MGAF-ConvNeXt intelligent evaluation method for frequency stability of renewable energy power systems under incomplete measurement

Abstract

The present invention discloses a MGAF-ConvNeXt intelligent evaluation method for frequency stability of a new energy power system under incomplete measurement. First, the frequency response of the new energy power system is partitioned; then, a MGAF-ConvNeXt lightweight deep learning model for intelligent frequency stability evaluation is constructed; simulation data under multiple operating modes and multiple disturbance scenarios are obtained, and according to the system partition results, the generators with the largest inertia or unit capacity in each area and the closest distance to the system inertia center are selected as measurement points, a historical data set is constructed, and the model is trained. The test set data is used to predict the frequency stability of the power system; finally, the frequency stability prediction accuracy under different measurements is evaluated. The present invention maintains a high frequency stability evaluation accuracy even when the system PMU coverage rate is low, providing an efficient solution for the economic deployment of PMUs and real-time frequency stability evaluation in new energy power systems.

Description

Technical Field

[0001] The present invention belongs to the technical field of power system security and stability analysis and control, and in particular relates to a hybrid Gram angular field-next-generation convolutional neural network (MGAF-ConvNeXt) intelligent evaluation method for frequency stability of a renewable energy power system under incomplete measurement. Background Art

[0002] Frequency stability assessment of new energy power systems is a core link in ensuring the safe operation of power grids. In recent years, with the large-scale grid connection of low-inertia power sources such as wind power and photovoltaics, the proportion of traditional power sources has continued to decline, the system inertia support capacity has significantly decreased, and frequency fluctuations have intensified. The power grid's requirements for the real-time and accuracy of frequency stability assessments have reached unprecedented levels.

[0003] Existing frequency stability assessment methods are mainly divided into three categories: simplified model methods based on the center of inertia, full-time domain simulation methods, and data-driven intelligent methods. Simplified model methods ignore the spatiotemporal distribution characteristics of system frequency and cannot reflect the complex dynamics after the integration of new energy sources; full-time domain simulation methods have problems such as complex modeling and low computational efficiency; although data-driven methods have achieved breakthroughs in accuracy and speed, they rely on complete system measurement data. However, actual power grids are limited by the cost of PMU deployment and generally suffer from insufficient measurement node coverage, resulting in missing feature information and seriously affecting the reliability of assessments. Among current deep learning methods, convolutional neural networks have problems such as weak global feature capture and insufficient local feature extraction of attention mechanism models. Hybrid architectures also face bottlenecks such as parameter redundancy and time-consuming training.

[0004] While deep learning methods have shown promise in frequency stability assessment, existing models such as CNNs and Transformers still have significant limitations: CNNs are inadequate at modeling global correlations in time series data, while Transformers lack the ability to extract local features. While the hybrid ConvNeXt architecture combines the strengths of both, its large convolution kernels lead to a bloated parameter space and inefficient traditional 32-bit floating-point precision training. Against this backdrop, developing novel spatiotemporal feature fusion mechanisms and lightweight deep learning architectures is crucial to overcome the data dimensionality problem in incomplete measurement scenarios and improve the robustness and timeliness of frequency stability assessment. Summary of the Invention

[0005] Purpose of the Invention: To address the shortcomings and deficiencies of existing technologies, this invention proposes a MGAF-ConvNeXt intelligent frequency stability assessment method for renewable energy power systems with incomplete measurements. When partitioning the power system frequency response, a frequency response partitioning algorithm based on Hausdorff's Kmeans++ is used. By constructing a multidimensional spatiotemporal feature vector, the dynamic response correlation characteristics of the inertia release phase and the primary frequency modulation phase are fully characterized, effectively addressing the partitioning inaccuracy problem caused by the lack of local features in traditional methods.

[0006] Technical solution: The present invention provides an MGAF-ConvNeXt intelligent evaluation method for frequency stability of a new energy power system under incomplete measurement, comprising the following steps:

[0007] Step 1: Perform frequency response partitioning of the new energy power system to obtain frequency response partitioning results;

[0008] Step 2: Construct a hybrid Grammar angular field-next-generation convolutional neural network MGAF-ConvNeXt frequency response intelligent evaluation model;

[0009] Step 3: Simulate and obtain the response data of the example system under multiple operating modes and multiple disturbance scenarios. Based on the frequency response partitioning results of Step 1, select the generator with the largest inertia or unit capacity in each area or the closest distance to the system inertia center as the measurement point to construct a historical data set.

[0010] Step 4: Accelerate the training of the MGAF-ConvNeXt frequency response intelligent evaluation model based on the mixed precision quantization method;

[0011] Step 5: Establish an incomplete measurement case library, train the MGAF-ConvNeXt frequency response intelligent evaluation model using historical data sets, and perform error analysis on the frequency stability prediction accuracy under incomplete measurements based on the trained intelligent evaluation model.

[0012] Furthermore, step 1 specifically includes the following steps:

[0013] Step 1.1: Establish a new energy power system case based on the PSASP platform, construct scenarios with different outputs of new energy and different load disturbances, and simulate the frequency disturbance trajectory of each node in the power system based on each constructed scenario;

[0014] Step 1.2: Based on the frequency-disturbed trajectory of each node in the system, calculate the similarity matrix of the frequency response curve of each node based on the Hausdorff distance; the Hausdorff distance calculation formula between the frequency-disturbed trajectory of node m and the frequency-disturbed trajectory of node n is:

[0015] H ( m , n ) = max[ h ( m , n ), h ( n , m )] (1)

[0016] in: h ( m , n ) = max[min ‖ m u − n v ‖ ] , h ( n , m ) = max[min ‖ n v − m u ‖ ] ; The data set of frequency disturbed trajectory m=[m1, m2, …, m a ] and n=[n1, n2, …, n b ] are two curves of unequal lengths, a and b are the total lengths of data in sets m and n respectively; h(n,m) is the one-way Hausdorff distance from set n to set m, ||m u -n v || is m u to n v Euclidean distance; u=[1, 2,…, a], v=[1,2,…, b], for a system with N nodes, the dimension of the similarity matrix H is N×N;

[0017] Step 1.3: Use the K-means++ clustering algorithm to classify the similarity matrix and obtain the cluster centers required for power system partitioning.

[0018] Furthermore, step 1.3 is specifically as follows:

[0019] Randomly select the i-th row h in the Hausdorff similarity matrix H i As the kth initial cluster center h k , i=[1, 2,…,N]; k=[1, 2,…, l], where l is the number of cluster centers;

[0020] Calculate the shortest distance D(H) between each node and the current cluster center:

[0021] (2)

[0022] Where j = [1, 2, …, N]; h ij represents the element in the i-th row and j-th column of the similarity matrix H, h kj represents the k-th row and j-th column element in the similarity matrix H;

[0023] Calculate the probability P(H) of each node being selected as the next cluster center, and select the node corresponding to the maximum probability value as the next cluster center;

[0024] (3)

[0025] Repeat equations (2) and (3) until the l cluster centers required for power system partitioning are selected, and based on D(H), find the minimum distance between each node and each cluster center, divide the node with the shortest distance to the cluster center into a region, and finally obtain the system partitioning result.

[0026] Furthermore, step 2 specifically includes the following steps:

[0027] Step 2.1: The Mixed Gram Angular Field (MGAF) method converts one-dimensional data into a two-dimensional matrix that can be recognized by the deep learning model, including:

[0028] The one-dimensional time series data g of length L is mapped to the interval [-1, 1] through maximum and minimum normalization. The formula is:

[0029] (4)

[0030] Where g ζ is the ζ-th variable in sample g, is the normalized sample, ζ=[0, 1,…, L];

[0031] Convert the normalized data to the polar coordinate system using the following formula:

[0032] { Φ ζ = arccos( g ˜ ζ ) 0 ≤ g ˜ ζ ≤ 1 , g ˜ ζ ∈ g r ζ = ζ / L ζ ∈ [ 1 , L ] (5)

[0033] Where: Φ ζ is the ζth polar angle in the sample, r ζ is the ζth polar diameter in the sample;

[0034] Step 2.2: Fuse the Gram sum angular field GASF and the Gram difference angular field GADF to generate the asymmetric MGAF matrix. The formula is:

[0035] { MGAF ( ζ , β ) = GASF ( ζ , β ) ζ ≥ β MGAF ( ζ , β ) = GADF ( ζ , β ) ζ < β ， GASF ( ζ , β ) = [cos( Φ ζ + Φ β )] GADF ( ζ , β ) = [sin( Φ ζ −Φ β )] (6)

[0036] Among them, Φ β is the βth polar angle in the sample, r β is the βth pole in the sample, β=[0, 1,…, L];

[0037] Step 2.3, ConvNeXt deep learning model, the structure includes input layer, intermediate layer and output layer;

[0038] The input layer performs spatial downsampling and feature dimension expansion through 4×4 convolution;

[0039] The middle layer consists of three cascaded ConvNeXt Block and MaxPooling composite units. The ConvNeXtBlock contains: a 7×7 depthwise separable convolution layer; a Layer Normalization layer; a bottleneck structure consisting of two levels of 1×1 convolution layers, with a GELU activation function embedded between the two levels of 1×1 convolution layers.

[0040] The output layer generates a five-dimensional frequency stability index through global average pooling and a fully connected layer.

[0041] Furthermore, step 3 specifically includes the following steps:

[0042] Step 3.1. Input data: voltage amplitude V, current amplitude I, active power P, and reactive power Q of the measurement node;

[0043] Step 3.2, output data: the maximum frequency change rate of the system inertia center ROCOF max , maximum frequency offset Δf max , quasi-steady-state frequency f ss , maximum frequency deviation delay t ir and quasi-steady-state frequency delay t ss , specifically defined as follows:

[0044] (7)

[0045] (8)

[0046] (9)

[0047] (10)

[0048] (11)

[0049] Among them, f COI is the system inertia center frequency; f0 is the system steady-state frequency; ε is the threshold; d represents the differential operator, and t represents time.

[0050] Step 3.3, the disturbance scenario is set as follows: at t=0s, each load node in the system increases the load by 10-100MW in steps of 10MW, and the number of disturbance scenarios is 10 times the number of load nodes; the scenario set is a combination of different renewable energy output scenarios and different disturbance scenarios. According to the simulation of different scenario sets, the disturbed trajectories of the frequency f, voltage amplitude V, current amplitude I, active power P and reactive power Q of each node are obtained. The time window of the measurement data is selected as [0.1, 0.3]s, and the sampling frequency is 100Hz. The disturbed trajectories obtained under the scenario set are randomly divided into training set and test set according to proportion.

[0051] Furthermore, step 4 specifically includes the following steps:

[0052] Step 4.1: During the forward and backward propagation of training, the activation, weight, and gradient data are stored in FP16 format, and a copy of the weights in FP32 format is retained for gradient updates.

[0053] Step 4.2: Weight update retains the 32-bit floating point precision master copy, and the gradient update formula is:

[0054] (12)

[0055] Among them, W FP32 is a 32-bit floating point weight, G FP16 is a 16-bit weight gradient update, is the scaling factor for gradient updates; this is done by backing up the weights in FP32 format and ensuring that the update process is performed in FP32.

[0056] Furthermore, step 5 specifically includes the following steps:

[0057] Step 5.1: The established system incomplete measurement case library includes:

[0058] Case 0: All power nodes are measurable;

[0059] Case 1: A single inertia power node can be measured;

[0060] Case 2: All inertia power nodes in the specified partition can be measured;

[0061] Case 3: A single inertia power node in each partition can be measured, specifically:

[0062] Case 3-1 to Case 3-3: Select the node with the largest inertia time constant, the closest distance to the inertia center, or the largest unit capacity in each partition;

[0063] Case 3-4 to Case 3-15: Optimize the measurement configuration by replacing suboptimal nodes within the partition or adding redundant nodes;

[0064] Step 5.2: Determine the measurable nodes based on the incomplete measurement case library, determine the data set accordingly, train the MGAF-ConvNeXt model based on the training set data, and further perform frequency stability prediction on the test set data based on the trained model to obtain the system's frequency stability index, specifically including the maximum frequency change rate of the system's center of inertia, ROCOF. max , maximum frequency offset Δf max , quasi-steady-state frequency f ss , maximum frequency deviation delay t irand quasi-steady-state frequency delay t ss ;

[0065] The errors of frequency stability indicators under different case libraries are compared, and the accuracy of intelligent assessment of power system frequency stability under incomplete measurements is analyzed.

[0066] The present invention also discloses a MGAF-ConvNeXt intelligent evaluation system for frequency stability of a new energy power system under incomplete measurement, comprising:

[0067] Modeling unit: built the MGAF-ConvNeXt frequency response intelligent evaluation model;

[0068] The design unit calculates similarity based on the frequency response data of all nodes in the system, and clusters the similarity matrix using the K-means++ method to obtain the frequency response partitioning result of the system;

[0069] Computing unit: Convert one-dimensional time series data into two-dimensional graphics data according to the MGAF method; further train the ConvNeXt network model based on the training set data, and adopt a mixed precision quantization model training acceleration method.

[0070] The present invention further discloses a computer device, comprising a memory, a processor, and a computer program stored in the memory, wherein the processor executes the computer program to implement the steps of the method of the present invention.

[0071] The present invention further discloses a computer-readable storage medium having a computer program / instruction stored thereon, which implements the steps of the method of the present invention when the computer program / instruction is executed by a processor.

[0072] Beneficial effects: Compared with the prior art, the present invention has the following significant advantages:

[0073] (1) The multi-dimensional spatiotemporal partitioned frequency stability intelligent evaluation method proposed in this paper can maintain high evaluation accuracy even when the PMU coverage is low, and has a certain improvement over the traditional CNN and Transformer models; the MGAF data conversion strategy effectively reduces the feature dimension while retaining the temporal correlation; the AMP training strategy greatly improves the model training efficiency; the partitioned measurement node selection criterion proposed in this paper can reduce the PMU deployment cost by 63.8%.

[0074] (2) This paper proposes to convert time series data into spatial matrix data based on the MGAF method, which can significantly reduce the data dimension while retaining the data features. A frequency stability assessment method based on the ConvNeXt model is proposed. This method combines the local feature extraction capability of CNN with the global correlation advantage of the Attention mechanism, significantly reduces the computational complexity of the model and improves the accuracy of frequency stability assessment. A ConvNeXt model training strategy based on AMP is proposed. This strategy achieves a significant improvement in model training efficiency while maintaining the accuracy of FP32 gradient calculation. A configuration suggestion for the system PMU under incomplete measurement is proposed, which ensures the accuracy of frequency stability assessment while significantly reducing the input data dimension.

[0075] (3) When conducting power system stability assessment, a multi-dimensional space-time frequency feature vector is constructed to fully characterize the dynamic response correlation characteristics of the inertia release stage and the primary frequency regulation stage, effectively solving the partition inaccuracy problem caused by the lack of local features in traditional methods. BRIEF DESCRIPTION OF THE DRAWINGS

[0076] Figure 1 Flow chart of the method of the present invention;

[0077] Figure 2 It is a 39-node system in New England;

[0078] Figure 3 The similarity matrix heat map of nodes in the New England 39-node system. (a) is the Hausdorff similarity matrix heat map of all generator nodes; (b) is the Hausdorff similarity matrix heat map of all nodes;

[0079] Figure 4 The results of the generator node partitioning in the New England 39-node system are shown;

[0080] Figure 5 This is an example of MGAF data conversion;

[0081] Figure 6 This is the ConvNeXt network structure diagram;

[0082] Figure 7 ConvNeXt Block model structure

[0083] Figure 8 Training strategy framework for AMP;

[0084] Figure 9 Test the loss value for the model with various data inputs;

[0085] Figure 10 Calculate the evaluation graph for the model under Case 1;

[0086] Figure 11 Calculate the evaluation graph for the model of Case 2 where the system is not fully measurable;

[0087] Figure 12 Calculate the evaluation graph for the model of Case 3 where the system is not fully measurable. DETAILED DESCRIPTION

[0088] The technical solution of the present invention will be further described below with reference to the accompanying drawings.

[0089] like Figure 1 As shown, the MGAF-ConvNeXt intelligent evaluation method for frequency stability of a new energy power system under incomplete measurement of the present invention includes the following steps:

[0090] Step 1: Perform frequency response zoning of the new energy power system;

[0091] Step 2: Construct a hybrid Grammar angular field-next-generation convolutional neural network MGAF-ConvNeXt frequency response intelligent evaluation model;

[0092] Step 3: Simulate and obtain the response data of the example system under multiple operating modes and multiple disturbance scenarios. Based on the frequency response partitioning results of Step 1, select the generator with the largest inertia or unit capacity in each area or the closest distance to the system inertia center as the measurement point to construct a historical data set.

[0093] Step 4: Accelerate the training of the MGAF-ConvNeXt frequency response intelligent evaluation model based on the mixed precision quantization method;

[0094] Step 5: Establish an incomplete measurement case library, train the MGAF-ConvNeXt frequency response intelligent evaluation model using historical data sets, and perform error analysis on the frequency stability prediction accuracy under incomplete measurements based on the trained intelligent evaluation model.

[0095] Example 1

[0096] Step 1: Frequency response zoning of new energy power system.

[0097] Based on PSASP (Power System Analysis Software Package) platform to build the attached Figure 2 In the New England 10-unit, 39-node system shown, new energy sources were added to nodes G3 to G9 of the original system, partially or completely replacing the original synchronous power source power while maintaining the total power generation capacity of the original nodes. The new energy penetration rate range was set to 0-70% of the total load. The layout of the new energy units under different new energy penetration rates is shown in Table 1.

[0098] Table 1 New energy units with different new energy penetration rates

[0099]

[0100] The disturbance is set as a sudden load increase fault of 10-100MW (step size of 10MW) at each load node in the system at t=0s, and the simulation step size is 0.01s. The simulation obtains the {frequency f, voltage amplitude V, current amplitude I, active power P and reactive power Q} of each generator node under different renewable energy penetration and load disturbance scenarios.

[0101] According to the different scenarios obtained by different new energy penetration rates and different disturbance combinations in Table 1, the frequency disturbance trajectory of each node in the system under each scenario is simulated, and the similarity matrix of the frequency response characteristics of each node in the system is calculated based on the Hausdorff similarity algorithm. The Hausdorff distance calculation method is as follows:

[0102] Assume that the frequency disturbed trajectory data set m=[m1, m2, …, m a ] and n=[n1, n2, …, n b ] are two curves of unequal lengths, a and b are the total lengths of the data in sets m and n respectively; first, the one-way Hausdorff distance h(m, n) from set m to set n is calculated, which is defined as the minimum distance from any point on set m to any point in set n, and then the maximum value is taken, as shown in formula (1).

[0103] h ( m , n ) = max[min ‖ m u − n v ‖ ] (1)

[0104] where ||m u -n v || is m u to n v The Euclidean distance of set n to set m is u=[1, 2,…, a], v=[1,2,…, b]. Similarly, the one-way Hausdorff distance h(n,m) from set n to set m can be obtained as shown in formula (2).

[0105] h ( n , m ) = max[min ‖ n v − m u ‖ ] (2)

[0106] Finally, the maximum value of h(m, n) and h(n, m) is the bidirectional Hausdorff distance between sets m and n, as shown in formula (3).

[0107] H ( n , m ) = max[ h ( m , n ), h ( n , m )] (3)

[0108] For a system with N nodes, the dimension of its similarity matrix H is N×N. Figure 3 The Hausdorff similarity matrix of the frequency response of each node under the scenario of 30% renewable energy penetration and a sudden increase of 10MW load at BUS16 is given, where Figure 3 (a) in the figure is the similarity matrix of 39 nodes. Figure 3 (b) in the figure is the similarity matrix of 10 generator sets.

[0109] The K-means++ clustering algorithm is further used to cluster the similarity matrix and partition it according to the clustering results. The frequency response partitioning steps of the K-means++ algorithm are as follows:

[0110] Randomly select the i-th row h in the Hausdorff similarity matrix H i As the kth initial cluster center h k , i=[1, 2,…,N], k=[1, 2,…, l], where l is the number of cluster centers.

[0111] Calculate the shortest distance D(H) between each generator node and the current cluster center,

[0112] (4)

[0113] Where j=[1,2, …, N]. ij represents the element in the i-th row and j-th column of the similarity matrix H, h kj represents the k-th row and j-th column element in the similarity matrix H;

[0114] Calculate the probability P(H) of each node being selected as the next cluster center, and select the generator node corresponding to the maximum probability value as the next cluster center;

[0115] (5)

[0116] Repeat formula (4) and formula (5) until the l cluster centers required for power system partitioning are selected, and based on D(H), find the minimum distance between each generator node and each cluster center, divide the node and the cluster center with the shortest distance into a region, and finally obtain the system partitioning result.

[0117] The partition results of the system under different node load disturbances and with a new energy penetration rate between 0% and 70% are slightly different. Based on the frequency of occurrence of each partition, the scenario with the highest probability is taken as the final partition result. Figure 4 , specifically area 1 (G01), area 2 (G02, G03), area 3 (G04, G05, G06, G07) and area 4 (G08, G09, G10). Figure 2 The dotted line in is the partition result of all nodes.

[0118] Step 2: Build the MGAF-ConvNeXt frequency stability intelligent evaluation model.

[0119] A ConvNeXt model for frequency stability assessment was designed. This model combines the local feature extraction capabilities of CNNs with the global correlation advantages of the Attention mechanism, significantly improving the accuracy of frequency stability assessment. Furthermore, to convert time series data into spatial data suitable for CNN analysis, a data dimension conversion method based on MGAF was proposed. This method converts time series features into two-dimensional spatial features while preserving them. This method also reduces the dimensionality of the matrix obtained by the GAF method, improving data feature extraction capabilities and accelerating frequency stability assessment calculations.

[0120] (1) Converting one-dimensional time series data into two-dimensional image data based on the MGAF method

[0121] The Gramian Angular Field (GAF) includes the Gramian angular summation field (GASF) and the Gramian angular difference field (GADF). The transformation principle is as follows:

[0122] Based on the maximum and minimum normalization method, the one-dimensional time series data g with a length of L is mapped to the interval [-1, 1], as shown in formula (6):

[0123] (6)

[0124] Where g ζ is the ζ-th variable in sample g, is the normalized sample, ζ=[0, 1,…, L];

[0125] Convert the normalized data to the polar coordinate system using the following formula:

[0126] { Φ ζ = arccos( g ˜ ζ ) 0 ≤ g ˜ ζ ≤ 1 , g ˜ ζ ∈ g r ζ = ζ / L ζ ∈ [ 1 , L ] (7)

[0127] Where: Φ ζ is the ζth polar angle in the sample, r ζ is the ζth polar diameter in the sample;

[0128] After converting the data into polar coordinates, GASF and GADF are used to convert the data into two-dimensional image data, as shown in equations (8) and (9):

[0129] GASF ( ζ , β ) = [cos( Φ ζ + Φ β )] (8)

[0130] GADF ( ζ , β ) = [sin( Φ ζ −Φ β )] (9)

[0131] Among them, Φ β is the βth polar angle in the sample, r β is the βth pole in the sample, β=[0, 1,…, L];

[0132] As shown in Equations (8) and (9), the Gram sum angular field transforms one-dimensional structured data into a symmetric matrix whose elements store the original data information. The matrix obtained by the Gram difference angular field is a matrix with all diagonal elements set to zero, and the elements symmetrical on both sides of the main diagonal are opposite to each other. Therefore, the numerical matrices after the two types of angular field transformations have symmetrical elements on both sides of the diagonal that are redundant with each other. To solve the problem of matrix element redundancy, the present invention proposes a MGAF method, which fuses the lower triangular and focus elements of GASF with the upper triangular elements of GADF to obtain a new MGAF matrix, as shown in Equation (10).

[0133] (10)

[0134] Figure 5 An example of applying MGAF transformation to frequency response time series data is shown.

[0135] (2) ConvNeXt network

[0136] A hybrid architecture ConvNeXt that combines the local perception advantages of CNN and the global modeling capabilities of Transformer. The specific structure is as follows Figure 6 As shown in the figure, the model first implements spatial downsampling and feature dimension expansion through 4×4 convolution to increase the channel dimension of the input data. The core feature extraction module consists of three cascaded ConvNeXt Block-MaxPooling composite units, each containing three ConvNeXt Blocks and one maximum pooling layer. Through three iterations, the feature space is transformed from local details to global associations. Finally, a global average pooling layer compresses the high-dimensional feature tensor into a channel feature vector, eliminating redundant information in the spatial dimension while retaining the core feature response pattern. The terminal fully connected layer maps the abstract features to a 5-dimensional frequency stability evaluation indicator using a learnable weight matrix.

[0137] The ConvNeXt Block structure is as follows Figure 7As shown in the figure. This module inherits the basic residual connection framework of ResNet and draws on the hierarchical feature extraction paradigm of the Swin Transformer. It moves the depthwise separable convolution from the second level of the traditional residual block to the initial stage of the feature processing flow. This structure advances the spatial feature extraction operation to before the channel dimension transformation, improving the model's hierarchical perception capabilities. Secondly, the block refers to the receptive field range of the self-attention mechanism in the Vision Transformer and expands the traditional 3×3 small convolution kernel to 7×7. By increasing the local perception area, it achieves cross-pixel long-range dependency modeling and constructs a self-attention-like feature mixing weighting mechanism in the single-channel feature space, effectively improving the ability to capture global contextual information. The module introduces Layer Normalization after the deep convolution layer to calibrate the feature distribution along the channel dimension. In the feature transformation stage, a two-level 1×1 point convolution is used to construct a bottleneck structure. The first-level convolution achieves 4-fold channel expansion to enhance the nonlinear representation space, and the secondary convolution performs dimensionality reduction to control the parameter scale. At the same time, the GELU activation function is embedded between the two-level 1×1 convolution layers to replace the traditional ReLU. Its continuous differentiable mathematical properties are more conducive to the stability of gradient propagation.

[0138] Step 3: Build a historical dataset;

[0139] The scenario set consists of a combination of the different renewable energy output scenarios shown in Table 1 and the different disturbance scenarios described above. Simulating all scenarios in the multi-scenario set yields disturbed trajectories for each node's frequency f, voltage amplitude V, current amplitude I, active power P, and reactive power Q. The measurement data is stored in a time window of [0.1, 0.3] s, with a sampling frequency of 100 Hz. The disturbed trajectories obtained from each scenario set are randomly divided into training and test sets in a 7:3 ratio.

[0140] Input data: {voltage amplitude V, current amplitude I, active power P, and reactive power Q} of the measurement node;

[0141] Output data: 5 indicators that characterize the system frequency stability, namely {maximum frequency change rate of the system inertia center ROCOF max , maximum frequency offset Δf max , quasi-steady-state frequency f ss , maximum frequency deviation delay t ir and quasi-steady-state frequency delay t ss}. The specific definitions are as follows:

[0142] (11)

[0143] (12)

[0144] (13)

[0145] (14)

[0146] (15)

[0147] Among them, f COI is the system inertia center frequency; f0 is the system steady-state frequency; ε is a small threshold, which is 0.005 in the present invention; d is the differential operator, and t is time.

[0148] Step 4: Model training acceleration method based on mixed precision quantization, see Figure 8 As shown;

[0149] During the forward and backward propagation of training, data such as activations, weights, and gradients are stored in FP16 format, and a copy of the weights in FP32 format is retained for gradient updates;

[0150] The weight update retains the 32-bit floating point precision master copy, and the gradient update formula is:

[0151] (16)

[0152] Where W FP32 is a 32-bit floating point weight, G FP16 is a 16-bit weight gradient update, is the scaling factor for gradient updates.

[0153] By backing up the weights in FP32 format and ensuring that updates are performed in FP32, rounding errors in half-precision training can be avoided.

[0154] Step 5: Intelligent frequency stability evaluation and error analysis under incomplete measurement.

[0155] The system incomplete measurement case library includes Case 0-Case 3, which are:

[0156] Case 0: All power nodes are measurable;

[0157] Case 1: A single inertia power node can be measured;

[0158] Case 2: All inertia power nodes in the specified partition can be measured;

[0159] Case 3: A single inertia power node in each partition can be measured, specifically:

[0160] Case 3-1 to Case 3-3: Select the node with the largest inertia time constant, the closest distance to the inertia center, or the largest unit capacity in each partition;

[0161] Case 3-4 to Case 3-15: Optimize the measurement configuration by replacing suboptimal nodes within the partition or adding redundant nodes.

[0162] The measurable units in each case library are shown in Table 2.

[0163] Table 2 Measurable units for Case 0-3

[0164]

[0165] The measurable nodes in Table 2 are analyzed to determine the data set. The MGAF-ConvNeXt model is trained based on the training set data. The frequency stability prediction of the test set data is further performed based on the trained model to obtain the system frequency stability index {maximum frequency change rate of the system inertia center ROCOF max , maximum frequency offset Δf max , quasi-steady-state frequency f ss , maximum frequency deviation delay t ir and quasi-steady-state frequency delay t ss}.

[0166] The errors of frequency stability indicators under different case libraries are compared, and the accuracy of intelligent assessment of power system frequency stability under incomplete measurements is analyzed.

[0167] The mean absolute error (MAE) and mean square error (MSE) are used to evaluate the forecast accuracy of frequency stability indicators. MAE can intuitively represent the average level of forecast error, while MSE can reflect the overall deviation between the predicted value and the actual value, giving higher weight to larger errors and reducing its sensitivity to outliers.

[0168] (17)

[0169] (18)

[0170] Where: E MA is the mean absolute error; E MS is the mean square error; n s is the number of samples; y i is the true value of the i-th sample; is the predicted value of the i-th sample.

[0171] To analyze the evaluation accuracy E of frequency stability under incomplete measurement rp , the evaluation accuracy under incomplete measurement is compared with that under complete measurement, as shown in formula (19):

[0172] (19)

[0173] Among them, E c is the MAE average error of frequency stability evaluation under Cases 1 to Case 3, E Case0 is the MAE average error of the frequency stability evaluation under full measurement.

[0174] First, for Case 0, where all power node data of the system can be measured, the ConvNeXt model is trained. The calculation accuracy of the five frequency stability indicators is shown in Table 3. To eliminate dimensional differences and improve the comparability of the results, all indicators are preprocessed using the maximum and minimum normalization method shown in Equation (6).

[0175] Table 3 Frequency stability index error

[0176]

[0177] Table 3 shows that the proposed method has high calculation accuracy for all five key frequency stability indicators, and the MAE of all indicators is controlled within the relative error range of 1.5%, which verifies the effectiveness of the proposed method in power system frequency stability analysis. ss and t ir The calculation error of ROCOF is significantly lower than that of other indicators, and its MAE and MSE are better than those of ROCOF. max , Δf max and t ss The three indicators decreased by an average of 48.7%.

[0178] To verify the superiority of the ConvNeXt model in predicting stable frequencies, we compared our model with CNN, ResNet, VGG, Transformer, recurrent neural networks, LSTM, and GRU. CNN is the ConvNeXt model without the attention mechanism encoder layer. The mean error values ​​for the five metrics on the validation set are shown in Table 4.

[0179] Table 4 Model size, training time, and MAE and MSE for power system frequency stability indicators

[0180]

[0181] From Table 4 we can see that:

[0182] 1) Compared to models like GRU, LSTM, and Transformer, which use measured time series data directly as input, the ConvNeXt model, thanks to the CNN, can more quickly learn the characteristics of frequency dynamic response, enabling rapid model convergence. Furthermore, compared to pure convolutional neural networks like ResNet and VGG, the model of this invention achieves higher computational accuracy due to the presence of the attention mechanism.

[0183] 2) The ConvNeXt model has higher computational accuracy than CNN models without attention mechanisms, but compared to the Transformer model with only attention mechanisms, the model of the present invention requires less storage space and takes less time to train. Compared to other models, the model of the present invention has higher model accuracy and faster convergence speed, and with shorter training time and less storage space, it is an excellent choice for frequency stability analysis and calculations.

[0184] To further verify the ConvNeXt model's accuracy in calculating the frequency stability index at different renewable energy penetration rates, it is compared with the CNN and ResNet models (Table 3), which have stronger computational capabilities and shorter training times. The average errors of the frequency stability index at renewable energy penetration rates from 0% to 70% are shown in Table 5, where NPR represents renewable energy penetration rate (%).

[0185] Table 5. Test MAE of the model under different new energy penetration rates

[0186]

[0187] As shown in Table 5, the ConvNeXt model shows significant performance advantages under different new energy penetration rates. In terms of absolute error, its MAE value is always lower than the comparison model: in the stage of NPR%≤30, the MAE of ConvNeXt is stable at 1.02×10 -2 to 1.08×10 -2 In the range of 100% to 200% new energy penetration, the average error reduction was 4.7% compared to the CNN model and 51.3% compared to the ResNet model. When the new energy penetration rate exceeded 60%, while the errors of all models showed an upward trend, ConvNeXt remained relatively stable. Its MAE at NPR% = 70 was 45.4% and 48.1% lower than those of the CNN and ResNet models, respectively, demonstrating greater robustness. Analysis of the error growth rate showed that the MAE of ConvNeXt increased by approximately 12.8% for every 10% increase in penetration, significantly lower than the 19.6% of the CNN model and the 24.3% of the ResNet model, further validating its superior generalization ability. Although increasing new energy penetration will lead to a gradual increase in model error, the impact is significantly smaller than the performance fluctuations caused by differences in model architecture, so the impact of new energy penetration on model computing power can be ignored.

[0188] In order to verify the effectiveness of the proposed MGAF in transforming one-dimensional frequency dynamic response data into a two-dimensional matrix, the frequency data were transformed using MGAF, GADF, GASF, Continuous Wavelet Transform (CWT) and Short Time Fourier Transform (STFT) methods respectively to obtain different types of spatial matrix samples, which were then input into ConvNeXt for performance testing. The original data was repeated 20 times to form a matrix that the model can recognize. The model test loss value is as follows: Figure 9 As shown. Figure 9 As can be seen, data that has undergone dimensionality increase significantly improves classification accuracy compared to one-dimensional data. The MGAF method proposed in this paper combines the features of GASF and GADF, achieving higher computational accuracy than other dimensionality transformation methods. Therefore, in subsequent numerical analysis experiments, the matrix obtained by MGAF data dimensionality transformation was used as the model input.

[0189] In order to verify the feasibility of the AMP method used to accelerate the model training process, the ConvNeXt model was trained in FP32 mode, FP16 mode, and AMP mode respectively. The training time and the average test accuracy of the model are shown in Table 6.

[0190] Table 6 Model calculation accuracy and time consumption based on AMP method

[0191]

[0192] Table 6 shows that while the AMP method results in a very slight decrease in model accuracy, it still outperforms the other models in Table 4. Furthermore, the AMP-based model training time is only 70.93% of that of the FP32 method. In particular, the model trained with FP16 precision, while taking only 62.19% of the FP32 method, experienced an overflow error during training, making it impossible to calculate the model's test MAE.

[0193] To compare the evaluation accuracy of frequency stability under incomplete measurement, Figure 10 The mean prediction error of the frequency stability index under Case 0 (full measurement) and Case 1 is given. Figure 11 is the mean prediction error of the frequency stability index of Case0 and Case2, Figure 12 is the mean of the frequency stable prediction errors of Case0 and Case3.

[0194] from Figure 10 It can be seen that:

[0195] 1) In the full measurement case 0, the average MAE and MSE are 1.03×10 -2 and 2.471×10 -4 Compared to Case 0, the frequency stability assessment accuracy for incomplete measurements (Case 1–Case 3) decreased. Generally speaking, fewer measurements lead to lower accuracy. The assessment accuracy for a small number of cases improved slightly, due to a slight improvement in the MGAF-ConvNeXt model's prediction accuracy, resulting from a reduction in the input data dimension.

[0196] 2) Figure 10 When only one generator node is measurable, Case 1-3 (generator G08 is measurable) and Case 1-10 (generator G01 is measurable) achieve the highest frequency stability assessment accuracy. This is because generators G08 and G01 are close to the system's inertia center, and G01 has a larger inertia. However, when only one generator node is measurable, the frequency stability assessment error is still large. For example, in Case 1-2, when only generator G09 is measurable, its E rp It is 277.18%.

[0197] 3) Figure 11 When only one region's generator nodes are measurable, Case 2-1 (generator G01 is measurable) and Case 2-4 (generators G08, G09, and G10 are measurable) achieve the highest frequency stability assessment accuracy. Furthermore, Case 2-4 achieves higher accuracy than Case 1-3, indicating that frequency stability assessment accuracy is higher when the region is closer to the system's inertia center and the inertia of the generators within the region is greater. Furthermore, increasing system measurements generally improves frequency stability assessment accuracy.

[0198] 4) Figure 12 In Case 3-1 (G01, G03, G06 and G10 can be measured, and the area corresponding to the maximum inertia can be measured. At this time, E rp =6.22%), Case3-2 (G01, G03, G04 and G08 can be measured, the area closest to the system inertia center can be measured, E rp =-10.15%), Case3-3 (G01, G02, G06, G09 can be measured, corresponding to the area with the maximum generator capacity, E rp =1.62%) has the lowest frequency stability prediction error. This indicates that the larger the inertia of the region, the larger the unit capacity, or the closer the distance to the system inertia center, the higher the frequency stability assessment accuracy.

[0199] Under incomplete system measurement, the generator nodes with large unit capacity, large unit inertia time constant and small electrical distance from the system inertia center should be selected as data collection points; or when deploying PMUs, the PMU deployment redundancy at the above nodes should be increased to improve the assessment accuracy of power system frequency stability.

[0200] In a second aspect, the present invention further discloses a MGAF-ConvNeXt intelligent evaluation system for frequency stability of a new energy power system under incomplete measurement, comprising:

[0201] Modeling unit: Constructed the MGAF-ConvNeXt frequency response intelligent evaluation model, see Figure 5-Figure 7 ;

[0202] The design unit designed a similarity calculation method based on Hausdorff distance according to the frequency response data of all nodes in the system, and further clustered the similarity matrix using the K-means++ method to obtain the frequency response partitioning result of the system;

[0203] Computing unit: Convert one-dimensional time series data into two-dimensional graphics data according to the MGAF method; further train the ConvNeXt network model based on the training set data, and adopt a mixed precision quantization model training acceleration method.

Claims

1. A MGAF-ConvNeXt intelligent evaluation method for frequency stability of renewable energy power systems under incomplete measurement, characterized by: The steps include: Step 1: Perform frequency response partitioning of the new energy power system to obtain frequency response partitioning results; Step 1 specifically includes the following steps: Step 1.1: Establish a new energy power system case based on the PSASP platform, construct scenarios with different outputs of new energy and different load disturbances, and simulate the frequency disturbance trajectory of each node in the power system based on each constructed scenario; Step 1.2: Based on the frequency-disturbed trajectory of each node in the system, calculate the similarity matrix of the frequency response curve of each node based on the Hausdorff distance; the Hausdorff distance calculation formula between the frequency-disturbed trajectory of node m and the frequency-disturbed trajectory of node n is: (1) in: , ; The data set of frequency disturbed trajectory m=[m1, m2, …, m a ] and n=[n1, n2, …, n b ] are two curves of unequal lengths, a and b are the total lengths of data in sets m and n respectively; h(n,m) is the one-way Hausdorff distance from set n to set m, ||m u -n v || is m u to n v Euclidean distance; u=[1, 2,…, a], v=[1,2,…, b], for a system with N nodes, the dimension of the similarity matrix H is N×N; Step 1.3: Use the K-means++ clustering algorithm to classify the similarity matrix and obtain the cluster centers required for power system partitioning; Step 1.3 is as follows: Randomly select the i-th row h in the Hausdorff similarity matrix H i As the kth initial cluster center h k , i=[1, 2,…, N]; k=[1, 2,…, l], where l is the number of cluster centers; Calculate the shortest distance D(H) between each node and the current cluster center: (2) Where, j = [1, 2, …, N]; h ij represents the element in the i-th row and j-th column of the similarity matrix H, h kj represents the k-th row and j-th column element in the similarity matrix H; Calculate the probability P(H) of each node being selected as the next cluster center, and select the node corresponding to the maximum probability value as the next cluster center; (3) Repeat equations (2) and (3) until the l cluster centers required for power system partitioning are selected, and based on D(H), find the minimum distance between each node and each cluster center, divide the node with the shortest distance to the cluster center into a region, and finally obtain the system partition result; Step 2: Construct a hybrid Grammar angular field-next-generation convolutional neural network MGAF-ConvNeXt frequency response intelligent evaluation model; Step 2 specifically includes the following steps: Step 2.1: The Mixed Gram Angular Field (MGAF) method converts one-dimensional data into a two-dimensional matrix that can be recognized by the deep learning model, including: The one-dimensional time series data g of length L is mapped to the interval [-1, 1] through maximum and minimum normalization. The formula is: (4) Where g ζ is the ζ-th variable in sample g, is the normalized sample, ζ=[0, 1,…, L]; Convert the normalized data to the polar coordinate system using the following formula: (5) Where: Φ ζ is the ζth polar angle in the sample, r ζ is the ζth polar diameter in the sample; Step 2.2: Fuse the Gram sum angular field GASF and the Gram difference angular field GADF to generate the asymmetric MGAF matrix. The formula is: (6) Among them, Φ β is the βth polar angle in the sample, r β is the βth pole in the sample, β=[0, 1,…, L]; Step 2.3, ConvNeXt deep learning model, the structure includes input layer, intermediate layer and output layer; The input layer performs spatial downsampling and feature dimension expansion through 4×4 convolution; The middle layer consists of three cascaded ConvNeXt Block and MaxPooling composite units. The ConvNeXt Block includes: a 7×7 depthwise separable convolutional layer; a Layer Normalization layer; and a bottleneck structure consisting of two levels of 1×1 convolutional layers, with a GELU activation function embedded between the two levels of 1×1 convolutional layers. The output layer generates a five-dimensional frequency stability index through global average pooling and a fully connected layer; Step 3: Simulate and obtain the response data of the example system under multiple operating modes and multiple disturbance scenarios. Based on the frequency response partitioning results of Step 1, select the generator with the largest inertia or unit capacity in each area or the closest distance to the system inertia center as the measurement point to construct a historical data set. Step 4: Accelerate the training of the MGAF-ConvNeXt frequency response intelligent evaluation model based on the mixed precision quantization method; Step 5: Establish an incomplete measurement case library, train the MGAF-ConvNeXt frequency response intelligent evaluation model using historical data sets, and perform error analysis on the frequency stability prediction accuracy under incomplete measurements based on the trained intelligent evaluation model.

2. The MGAF-ConvNeXt intelligent evaluation method for frequency stability of a new energy power system under incomplete measurement according to claim 1 is characterized in that: Step 3 specifically includes the following steps: Step 3.

1. Input data: voltage amplitude V, current amplitude I, active power P, and reactive power Q of the measurement node; Step 3.2, output data: the maximum frequency change rate of the system inertia center ROCOF max , maximum frequency offset Δf max , quasi-steady-state frequency f ss , maximum frequency deviation delay t ir and quasi-steady-state frequency delay t ss , specifically defined as follows: (7) (8) (9) (10) (11) Among them, f COI is the system inertia center frequency; f0 is the system steady-state frequency; ε is the threshold; d represents the differential operator, and t represents time; Step 3.3, the disturbance scenario is set as follows: at t=0s, each load node in the system increases the load by 10-100MW in steps of 10MW, and the number of disturbance scenarios is 10 times the number of load nodes; the scenario set is a combination of different renewable energy output scenarios and different disturbance scenarios. According to the simulation of different scenario sets, the disturbed trajectories of the frequency f, voltage amplitude V, current amplitude I, active power P and reactive power Q of each node are obtained. The time window of the measurement data is selected as [0.1, 0.3]s, and the sampling frequency is 100Hz. The disturbed trajectories obtained under the scenario set are randomly divided into training set and test set according to proportion.

3. The MGAF-ConvNeXt intelligent evaluation method for frequency stability of a new energy power system under incomplete measurement according to claim 1 is characterized in that: Step 4 specifically includes the following steps: Step 4.1: During the forward and backward propagation of training, the activation, weight, and gradient data are stored in FP16 format, and a copy of the weights in FP32 format is retained for gradient updates. Step 4.2: Weight update retains the 32-bit floating point precision master copy, and the gradient update formula is: (12) Among them, W FP32 is a 32-bit floating point weight, G FP16 is a 16-bit weight gradient update, is the scaling factor for gradient updates; this is done by backing up the weights in FP32 format and ensuring that the update process is performed in FP32.

4. The MGAF-ConvNeXt intelligent evaluation method for frequency stability of a new energy power system under incomplete measurement according to claim 1 is characterized in that: Step 5 specifically includes the following steps: Step 5.1: The established system incomplete measurement case library includes: Case 0: All power nodes are measurable; Case 1: A single inertia power node can be measured; Case 2: All inertia power nodes in the specified partition can be measured; Case 3: A single inertia power node in each partition can be measured, specifically: Case 3-1 to Case 3-3: Select the node with the largest inertia time constant, the closest distance to the inertia center, or the largest unit capacity in each partition; Case 3-4 to Case 3-15: Optimize the measurement configuration by replacing suboptimal nodes within the partition or adding redundant nodes; Step 5.2: Determine the measurable nodes based on the incomplete measurement case library, determine the data set accordingly, train the MGAF-ConvNeXt model based on the training set data, and further perform frequency stability prediction on the test set data based on the trained model to obtain the system's frequency stability index, specifically including the maximum frequency change rate of the system's center of inertia, ROCOF. max , maximum frequency offset Δf max , quasi-steady-state frequency f ss , maximum frequency deviation delay t ir and quasi-steady-state frequency delay t ss ; The errors of frequency stability indicators under different case libraries are compared, and the accuracy of intelligent assessment of power system frequency stability under incomplete measurements is analyzed.

5. A MGAF-ConvNeXt intelligent evaluation system for frequency stability of a new energy power system under incomplete measurement, used to implement the method according to claim 1, characterized in that: include: Modeling unit: built the MGAF-ConvNeXt frequency response intelligent evaluation model; The design unit calculates similarity based on the frequency response data of all nodes in the system, and clusters the similarity matrix using the K-means++ method to obtain the frequency response partitioning result of the system; Computing unit: Convert one-dimensional time series data into two-dimensional graphics data according to the MGAF method; further train the ConvNeXt network model based on the training set data, and adopt a mixed precision quantization model training acceleration method.

6. A computer device comprising a memory, a processor, and a computer program stored in the memory, wherein: The processor executes the computer program to implement the steps of the method according to claim 1.

7. A computer-readable storage medium having a computer program / instruction stored thereon, characterized in that: When the computer program / instructions are executed by a processor, the steps of the method according to claim 1 are implemented.

Images