A strategy for evaluating the nodal inertia time constant of a power system based on continuous evolution feature learning.

Abstract

This invention discloses a power system node inertia time constant assessment strategy based on continuous evolution feature learning. The steps include: First, collecting frequency response data of the power system during active power disturbances, extracting statistical and time-frequency features from the frequency response data, constructing a feature vector, and obtaining coded features through a multilayer perceptron encoding module; Second, gating and weighting the coded features through a hybrid expert module to form a fused feature vector, which is then input as a state variable into a neural frequent differential equation model. In the state space adaptively adjusted by the hybrid expert module, a numerical step-by-step solution algorithm is used to continuously evolve the state variable, obtaining a dynamic feature vector; Finally, the dynamic feature vector is input into a decoding module based on a multilayer perceptron, outputting the power system node inertia time constant assessment result. This invention achieves accurate assessment of the power system node inertia time constant through encoding, continuous evolution, and decoding processes.

Description

Technical Field

[0001] This invention relates to the field of power systems, and more specifically to a strategy for evaluating the inertial time constant of power system nodes based on continuous-time evolution feature learning. Background Technology

[0002] The inertial time constant of a power system is a key parameter reflecting the system's operational stability, characterizing its ability to suppress the rate of frequency change and maintain operational stability. When active power disturbances occur, a larger inertial time constant indicates a smoother frequency change, thus providing more response time for subsequent frequency regulation control. Therefore, it is crucial for ensuring the safe and stable operation of the power grid. Existing inertial time constant assessment methods are mostly based on physical models or disturbance response curves for parameter identification, which struggles to accurately characterize the complex nonlinear dynamic characteristics of power systems. Furthermore, they typically rely on large disturbance events, making continuous estimation difficult under normal operating conditions. They lack continuous-time modeling mechanisms for the frequency dynamic evolution process, have insufficient node-level assessment capabilities, and their estimation accuracy still has room for improvement. Therefore, this invention proposes a power system node inertial time constant assessment strategy based on continuous-time evolution feature learning, which can characterize the continuous dynamic evolution characteristics of frequency and achieve accurate assessment of the node-level inertial time constant. Summary of the Invention

[0003] To achieve the above objectives, the technical solution provided by this invention is as follows:

[0004] A power system nodal inertia time constant evaluation strategy based on continuous evolution feature learning is characterized by the following steps: S1: Collect frequency response data when active power disturbance occurs in the power system, extract statistical and time-frequency features of the frequency response data, construct feature vectors, and obtain coded features through the multilayer perceptron coding module; S2: The encoded features are gated and weighted by the hybrid expert module to form a fused feature vector, which is then used as a state variable to input the neural ordinary differential equation model. In the state space adaptively adjusted by the hybrid expert module, the state variable is continuously evolved using a numerical step-by-step solution algorithm to obtain the dynamic feature vector. S3: Input the dynamic feature vector into the decoding module built on the multilayer perceptron, and output the evaluation result of the power system node inertial time constant; Furthermore, the specific steps of step S1 are as follows: S1-1: When an active power disturbance occurs in the power system, at the system node i Frequency response data is collected at the location, and the frequency response signal is discretely sampled to obtain the node. i Corresponding frequency response sampling sequence The specific form is: (1) In equation (1), Represents a node i In the k Frequency values ​​corresponding to each sampling time. , K This represents the total number of sampling points in the frequency response during a single active power disturbance. S1-2: Frequency response sampling sequence obtained in step S1-1 Extract statistical features and time-frequency features, including nodes. i First n Statistical characteristic components Defined as: (2) In equation (2), Represents the statistical feature mapping function, symbol s Represents statistical characteristic type, n For indexes of statistical characteristic components, , The number of statistical characteristic components; node i First m Each time-frequency characteristic component Defined as: (3) In equation (3), Represents the time-frequency feature mapping function, symbol tf Represents time-frequency characteristic type, m This is the index of the time-frequency characteristic components. , This represents the number of time-frequency characteristic components; S1-3: Based on the extracted statistical features and time-frequency features extracted in step S1-2, adaptive weighted fusion is performed to construct nodes. i The frequency response eigenvector at that location; Define the disturbance intensity sensing coefficient : (4) In equation (4), Represents a node i Maximum frequency deviation, Indicates the system's rated frequency. Indicates the adjustment parameter. Represents an exponential function; Therefore, nodes are constructed. i Frequency response eigenvector at point : (5) S1-4: Construct an encoding module based on a multilayer perceptron. The encoding module consists of multilayer fully connected mappings and nonlinear activation functions, which are applied to the frequency response feature vector obtained in step S1-3. The encoding process is as follows: (6) In equation (6), To encode feature vectors, Represents the real number field. d The dimension of the latent feature representation space output by the encoding module is represented by . This represents the encoding mapping function of a multilayer perceptron. These are the network parameters of the encoding module, where The specific form is: (7) In equation (7), Indicates the first l The weight matrix of the layer, Indicates the first l The layer's bias vector, , Indicates the number of layers in the encoding module; The encoding mapping function is represented as a layer-by-layer nonlinear transformation: (8) In equation (8), Indicates the first layer of the multilayer perceptron coding module l The hidden state vector of the layer, Indicates the first layer of the multilayer perceptron coding module l The hidden state vector of layer -1 is an intermediate feature representation obtained after the previous layer of the neural network performs a nonlinear transformation on the input features. It is a non-linear activation function.

[0005] Furthermore, the specific steps of step S2 are as follows: S2-1: The nodes obtained in step S1-4 i Encoded feature vector Input hybrid expert module, which includes J One expert network and one gated network; No. j A network of experts for nodes i Output vector of encoded features Represented as: (9) In equation (9), Indicates the first j Nonlinear mapping function of an expert network, ,J This indicates the number of expert networks in the hybrid expert module; The process by which a temporal dynamic gating network generates the weight coefficients of each expert based on coding features is represented as follows: (10) In equation (10), Represents a node i In time t Next j The original gated response values ​​of the expert network are generated by the time-series dynamic gated network. and This represents the weight matrix and bias vector of a time-series dynamic gating network. This represents the dynamic expert weight coefficients generated by the time-series dynamic gating network. This represents the normalization function, used to map the output of the gated network to a non-negative weight distribution that sums to 1; Output the fused feature vector of the hybrid expert module Represented as: (11) S2-2: Fusing feature vectors As the initial state variables input into the neural ordinary differential equation model, by introducing a hybrid dynamics function weighted by expert weight coefficients generated by a time-series dynamic gating network, the evolution process of the state variables is expressed as follows: (12) In equation (12), Represents continuous-time state variables Regarding time t The first derivative, Represents a node i In time t The continuous-time state variables, and the initial conditions , This represents a node generated by a time-series dynamic gating network. i Corresponding to the j A dynamic expert weighting coefficient Indicates the first j The dynamic function corresponding to each expert network This represents the implicit parameter vector used to characterize the nodal inertial time constant of the system; S2-3: Based on the ordinary differential equation model constructed in step S2-2, perform node... i The state variables undergo continuous-time evolution, wherein the neural ordinary differential equations are solved using a numerical step-by-step approach to approximate the continuous-time evolution process of the state variables in a discrete computing environment, and a predetermined termination time is reached. T Obtain Nodei Corresponding node-level dynamic feature vector .

[0006] Furthermore, the specific steps of step S3 are as follows: S3-1: The nodes obtained in step S2-3 i At the preset termination time T Node-level dynamic feature vectors The input is a decoding module built on a multilayer perceptron, which consists of multilayer fully connected mappings and nonlinear activation functions. This module decodes the dynamic feature vector through layer-by-layer nonlinear mappings, and outputs nodes. i The nodal inertial time constant parameter corresponding to the active disturbance condition. The decoding process is represented as follows: (13) In equation (13), This represents the decoding mapping function for constructing a multilayer perceptron. These are the network parameters for the decoding module; S3-2: Based on the nodal inertia time constant parameters obtained in step S3-1 As a node i The evaluation results of the nodal inertial time constant under the active disturbance conditions are output or stored.

[0007] Compared with existing technologies, the principles and advantages of this solution are as follows: This invention discloses a power system node inertia time constant assessment strategy based on continuous evolution feature learning. The steps include: First, collecting frequency response data of the power system during active power disturbances, extracting statistical and time-frequency features from the frequency response data, constructing a feature vector, and obtaining coded features through a multilayer perceptron encoding module; Second, gating and weighting the coded features through a hybrid expert module to form a fused feature vector, which is then input as a state variable into a neural frequent differential equation model. In the state space adaptively adjusted by the hybrid expert module, a numerical step-by-step solution algorithm is used to continuously evolve the state variable, obtaining a dynamic feature vector; Finally, the dynamic feature vector is input into a decoding module based on a multilayer perceptron, outputting the power system node inertia time constant assessment result. This invention achieves accurate assessment of the power system node inertia time constant through encoding, continuous evolution, and decoding processes. Attached Figure Description

[0008] Figure 1 This is a flowchart of a power system node inertia time constant evaluation strategy based on continuous evolution feature learning in an embodiment of the present invention. Figure 2 This is a structural diagram of the continuous evolution feature learning model in the embodiments of the present invention; Figure 3 This is a comparison chart of the true and estimated values ​​of the inertial time constant estimated using a traditional multilayer perceptron model in an embodiment of the present invention. Figure 4 This is a residual plot showing the estimation of the inertial time constant using a traditional multilayer perceptron model in an embodiment of the present invention. Figure 5 This is a comparison chart of the true and estimated values ​​of the inertial time constant estimated using the continuous evolution feature learning model in this embodiment of the invention; Figure 6 This is a residual plot showing the estimation of the inertial time constant using a continuous evolution feature learning model in an embodiment of the present invention. Figure 7 This is a comparison chart of the error results between the continuous evolution feature learning model and the traditional multilayer perceptron model in this embodiment of the invention. Detailed Implementation

[0009] The present invention will be further described below with reference to specific embodiments: Figure 1 The diagram shows a flowchart of a power system node inertia time constant evaluation strategy based on continuous evolution feature learning. Figure 2 The diagram shown is a structural diagram of the continuous evolution feature learning model.

[0010] S1: Collect frequency response data when active power disturbances occur in the power system, extract statistical and time-frequency features from the frequency response data, construct a feature vector, and obtain coded features through a multilayer perceptron coding module. The specific steps of step S1 are as follows: S1-1: When an active power disturbance occurs in the power system, at the system node i Frequency response data is collected at the location, and the frequency response signal is discretely sampled to obtain the node. i Corresponding frequency response sampling sequence The specific form is: (14) In equation (14), Represents a node i In the k Frequency values ​​corresponding to each sampling time. , K This represents the total number of sampling points in the frequency response during a single active power disturbance. S1-2: Frequency response sampling sequence obtained in step S1-1 Extract statistical features and time-frequency features, including nodes. i First n Statistical characteristic components Defined as: (15) In equation (15), Represents the statistical feature mapping function, symbol s Represents statistical characteristic type, n For indexes of statistical characteristic components, , The number of statistical characteristic components; node i First m Each time-frequency characteristic component Defined as: (16) In equation (16), Represents the time-frequency feature mapping function, symbol tf Represents time-frequency characteristic type, m This is the index of the time-frequency characteristic components. , This represents the number of time-frequency characteristic components; S1-3: Based on the extracted statistical features and time-frequency features extracted in step S1-2, adaptive weighted fusion is performed to construct nodes. i The frequency response eigenvector at that location; Define the disturbance intensity sensing coefficient : (17) In equation (17), Represents a node i Maximum frequency deviation, Indicates the system's rated frequency. Indicates the adjustment parameter. Represents an exponential function; Therefore, nodes are constructed. i Frequency response eigenvector at point : (18) S1-4: Construct an encoding module based on a multilayer perceptron. The encoding module consists of multilayer fully connected mappings and nonlinear activation functions, which are applied to the frequency response feature vector obtained in step S1-3. The encoding process is as follows: (19) In equation (19), To encode feature vectors, Represents the real number field. d The dimension of the latent feature representation space output by the encoding module is represented by . This represents the encoding mapping function of a multilayer perceptron. These are the network parameters of the encoding module, where The specific form is: (20) In equation (20), Indicates the first l The weight matrix of the layer, Indicates the first l The layer's bias vector, , Indicates the number of layers in the encoding module; The encoding mapping function is represented as a layer-by-layer nonlinear transformation: (twenty one) In equation (21), Indicates the first layer of the multilayer perceptron coding module l The hidden state vector of the layer, Indicates the first layer of the multilayer perceptron coding module l The hidden state vector of layer -1 is an intermediate feature representation obtained after the previous layer of the neural network performs a nonlinear transformation on the input features. It is a non-linear activation function.

[0011] S2: The encoded features are gated and weighted by the hybrid expert module to form a fused feature vector, which is then input as a state variable into the neural ordinary differential equation model. Within the adaptively adjusted state space of the hybrid expert module, a numerical step-by-step solution algorithm is used to continuously evolve the state variable, yielding a dynamic feature vector. The specific steps of step S2 are as follows: S2-1: The nodes obtained in step S1-4 i Encoded feature vector Input hybrid expert module, which includes J One expert network and one gated network; No. j A network of experts for nodes i Output vector of encoded features Represented as: (twenty two) In equation (22), Indicates the first j Nonlinear mapping function of an expert network, , J This indicates the number of expert networks in the hybrid expert module; The process by which a temporal dynamic gating network generates the weight coefficients of each expert based on coding features is represented as follows: (twenty three) In equation (23), Represents a node iIn time t Next j The original gated response values ​​of the expert network are generated by the time-series dynamic gated network. and This represents the weight matrix and bias vector of a time-series dynamic gating network. This represents the dynamic expert weight coefficients generated by the time-series dynamic gating network. This represents the normalization function, used to map the output of the gated network to a non-negative weight distribution that sums to 1; Output the fused feature vector of the hybrid expert module Represented as: (twenty four) S2-2: Fusing feature vectors As the initial state variables input into the neural ordinary differential equation model, by introducing a hybrid dynamics function weighted by expert weight coefficients generated by a time-series dynamic gating network, the evolution process of the state variables is expressed as follows: (25) In equation (25), Represents continuous-time state variables Regarding time t The first derivative, Represents a node i In time t The continuous-time state variables, and the initial conditions , This represents a node generated by a time-series dynamic gating network. i Corresponding to the j A dynamic expert weighting coefficient Indicates the first j The dynamic function corresponding to each expert network This represents the implicit parameter vector used to characterize the nodal inertial time constant of the system; S2-3: Based on the ordinary differential equation model constructed in step S2-2, perform node... i The state variables undergo continuous-time evolution, wherein the neural ordinary differential equations are solved using a numerical step-by-step approach to approximate the continuous-time evolution process of the state variables in a discrete computing environment, and a predetermined termination time is reached. T Obtain Node i Corresponding node-level dynamic feature vector .

[0012] S3: Input the dynamic feature vector into the decoding module built on a multilayer perceptron, and output the evaluation result of the power system node inertial time constant. The specific steps of step S3 are as follows: S3-1: The nodes obtained in step S2-3 i At the preset termination time T Node-level dynamic feature vectors The input is a decoding module built on a multilayer perceptron, which consists of multilayer fully connected mappings and nonlinear activation functions. This module decodes the dynamic feature vector through layer-by-layer nonlinear mappings, and outputs nodes. i The nodal inertial time constant parameter corresponding to the active disturbance condition. The decoding process is represented as follows: (26) In equation (26), This represents the decoding mapping function for constructing a multilayer perceptron. These are the network parameters for the decoding module; S3-2: Based on the nodal inertia time constant parameters obtained in step S3-1 As a node i The evaluation results of the nodal inertial time constant under the active disturbance conditions are output or stored.

[0013] To verify the effectiveness of the proposed power system node inertia time constant assessment strategy based on continuous evolution feature learning, a simulation comparison experiment was designed: Based on the frequency response mechanism under active power disturbances in the power system, simulation scenarios of node frequency response under different inertia time constants and different disturbance intensities were constructed. Node frequency response sequences were generated through simulation and discrete sampling to obtain node frequency response data. An 8-dimensional key feature vector, including mean, variance, range, and difference statistics, was extracted from the frequency response data and input into both the traditional multilayer perceptron model and the continuous evolution feature learning model of this invention to assess the node inertia time constant. By comparing the error between the predicted results of each model and the actual values, the effectiveness and accuracy of the proposed method were verified.

[0014] The proposed continuous evolution feature learning model employs an "encoding-continuous evolution-decoding" network structure: the encoding module uses a multilayer perceptron to perform nonlinear mapping on the 8-dimensional input features and obtain hidden feature representations, with the hidden feature dimension set to 64; the continuous evolution module uses a neural network model based on a hybrid expert structure, with the number of hybrid expert networks set to 2 and the hidden parameter dimension set to 4, and uses the second-order Runge-Kutta midpoint method to numerically solve the neural network model, realizing the continuous-time evolution of state variables; simultaneously, a residual back-injection mechanism is introduced to enhance the feature representation capability under weak perturbation conditions. Model training and testing are implemented in the PyTorch framework, using the AdamW (Adaptive Moment Estimation) weight decay optimization algorithm optimizer for parameter updates, with a learning rate set to 10. -3The mean squared error was used as the loss function, the training epochs were set to 80, the batch size to 64, and the gradient clipping threshold was set to 1.0 to improve training stability. Finally, the estimated node inertial time constant was output through the decoding module and compared with the evaluation results of the traditional multilayer perceptron model.

[0015] Figure 3 , Figure 4 The figure shows the effect of using a traditional multilayer perceptron model to evaluate the nodal inertial time constant of a power system: where Figure 3 The curve showing the comparison between the true and estimated values ​​of the inertial time constant reveals a significant deviation between the model's estimated value and the true value. Figure 4 The residual plot shows a wide residual distribution, concentrated between -2 and 2, indicating poor stability. This model employs an "encoder-decoder" network structure, directly mapping the output inertial time constant through fully connected layers without modeling the temporal evolution of the frequency response. The test data originates from the node frequency response sequences generated in the simulation. Because traditional multilayer perceptrons cannot capture the continuous dynamic changes in frequency response over time, the fitting bias for samples with weak disturbances and small inertial ranges is significant. The overall average absolute error is 0.4191, and the coefficient of determination is 0.9205, reflecting a clear limitation in characterizing the nonlinear dynamic mapping relationship between frequency response and inertial time constant.

[0016] Figure 5 , Figure 6 The following shows the evaluation results of using the continuous evolution feature learning model of this invention: where Figure 5 The curve showing the comparison between the true and estimated values ​​of the inertial time constant reveals that the estimated and true values ​​almost completely overlap, and each sample point closely follows the trend of the true value. Figure 6 The model residual plot shows that the residuals are highly concentrated in... Between 0.1 and 0.15, the absolute value of the residuals for the vast majority of samples is less than 0.05, demonstrating the model's extremely high predictive stability and accuracy. On the same test set, the model employs an end-to-end architecture of "encoding-continuous evolution-decoding." First, the encoder extracts the static statistical features of the frequency response. Then, a hybrid expert-gated differential equation module mines the continuous evolution of the frequency response over time. Finally, the decoder maps the evolved latent parameters to the inertial time constant. This architecture effectively captures the dynamic correlation between the frequency response and the inertial time constant under different perturbation intensities, significantly reducing prediction bias in complex scenarios such as weak perturbations and small inertia. It accurately characterizes the dynamic mapping relationship between the frequency response and the inertial time constant, with an overall average absolute error of 0.0358 and a coefficient of determination of 0.9995.

[0017] Figure 7The comparison chart of the inertial time constant prediction performance of different models visually demonstrates the significant advantage of the proposed model in terms of evaluation accuracy compared to the traditional multilayer perceptron model. Compared to the traditional multilayer perceptron model, the continuous evolution feature learning model of the present invention achieves a significant improvement in evaluation accuracy: the mean absolute error decreases from 0.4191 to 0.0358, a relative improvement of 91.47%; the coefficient of determination increases from 0.9205 to 0.9995, enabling it to more accurately capture the complex nonlinear dynamic characteristics of the node inertial time constant of the power system and achieve high-precision evaluation of the node-level inertial time constant.

[0018] The above-described embodiments are merely preferred embodiments of the present invention and are not intended to limit the scope of the present invention. Therefore, any changes made in accordance with the shape and principle of the present invention should be covered within the protection scope of the present invention.

Claims

1. A power system nodal inertia time constant evaluation strategy based on continuous evolution feature learning, characterized in that, Includes the following steps: S1: Collect frequency response data when active power disturbance occurs in the power system, extract statistical and time-frequency features of the frequency response data, construct feature vectors, and obtain coded features through the multilayer perceptron coding module; S2: The encoded features are gated and weighted by the hybrid expert module to form a fused feature vector, which is then used as a state variable to input the neural ordinary differential equation model. In the state space adaptively adjusted by the hybrid expert module, the state variable is continuously evolved using a numerical step-by-step solution algorithm to obtain the dynamic feature vector. S3: Input the dynamic feature vector into the decoding module built on the multilayer perceptron, and output the evaluation result of the power system node inertial time constant; Step S1 includes: S1-1: When an active power disturbance occurs in the power system, at the system node i Frequency response data is collected at the location, and the frequency response signal is discretely sampled to obtain the node. i Corresponding frequency response sampling sequence The specific form is as follows: (1) In equation (1), Represents a node i In the k Frequency values ​​corresponding to each sampling time. , K This represents the total number of sampling points in the frequency response during a single active power disturbance. S1-2: Frequency response sampling sequence obtained in step S1-1 Extract statistical features and time-frequency features, including nodes. i First n Statistical characteristic components Defined as: (2) In equation (2), Represents the statistical feature mapping function, symbol s Represents statistical characteristic type, n For indexes of statistical characteristic components, , The number of statistical characteristic components; node i First m Each time-frequency characteristic component Defined as: (3) In equation (3), Represents the time-frequency feature mapping function, symbol tf Represents time-frequency characteristic type, m This is the index of the time-frequency characteristic components. , The number of time-frequency characteristic components; S1-3: Based on the extracted statistical features and time-frequency features extracted in step S1-2, adaptive weighted fusion is performed to construct nodes. i The frequency response eigenvector at that location; Define the disturbance intensity sensing coefficient : (4) In equation (4), Represents a node i Maximum frequency deviation, Indicates the system's rated frequency. Indicates the adjustment parameter. Represents an exponential function; Therefore, nodes are constructed. i Frequency response eigenvector at point : (5) S1-4: Construct an encoding module based on a multilayer perceptron. The encoding module consists of multilayer fully connected mappings and nonlinear activation functions, which are applied to the frequency response feature vector obtained in step S1-3. The encoding process is as follows: (6) In equation (6), To encode feature vectors, Represents the real number field. d The dimension of the latent feature representation space output by the encoding module is represented by . This represents the encoding mapping function of a multilayer perceptron. These are the network parameters of the encoding module, where The specific form is as follows: (7) In equation (7), Indicates the first l The weight matrix of the layer, Indicates the first l Layer bias vector, , Indicates the number of layers in the encoding module; The encoding mapping function is represented as a layer-by-layer nonlinear transformation: (8) In equation (8), Indicates the first layer of the multilayer perceptron coding module l The hidden state vector of the layer, Indicates the first l The hidden state vector of layer -1 is an intermediate feature representation obtained after the previous layer of the neural network performs a nonlinear transformation on the input features. It is a non-linear activation function.

2. The power system node inertia time constant evaluation strategy based on continuous evolution feature learning according to claim 1, characterized in that, Step S2 includes: S2-1: The nodes obtained in step S1-4 i Encoded feature vector Input hybrid expert module, which includes J One expert network and one gated network; No. j A network of experts for nodes i Output vector of encoded features Represented as: (9) In equation (9), Indicates the first j Nonlinear mapping function of an expert network, , J This indicates the number of expert networks in the hybrid expert module; The process by which a temporal dynamic gating network generates the weight coefficients of each expert based on coding features is represented as follows: (10) In equation (10), Represents a node i In time t Next j The original gated response values ​​of the expert network are generated by the time-series dynamic gated network. and This represents the weight matrix and bias vector of a time-series dynamic gating network. This represents the dynamic expert weight coefficients generated by the time-series dynamic gating network. This represents the normalization function, used to map the output of the gated network to a non-negative weight distribution that sums to 1; Output the fused feature vector of the hybrid expert module Represented as: (11) S2-2: Fusing feature vectors As the initial state variables input into the neural ordinary differential equation model, by introducing a hybrid dynamics function weighted by expert weight coefficients generated by a time-series dynamic gating network, the evolution process of the state variables is expressed as follows: (12) In equation (12), Represents continuous-time state variables Regarding time t The first derivative, Represents a node i In time t The continuous-time state variables, and the initial conditions , This represents a node generated by a time-series dynamic gating network. i Corresponding to the j A dynamic expert weighting coefficient Indicates the first j The dynamic function corresponding to each expert network This represents the implicit parameter vector used to characterize the nodal inertial time constant of the system; S2-3: Based on the ordinary differential equation model constructed in step S2-2, perform node... i The state variables undergo continuous-time evolution, wherein the neural ordinary differential equations are solved using a numerical step-by-step approach to approximate the continuous-time evolution process of the state variables in a discrete computing environment, and a predetermined termination time is reached. T Obtain Node i Corresponding node-level dynamic feature vector .

3. The power system node inertia time constant evaluation strategy based on continuous evolution feature learning according to claim 1, characterized in that, Step S3 includes: S3-1: The nodes obtained in step S2-3 i At the preset termination time T Node-level dynamic feature vectors The input is a decoding module built on a multilayer perceptron, which consists of multilayer fully connected mappings and nonlinear activation functions. This module decodes the dynamic feature vector through layer-by-layer nonlinear mappings, and outputs nodes. i The nodal inertia time constant parameter corresponding to the active disturbance condition. The decoding process is represented as follows: (13) In equation (13), This represents the decoding mapping function for constructing a multilayer perceptron. These are the network parameters for the decoding module; S3-2: Based on the nodal inertia time constant parameters obtained in step S3-1 As a node i The evaluation results of the nodal inertial time constant under the active disturbance conditions are output or stored.

Images