Power system transient stability prediction method and device based on Densection network

Abstract

The invention discloses a method and a device for predicting transient stability of a power system based on a Densection network. The method comprises the following steps of: constructing a plurality of sample sets covering steady-state operation conditions under different loads by using hyper-Latin sampling; setting a three-phase short-circuit fault for each steady-state point, and obtaining transient data and a transient stability index TSI; extracting voltage, power and the like of a plurality of sampling points as input features and forming a feature matrix; expanding the feature matrix into a three-channel two-dimensional time sequence feature map; training a prediction model by utilizing a DenseNet, Xception and a Denseption network of a full connection layer, wherein the DenseNet, the Xception and the full connection layer are fused with the Denseption network; real-time data is input during online prediction, and after feature conversion, TSI is output through the prediction model to judge the stability. According to the method, feature reuse and multi-scale extraction capabilities are fused through the Densection network, and the accuracy and reliability of transient stability prediction of the power system are remarkably improved in combination with full-connection layer regression output.

Description

TECHNICAL FIELD

[0001] The present application relates to the field of power system transient stability evaluation, and in particular to a power system transient stability prediction method and device based on a Denseception network. BACKGROUND

[0002] With the deepening of global energy transformation, the power system is rapidly developing towards high proportion of renewable energy and high power electronics. The access of a large number of intermittent power sources such as wind power and photovoltaic power and power electronic devices makes the dynamic characteristics of the power system more complex, and the operating state is also more variable, which poses unprecedented severe challenges to the safe and stable operation of the power system.

[0003] As one of the core indicators of the safe operation of the power system, transient stability is the ability of the power system to maintain synchronous operation between synchronous generators after being subjected to a large disturbance (such as short-circuit fault, generator trip, etc.). Once the system loses transient stability, it may cause a large-scale power outage accident, causing great losses to the society and economy. Therefore, accurate and rapid prediction of the transient stability of the power system is a key link to ensure the safe and reliable operation of the system.

[0004] Power system transient stability prediction can be defined as: through the analysis and processing of real-time operation data, fault information, etc. of the power system, using corresponding algorithms and models, predicting whether the system will lose transient stability and the development trend of the stability situation within a short time after the fault occurs, and providing timely and effective decision support for dispatchers to take control measures.

[0005] However, the time-varying nature of component parameters, the suddenness of fault disturbances and the dynamic reconstruction of network topology in the power system make the evolution mechanism of transient stability highly nonlinear and uncertain, increasing the difficulty of constructing the prediction model; at the same time, the high-frequency mass data generated by the wide-area measurement system (WAMS), the millisecond-level response characteristics of the transient process and the extreme diversity of fault scenarios also pose strict requirements on the real-time performance and generalization ability of transient stability prediction. Therefore, how to construct a transient stability prediction model that takes into account accuracy and efficiency to cope with the dynamic challenges of new power systems plays a decisive role in improving the system's fault defense capability and ensuring the continuous and reliable supply of electricity.

[0006] In essence, the above problem is a complex modeling problem of fusing data-driven and physical mechanism. In actual engineering, simplified empirical criteria or offline simulation results are often used to make decisions, such as rapid judgment based on power angle stability threshold, which is difficult to adapt to real-time needs of system dynamic changes. Although the traditional physical modeling method (such as electromechanical transient simulation, energy function method) can reflect the essence of transient state, it takes too long to calculate and cannot meet the timeliness requirements of online prediction. Compared with the traditional physical modeling method, data-driven algorithms represented by deep learning (Deep learning, DL), extreme learning machine (Extreme learning machine, ELM) and support vector machine (Support vector machine, SVM) have faster prediction speed and stronger nonlinear fitting ability. However, due to insufficient embedding of system physical laws and high sensitivity to the distribution characteristics of training data, the above algorithms often have large prediction deviation and weak robustness in small sample fault scenarios or out-of-distribution conditions. Therefore, it is urgent to develop a hybrid prediction method that combines physical mechanism and data wisdom for accurate and real-time prediction of power system transient stability. SUMMARY

[0007] The purpose of the present application is to overcome the shortcomings and deficiencies of the prior art, and to provide a power system transient stability prediction method and device based on Denseception network, which combines DenseNet and Xception structure to form a power system transient stability prediction model-Denseception. The model combines the dense connection mechanism of DenseNet and the depth-wise separable convolution (Depth-wise Separable Convolution, DSC) characteristics of Xception, fully utilizes the synergistic advantages of the two, and realizes efficient capture and accurate analysis of the complex nonlinear dynamic characteristics of the power system.

[0008] The purpose of the present application is achieved by the following technical solutions:

[0009] A power system transient stability prediction method based on Denseception network, comprising the following steps:

[0010] (1) Construct a sample set of multiple power system steady-state operating conditions covering different load conditions using hyper-Latin sampling method;

[0011] (2) For each power system steady-state operating point in step (1), set a three-phase short-circuit fault, and obtain the transient data and transient stability index TSI of each transient fault scenario;

[0012] (3) Select the node voltage, generator terminal voltage, generator rotor angular velocity, generator active power and generator reactive power of several sampling points in each transient data obtained in step (2) as input features and form a feature matrix;

[0013] (4) The feature matrix obtained in step (3) is copied and expanded along the channel dimension three times, and finally a three-channel two-dimensional time sequence feature map is formed;

[0014] (5) The DenseNet, Xception and fully connected layer are fused to construct a Denseception network, the three-channel two-dimensional time sequence feature map obtained in step (4) is input into the Denseception network for training, and finally a power system transient stability prediction model based on the Denseception network is obtained;

[0015] (6) Obtain transient stability real-time data from the power system; then extract input features from the transient stability real-time data; convert the input features into images; then input the images into the transient stability prediction model based on Denseception; and finally output TSI to obtain the transient stability of the power system after the fault.

[0016] Preferably, the calculation formula of the transient stability index TSI in step (2) is:

[0017] (1)

[0018] In the formula, is the maximum value of the angular difference of any two generators in the time domain simulation time; is the absolute value; max is the maximum value; when TSI> 0, the system is transiently stable; when TSI< 0, the system is transiently unstable.

[0019] Preferably, the feature matrix in step (3) is as follows:

[0020] (2)

[0021] In the formula, is the bus voltage amplitude; is the generator active power; is the generator reactive power; is the generator terminal voltage; is the generator rotor angular velocity; n is the number of generators; m is the number of buses; is the voltage amplitude of the first bus; is the voltage amplitude of the second bus; is the voltage amplitude of the nth bus; is the voltage amplitude of the n+1th bus; The voltage amplitude of the m-th bus; This represents the active power of the first generator. This represents the active power of the second generator. Let be the active power of the nth generator; This represents the reactive power of the first generator. This refers to the reactive power of the second generator; Let n be the reactive power of the nth generator; This is the terminal voltage of the first generator; This is the terminal voltage of the second generator; Let be the terminal voltage of the nth generator; The angular velocity of the first generator; The angular velocity of the second generator; Let be the angular velocity of the nth generator.

[0022] Preferably, the Denseception network structure in step (5) is as follows:

[0023] The Denseception network takes an image of size m×40×3 as input. It first extracts initial features through a 7×7 convolutional layer, then passes through a 3×3 max pooling layer to reduce spatial dimensionality and expand the receptive field. Next, it connects four DenseNet modules to achieve dense cross-layer feature reuse, then connects three Xception modules to achieve multi-scale convolutional feature extraction. Finally, it passes through a global average pooling layer to converge spatial features, and finally connects to a fully connected layer and outputs the transient stable state through a Softmax layer.

[0024] Preferably, the specific structure of the DenseNet module is as follows: the input first undergoes a 1×1 convolution, then is divided into four branches: branch 1 is a single 1×1 convolution, branch 2 is a single 1×1 convolution followed by a single 3×3 convolution, branch 3 is a single average pooling operation followed by a single 3×3 convolution, and branch 4 is a single 1×1 convolution followed by two consecutive 3×3 convolutions. The outputs of the four branches are then concatenated along the channel dimension and used as the output of the module. The specific structure of the Xception module is as follows: the input is also divided into four branches: branch 1 is a single 1×1 convolution, branch 2 is a single 1×1 convolution followed by a single 3×3 convolution, branch 3 is a single 1×1 convolution followed by a single 5×5 convolution, and branch 4 is a single max pooling operation followed by a single 1×1 convolution. The outputs of the four branches are then concatenated along the channel dimension and used as the output of the module.

[0025] A power system transient stability prediction device based on a Denseception network includes:

[0026] The steady-state operating condition sample set construction module is used to construct sample sets of various power system steady-state operating conditions covering different load conditions using the super-Latin sampling method;

[0027] The transient data and TSI acquisition module is used to set up a three-phase short-circuit fault for each power system steady-state operating point and acquire transient data and transient stability index (TSI) for each transient fault scenario.

[0028] The feature matrix extraction module is used to select node voltage, generator terminal voltage, generator rotor angular velocity, generator active power and generator reactive power from several sampling points in each transient data as input features and form a feature matrix;

[0029] The three-channel feature map generation module is used to copy and expand the feature matrix three times along the channel dimension, ultimately forming a three-channel two-dimensional time-series feature map.

[0030] The prediction model training module is used to fuse DenseNet, Xception and fully connected layers to construct the Denseception network. The three-channel two-dimensional time series feature map is input into the Denseception network for training, and finally the power system transient stability prediction model based on the Denseception network is obtained.

[0031] The transient stability online prediction module is used to acquire real-time transient stability data from the power system; then extract input features from the real-time transient stability data; convert the input features into an image; then input the image into a Denseception-based transient stability prediction model; finally output TSI to obtain the transient stability of the power system after a fault.

[0032] Preferably, the transient stability index (TSI) is calculated using the following formula:

[0033] (1);

[0034] In the formula, This represents the maximum value of the rotor angle difference between any two generators within the time-domain simulation period; To find the absolute value; max is the maximum value; when TSI>0, the system is transiently stable; when TSI<0, the system is transiently unstable.

[0035] Preferably, the feature matrix is ​​as follows:

[0036] (2);

[0037] In the formula, This refers to the amplitude of the bus voltage. This refers to the active power of the generator. This refers to the reactive power of the generator. This refers to the generator terminal voltage; ω is the angular velocity of the generator rotor; n is the number of generators; m is the number of busbars. This refers to the voltage amplitude of the first busbar. This refers to the voltage amplitude of the second busbar; The voltage amplitude of the nth bus; This represents the voltage amplitude of the (n+1)th bus. The voltage amplitude of the m-th bus; This represents the active power of the first generator. This represents the active power of the second generator. Let be the active power of the nth generator; This represents the reactive power of the first generator. This refers to the reactive power of the second generator; Let n be the reactive power of the nth generator; This is the terminal voltage of the first generator; This is the terminal voltage of the second generator; Let be the terminal voltage of the nth generator; The angular velocity of the first generator; The angular velocity of the second generator; Let be the angular velocity of the nth generator.

[0038] Preferably, the Denseception network structure is as follows:

[0039] The Denseception network takes an image of size m×40×3 as input. It first extracts initial features through a 7×7 convolutional layer, then passes through a 3×3 max pooling layer to reduce spatial dimensionality and expand the receptive field. Next, it connects four DenseNet modules to achieve dense cross-layer feature reuse, then connects three Xception modules to achieve multi-scale convolutional feature extraction. Finally, it passes through a global average pooling layer to converge spatial features, and finally connects to a fully connected layer and outputs the transient stable state through a Softmax layer.

[0040] Preferably, the specific structure of the DenseNet module is as follows: the input first undergoes a 1×1 convolution, then is divided into four branches: branch 1 is a single 1×1 convolution, branch 2 is a single 1×1 convolution followed by a single 3×3 convolution, branch 3 is a single average pooling followed by a single 3×3 convolution, and branch 4 is a single 1×1 convolution followed by two consecutive 3×3 convolutions. The outputs of the four branches are then concatenated along the channel dimension and used as the output of the module. Similarly, the specific structure of the Xception module is as follows: the input is also divided into four branches: branch 1 is a single 1×1 convolution, branch 2 is a single 1×1 convolution followed by a single 3×3 convolution, branch 3 is a single 1×1 convolution followed by a single 5×5 convolution, and branch 4 is a single max pooling followed by a single 1×1 convolution. The outputs of the four branches are then concatenated along the channel dimension and used as the output of the module. This invention has the following advantages and effects compared to the prior art:

[0041] (1) To address the complexity of fast and slow dynamic coupling in transient processes, this invention designs a heterogeneous network fusion architecture: it utilizes the dense connections of DenseNet to preserve low-frequency features, captures high-frequency oscillation details through Xception's DSC, and finally dynamically fuses global spatiotemporal information using FC. This design solves the problem of low efficiency in feature cross-scale transmission in traditional single network architectures.

[0042] (2) This invention abandons the traditional coarse-grained prediction mode of binary classification and constructs an end-to-end prediction framework for TSI based on continuous numerical regression, realizing the direct quantitative output of transient stability margin. Through an offline-online collaborative computing strategy, Denseception achieves real-time inference while ensuring prediction accuracy, and can provide stability warnings in large-scale fault scenarios, gaining a critical time window for the formulation of emergency control strategies and significantly reducing the risk of system instability.

[0043] (3) The transient data modeling method based on three-channel two-dimensional feature maps of the present invention reconstructs the time series data of the entire process before, during and after the fault into an image-like structure, and explicitly separates the dynamic characteristics of multiple stages. It not only adapts to the local perception advantage of convolutional networks, but also suppresses noise interference through the channel redundancy mechanism, and provides a unified data interface for cross-scenario migration of power grids of different scales. Attached Figure Description

[0044] Figure 1 This is a flowchart of the power system transient stability prediction method based on Denseception network according to the present invention;

[0045] Figure 2 This is a diagram of the Denseception network model.

[0046] Figure 3 This is a schematic diagram of the iteration error and training efficiency of Denseception. Detailed Implementation

[0047] To make the objectives, technical solutions, and advantages of the embodiments of the present invention clearer, the technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0048] This embodiment focuses on a power system transient stability prediction model. Please refer to [link / reference]. Figure 1 This embodiment presents a power system transient stability prediction method based on Denseception networks, comprising the following steps:

[0049] (1) A sample set of 8,000 steady-state operating conditions was constructed using the super Latin sampling method, covering the steady-state operating states of the system under different load conditions;

[0050] (2) For each steady-state operating point of the power system in step (1), a three-phase short-circuit fault is set up, and detailed transient data and transient stability index (TSI) for each transient fault scenario are obtained; the formula for calculating the transient stability index (TSI) is:

[0051] (1)

[0052] In the formula, This represents the maximum value of the rotor angle difference between any two generators within the time-domain simulation period; To find the absolute value; max is the maximum value; when TSI > 0, the system is transiently stable; when TSI < 0, the system is transiently unstable.

[0053] (3) Select the node voltage, generator terminal voltage, generator rotor angular velocity, generator active power, and generator reactive power of ten sampling points from each transient data obtained in step (2) as input features and form a feature matrix; the specific form of the feature matrix is ​​as follows:

[0054] (2)

[0055] In the formula, This refers to the amplitude of the bus voltage. This refers to the active power of the generator. This refers to the reactive power of the generator. This refers to the generator terminal voltage; ω is the angular velocity of the generator rotor; n is the number of generators; m is the number of busbars. This refers to the voltage amplitude of the first busbar. This refers to the voltage amplitude of the second busbar; The voltage amplitude of the nth bus; This represents the voltage amplitude of the (n+1)th bus. The voltage amplitude of the m-th bus; This represents the active power of the first generator. This represents the active power of the second generator. Let be the active power of the nth generator; This represents the reactive power of the first generator. This refers to the reactive power of the second generator; Let n be the reactive power of the nth generator; This is the terminal voltage of the first generator; This is the terminal voltage of the second generator; Let be the terminal voltage of the nth generator; The angular velocity of the first generator; The angular velocity of the second generator; Let be the angular velocity of the nth generator;

[0056] (4) The feature matrix is ​​copied and expanded three times along the channel dimension to finally form a three-channel two-dimensional time series feature map;

[0057] (5) Construct a Denseception network by fusing DenseNet, Xception, and fully connected layers (e.g., Figure 2As shown in the figure, the three-channel two-dimensional temporal feature map obtained in step (4) is input into the Denseception network for training, and finally the transient stability prediction model of the power system based on the Denseception network is obtained. The input of the Denseception network structure is an image with a size of m×40×3. First, it passes through a 7×7 convolutional layer for initial feature extraction, then through a 3×3 max pooling layer to reduce the spatial dimension and expand the receptive field, then through 4 DenseNet modules in sequence to achieve dense cross-layer feature reuse, then through 3 Xception modules in sequence to achieve multi-scale convolutional feature extraction, then through a global average pooling layer to converge the spatial features, and finally through a fully connected layer and through a Softmax layer to output the transient stable state. The DenseNet module has the following structure: the input first undergoes a 1×1 convolution, then splits into four branches: branch 1 is a single 1×1 convolution, branch 2 is a single 1×1 convolution followed by a 3×3 convolution, branch 3 is a single average pooling convolution followed by a 3×3 convolution, and branch 4 is a single 1×1 convolution followed by two consecutive 3×3 convolutions. The outputs of the four branches are then concatenated along the channel dimension to serve as the module's output. The Xception module also has the following structure: the input is similarly divided into four branches: branch 1 is a single 1×1 convolution, branch 2 is a single 1×1 convolution followed by a 3×3 convolution, branch 3 is a single 1×1 convolution followed by a 5×5 convolution, and branch 4 is a single max pooling convolution followed by a 1×1 convolution. The outputs of the four branches are then concatenated along the channel dimension to serve as the module's output.

[0058] (6) During online prediction, real-time transient stability data of the power system is first obtained; then input features are extracted from the real-time transient stability data; the input features are converted into images; then the images are input into the transient stability prediction model based on Denseception; finally, TSI is output to obtain the transient stability of the power system after the fault.

[0059] Figure 3 (a) shows that Denseception rapidly reduces the error in the first 200 iterations and maintains the lowest RMSE in the later stages of training with the smallest fluctuation. Compared with Xception and DenseNet201, Denseception is more stable, while Vgg16 has a significantly higher RMSE and more drastic fluctuations, resulting in poor convergence. Figure 3 (b) Denseception demonstrates excellent training efficiency while maintaining high-precision prediction. The training time of Denseception is significantly shorter than that of DenseNet201 and Vgg16. In particular, DenseNet201 takes nearly 5000 seconds to train, while Denseception only requires a fraction of that time, resulting in a significant improvement in computational efficiency.

[0060] This invention has the following features and effects:

[0061] 1. Heterogeneous network fusion innovation: DenseNet dense connections are used to preserve low-frequency features and Xception deep separable convolutions are used to capture high-frequency oscillation details, solving the problem of low cross-scale feature transfer efficiency of traditional single networks and realizing efficient capture of complex nonlinear dynamic characteristics of power systems;

[0062] 2. Innovation in quantitative prediction framework: Construct an end-to-end continuous numerical regression framework for TSI, abandon the binary classification coarse-grained prediction mode, realize the direct quantitative output of transient stability margin, and balance prediction accuracy and real-time performance through an offline-online collaborative strategy, so as to gain a critical time window for the formulation of emergency control strategies.

[0063] 3. Innovative Feature Modeling: Three-channel two-dimensional temporal feature map modeling adapts to the local perception advantages of convolutional networks, suppresses noise interference, and provides a unified data interface for cross-scenario migration of power grids of different scales, enhancing the model's generalization ability.

[0064] The above description is merely a specific embodiment of the present invention, but the scope of protection of the present invention is not limited thereto. Any variations or substitutions that can be easily conceived by those skilled in the art within the technical scope disclosed in the present invention should be included within the scope of protection of the present invention. Therefore, the scope of protection of the present invention should be determined by the scope of the claims.

Claims

1. A method for predicting transient stability of a power system based on a Denseception network, characterized in that, Includes the following steps: (1) Construct sample sets of various power system steady-state operating conditions covering different load conditions using the super-Latin sampling method; (2) For each steady-state operating point of the power system in step (1), set a three-phase short-circuit fault, and obtain transient data and transient stability index (TSI) for each transient fault scenario; (3) Select the node voltage, generator terminal voltage, generator rotor angular velocity, generator active power and generator reactive power of several sampling points from each transient data obtained in step (2) as input features and form a feature matrix; (4) The feature matrix obtained in step (3) is copied and expanded three times along the channel dimension to finally form a three-channel two-dimensional time series feature map; (5) DenseNet, Xception and fully connected layer are fused to construct Denseception network. The three-channel two-dimensional time series feature map obtained in step (4) is input into Denseception network for training, and finally the power system transient stability prediction model based on Denseception network is obtained. (6) Obtain real-time transient stability data from the power system; Then, input features are extracted from real-time transient stability data; the input features are converted into images; the images are then input into a transient stability prediction model based on Denseception; finally, TSI is output to obtain the transient stability of the power system after the fault.

2. The power system transient stability prediction method based on Denseception network according to claim 1, characterized in that, The formula for calculating the transient stability index (TSI) in step (2) is as follows: （1）； In the formula, This represents the maximum value of the rotor angle difference between any two generators within the time-domain simulation period; To find the absolute value; max is the maximum value; when TSI>0, the system is transiently stable; when TSI<0, the system is transiently unstable.

3. The power system transient stability prediction method based on Denseception network according to claim 1, characterized in that, The feature matrix in step (3) is as follows: （2） ； In the formula, This refers to the amplitude of the bus voltage. This refers to the active power of the generator. This refers to the reactive power of the generator. This refers to the generator terminal voltage; ω is the angular velocity of the generator rotor; n is the number of generators; m is the number of busbars. This refers to the voltage amplitude of the first busbar. This refers to the voltage amplitude of the second busbar. The voltage amplitude of the nth bus; This represents the voltage amplitude of the (n+1)th bus. The voltage amplitude of the m-th bus; This represents the active power of the first generator. This represents the active power of the second generator. Let be the active power of the nth generator; This represents the reactive power of the first generator. This refers to the reactive power of the second generator; Let n be the reactive power of the nth generator; This is the terminal voltage of the first generator; This is the terminal voltage of the second generator; Let be the terminal voltage of the nth generator; The angular velocity of the first generator; The angular velocity of the second generator; Let be the angular velocity of the nth generator.

4. The power system transient stability prediction method based on Denseception network according to claim 1, characterized in that, The Denseception network structure in step (5) is as follows: The Denseception network takes an image of size m×40×3 as input. It first extracts initial features through a 7×7 convolutional layer, then passes through a 3×3 max pooling layer to reduce spatial dimensionality and expand the receptive field. Next, it connects four DenseNet modules to achieve dense cross-layer feature reuse, then connects three Xception modules to achieve multi-scale convolutional feature extraction. Finally, it passes through a global average pooling layer to converge spatial features, and finally connects to a fully connected layer and outputs the transient stable state through a Softmax layer.

5. The power system transient stability prediction method based on Denseception network according to claim 4, characterized in that, The specific structure of the DenseNet module is as follows: the input first undergoes a 1×1 convolution, then is divided into four branches: branch 1 is a single 1×1 convolution, branch 2 is a single 1×1 convolution followed by a single 3×3 convolution, branch 3 is a single average pooling convolution followed by a single 3×3 convolution, and branch 4 is a single 1×1 convolution followed by two consecutive 3×3 convolutions. The outputs of the four branches are concatenated along the channel dimension and used as the output of the module. The specific structure of the Xception module is as follows: the input is also divided into four branches: branch 1 is a single 1×1 convolution, branch 2 is a single 1×1 convolution followed by a single 3×3 convolution, branch 3 is a single 1×1 convolution followed by a single 5×5 convolution, and branch 4 is a single max pooling convolution followed by a single 1×1 convolution. The outputs of the four branches are concatenated along the channel dimension and used as the output of the module.

6. A power system transient stability prediction device based on a Denseception network, characterized in that, include: The steady-state operating condition sample set construction module is used to construct sample sets of various power system steady-state operating conditions covering different load conditions using the super-Latin sampling method; The transient data and TSI acquisition module is used to set up a three-phase short-circuit fault for each power system steady-state operating point and acquire transient data and transient stability index (TSI) for each transient fault scenario. The feature matrix extraction module is used to select node voltage, generator terminal voltage, generator rotor angular velocity, generator active power and generator reactive power from several sampling points in each transient data as input features and form a feature matrix; The three-channel feature map generation module is used to copy and expand the feature matrix three times along the channel dimension, ultimately forming a three-channel two-dimensional time-series feature map. The prediction model training module is used to fuse DenseNet, Xception and fully connected layers to construct the Denseception network. The three-channel two-dimensional time series feature map is input into the Denseception network for training, and finally the power system transient stability prediction model based on the Denseception network is obtained. The transient stability online prediction module is used to acquire real-time transient stability data from the power system; then extract input features from the real-time transient stability data; convert the input features into an image; then input the image into a Denseception-based transient stability prediction model; finally output TSI to obtain the transient stability of the power system after a fault.

7. A power system transient stability prediction device based on a Denseception network according to claim 6, characterized in that, The formula for calculating the Transient Stability Index (TSI) is as follows: （1）； In the formula, This represents the maximum value of the rotor angle difference between any two generators within the time-domain simulation period; To find the absolute value; max is the maximum value; when TSI>0, the system is transiently stable; when TSI<0, the system is transiently unstable.

8. A power system transient stability prediction device based on a Denseception network according to claim 6, characterized in that, The feature matrix is ​​as follows: （2） ； In the formula, This refers to the amplitude of the bus voltage. This refers to the active power of the generator. This refers to the reactive power of the generator. This refers to the generator terminal voltage; ω is the angular velocity of the generator rotor; n is the number of generators; m is the number of busbars. This refers to the voltage amplitude of the first busbar. This refers to the voltage amplitude of the second busbar. The voltage amplitude of the nth bus; This represents the voltage amplitude of the (n+1)th bus. The voltage amplitude of the m-th bus; This represents the active power of the first generator. This represents the active power of the second generator. Let be the active power of the nth generator; This represents the reactive power of the first generator. This refers to the reactive power of the second generator; Let n be the reactive power of the nth generator; This is the terminal voltage of the first generator; This is the terminal voltage of the second generator; Let be the terminal voltage of the nth generator; The angular velocity of the first generator; The angular velocity of the second generator; Let be the angular velocity of the nth generator.

9. A power system transient stability prediction device based on a Denseception network according to claim 6, characterized in that, The Denseception network structure is as follows: The Denseception network takes an image of size m×40×3 as input. It first extracts initial features through a 7×7 convolutional layer, then passes through a 3×3 max pooling layer to reduce spatial dimensionality and expand the receptive field. Next, it connects four DenseNet modules to achieve dense cross-layer feature reuse, then connects three Xception modules to achieve multi-scale convolutional feature extraction. Finally, it passes through a global average pooling layer to converge spatial features, and finally connects to a fully connected layer and outputs the transient stable state through a Softmax layer.

10. A power system transient stability prediction device based on a Denseception network according to claim 9, characterized in that, The specific structure of the DenseNet module is as follows: the input first undergoes a 1×1 convolution, then is divided into four branches: branch 1 is a single 1×1 convolution, branch 2 is a single 1×1 convolution followed by a single 3×3 convolution, branch 3 is a single average pooling convolution followed by a single 3×3 convolution, and branch 4 is a single 1×1 convolution followed by two consecutive 3×3 convolutions. The outputs of the four branches are concatenated along the channel dimension and used as the output of the module. The specific structure of the Xception module is as follows: the input is also divided into four branches: branch 1 is a single 1×1 convolution, branch 2 is a single 1×1 convolution followed by a single 3×3 convolution, branch 3 is a single 1×1 convolution followed by a single 5×5 convolution, and branch 4 is a single max pooling convolution followed by a single 1×1 convolution. The outputs of the four branches are concatenated along the channel dimension and used as the output of the module.

Images