A continuous evaluation method for power system node inertia based on convolutional neural network

Abstract

The present invention relates to the technical field of power system inertia, and specifically refers to a method for continuously evaluating power system node inertia based on a convolutional neural network, comprising: obtaining different combinations of power scenarios based on a power system simulation model, obtaining voltage spectrum data of different frequency domain sampling points of each node under each group of power scenarios based on voltage data of different time domain sampling points of each node under each group of power scenarios within a preset sampling period, and then determining the target frequency band of each node under each group of power scenarios; calculating the theoretical inertia value of each node under each group of power scenarios; using the amplitude modulus of the voltage spectrum data within the target frequency band of each node under each group of power scenarios as input data and the theoretical inertia value of each node under each group of power scenarios as output data, training a convolutional neural network, and constructing a node inertia evaluation model based on a convolutional neural network. The present invention improves the accuracy of power system inertia estimation and enhances the stability and safety of power system operation.

Description

Technical Field

[0001] The present invention relates to the technical field of power system inertia, and in particular to a method for continuously evaluating power system node inertia based on a convolutional neural network. Background Art

[0002] With the rapid development of renewable energy worldwide, traditional synchronous generator-based power grids are gradually being replaced by inverter-based resources (IBRs). Inverter-based resources refer to various power generation, energy storage and other related resources and supporting facilities, technologies, and information in the power system that are connected to the power system through power electronics interfaces by converting direct current into alternating current.

[0003] Synchronous generators can maintain their speed through the inertia provided by their rotating machinery when system loads change or disturbances occur, quickly resisting frequency fluctuations and thus smoothing system frequency changes. Furthermore, synchronous generators can effectively regulate system voltage by adjusting their own inertia and flux linkage, for example, through methods such as adjusting excitation. However, because inverter-based resources lack physical inertia, they cannot provide inertia support to smooth power changes when system loads suddenly change or when generated power suddenly changes. This can lead to increased frequency fluctuations in the power system, slowing the rate at which the system frequency returns to a stable value, and even causing persistent frequency excursions, making it difficult to return to normal operating ranges. This affects the stable operation of the power system and can easily cause rapid fluctuations in grid connection point voltage. Their ability to regulate power system voltage is weak, making it impossible to maintain voltage stability in a timely and effective manner when system reactive power demand fluctuates significantly, increasing the risk of voltage instability. Furthermore, the large-scale penetration of IBRs has led to a decrease in the overall inertia level of the power grid and a more dynamic inertia distribution. This makes the power system susceptible to low-frequency oscillations after disturbances, increasing the threat to grid stability and posing a significant security risk to the power system. Therefore, it is necessary to evaluate the inertia of the power system to clarify the changing trend of the system inertia under the resource access based on the inverter, identify the dangerous areas of system inertia in advance, and then prepare countermeasures to ensure the safety of system frequency.

[0004] Existing grid inertia estimation methods typically rely on frequency information, particularly by leveraging the relationship between frequency and active power after a significant disturbance. However, these methods suffer from insufficient real-time performance and rely on the significant disturbance's duration as a trigger, making continuous inertia assessment difficult. Furthermore, while some existing technologies can estimate inertia based on normal operating data, these rely on the assumptions of injected disturbance signals or system identification models, presenting significant limitations in practical grid applications. This results in low inertia estimation accuracy, impacting the stable operation of the power system. Summary of the Invention

[0005] To this end, the technical problem to be solved by the present invention is to overcome the problems in the existing technology of using frequency information and relying on disturbance event triggering, or using operation data and relying on injected disturbance signals or system identification models to perform inertia assessment, which have insufficient real-time performance, poor adaptability in practical applications, long power system recovery time, low reliability and safety, and affect the stable operation efficiency of the power system.

[0006] To solve the above technical problems, the present invention provides a method for continuous evaluation of power system node inertia based on convolutional neural networks, comprising:

[0007] Based on the power system simulation model, different combinations of power scenarios are obtained. Based on the voltage data of different time domain sampling points of each node in each power scenario within a preset sampling period, the voltage spectrum data of different frequency domain sampling points of each node in each power scenario are obtained. The nodes include each generator node, each load node, and each transmission line node.

[0008] Based on the voltage spectrum data of each node in each power scenario whose amplitude modulus energy ratio is greater than the energy threshold, the frequency domain sampling point value range of each node in each power scenario is obtained to determine the corresponding discretized frequency value range as the target frequency band of each node in each power scenario;

[0009] Calculate the theoretical inertia value of each node in each power scenario based on the inertia constant and rated capacity of each generator within the preset range of each node in each power scenario;

[0010] The amplitude modulus of the voltage spectrum data within the target frequency band of each node in each power scenario is used as input data, and the theoretical inertia value of each node in each power scenario is used as output data. A convolutional neural network is trained to construct a node inertia evaluation model based on the convolutional neural network.

[0011] Preferably, the obtaining of different combinations of power scenarios based on the power system simulation model includes:

[0012] Build an IEEE 39-node power system simulation model that includes various generators, loads, transmission lines, and line faults. Generators include synchronous generators and renewable energy inverter-type power generation equipment; loads include industrial loads, residential loads, and random load fluctuations.

[0013] Set the random distribution range of parameters for various generators, various loads, transmission lines, and line faults;

[0014] A generative adversarial network is used to randomly generate parameters of various types of generators, loads, transmission lines, and line faults to obtain power scenarios with different combinations.

[0015] Preferably, obtaining voltage data of different sampling points of each node in each group of power scenarios includes:

[0016] The randomly generated parameters of various generators, loads, transmission lines, and line faults are automatically imported into the IEEE 39-node power system simulation model. Power flow analysis and dynamic stability analysis methods are used to record the voltage data of different sampling points of each node in each power scenario in the power system simulation model.

[0017] Preferably, obtaining voltage spectrum data of different frequency domain sampling points of each node in each group of power scenarios based on voltage data of different time domain sampling points of each node in each group of power scenarios within a preset sampling period includes:

[0018] The voltage data of different time domain sampling points of each node in each power scenario within the preset sampling period are processed by outlier removal, low-pass filtering, spectrum normalization and frequency domain transformation to obtain the voltage spectrum data of different frequency domain sampling points of each node in each power scenario, namely:

[0019] The Z-score normalization method is used to calculate the Z-score value of the voltage data of each node and each time domain sampling point in each group of power scenarios;

[0020] Perform an outlier removal operation on the voltage data of each time-domain sampling point of each node in each power scenario, remove the voltage data with a Z-score value greater than the abnormal threshold, and obtain the normal voltage data of the remaining time-domain sampling points of each node in each power scenario;

[0021] Using a Butterworth low-pass filter, low-pass filter processing is performed on the normal voltage data of each remaining time domain sampling point of each node in each power scenario to obtain the filtered normal voltage data of each remaining time domain sampling point of each node in each power scenario;

[0022] Performing spectrum normalization processing on the filtered normal voltage data of the remaining time domain sampling points of each node in each group of power scenarios to obtain the normalized normal voltage data of the remaining time domain sampling points of each node in each group of power scenarios;

[0023] By using fast Fourier transform, the normalized voltage data of the remaining time domain sampling points of each node in each power scenario are transformed into the frequency domain to obtain the voltage spectrum data of each frequency domain sampling point of each node in each power scenario.

[0024] Preferably, the frequency domain sampling point value range of each node in each group of power scenarios is obtained based on the voltage spectrum data of each node in each group of power scenarios whose amplitude modulus energy ratio is greater than the energy threshold, including:

[0025] Based on the The first group of power scenarios Node No. The voltage spectrum data of the frequency domain sampling points is extracted The first group of power scenarios Node No. The amplitude modulus of the voltage spectrum data of the frequency domain sampling point is expressed as:

[0026] ;

[0027] in, 、 Respectively represent The first group of power scenarios Node No. Voltage spectrum data of frequency domain sampling points The real and imaginary parts of Indicates the The first group of power scenarios Node No. The amplitude modulus of the voltage spectrum data at each frequency domain sampling point;

[0028] According to The first group of power scenarios Node No. The voltage spectrum data of the frequency domain sampling point is calculated The first group of power scenarios Node No. The amplitude modulus energy ratio of the voltage spectrum data at the frequency domain sampling point , whose expression is:

[0029] ;

[0030] Set the energy threshold to , retain the voltage spectrum data whose energy ratio is greater than the energy threshold, and get the The first group of power scenarios The target set of voltage spectrum data for all frequency domain sampling points of a node , whose expression is: ;

[0031] Based on the The first group of power scenarios The target set of voltage spectrum data of all frequency domain sampling points of the node is obtained The first group of power scenarios The frequency domain sampling point value range of each node is: , and then obtain the frequency domain sampling value range of each node in each group of power scenarios.

[0032] Preferably, based on the frequency domain sampling point value range of each node in each group of power scenarios, the corresponding discretized frequency value range is determined as the target frequency band of each node in each group of power scenarios, including:

[0033] The relationship between frequency and frequency domain sampling point is expressed as:

[0034] ;

[0035] in, Indicates the The frequency corresponding to the frequency domain sampling point; Indicates the total number of sampling points of fast Fourier transform; The index of the frequency domain sampling point representing the fast Fourier transform; Indicates the sampling frequency;

[0036] The first The first group of power scenarios Substitute the minimum and maximum values ​​of the frequency domain sampling point value range of each node into the relationship expression between frequency and frequency domain sampling point to obtain the The first group of power scenarios The range of the frequency domain sampling points of each node corresponds to the discretized frequency range: , as the first The first group of power scenarios The target frequency band of each node.

[0037] Preferably, the inertia theoretical value of each node in each power scenario is calculated based on the inertia constant and rated capacity of each generator within a preset range of each node in each power scenario, and the expression is:

[0038] ;

[0039] in, Indicates the The first group of power scenarios Theoretical value of inertia of each node; Indicates the The first group of power scenarios The first node within the preset range Rated capacity of each generator; Indicates the The first group of power scenarios The first node within the preset range The inertia constant of the generator.

[0040] Preferably, the network structure of the convolutional neural network includes: a convolution module, a pooling module, a channel attention module, a fully connected module and an output layer;

[0041] The convolution module includes a three-layer one-dimensional convolution structure, a ReLU corrected linear unit activation function and a random inactivation suppression technology; wherein each convolution layer in the three-layer one-dimensional convolution has 32 convolution kernels and uses a scale of 、 、 Multi-scale convolution kernel;

[0042] The pooling module includes a layer of maximum pooling operation and a layer of global average pooling operation; wherein the pooling window size is 4;

[0043] The channel attention module includes a Sigmoid activation function;

[0044] The fully connected module includes two fully connected layers, the first fully connected layer contains 108 neurons, and the second fully connected layer contains 54 neurons;

[0045] The output layer outputs the inertia prediction value of each node in the power system.

[0046] Preferably, the training process of the node inertia evaluation model based on the convolutional neural network includes:

[0047] Based on the The first group of power scenarios The amplitude modulus of the voltage spectrum data within the target frequency band of each node constitutes the The first group of power scenarios The spectral feature matrix of nodes is input into the convolutional neural network; wherein the dimension of the spectral feature matrix is , Indicates the The total number of nodes in the group power scenario, Indicates the total number of frequency sampling points;

[0048] A dataset is constructed based on the spectrum feature matrices of different nodes under different power scenarios and their corresponding theoretical values ​​of node inertia. The dataset is randomly divided into training and test sets according to a preset ratio.

[0049] Using the training set, a convolutional neural network-based node inertia estimation model was trained. The Adam optimizer was used to iteratively update the parameters of the convolutional neural network-based node inertia estimation model. The root mean square error metric was used to evaluate the performance of the convolutional neural network-based node inertia estimation model. The parameters of the convolutional neural network-based node inertia estimation model included the convolutional neural network's convolution kernel size, multi-scale design, and attention module parameters.

[0050] After each training iteration, the test set is used to evaluate the performance of the node inertia evaluation model based on the convolutional neural network, including error evaluation and convergence judgment, until the error index is met, the optimal parameters of the node inertia evaluation model based on the convolutional neural network are obtained, and the trained inertia evaluation model based on the CNN neural network is obtained.

[0051] Preferably, after building the node inertia evaluation model based on the convolutional neural network, the method further includes:

[0052] A synchronized phasor measurement unit is used to collect voltage data of each node in the actual power system in real time, and the amplitude modulus of the voltage spectrum data of each node is extracted in real time. The amplitude modulus of the voltage spectrum data of each node is input into a trained node inertia assessment model based on a convolutional neural network, and the inertia prediction value of each node in the power system is output in real time.

[0053] The above technical solution of the present invention has the following beneficial effects compared with the prior art:

[0054] (1) The present invention discloses a method for continuous inertia evaluation of power system nodes based on convolutional neural networks. The method uses the voltage time series signal of the power system node as the core input for inertia evaluation and extracts the features related to inertia through spectrum analysis, thereby realizing high-precision and real-time evaluation of inertia. The voltage signal can be obtained in real time and is not restricted by the time of significant disturbance. It can timely reflect the changes in the operation status of the power grid and provide a timely data source for real-time evaluation of inertia. Moreover, the voltage signal can be used as input to reflect not only the dynamic characteristics of power fluctuations but also the spectrum characteristics of the inertia of the equipment inside the power system, providing a basis for more comprehensive inertia analysis. At the same time, the method does not rely on the injection of additional disturbance signals or the installation of special equipment. The method can be applied to power grids of different scales, topologies and operating modes, including scenarios with a high proportion of renewable energy access. In addition, based on the existing voltage measurement devices and data, the method avoids complex hardware transformation and high cost investment and can be directly integrated into the existing power grid monitoring system, which has good economic benefits and promotion potential.

[0055] (2) The present invention discloses a method for continuous evaluation of power system node inertia based on convolutional neural network. It is the first time that convolutional neural network is used to perform data-driven modeling of power system inertia, which is independent of the mathematical model of a specific system. At the same time, the convolutional neural network is used to process the grid node voltage signal online to achieve continuous evaluation of inertia, thus breaking through the technical bottleneck of traditional methods that can only perform intermittent evaluation. That is, the model learns the relationship between voltage spectrum characteristics and inertia through preliminary training, and can quickly process input data. Each time new voltage spectrum characteristic data is collected, the model can output the evaluated node inertia value in real time, thus achieving continuous evaluation of inertia, overcoming the defect that traditional methods can only perform intermittent evaluation, and improving the efficiency of stable operation of the power system. In addition, by continuously tracking the grid inertia, the risk of inertia fluctuation can be discovered in a timely manner, and data support can be provided for dynamic safety analysis and system stability control. It has important application value, especially in complex power grids with high penetration of new energy.

[0056] (3) The present invention describes a method for continuous inertia assessment of power system nodes based on convolutional neural networks. This method fully considers the differences in the characteristic responses of inertia to voltage spectra in different frequency bands. Through the adaptive adjustment of the frequency band of the CNN layer, it can accurately distinguish the inertia characteristics of synchronous generators, synchronous compensators, and inverter-based equipment. At the same time, through spectrum feature extraction and automatic learning, it can adapt to different types of power equipment and changing grid operating conditions. In addition, it can also adapt to changes in the characteristics of different equipment and network topology, and its estimation accuracy is significantly better than traditional methods, providing an innovative technical path for grid operators to conduct stability analysis in the context of a high proportion of IBRs access. BRIEF DESCRIPTION OF THE DRAWINGS

[0057] In order to make the content of the present invention more clearly understood, the present invention is further described in detail below based on specific embodiments of the present invention in conjunction with the accompanying drawings, wherein:

[0058] Figure 1 This is a flow chart of a method for continuous evaluation of power system node inertia based on convolutional neural networks provided by the present invention. DETAILED DESCRIPTION

[0059] The present invention will be further described below with reference to the accompanying drawings and specific embodiments so that those skilled in the art can better understand the present invention and implement it. However, the embodiments are not intended to limit the present invention.

[0060] Example 1

[0061] Reference Figure 1 As shown, Figure 1 This is a flow chart of a method for continuous evaluation of power system node inertia based on a convolutional neural network provided by the present invention; specifically comprising:

[0062] S1: Based on the power system simulation model, obtain different combinations of power scenarios, including:

[0063] Build an IEEE 39-node power system simulation model that includes various generators, loads, transmission lines, and line faults. Generators include synchronous generators and renewable energy inverter-type power generation equipment; loads include industrial loads, residential loads, and random load fluctuations.

[0064] Set the random distribution range of parameters for various generators, various loads, transmission lines, and line faults;

[0065] Among them, the random distribution range of parameters of various generators includes the random distribution range of inertia constant and rated power of synchronous generators, as well as the random distribution range of power generation of renewable energy inverter-type power generation equipment; the random distribution range of inertia constant of synchronous generators is , in seconds; the random distribution range of the rated power of the synchronous generator is , the unit is MVA; renewable energy inverter type power generation equipment is a non-inertia device, and its power generation random distribution range is , unit is MW;

[0066] The random distribution ranges of various loads are the random distribution range of industrial load power, the random distribution range of residential load power and random load fluctuation; the random distribution range of industrial load power is , in MW; the random distribution range of residential load power is , unit MW; the dynamic model of random load fluctuation is: ;in, Indicates time interval Change in internal load; represents the load mean, represents the regression rate; Indicates the actual load power at the current moment; Indicates the fluctuation amplitude; represents a random variable with a standard normal distribution; MW represents a standard normally distributed random variable with a mean of 0 and a standard deviation of 10MW;

[0067] The random distribution range of transmission line parameters includes the random distribution range of resistance and inductive reactance between each node; the random distribution range of resistance between any two nodes is , the unit is The random distribution range of the inductive reactance of the transmission line between any two nodes is , the unit is ;

[0068] The random distribution range of the parameters of the line fault is the random distribution range of the power of the line fault, and the random distribution range of the power of the line fault is , unit is MW;

[0069] Generative adversarial networks are used to randomly generate parameters for various types of generators, loads, transmission lines, and line faults, resulting in different combinations of power scenarios.

[0070] S2: Voltage data acquisition process, including:

[0071] Randomly generated parameters for various generators, loads, transmission lines, and line faults are automatically imported into an IEEE 39-node power system simulation model. Power flow analysis and dynamic stability analysis methods are then used to record voltage data at different sampling points for each node in each power scenario within the power system simulation model. Each node includes each generator node, each load node, and each transmission line node.

[0072] S3: Voltage data processing, including:

[0073] S31: Based on the voltage data of different time domain sampling points of each node in each power scenario in a preset sampling period, voltage spectrum data of different frequency domain sampling points of each node in each power scenario are obtained, including:

[0074] The voltage data of different time domain sampling points of each node in each power scenario within the preset sampling period are processed by outlier removal, low-pass filtering, spectrum normalization and frequency domain transformation to obtain the voltage spectrum data of different frequency domain sampling points of each node in each power scenario. The specific operations include:

[0075] The Z-score normalization method is used to calculate the Z-score value of the voltage data of each node and each time domain sampling point in each group of power scenarios. The first group of power scenarios Node No. The expression of the Z-score value of the voltage data of a time domain sampling point is:

[0076] ;

[0077] in, Indicates the The first group of power scenarios Node No. The Z-score value of the voltage data at each time domain sampling point; Indicates the The first group of power scenarios Node No. Voltage data of time domain sampling points; , Indicates the index of the sampling point, Indicates the total number of sampling points; Indicates the The first group of power scenarios nodes The mean value of the voltage data of the time domain sampling points; Indicates the The first group of power scenarios nodes The standard deviation of the voltage data at each time domain sampling point;

[0078] The voltage data of each time domain sampling point of each node in each power scenario are subjected to abnormal point removal operation, and the voltage data with Z-score value greater than the abnormal threshold are eliminated to obtain the normal voltage data of each remaining time domain sampling point of each node in each power scenario. The first group of power scenarios The sampling order of the voltage data of the remaining time domain sampling points of the node is obtained The first group of power scenarios nodes Normal voltage data of time domain sampling points; Represents the total number of all remaining time domain sampling points;

[0079] Using Butterworth low-pass filter, the normal voltage data of each remaining time domain sampling point of each node under each power scenario is low-pass filtered to obtain the filtered normal voltage data of each remaining time domain sampling point of each node under each power scenario. The first group of power scenarios Node No. The expression of the filtered normal voltage data of the time domain sampling point is:

[0080] ;

[0081] in, Indicates the The first group of power scenarios Node No. Normal voltage data of time domain sampling points; Indicates the The first group of power scenarios Node No. Normal voltage data of time domain sampling points; Indicates the The first group of power scenarios Node No. Normal voltage data after filtering at time domain sampling points; Indicates the The first group of power scenarios Node No. Normal voltage data after filtering at time domain sampling points; represents the first coefficient of the low-pass filter; represents the second coefficient of the low-pass filter; Indicates the input order of the low-pass filter; Indicates the output order of the low-pass filter; Indicates the operation number of the low-pass filter;

[0082] Perform spectrum normalization on the filtered normal voltage data of the remaining time domain sampling points of each node under each group of power scenarios to obtain the normalized normal voltage data of the remaining time domain sampling points of each node under each group of power scenarios. The first group of power scenarios Node No. The expression of the normalized normal voltage data of the time domain sampling point is:

[0083] ;

[0084] in, Indicates the The first group of power scenarios Node No. Normal voltage data after normalization of time domain sampling points; Indicates the The first group of power scenarios nodes The mean value of the filtered normal voltage data of the time domain sampling points; Indicates the The first group of power scenarios nodes The standard deviation of the filtered normal voltage data of the time domain sampling points;

[0085] By using fast Fourier transform, the normal voltage data of the remaining time domain sampling points of each node in each group of power scenarios are transformed into frequency domain to obtain the voltage spectrum data of each frequency domain sampling point of each node in each group of power scenarios. The first group of power scenarios Node No. The expression of the voltage spectrum data of a frequency domain sampling point is:

[0086] ;

[0087] in, Indicates the The first group of power scenarios Node No. Voltage spectrum data of frequency domain sampling points; Indicates the total number of frequency domain sampling points of the fast Fourier transform; The index of the frequency domain sampling point representing the fast Fourier transform;

[0088] S22: Based on the voltage spectrum data of each node in each power scenario whose amplitude modulus energy ratio is greater than the energy threshold, obtain the frequency domain sampling point value range of each node in each power scenario, so as to determine the corresponding discretized frequency value range as the target frequency band of each node in each power scenario, including:

[0089] Based on the The first group of power scenarios Node No. The voltage spectrum data of the frequency domain sampling points is extracted The first group of power scenarios Node No. The amplitude modulus of the voltage spectrum data of the frequency domain sampling point is expressed as:

[0090] ;

[0091] in, 、 Respectively represent The first group of power scenarios Node No. Voltage spectrum data of frequency domain sampling points The real and imaginary parts of Indicates the The first group of power scenarios Node No. The amplitude modulus of the voltage spectrum data at each frequency domain sampling point;

[0092] According to The first group of power scenarios Node No. The voltage spectrum data of the frequency domain sampling point is calculated The first group of power scenarios Node No. The amplitude modulus energy ratio of the voltage spectrum data at the frequency domain sampling point , whose expression is:

[0093] ;

[0094] Set the energy threshold to , retain the voltage spectrum data whose energy ratio is greater than the energy threshold, and get the The first group of power scenarios The target set of voltage spectrum data for all frequency domain sampling points of a node , whose expression is: ;

[0095] Based on the The first group of power scenarios The target set of voltage spectrum data of all frequency domain sampling points of the node is obtained The first group of power scenarios The frequency domain sampling point value range of each node is: , and then obtain the frequency domain sampling value range of each node in each group of power scenarios;

[0096] The relationship between frequency and frequency domain sampling point is expressed as:

[0097] ;

[0098] in, Indicates the The frequency corresponding to the frequency domain sampling point; Indicates the total number of sampling points of fast Fourier transform; The index of the frequency domain sampling point representing the fast Fourier transform; Indicates the sampling frequency;

[0099] The first The first group of power scenarios Substitute the minimum and maximum values ​​of the frequency domain sampling point value range of each node into the relationship expression between frequency and frequency domain sampling point to obtain the The first group of power scenarios The range of the frequency domain sampling points of each node corresponds to the discretized frequency range: , as the first The first group of power scenarios The target frequency band of each node;

[0100] S4: Based on the inertia constant and rated capacity of each generator within the preset range of each node in each power scenario, calculate the theoretical inertia value of each node in each power scenario. The expression is:

[0101] ;

[0102] in, Indicates the The first group of power scenarios Theoretical value of inertia of each node; Indicates the The first group of power scenarios The first node within the preset range Rated capacity of each generator; Indicates the The first group of power scenarios The first node within the preset range The inertia constant of each generator; the preset range of each node refers to the neighborhood of each node;

[0103] S5: Using the amplitude modulus of the voltage spectrum data within the target frequency band of each node in each power scenario as input data and the theoretical inertia value of each node in each power scenario as output data, a convolutional neural network is trained to construct a node inertia assessment model based on the convolutional neural network.

[0104] The network structure of the convolutional neural network includes: a convolution module, a pooling module, a channel attention module, a fully connected module and an output layer;

[0105] The convolution module includes a three-layer one-dimensional convolution structure, a ReLU corrected linear unit activation function and a random inactivation suppression technology; wherein each convolution layer in the three-layer one-dimensional convolution has 32 convolution kernels and uses a scale of 、 、 The multi-scale convolution kernel is used to extract local features of the input data; the ReLU corrected linear unit activation function is used to extract spectral dynamic features of different scales; in addition, random deactivation technology is introduced in the convolution module to suppress the overfitting phenomenon of the CNN-LSTM neural network; in a specific embodiment of the present invention, the Dropout rate is set to 0.5;

[0106] The pooling module includes a layer of maximum pooling operation and a layer of global average pooling operation; wherein the pooling window size is 4; it is used to reduce the feature scale and retain key features;

[0107] The channel attention module includes a Sigmoid activation function, which is used to dynamically adjust the channel weights to improve the quality of the features that the model focuses on.

[0108] The fully connected module includes two fully connected layers, the first fully connected layer includes 108 neurons, and the second fully connected layer includes 54 neurons; it is used to obtain the inertia value of each node;

[0109] The output layer outputs the inertia prediction value of each node in the power system.

[0110] The training process of the node inertia evaluation model based on convolutional neural network includes:

[0111] Based on the The first group of power scenarios The amplitude modulus of the voltage spectrum data within the target frequency band of each node constitutes the The first group of power scenarios The spectral feature matrix of nodes is input into the convolutional neural network; wherein the dimension of the spectral feature matrix is , Indicates the The total number of nodes in the group power scenario, Indicates the total number of frequency sampling points;

[0112] A data set is constructed based on the spectrum feature matrices of different nodes under different power scenarios and their corresponding theoretical values ​​of node inertia. The data set is randomly divided into a training set and a test set according to a preset ratio. The preset ratio is 8:2, meaning that 80% of the random data in the data set constitutes the training set and 20% of the random data constitutes the test set.

[0113] Using the training set, a convolutional neural network-based node inertia estimation model was trained. The Adam optimizer was used to iteratively update the parameters of the convolutional neural network-based node inertia estimation model. The root mean square error metric was used as the loss function to evaluate the performance of the convolutional neural network-based node inertia estimation model. The parameters of the convolutional neural network-based node inertia estimation model included the convolution kernel size, multi-scale design, and attention module parameters of the convolutional neural network.

[0114] After each training iteration, the test set is used to evaluate the performance of the node inertia evaluation model based on the convolutional neural network, including error evaluation and convergence judgment, until the error index is met, the optimal parameters of the node inertia evaluation model based on the convolutional neural network are obtained, and the trained inertia evaluation model based on the CNN neural network is obtained.

[0115] After building the node inertia evaluation model based on convolutional neural network, it also includes:

[0116] A synchronized phasor measurement unit is used to collect voltage data of each node in the actual power system in real time, and the amplitude modulus of the voltage spectrum data of each node is extracted in real time. The amplitude modulus of the voltage spectrum data of each node is input into a trained node inertia assessment model based on a convolutional neural network, and the inertia prediction value of each node in the power system is output in real time.

[0117] Example 2

[0118] Based on the process from S1 to S5 in Example 1, the implementation process and technical effects of the method are described in detail with a regional power grid as the background; specifically, the following are described:

[0119] Step 1: Construct an IEEE 39-bus power system model, which includes 10 synchronous generators, several wind power and photovoltaic power stations, and industrial and residential loads. The inertia constant of the synchronous generators is randomly distributed between 2 and 10 seconds, and the rated power is randomly distributed between 100 and 500 MVA. The power generation of renewable energy inverter-type power generation equipment is randomly distributed between 10 and 200 MVA. The power of industrial loads is randomly distributed between 100 and 300 MW. The power of residential loads is randomly distributed between 50 and 150 MW. The fluctuation amplitude of random load fluctuations is randomly distributed between 10 and 10 seconds. The resistance parameters of the transmission line between any two nodes are randomly distributed in the range of 0.01 to 0.05. ; The reactance of the transmission line between any two nodes is randomly distributed in the range of 0.1 to 0.5 ; and used generative adversarial networks to generate 2,000 sets of power scenario data;

[0120] Step 2: Based on 2000 randomly generated power scenario data sets, power flow analysis and dynamic stability analysis methods were used to obtain the voltage time series signal within 1 second of each power scenario simulation. The sampling rate was 1000 Hz, and the voltage data at different time domain sampling points of each node in each power scenario were obtained.

[0121] Step 3: Process the voltage data. The Z-score method is used to detect and remove outliers from all voltage data, with the outlier threshold set at 2.5. A Butterworth filter is used to filter all voltage data excluding outliers, with a filter cutoff frequency of 5 Hz. All filtered voltage data are normalized. Fast Fourier transform (FFT) is used to extract the voltage spectrum features in the 0.5 Hz to 3 Hz frequency band from all normalized voltage data.

[0122] Step 4: Based on the inertia formula, calculate the theoretical inertia value of each node in each power scenario;

[0123] Step 5: Build a node inertia evaluation model based on convolutional neural network.

[0124] The grid was configured with synchronized phasor measurement units (PMUs) to collect node voltage signals in real time. The wind and photovoltaic penetration rates in the test scenario gradually increased to 80%. A trained convolutional neural network-based node inertia assessment model was used to assess node inertia in real time. Table 1 shows a comparison between the theoretical and predicted inertia values ​​for some nodes.

[0125] Table 1 Comparison of theoretical inertia values ​​and predicted inertia values ​​of nodes

[0126] .

[0127] With the continuous increase in the proportion of new energy access, the traditional power system based on synchronous generators is transitioning to a system based on inverters; however, since the inverter does not provide physical inertia, the overall inertia level of the power grid has dropped significantly, and the frequency stability is facing challenges; to address this problem, the present invention proposes a node inertia continuous evaluation method based on convolutional neural networks, which accurately and dynamically evaluates the inertia value of each node in the power grid by real-time analysis of voltage time series data. The present invention uses voltage time series signals as the core input for inertia evaluation. Different from the traditional inertia calculation method that relies on frequency and power response, the voltage signal can not only reflect the dynamic characteristics of power fluctuations, but also includes the spectral characteristics of the inertia of the internal equipment of the system, providing a basis for more comprehensive inertia analysis. Traditional inertia estimation methods are mostly based on triggering of significant disturbance events and cannot achieve real-time update of inertia; the present invention uses convolutional neural networks to process the voltage signals of power grid nodes online to achieve continuous evaluation of inertia, thus breaking through the technical bottleneck that traditional methods can only perform intermittent evaluation. This invention utilizes a convolutional neural network (CNN) for the first time to perform data-driven modeling of power system inertia, independent of a specific system mathematical model. Through spectral feature extraction and automatic learning, this method adapts to different types of power equipment and varying grid operating conditions. The design of this invention fully considers the differences in inertia's response to voltage spectra across different frequency bands. By adaptively adjusting the frequency band of the CNN layer, it can accurately distinguish the inertia characteristics of synchronous generators, synchronous compensators, and inverter-based equipment.

[0128] The present invention only requires the voltage signals of a small number of nodes as input, does not rely on additional disturbance signal injection or special equipment installation, and is applicable to power grids of different scales, topologies, and operating modes, including scenarios with a high proportion of renewable energy access. By continuously tracking the inertia of the power grid, the present invention can promptly detect the risk of inertia fluctuations and provide data support for dynamic safety analysis and system stability control. It has important application value, especially in complex power grids with a high penetration rate of new energy. Since the present invention is based on existing voltage measurement devices and data considerations, it avoids complex hardware modifications and high cost investment, can be directly integrated into existing power grid monitoring systems, and has good economic benefits and promotion potential.

[0129] Obviously, the above embodiments are merely examples for clarity of explanation and are not intended to limit the implementation methods. Those skilled in the art will appreciate that other variations or modifications can be made based on the above description. It is not necessary and impossible to enumerate all implementation methods here. Obvious variations or modifications arising therefrom remain within the scope of protection of the present invention.

Claims

1. A method for continuous evaluation of power system node inertia based on convolutional neural network, characterized in that: include: Based on the power system simulation model, different combinations of power scenarios are obtained. Based on the voltage data of different time domain sampling points of each node in each power scenario within a preset sampling period, the voltage spectrum data of different frequency domain sampling points of each node in each power scenario are obtained. The nodes include each generator node, each load node, and each transmission line node. Based on the voltage spectrum data of each node in each power scenario whose amplitude modulus energy ratio is greater than the energy threshold, the frequency domain sampling point value range of each node in each power scenario is obtained to determine the corresponding discretized frequency value range as the target frequency band of each node in each power scenario; Calculate the theoretical inertia value of each node in each power scenario based on the inertia constant and rated capacity of each generator within the preset range of each node in each power scenario; The amplitude modulus of the voltage spectrum data within the target frequency band of each node in each power scenario is used as input data, and the theoretical inertia value of each node in each power scenario is used as output data. A convolutional neural network is trained to construct a node inertia evaluation model based on the convolutional neural network.

2. The method for continuous evaluation of power system node inertia based on convolutional neural network according to claim 1, characterized in that: The power system simulation model is used to obtain different combinations of power scenarios, including: Build an IEEE 39-node power system simulation model that includes various generators, loads, transmission lines, and line faults. Generators include synchronous generators and renewable energy inverter-type power generation equipment; loads include industrial loads, residential loads, and random load fluctuations. Set the random distribution range of parameters for various generators, various loads, transmission lines, and line faults; A generative adversarial network is used to randomly generate parameters of various types of generators, loads, transmission lines, and line faults to obtain power scenarios with different combinations.

3. The method for continuous evaluation of power system node inertia based on convolutional neural network according to claim 2, characterized in that: Obtaining voltage data at different sampling points of each node in each power scenario includes: The randomly generated parameters of various generators, loads, transmission lines, and line faults are automatically imported into the IEEE 39-node power system simulation model. Power flow analysis and dynamic stability analysis methods are used to record the voltage data of different sampling points of each node in each power scenario in the power system simulation model.

4. The method for continuous evaluation of power system node inertia based on convolutional neural network according to claim 1, characterized in that: The voltage spectrum data of different frequency domain sampling points of each node in each power scenario are obtained based on the voltage data of different time domain sampling points of each node in each power scenario within the preset sampling period, including: The voltage data of different time domain sampling points of each node in each power scenario within the preset sampling period are processed by outlier removal, low-pass filtering, spectrum normalization and frequency domain transformation to obtain the voltage spectrum data of different frequency domain sampling points of each node in each power scenario, namely: The Z-score normalization method is used to calculate the Z-score value of the voltage data of each node and each time domain sampling point in each group of power scenarios; Perform an outlier removal operation on the voltage data of each time-domain sampling point of each node in each power scenario, remove the voltage data with a Z-score value greater than the abnormal threshold, and obtain the normal voltage data of the remaining time-domain sampling points of each node in each power scenario; Using a Butterworth low-pass filter, low-pass filter processing is performed on the normal voltage data of each remaining time domain sampling point of each node in each power scenario to obtain the filtered normal voltage data of each remaining time domain sampling point of each node in each power scenario; Performing spectrum normalization processing on the filtered normal voltage data of the remaining time domain sampling points of each node in each group of power scenarios to obtain the normalized normal voltage data of the remaining time domain sampling points of each node in each group of power scenarios; By using fast Fourier transform, the normalized voltage data of the remaining time domain sampling points of each node in each power scenario are transformed into the frequency domain to obtain the voltage spectrum data of each frequency domain sampling point of each node in each power scenario.

5. The method for continuous evaluation of power system node inertia based on convolutional neural network according to claim 1, characterized in that: The frequency domain sampling point value range of each node in each power scenario is obtained based on the voltage spectrum data of each node in each power scenario whose amplitude modulus energy ratio is greater than the energy threshold. Based on the The first group of power scenarios Node No. The voltage spectrum data of the frequency domain sampling points is extracted The first group of power scenarios Node No. The amplitude modulus of the voltage spectrum data of the frequency domain sampling point is expressed as: ； in, 、 Respectively represent The first group of power scenarios Node No. Voltage spectrum data of frequency domain sampling points The real and imaginary parts of Indicates the The first group of power scenarios Node No. The amplitude modulus of the voltage spectrum data of the frequency domain sampling point; According to The first group of power scenarios Node No. The voltage spectrum data of the frequency domain sampling point is calculated The first group of power scenarios Node No. The amplitude modulus energy ratio of the voltage spectrum data at the frequency domain sampling point , whose expression is: ； Set the energy threshold to , retain the voltage spectrum data whose energy ratio is greater than the energy threshold, and get the The first group of power scenarios The target set of voltage spectrum data for all frequency domain sampling points of a node , whose expression is: ; Based on the The first group of power scenarios The target set of voltage spectrum data of all frequency domain sampling points of the node is obtained The first group of power scenarios The frequency domain sampling point value range of each node is: , and then obtain the frequency domain sampling value range of each node in each group of power scenarios.

6. The method for continuous evaluation of power system node inertia based on convolutional neural network according to claim 1, characterized in that: Based on the frequency domain sampling point value range of each node in each power scenario, the corresponding discretized frequency value range is determined. The target frequency bands for each node in each power scenario include: The relationship between frequency and frequency domain sampling point is expressed as: ； in, Indicates the The frequency corresponding to the frequency domain sampling point; Indicates the total number of sampling points of fast Fourier transform; The index of the frequency domain sampling point representing the fast Fourier transform; Indicates the sampling frequency; The first The first group of power scenarios Substitute the minimum and maximum values ​​of the frequency domain sampling point value range of each node into the relationship expression between frequency and frequency domain sampling point to obtain the The first group of power scenarios The range of the frequency domain sampling points of each node corresponds to the discretized frequency range: , as the first The first group of power scenarios The target frequency band of each node.

7. The method for continuous evaluation of power system node inertia based on convolutional neural network according to claim 1, characterized in that: Based on the inertia constant and rated capacity of each generator within the preset range of each node in each power scenario, the theoretical inertia value of each node in each power scenario is calculated, and the expression is: ； in, Indicates the The first group of power scenarios Theoretical value of inertia of each node; Indicates the The first group of power scenarios The first node within the preset range Rated capacity of each generator; Indicates the The first group of power scenarios The first node within the preset range The inertia constant of the generator.

8. The method for continuous evaluation of power system node inertia based on convolutional neural network according to claim 1, characterized in that: The network structure of the convolutional neural network includes: a convolution module, a pooling module, a channel attention module, a fully connected module and an output layer; The convolution module includes a three-layer one-dimensional convolution structure, a ReLU corrected linear unit activation function and a random inactivation suppression technology; wherein each convolution layer in the three-layer one-dimensional convolution has 32 convolution kernels and uses a scale of 、 、 Multi-scale convolution kernel; The pooling module includes a layer of maximum pooling operation and a layer of global average pooling operation; wherein the pooling window size is 4; The channel attention module includes a Sigmoid activation function; The fully connected module includes two fully connected layers, the first fully connected layer contains 108 neurons, and the second fully connected layer contains 54 neurons; The output layer outputs the inertia prediction value of each node in the power system.

9. The method for continuous evaluation of power system node inertia based on convolutional neural network according to claim 1, characterized in that: The training process of the node inertia evaluation model based on convolutional neural network includes: Based on the The first group of power scenarios The amplitude modulus of the voltage spectrum data within the target frequency band of each node constitutes the The first group of power scenarios The spectral feature matrix of nodes is input into the convolutional neural network; wherein the dimension of the spectral feature matrix is , Indicates the The total number of nodes in the group power scenario, Indicates the total number of frequency sampling points; A dataset is constructed based on the spectrum feature matrices of different nodes under different power scenarios and their corresponding theoretical values ​​of node inertia. The dataset is randomly divided into training and test sets according to a preset ratio. Using the training set, a convolutional neural network-based node inertia estimation model was trained. The Adam optimizer was used to iteratively update the parameters of the convolutional neural network-based node inertia estimation model. The root mean square error metric was used to evaluate the performance of the convolutional neural network-based node inertia estimation model. The parameters of the convolutional neural network-based node inertia estimation model included the convolutional neural network's convolution kernel size, multi-scale design, and attention module parameters. After each training iteration, the test set is used to evaluate the performance of the node inertia evaluation model based on the convolutional neural network, including error evaluation and convergence judgment, until the error index is met, the optimal parameters of the node inertia evaluation model based on the convolutional neural network are obtained, and the trained inertia evaluation model based on the CNN neural network is obtained.

10. The method for continuous evaluation of power system node inertia based on convolutional neural network according to claim 1, characterized in that: After building the node inertia evaluation model based on convolutional neural network, it also includes: A synchronized phasor measurement unit is used to collect voltage data of each node in the actual power system in real time, and the amplitude modulus of the voltage spectrum data of each node is extracted in real time. The amplitude modulus of the voltage spectrum data of each node is input into a trained node inertia assessment model based on a convolutional neural network, and the inertia prediction value of each node in the power system is output in real time.

Images