424: Correlate the operating characteristic matrix with the physical parameters of the power station to obtain a physical characteristic sequence; perform feature fusion on the physical characteristic sequence to generate a fusion feature set; construct a feature expression model based on the fusion feature set to obtain a centralized new energy characteristic matrix.

[0185] Among them, the spectrum analysis method: the spectral characteristic matrix S(f,t) = |STFT(x(t))| 2 + Ψ(f,t); the time-frequency compensation term Ψ(f,t) = α·|W(f,t)|·exp(-β·|f - f0|); the wavelet transform W(f,t) = ∫x(τ)·ψ((τ - t)·f)dτ; where: STFT(·) is the short-time Fourier transform; ψ(·) is the wavelet basis function; α, β are adjustment parameters; f0 is the reference frequency; x(t) is the input signal; f is the frequency variable; t is the time variable; τ is the integration variable.

[0186] It realizes the in-depth mining and expression of the characteristics of centralized new energy power stations and establishes a complete power station characteristic analysis system. Through operating status classification and characteristic parameter extraction, the operating characteristics of the power station are accurately characterized; based on time series decomposition and spectrum analysis, key frequency characteristics are identified; through characteristic model optimization and physical parameter mapping, the practical significance of the characteristics is ensured. This innovative characteristic analysis method solves the limitations of traditional methods in dealing with the characteristics of large-scale new energy power stations. Especially under complex meteorological conditions, it can accurately grasp the operating rules and output characteristics of the power station, providing strong support for the dispatching management of the power station and significantly improving the controllability and predictability of the power station operation.

[0187] According to one aspect of the present invention, step S43 is specifically as follows:

[0188] Step S431: Align the features of the distributed new energy feature matrix and the centralized new energy feature matrix to generate an aligned feature sequence; calculate the correlation between features in the aligned feature sequence to obtain a correlation matrix; screen key features based on the correlation matrix to generate a key feature set.

[0189] Step S432: Construct multiple basic prediction models according to the key feature set to obtain a prediction model sequence; train the parameters of the models in the prediction model sequence to generate a model parameter set; calculate the prediction results of each model based on the model parameter set to obtain an initial prediction sequence.

[0190] Step S433: Conduct error analysis on the prediction results of each model in the initial prediction sequence to generate an error index set; calculate the model weights according to the error index set to obtain a weight coefficient matrix; use the weight coefficient matrix to integrate the prediction results to obtain a new energy prediction matrix.

[0191] Among them, the integrated prediction algorithm is specifically: prediction result Y(t) = ∑(w(i,t)·y(i,t))·B(t); dynamic weight w(i,t) = exp(-λ·e(i,t)) / ∑exp(-λ·e(i,t)); error function e(i,t) = ||y(i,t)-y*(t)||2·(1 + γ·|dy(i,t) / dt|); confidence level B(t) = 1 / (1 + exp(-κ·(C(t)-C0))); correlation index C(t) = ∑(ρ(i,j,t)·w(i,t)·w(j,t)); where: y(i,t) is the prediction value of the i-th model; y*(t) is the actual value; ρ(i,j,t) is the correlation coefficient between models; λ, γ, κ are adjustment parameters; C0 is the reference confidence level; i, j are model indices; t is the time index.

[0192] A collaborative prediction framework for distributed and centralized new energy is constructed, realizing accurate prediction of new energy output. Through feature alignment and correlation analysis, different types of new energy features are fused; based on multi-model integration and parameter optimization, the prediction accuracy is improved; through error analysis and weight adjustment, the dynamic optimization of prediction results is realized. This innovative prediction method solves the problem that traditional methods are difficult to uniformly handle the prediction of different types of new energy. Especially in the case of a high proportion of new energy, it can accurately predict the change of the overall new energy output of the system, providing reliable prediction support for system dispatching decisions, and significantly improving the efficiency of new energy consumption and the stability of system operation.

[0193] According to one aspect of the present invention, step S51 is specifically:

[0194] Step S511: Align the system inertia feature matrix, distributed new energy feature matrix, and centralized new energy feature matrix in the time dimension to generate a time-aligned sequence; perform data standardization on the time-aligned sequence to obtain a standardized data set; calculate the signal energy distribution based on the standardized data set to generate an energy distribution sequence.

[0195] Step S512: Determine the decomposition scale number according to the energy distribution sequence to obtain a set of scale parameters; perform wavelet decomposition on the standardized data set according to the set of scale parameters to generate a wavelet component sequence; calculate the frequency characteristics of each component in the wavelet component sequence to obtain a frequency characteristic matrix.

[0196] Step S513: Perform modal identification on the components in the frequency characteristic matrix to generate a modal sequence; group the modal sequence according to the frequency bandwidth to obtain a grouped feature set; perform feature reconstruction on the grouped feature set to generate a multi-scale component matrix.

[0197] In the wavelet decomposition algorithm: the decomposition coefficient D(j,k) = ∑(x(t)·ψ(j,k)(t)) + Φ(j,k); the adaptive wavelet basis ψ(j,k)(t) = 2^(j / 2)·ψ(2^j·t - k)·exp(-α·|E(j,k)|); the energy compensation term Φ(j,k) = β·|W(j,k)|·exp(-μ·|j - j0|); the energy feature E(j,k) = |D(j,k)| 2 / ∑|D(j,·)| 2 ; where: x(t) is the input signal; W(j,k) is the wavelet transform coefficient; α, β, μ are adjustment parameters; j0 is the reference scale; j is the scale index; k is the translation index; t is the time variable.

[0198] A time-scale decomposition framework for multi-source features is established, realizing the multi-scale expression of system characteristics. Through time alignment and energy distribution analysis, the accuracy of feature decomposition is ensured; based on wavelet decomposition and modal identification, feature components at different time scales are extracted; through frequency feature analysis and grouped reconstruction, the integrity of features is guaranteed. This innovative decomposition method solves the limitations of traditional methods in dealing with multi-time-scale coupling features. Especially in the case of complex system dynamic characteristics, it can accurately separate and identify features at different time scales, providing multi-dimensional feature support for system analysis and significantly improving the depth and accuracy of analysis.

[0199] According to one aspect of the present invention, step S52 is specifically as follows:

[0200] Step S521: Calculate the information entropy for each scale component in the multi-scale component matrix to generate an information entropy sequence; divide the information entropy sequence by time windows to obtain a window entropy value matrix; calculate the entropy value change rate for each window in the window entropy value matrix to generate an entropy value change sequence.

[0201] Step S522: Establish an inter-scale transfer model based on the entropy value change sequence to obtain a set of transfer model parameters; calculate the information flow between scales according to the set of transfer model parameters to generate an information flow matrix; perform normalization processing on the information flow matrix to obtain a normalized flow sequence.

[0202] Step S523: Convert the normalized flow sequence into a correlation strength index to generate a correlation strength sequence; perform matrix processing on the correlation strength sequence to obtain a time-scale correlation degree matrix.

[0203] The accurate evaluation of the dynamic correlation relationship between different time scales is realized, and a complete scale correlation analysis system is established. Through information entropy calculation and window division, the information characteristics within the scale are quantified; based on the inter-scale transfer model and information flow analysis, the action mechanism between scales is revealed; through correlation strength calculation and matrix processing, the systematic expression of correlation characteristics is realized. This innovative correlation analysis method solves the problem that traditional methods are difficult to accurately describe the complex correlations between multiple time scales. Especially in the case of rapid changes in system characteristics, it can accurately identify and quantify the mutual influences between different time scales, providing reliable feature support for system stability analysis and significantly improving the comprehensiveness and accuracy of the analysis.

[0204] According to one aspect of the present invention, step S53 is specifically as follows:

[0205] Step S531: Reorganize the multi-scale component matrix according to the time-scale correlation degree matrix to generate a reorganized feature sequence; construct a tensor structure for the reorganized feature sequence to obtain an initial feature tensor; calculate the core tensor of the initial feature tensor to generate a core feature set.

[0206] Step S532: Extract the main modes based on the core feature set to obtain a main mode sequence; perform feature enhancement on the main mode sequence to generate an enhanced feature matrix; combine the enhanced feature matrix with the time dimension information to obtain a spatio-temporal feature sequence.

[0207] Step S533: Reconstruct the spatio-temporal feature sequence into a tensor form to generate a reconstructed tensor sequence; perform dimensionality reduction optimization on the reconstructed tensor sequence to obtain an optimized feature set; construct a final feature expression based on the optimized feature set to generate a multi-scale feature tensor.

[0208] In the tensor decomposition algorithm: the feature tensor The core tensor G = T ×1 U ×2 V ×3 W; the residual correction R(t) = η·||T - T'||_F·exp(-γ·t); the reconstruction error E = ||T - G ×1 U ×2 V ×3 W||_F; where: λ k is the singular value; u(k), v(k), w(k) are the eigenvectors; is the tensor outer product; × n is the n-mode tensor product; T' is the reconstructed tensor; η, γ are the adjustment parameters; t is the time index; k is the component index; ||·||_F is the Frobenius norm.

[0209] A tensor fusion framework with multi-scale features is established to achieve a unified expression of system characteristics. Through feature recombination and tensor construction, the structural integrity of features is ensured; based on principal mode extraction and feature enhancement, the expression ability of features is improved; through tensor dimensionality reduction and optimized reconstruction, efficient expression of features is achieved. This innovative feature fusion method solves the curse of dimensionality problem in traditional methods when dealing with high-dimensional features. Especially in the case of needing to comprehensively analyze features in multiple dimensions, it can effectively reduce the dimension of features and retain information, providing a compact and comprehensive feature expression for system analysis, and significantly improving the efficiency and usability of feature processing.

[0210] According to one aspect of the present invention, step S61 is specifically as follows:

[0211] Step S611: Expand the multi-scale feature tensor in the time dimension to generate a time feature sequence; reconstruct the time feature sequence according to the time scale correlation matrix to obtain a reconstructed feature set; associate the inertia change trend index with the reconstructed feature set to generate a combined feature matrix.

[0212] Step S612: Perform sensitivity analysis on the features in the combined feature matrix to obtain a sensitivity sequence; construct a risk assessment index based on the sensitivity sequence to generate an assessment index set; classify the assessment index set according to the influence degree to obtain a hierarchical index matrix.

[0213] Step S613: Establish a vulnerability assessment model according to the hierarchical index matrix to generate an assessment model parameter set; calculate the system risk degree using the assessment model parameter set to obtain a risk degree sequence; perform quantization processing on the risk degree sequence to generate a system risk assessment matrix.

[0214] The process of risk assessment is specifically as follows: Risk index R(i,t) = θ(t)·P(i,t)·S(i,t)·C(i,t); Probability term P(i,t) = 1 / (1 + exp(-α·(V(i,t) - V0))); Severity S(i,t) = exp(β·|ΔF(i,t)| / F0)·(1 + μ·|dF(i,t) / dt|); Correlation degree C(i,t) = ∑(w(i,j)·exp(-d(i,j) / r0)·ρ(i,j,t)); where: θ(t) is the time-varying weight; V(i,t) is the voltage value; ΔF(i,t) is the frequency deviation; w(i,j) is the node weight; d(i,j) is the electrical distance; ρ(i,j,t) is the correlation coefficient; α, β, μ are adjustment parameters; V0, F0, r0 are reference values; i, j are node indices; t is the time index.

[0215] A system risk assessment system based on multi-scale features is constructed to achieve precise quantification of risks. Through feature reconstruction and sensitivity analysis, key risk factors are identified; based on the construction and hierarchical processing of evaluation indicators, risk grading is achieved; through vulnerability model optimization and risk quantification, the accuracy of the assessment is ensured. This innovative risk assessment method solves the problem of one-sidedness of traditional methods in complex system risk assessment. Especially in the case of coupling of multiple risk factors, it can accurately identify and evaluate the overall risk level of the system, providing a reliable risk assessment basis for control decisions and significantly improving the effectiveness and reliability of risk management.

[0216] According to one aspect of the present invention, step S62 is specifically as follows:

[0217] Step S621: Corresponding the risk indicators in the system risk assessment matrix with the new energy prediction matrix to generate a risk prediction sequence; performing spatio-temporal analysis on the risk prediction sequence to obtain a spatio-temporal distribution matrix; identifying key risk regions based on the spatio-temporal distribution matrix to generate a risk region set.

[0218] Step S622: Constructing a load assessment model according to the risk region set to obtain a load assessment parameter set; calculating the importance of each load node using the load assessment parameter set to generate an importance sequence; matching the importance sequence with the system constraint conditions to obtain a constraint matching matrix.

[0219] Step S623: Establishing a priority ranking model based on the constraint matching matrix to generate a ranking model sequence; performing multi-objective optimization on the ranking model sequence to obtain a load shedding priority sequence.

[0220] A load priority evaluation system based on risk prediction is established to optimize the determination of the shedding sequence. Through risk prediction and spatio-temporal analysis, critical risk areas are accurately identified; based on load assessment and constraint matching, the importance of load nodes is quantified; through multi-objective optimization and sorting model construction, the rationality of the shedding sequence is ensured. This innovative priority evaluation method solves the problem of incomplete consideration of factors in the determination of load shedding priority by traditional methods. Especially in the case of complex system conditions, it can comprehensively consider various influencing factors, provide an optimized execution sequence for load shedding, and significantly improve the pertinence and effectiveness of shedding control.

[0221] According to one aspect of the present invention, step S63 is specifically as follows:

[0222] Step S631: Convert the load shedding priority sequence into a load distribution feature to generate a load distribution sequence; calculate the shedding demand of each region according to the system risk assessment matrix to obtain a shedding demand matrix; match the shedding demand matrix with the load distribution sequence to generate an initial allocation scheme set.

[0223] Step S632: Construct an optimization objective function for the initial allocation scheme set to obtain an objective function sequence; design constraint conditions based on the objective function sequence to generate a constraint condition set; convert the constraint condition set into optimization parameters to obtain an optimization parameter matrix.

[0224] Step S633: Perform iterative optimization calculations according to the optimization parameter matrix to generate an iterative result sequence; perform convergence analysis on the iterative result sequence to obtain a convergence solution set; convert the convergence solution set into a shedding power allocation scheme to generate a shedding power quantity matrix.

[0225] Among them, the objective function J = ∑(w1·J1 + w2·J2 + w3·J3); the frequency objective J1 = ∑(|ΔF(t)| 2 ·exp(-α·t)); the power objective J2 = ∑(|ΔP(t)| 2 ·exp(-β·t)); the stability objective J3 = ∑(|dV(t) / dt| 2 ·exp(-γ·t)); constraint conditions: g1(x) = ΔP(t) - ΔP_max ≤ 0; g2(x) = |dF(t) / dt| - k_max ≤ 0; where: ΔF(t) is the frequency deviation; ΔP(t) is the power change; V(t) is the voltage value; w1, w2, w3 are weight coefficients; α, β, γ are time decay coefficients; ΔP_max is the maximum power change; k_max is the maximum change rate; t is the time index.

[0226] Intelligent optimization of load shedding power distribution is achieved, and a complete optimization system for the shedding scheme is established. Through load distribution analysis and shedding demand calculation, the shedding capacity of each area is accurately determined; based on the construction of the objective function and the design of constraint conditions, global optimization of the shedding scheme is realized; through iterative calculation and convergence analysis, the reliability of the optimization result is ensured. This innovative power distribution method solves the problem that traditional methods are difficult to balance multiple objectives in shedding capacity distribution. Especially in the case of complex system constraints, it can find the optimal shedding scheme that meets multiple constraints, provides a reliable execution scheme for the stable control of the system, and significantly improves the accuracy and coordination of shedding control.

[0227] According to one aspect of the present invention, step S71 is specifically as follows:

[0228] Step S711: Group the load shedding priority sequence according to voltage levels to generate a voltage grouping sequence; calculate the shedding capacity of each group according to the shedding power quantity matrix to obtain a capacity distribution matrix; correspond the voltage level coupling degree index with the capacity distribution matrix to generate a coupling distribution sequence.

[0229] Step S712: Establish a hierarchical execution model for the data in the coupling distribution sequence to obtain an execution model sequence; calculate the execution time sequence of each level according to the execution model sequence to generate an execution time sequence matrix; determine the control priority based on the execution time sequence matrix to obtain a control priority set.

[0230] Step S713: Convert the control priority set into specific control strategies to generate a control strategy sequence; perform time sequence coordination on the control strategy sequence to obtain a coordination strategy matrix; generate an execution instruction based on the coordination strategy matrix to obtain a hierarchical control instruction.

[0231] In S711, the hierarchical grouping control is specifically as follows: the control instruction C(i,t) = K(i,t)·E(i,t)·M(i,t) + Φ(i,t); the dynamic gain K(i,t) = K0·exp(-α·|dF(t) / dt|) / (1 + β·||V(t) - V0||2); the execution priority E(i,t) = exp(-μ·R(i,t))·(1 + γ·|dP(i,t) / dt|); the coordination term M(i,t) = ∑(w(i,j)·ρ(i,j,t)·exp(-d(i,j) / r0)); the compensation term Φ(i,t) = η·∑(ε(t - k)·exp(-λ·k)); where: F(t) is the frequency; V(t) is the voltage vector; P(i,t) is the power; R(i,t) is the risk degree; ρ(i,j,t) is the inter-layer coupling degree; ε(t) is the control error; w(i,j) is the coupling weight; d(i,j) is the electrical distance; α, β, μ, γ, η, λ are adjustment parameters; K0, V0, r0 are reference values; i, j are hierarchical indexes; t is the time index; k is the historical time index.

[0232] A hierarchical control strategy system based on voltage levels is constructed to achieve the coordinated execution of control instructions. Through voltage grouping and capacity allocation, the control levels are reasonably divided; based on the construction of the execution model and time sequence coordination, the execution order of the control strategy is optimized; through the setting of control priorities and strategy conversion, the orderliness of control is ensured. This innovative hierarchical control method solves the limitations of traditional methods in the coordinated control of multiple voltage levels. Especially in the case of a complex system structure, it can achieve the coordinated control between different voltage levels, providing a reliable execution mechanism for the stable control of the system and significantly improving the coordination and effectiveness of control.

[0233] According to one aspect of the present invention, step S72 is specifically as follows:

[0234] Step S721: Decompose the hierarchical control instruction into an execution sequence to generate an execution sequence set; extract the system state at the corresponding moment from the multi-source preprocessed dataset to obtain a state snapshot sequence; correspond the execution sequence set with the state snapshot sequence to generate an execution state matrix.

[0235] Step S722: Calculate the control response characteristics based on the execution state matrix to obtain a response characteristic sequence; perform time dimension analysis on the response characteristic sequence to generate a time response matrix; calculate the response deviation in the time response matrix to obtain a deviation index set.

[0236] Step S723: Convert the deviation index set into evaluation parameters to generate an evaluation parameter sequence; perform a comprehensive evaluation on the evaluation parameter sequence to obtain a control effect evaluation index.

[0237] Among them, the analysis process of the response characteristics is specifically as follows: the response characteristic R(t) = H(t)·D(t)·Q(t); the transfer characteristic H(t) = (1 + sT1) / (1 + sT2)·exp(-τ·s)·[1 + ψ(t)·G(s)]; the dynamic characteristic D(t) = exp(-α·|d 2 F(t) / dt 2 |); the quality factor Q(t) = (1 - exp(-γ·t))·exp(-μ·||ε(t)||2); where: F(t) is the frequency response; ε(t) is the control error; ψ(t) is the time-varying gain; G(s) is the compensation link; T1, T2 are time constants; τ is the time delay; α, β, γ, μ are adjustment parameters; t is the time variable; s is the Laplace operator.

[0238] A real-time evaluation system for the control effect is established to achieve dynamic supervision of the control execution. By performing execution sequence decomposition and state snapshot acquisition, the control execution process is accurately tracked; based on response characteristic analysis and time dimension evaluation, the control effect is quantified; through deviation calculation and comprehensive evaluation, an objective evaluation of the control effect is achieved. This innovative evaluation method solves the problem of lack of real-time performance in the traditional method for evaluating the control effect. Especially in the case of dynamic changes in the control process, it can timely detect and evaluate the control deviation, provide a reliable basis for the adjustment of the control strategy, and significantly improve the adaptability and reliability of the control.

[0239] According to one aspect of the present invention, step S73 is specifically as follows:

[0240] Step S731: Compare the control effect evaluation index with the system risk evaluation matrix to generate an effect deviation sequence; perform factor decomposition on the effect deviation sequence to obtain a deviation factor matrix; establish an optimization objective based on the deviation factor matrix to generate an optimization objective set.

[0241] Step S732: Construct a strategy adjustment model according to the optimization objective set to obtain an adjustment model sequence; perform parameter optimization on the adjustment model sequence to generate an optimization parameter set; convert the optimization parameter set into an adjustment strategy to obtain the control strategy adjustment parameter.

[0242] Step S733: Modify the original control instruction based on the control strategy adjustment parameter to obtain a modified instruction sequence; perform feasibility verification on the modified instruction sequence to generate a verification result set; generate a final control instruction according to the verification result set to obtain the updated control instruction.

[0243] Among them, the optimization objective \(L(\theta)=E[R(t)+\lambda\cdot H(a)-\beta\cdot D_{KL}(\pi_{\theta}||\pi_{\theta_{old}})]\); the reward function \(R(t)=\sum(w_1\cdot r_f + w_2\cdot r_p + w_3\cdot r_v)\cdot\exp(-\alpha\cdot t)\); the frequency reward \(r_f = -|\Delta F(t)|\) 2 \(\cdot(1 + \mu\cdot|dF(t) / dt|)\); the power reward \(r_p = -|\Delta P(t)|\) 2 \(\cdot\exp(-\gamma\cdot|t - t_0|)\); the voltage reward \(r_v = -||V(t)-V_0||^2\) 2 ; the entropy term \(H(a)=-\sum\pi_{\theta}(a|s)\cdot\log(\pi_{\theta}(a|s))\); where: \(\pi_{\theta}\) is the policy function; \(D_{KL}\) is the KL divergence; \(\Delta F(t)\) is the frequency deviation; \(\Delta P(t)\) is the power change; \(V(t)\) is the voltage vector; \(w_1, w_2, w_3, \lambda, \beta, \alpha, \mu, \gamma\) are weight coefficients; \(V_0\) is the rated voltage; \(t_0\) is the reference time; \(t\) is the time index.

[0244] Optimization parameters Objective gradient The advantage function \(A(s,a)=Q(s,a)-V(s)\); Dynamic correction Momentum term where: \(\pi_{\theta}\) is the policy function; \(Q(s,a)\) is the action value function; \(V(s)\) is the state value function; \(\eta\) is the learning rate; \(\alpha, \beta, \gamma\) are adjustment parameters; \(\theta_0\) is the initial parameter; \(t\) is the time index; \(s\) is the state; \(a\) is the action.

[0245] The online optimization and dynamic adjustment of the control strategy are realized, and a complete control optimization system is established. Through the analysis of the effect deviation and factor decomposition, the reasons for the control deviation are accurately identified; based on the policy adjustment model and parameter optimization, the dynamic optimization of the control strategy is realized; through the instruction correction and feasibility verification, the reliability of the adjustment scheme is guaranteed. This innovative strategy optimization method solves the problem of the lag in the response of traditional methods in the control strategy adjustment. Especially in the case of rapid changes in system conditions, it can realize the rapid adjustment and optimization of the control strategy, providing a flexible adjustment mechanism for the stable control of the system, and significantly improving the adaptability and reliability of the control.

[0246] In another embodiment of the present invention,

[0247] Step S12 can also be: Obtain a time-series sampling data set from power system equipment, analyze the time-series sampling data set using a multi-dimensional anomaly detection method based on deep learning to obtain a data quality index and an outlier marker set; According to the data quality index, use an adaptive threshold method to evaluate the data reliability of the time-series sampling data set to obtain a data reliability score matrix; Based on the data reliability score matrix, use a weighted fusion method to adaptively adjust the weight parameters of subsequent data processing steps to obtain a set of processing weight parameters.

[0248] Step S23 can also be: Obtain the dynamic influence feature matrix obtained in step S22 and the set of processing weight parameters obtained in step S12, analyze the coupling relationship between voltage levels using an improved fuzzy comprehensive evaluation method to obtain a voltage level influence feature matrix and a voltage level coupling degree index; Based on the voltage level coupling degree index, use the analytic hierarchy process to calculate the dynamic stability evaluation values of each voltage level to obtain a voltage level stability matrix; According to the voltage level stability matrix, use a dynamic programming algorithm to optimize the system operation parameters to obtain a set of system operation optimization parameters.

[0249] Step S31 can also be: Obtain the multi-source preprocessed data set obtained in step S1 and the set of system operation optimization parameters obtained in step S23, analyze the system frequency response characteristics using an improved Kalman filter algorithm to obtain a physical inertia estimation value; Based on the response characteristics of power electronic devices, use a deep neural network method to evaluate the virtual inertia contribution of distributed energy to obtain a virtual inertia estimation value; Input the physical inertia estimation value and the virtual inertia estimation value into an adaptive weighted model to obtain a comprehensive inertia estimation value.

[0250] Step S32 can also be: Obtain the comprehensive inertia estimation value obtained in step S31 and the voltage level influence feature matrix obtained in step S2, use a long short-term memory network algorithm to predict the system inertia change trend to obtain an inertia change trend index; According to the inertia change trend index and the voltage level coupling degree index obtained in step S23, use a collaborative optimization algorithm to calculate the system stability margin to obtain a system stability margin matrix.

[0251] The preferred embodiments of the present invention have been described in detail above. However, the present invention is not limited to the specific details in the above embodiments. Within the technical concept of the present invention, various equivalent transformations can be made to the technical solutions of the present invention, and these equivalent transformations all fall within the protection scope of the present invention.

Claims

1. A dynamic adaptive low-frequency load shedding method for a high-proportion distributed new energy power grid, characterized in that, It includes the following steps: Step S1: Obtain the original frequency data, original power data, original voltage data, original current data, original new energy output data, and original energy storage state data from the power system equipment, perform real-time sampling on them to obtain a time-series sampling data set, conduct data quality assessment on the time-series sampling data set to obtain data quality indicators and outlier marker sets, use the outlier marker sets to preprocess the time-series sampling data set to obtain a multi-source preprocessed data set, and extract voltage level features based on the multi-source preprocessed data set to obtain a voltage level feature matrix; Step S2: Obtain the multi-source preprocessed data set and the voltage level feature matrix, analyze the correlation characteristics between different voltage levels to obtain a voltage level correlation matrix, evaluate the dynamic influence characteristics based on the voltage level correlation matrix and the multi-source preprocessed data set to obtain a dynamic influence feature matrix, conduct coupling characteristic analysis based on the dynamic influence feature matrix to obtain a voltage level influence feature matrix and a voltage level coupling degree index; Step S3: Obtain the multi-source preprocessed data set and the voltage level influence feature matrix, perform inertia estimation based on the frequency response characteristic to obtain an instantaneous inertia estimation value, analyze the inertia change trend based on the instantaneous inertia estimation value and the voltage level influence feature matrix to obtain an inertia change trend index, conduct feature extraction based on the instantaneous inertia estimation value and the inertia change trend index to obtain a system inertia feature matrix; Step S4: Obtain the multi-source preprocessed data set, use a distributed feature extraction method to obtain a distributed new energy feature matrix, use a centralized feature extraction method to obtain a centralized new energy feature matrix, and conduct new energy output prediction based on the distributed new energy feature matrix and the centralized new energy feature matrix to obtain a new energy prediction matrix; Step S5: Obtain the system inertia feature matrix, the distributed new energy feature matrix, and the centralized new energy feature matrix, use a multi-scale decomposition method to obtain a multi-scale component matrix, analyze the time-scale correlation relationship based on the multi-scale component matrix to obtain a time-scale correlation degree matrix, and conduct feature fusion based on the multi-scale component matrix and the time-scale correlation degree matrix to obtain a multi-scale feature tensor; Step S6: Obtain the multi-scale feature tensor, the time-scale correlation degree matrix, the inertia change trend index, and the new energy prediction matrix, use a multi-dimensional risk assessment method to obtain a system risk assessment matrix, conduct load priority assessment based on the system risk assessment matrix and the new energy prediction matrix to obtain a load shedding priority sequence, and optimize the shedding strategy based on the load shedding priority sequence and the system risk assessment matrix to obtain a shedding power quantity matrix; Step S7: Obtain the load shedding priority sequence, the shedding power quantity matrix, and the voltage level coupling degree index, use a hierarchical control method to generate hierarchical control instructions, evaluate the control effect based on the hierarchical control instructions and the multi-source preprocessed data set to obtain a control effect assessment index, and conduct strategy optimization based on the control effect assessment index and the system risk assessment matrix to obtain control strategy adjustment parameters and updated control instructions.

2. The dynamic adaptive low-frequency load shedding method for a high-proportion distributed new energy power grid according to claim 1, wherein Specifically, step S1 is as follows Step S11: Real-time collect the original frequency data, original power data, original voltage data, original current data, original new energy output data, and original energy storage state data from power system equipment, and process the data using an adaptive time-series sampling method to obtain a time-series sampling data set; Step S12: Obtain the time-series sampling data set, and analyze the time-series sampling data set using an improved multi-dimensional anomaly detection method to generate a data quality index and an outlier label set; Step S13: Obtain the time-series sampling data set and the outlier label set, process the labeled outliers using an adaptive filtering algorithm, and then perform noise reduction using a multi-level signal decomposition method to obtain a multi-source preprocessed data set; Step S14: Obtain the multi-source preprocessed data set, and analyze the data of different voltage levels in the multi-source preprocessed data set using an adaptive feature extraction method to obtain a voltage level feature matrix.

3. The dynamic adaptive low-frequency load shedding method for a high-proportion distributed new energy power grid according to claim 1, characterized in that Step S2 specifically includes: Step S21: Obtain the multi-source preprocessed data set and the voltage level feature matrix, and construct a voltage level association model using a multi-level network analysis method to obtain a voltage level association matrix; Step S22: Obtain the voltage level association matrix and the multi-source preprocessed data set, and evaluate the dynamic influence relationship between voltage levels using an improved time-varying causal analysis method to obtain a dynamic influence feature matrix; Step S23: Obtain the dynamic influence feature matrix, and analyze the coupling relationship between voltage levels using an adaptive coupling evaluation method to obtain a voltage level influence feature matrix and a voltage level coupling degree index.

4. The dynamic adaptive low-frequency load shedding method for a high-proportion distributed new energy power grid according to claim 1, wherein Step S3 specifically includes: Step S31: Obtain the multi-source preprocessed data set, and analyze the system frequency response characteristics using an improved state estimation algorithm to obtain an instantaneous inertia estimation value; Step S32: Obtain the instantaneous inertia estimation value and the voltage level influence feature matrix, and predict the system inertia change trend using a deep learning optimization algorithm to obtain an inertia change trend index; Step S33: Obtain the instantaneous inertia estimation value and the inertia change trend index, and construct an inertia feature space using a multi-dimensional feature extraction method to obtain a system inertia feature matrix.

5. The dynamic adaptive low-frequency load shedding method for a high-proportion distributed new energy power grid according to claim 1, characterized in that, Step S4 specifically includes: Step S41: Obtain the multi-source preprocessed data set, and analyze the distributed new energy characteristics using an improved deep learning feature extraction method to obtain a distributed new energy feature matrix; Step S42: Obtain the multi-source preprocessed data set, and extract the centralized new energy features using an enhanced neural network method to obtain a centralized new energy feature matrix; Step S43: Obtain the distributed new energy feature matrix and the centralized new energy feature matrix, and perform multi-step prediction using an integrated learning prediction method to obtain a new energy prediction matrix.

6. The method for dynamically adaptively shedding load at low frequency in a high-proportion distributed new energy power grid according to claim 1, wherein Step S5 specifically includes: Step S51: Obtain the system inertia feature matrix, the distributed new energy feature matrix, and the centralized new energy feature matrix, and perform multi-time scale decomposition using an improved modal decomposition method to obtain a multi-scale component matrix; Step S52: Obtain the multi-scale component matrix, and calculate the dynamic correlation relationship between different time scales using an improved information entropy analysis method to obtain a time scale correlation degree matrix. Step S53: Obtain the multi-scale component matrix and the time-scale correlation matrix, and perform feature fusion using the adaptive tensor decomposition method to obtain the multi-scale feature tensor.

7. The method for dynamically adaptively shedding load at low frequency in a high-proportion distributed new energy power grid according to claim 1, wherein Step S6 is specifically as follows: Step S61: Obtain the multi-scale feature tensor, the time-scale correlation matrix, and the inertia change trend index, and construct a system vulnerability model using the improved multi-dimensional evaluation method to obtain the system risk assessment matrix; Step S62: Obtain the system risk assessment matrix and the new energy prediction matrix, and determine the load shedding priority order using the comprehensive evaluation optimization method to obtain the load shedding priority sequence; Step S63: Obtain the load shedding priority sequence and the system risk assessment matrix, and optimize the cut-off power allocation scheme using the improved swarm intelligence algorithm to obtain the cut-off power quantity matrix.

8. The dynamic adaptive low-frequency load shedding method for a high-proportion distributed new energy power grid according to claim 1, characterized in that, Step S7 is specifically as follows: Step S71: Obtain the load shedding priority sequence, the cut-off power quantity matrix, and the voltage level coupling degree index, and construct a hierarchical control strategy using the improved intelligent learning method to obtain the hierarchical control instruction; Step S72: Obtain the hierarchical control instruction and the multi-source preprocessed dataset, and analyze the control execution effect using the adaptive evaluation method to obtain the control effect evaluation index; Step S73: Obtain the control effect evaluation index and the system risk assessment matrix, and perform control strategy adjustment using the online learning optimization method to obtain the control strategy adjustment parameter and the updated control instruction.

9. The dynamic adaptive low-frequency load shedding method for a high-proportion distributed new energy power grid according to claim 1, wherein Step S21 is specifically as follows: Step S211: Stratify the feature data in the voltage level feature matrix according to the voltage level to generate a voltage stratification sequence; calculate the time distribution characteristics of each layer of data in the voltage stratification sequence to obtain a time feature set; construct an inter-layer time series correlation graph according to the time feature set to generate a time series correlation table; Step S212: Extract the topological connection relationship of each voltage level node from the multi-source preprocessed dataset to obtain a topological connection sequence; Quantify the connection relationship in the topological connection sequence into a connection strength value to generate a connection strength matrix; Calculate the electrical distance between nodes based on the connection strength matrix to obtain an electrical distance sequence; Step S213: Construct an inter-layer coupling strength calculation model according to the time series correlation table and the electrical distance sequence to obtain a coupling strength matrix; Perform normalization processing on the coupling strength matrix to generate a normalized strength sequence; Calculate the inter-layer influence weight based on the normalized strength sequence to obtain an inter-layer weight matrix; Step S214: Convert the inter-layer weight matrix into a network connection structure to generate a network structure sequence; Calculate the centrality index of each node in the network structure sequence to obtain a centrality index set; determine the key connection path according to the centrality index set to generate a key path sequence; Step S215: Establish a hierarchical propagation model based on the key path sequence, calculate the propagation coefficient to obtain a propagation coefficient matrix; combine the propagation coefficient matrix with the coupling strength matrix to generate a comprehensive correlation sequence; perform feature extraction on the comprehensive correlation sequence to obtain a voltage level correlation matrix.

10. A high-proportion distributed new energy power grid dynamic adaptive low-frequency load shedding system, characterized in that, Including: At least one processor; and a memory communicatively connected to at least one of the processors; wherein the memory stores instructions executable by the processors, and the instructions are used to be executed by the processors to implement the high-proportion distributed new energy grid dynamic adaptive low-frequency load shedding method according to any one of claims 1 to 9.

Images