A new energy power system transient stability evaluation method and device

Abstract

The application discloses a new energy power system transient stability evaluation method and device, belongs to the field of power grid transient stability evaluation technology, and comprises the following steps: constructing a transient dynamics model of a new energy power system; including establishing a transient dynamics model of a synchronous machine, establishing a transient dynamics model of a virtual synchronous machine, and establishing a transient dynamics model of the whole system based on the transient dynamics model of the synchronous machine and the transient dynamics model of the virtual synchronous machine; adopting an adjoint gradient physical neural network to predict the transient characteristics of the system; including: constructing a transient dynamics dataset, calculating a data loss function, a physical loss function and a total loss function, calculating an adjoint state and its gradient, and updating the parameters of the adjoint gradient physical neural network; and adopting a self-attention network to evaluate the transient stability performance of the system. The application realizes quantitative evaluation of the transient stability of a power system, and solves the problems of poor interpretability and low precision of existing data-driven artificial intelligence methods.

Description

TECHNICAL FIELD

[0001] The application belongs to the technical field of power grid transient stability evaluation, and particularly relates to a new energy power system transient stability evaluation method and device. BACKGROUND

[0002] Transient stability evaluation technology is an important guarantee for the safe and stable operation of a power system. The transient characteristics of a traditional power system are determined by the rotor inertia of a small number of synchronous generators, but as the penetration rate of renewable energy sources such as wind and solar energy continues to increase, the transient characteristics are affected by a large number of new energy power electronic converters, and increasingly show high-dimensional and nonlinear characteristics, which poses new challenges to the transient stability evaluation technology of new energy power systems.

[0003] At present, there have been a large number of studies on the transient stability evaluation method of a power system, and common methods include time domain simulation method, Lyapunov energy function method, and data-driven methods represented by supervised learning. However, the existing methods have the following defects and deficiencies: the time domain simulation method has large calculation amount and long time consumption, and is difficult to meet the online evaluation requirements; the Lyapunov energy function method relies on expert experience to construct the energy function, and it is difficult to construct the energy function of a complex multi-machine system, and such methods have the disadvantage of strong conservatism; the existing data-driven artificial intelligence methods belong to "black box" learning, and are extremely dependent on prior data sets, and have the problems of poor interpretability and low evaluation accuracy; and the existing physical neural network methods only simply embed the physical mechanism in the loss function, and severely rely on initial values when solving complex power system dynamics equations, and have weak generalization ability. SUMMARY

[0004] To solve the above technical problems, the application adopts the following technical solutions:

[0005] A new energy power system transient stability evaluation method, comprising:

[0006] Step 1, constructing a transient dynamics model of a new energy power system; including establishing a transient dynamics model of a synchronous machine, establishing a transient dynamics model of a virtual synchronous machine, and establishing a transient dynamics model of the whole system based on the transient dynamics model of the synchronous machine and the transient dynamics model of the virtual synchronous machine;

[0007] Step 2, predicting the transient characteristics of the system by using an adjoint gradient physical neural network; including: constructing a transient dynamics data set, calculating a data loss function, a physical loss function and a total loss function, calculating an adjoint state and its gradient, and updating the parameters of the adjoint gradient physical neural network;

[0008] Step 3, evaluating the transient stability performance of the new energy power system by using a self-attention network.

[0009] A new energy power system transient stability evaluation device, comprising:

[0010] A transient dynamics model construction module constructs a transient dynamics model of the new energy power system, including establishing a transient dynamics model of a synchronous machine, establishing a transient dynamics model of a virtual synchronous machine, and establishing a transient dynamics model of the whole system based on the transient dynamics model of the synchronous machine and the transient dynamics model of the virtual synchronous machine;

[0011] A transient characteristic prediction module predicts the transient characteristics of the system using an adjoint gradient physical neural network, including: constructing a transient dynamics dataset, calculating a data loss function, a physical loss function and a total loss function, calculating an adjoint state and its gradient, and updating the parameters of the adjoint gradient physical neural network;

[0012] An evaluation module evaluates the transient stability performance of the new energy power system using a self-attention network.

[0013] The present application has the following beneficial effects:

[0014] The present application combines the adjoint gradient physical neural network and the Transformer network, uses the transient characterization capability of the adjoint gradient physical neural network to provide interpretable auxiliary time series features for the Transformer network, realizes quantitative evaluation of the transient stability of the power system, solves the problems of poor interpretability and low precision of existing data-driven artificial intelligence methods, and overcomes the drawbacks of strong conservatism of Lyapunov energy function method and long time consumption of time domain simulation method.

[0015] The present application realizes accurate characterization of the transient dynamics of the power system in a data-physical fusion driven manner by embedding physical mechanisms and introducing an adjoint gradient mechanism, has stronger transient recognition capability and higher evaluation precision, and solves the initial value dependence problem of existing physical neural network methods. BRIEF DESCRIPTION OF DRAWINGS

[0016] Figure 1 The flowchart of the new energy power system transient stability evaluation method based on the adjoint gradient physical neural network of the present application. DETAILED DESCRIPTION

[0017] In order to make the purpose, technical scheme and advantages of the present application clearer, the present application will be further described in detail below in combination with the drawings and examples. It should be understood that the specific examples described herein are only used to explain the present application and do not limit the present application. In addition, the technical features involved in each embodiment of the present application described below can be combined with each other as long as they do not conflict with each other.

[0018] As Figure 1As shown, the present application provides a new energy power system transient stability evaluation method, comprising:

[0019] Step 1, constructing a transient dynamics model of the new energy power system;

[0020] Step 2, using an adjoint gradient physical neural network to predict the transient characteristics of the system;

[0021] Step 3, using a self-attention network (Transformer) network to evaluate the transient stability performance of the system.

[0022] Preferably, the step 1 comprises:

[0023] Step 11, establishing a transient dynamics model of the synchronous machine:

[0024] The following 4-order transient dynamics equation of the synchronous machine is established:

[0025] (1)

[0026] In the formula, is the phase of the i-th synchronous machine (θi) is the first-order derivative with respect to time, which is well known in the art, and similar forms of other parameters have similar meanings, which will not be described in detail), is the reference angular frequency, is the angular frequency of the i-th synchronous machine, is the synchronous angular frequency, is the inertia constant of the i-th synchronous machine, is the mechanical input power of the i-th synchronous machine, is the electrical output power of the i-th synchronous machine, is the damping coefficient of the i-th synchronous machine, is the q-axis transient electromotive force of the i-th synchronous machine, is the excitation electromotive force of the i-th synchronous machine, is the d-axis open-circuit transient time constant of the i-th synchronous machine, is the d-axis synchronous reactance of the i-th synchronous machine, is the d-axis transient reactance of the i-th synchronous machine, is the component of the stator current in the d-axis, is the d-axis transient electromotive force of the i-th synchronous machine, is the q-axis open-circuit transient time constant of the i-th synchronous machine, is the q-axis synchronous reactance of the i-th synchronous machine, is the q-axis transient reactance of the i-th synchronous machine, is the component of the stator current in the q-axis.

[0027] ​The stator algebraic equation is:

[0028] (2)

[0029] where, is the stator d-axis voltage of the i-th synchronous machine, is the stator q-axis voltage of the i-th synchronous machine, is the stator winding resistance.

[0030] Step 12, establish the transient dynamics model of the virtual synchronous machine:

[0031] The transient dynamics equation of the following 2nd-order virtual synchronous machine is established:

[0032] (3)

[0033] where, is the phase of the k-th virtual synchronous machine, is the angular frequency of the k-th virtual synchronous machine, is the virtual inertia constant of the k-th virtual synchronous machine, is the input power of the i-th virtual synchronous machine, is the output power of the i-th virtual synchronous machine, is the virtual damping coefficient of the i-th virtual synchronous machine.

[0034] Step 13, establish the transient dynamics model of the whole system:

[0035] Based on the transient dynamics model of the synchronous machine in step 11 and the transient dynamics model of the virtual synchronous machine in step 12, the following transient dynamics model of the whole system is established:

[0036] (4)

[0037] where, is the number of synchronous machine units, is the number of virtual synchronous machine units, is is the unit matrix of dimension n, is the diagonal matrix composed of the inertia constants of all synchronous machines, is the diagonal matrix composed of the d-axis open-loop transient time constants of all synchronous machines, is the diagonal matrix composed of the q-axis open-loop transient time constants of all synchronous machines, is is the unit matrix of dimension n, is the diagonal matrix composed of the virtual inertia constants of all virtual synchronous machines;

[0038] is the angular frequency of all synchronous machines, mechanical input power of all synchronous machines, output power of all synchronous machines, inertia coefficient of all synchronous machines, field voltage of all synchronous machines, q-axis transient electromotive force of all synchronous machines, d-axis synchronous reactance of all synchronous machines, d-axis transient reactance of all synchronous machines, represents Hadamard product, d-axis current of all synchronous machines, d-axis transient electromotive force of all synchronous machines, q-axis synchronous reactance of all synchronous machines, q-axis transient reactance of all synchronous machines, q-axis current of all synchronous machines, angular frequency deviation of all virtual synchronous machines, input power of all virtual synchronous machines, output power of all virtual synchronous machines, virtual damping coefficient of all virtual synchronous machines;

[0039] state variables of the system, specifically including phase angle of all synchronous machines angular frequency deviation of all synchronous machines q-axis transient electromotive force of all synchronous machines d-axis transient electromotive force of all synchronous machines phase angle of all virtual synchronous machines angular frequency deviation of all virtual synchronous machines ; for all parameters, -1 in the upper right corner represents inversion operation of the matrix.

[0040] Further, the above transient dynamics model of the whole system is rewritten into a general form:

[0041] (5)

[0042] wherein, state variables of the system, nonlinear dynamics function satisfying formula (4).

[0043] Preferably, the step 2 comprises:

[0044] Step 21, constructing a transient dynamics dataset:

[0045] In combination with simulation and actual data, the following dataset is constructed:

[0046] (6)

[0047] (7)

[0048] In the formula, For the labeled dataset, For the accompanying dataset, For the sampling time in the labeled dataset, For the label data to be concentrated The values ​​of the state variables at each time step are: For the sampling time in the accompanying dataset, For accompanying data to be concentrated The value of the system's nonlinear dynamic function at time t is... The total number of samples in the labeled dataset. This represents the total number of samples in the accompanying dataset.

[0049] Step 22, calculate the data loss function:

[0050] Sample from the labeled dataset and calculate the data loss function using the following formula:

[0051] (8)

[0052] In the formula, For data loss function, The number of samples drawn from the labeled dataset. for The actual value of the state variable at time t. for The predicted value of the state variable at time t.

[0053] Step 23, calculate the physical loss function:

[0054] The following physical loss function is calculated based on all data in the accompanying dataset:

[0055] (9)

[0056] In the formula, For physical loss function, For the j-th sampling time in the accompanying dataset, For the system dynamics function learned by the accompanying gradient physical neural network, These are the parameters of the physical neural network with accompanying gradients.

[0057] Step 24, calculate the total loss function:

[0058] Calculate the total loss function as follows:

[0059] (10)

[0060] wherein, is the total loss function, is the weight coefficient of the data loss function, is the weight coefficient of the physical loss function.

[0061] Step 25, calculate the adjoint state and its gradient:

[0062] For all samples in the labeled data set, its adjoint state is calculated according to the following formula:

[0063] (11)

[0064] wherein, is the state variable The adjoint state of ;

[0065] The gradient of the adjoint state is calculated at the same time:

[0066] (12)

[0067] wherein, is the gradient of the adjoint state, represents the partial differential operator, is the system dynamics function of the adjoint gradient physical neural network learning.

[0068] Step 26, update the adjoint gradient physical neural network parameters:

[0069] The change amount of the loss function gradient is calculated according to the following formula:

[0070] (13)

[0071] wherein, is the change amount of the loss function gradient; is the system dynamics function of the adjoint gradient physical neural network learning.

[0072] Then update the parameters of the adjoint gradient physical neural network according to the following formula:

[0073] (14)

[0074] wherein, is the network parameter update weight;

[0075] Repeat steps 22 to 26 until the maximum number of training rounds is reached.

[0076] Preferably, the step 3 comprises:

[0077] Step 31, prepare transient stability discrimination data set:

[0078] Combined with simulation and actual data, the following transient stability discrimination data set is constructed:

[0079] (15)

[0080] wherein, is the transient stability discrimination data set, is the total number of samples in the transient stability discrimination data set, is the system state variable sample of the i-th initial time, is the label true value corresponding to the i-th sample.

[0081] Step 32, extract system initial state variable:

[0082] The system initial state variable in the data set is extracted by using a fully connected neural network:

[0083] (16)

[0084] (17)

[0085] (18)

[0086] wherein, represents the state variable initial value information after feature extraction, represents the output of the fully connected neural network, is the output of the k-th fully connected neural network, is the k-th fully connected neural network, is the total number of fully connected neural networks, is the weight term of the k-th fully connected neural network, is the bias term of the k-th fully connected neural network, is an activation function operator, and its expression is wherein, is the independent variable of the function.

[0087] Step 33, calculate transient process:

[0088] Based on the aforementioned well-trained (i.e., repeatedly perform steps 22 to 26 until the maximum number of training rounds is reached) adjoint gradient physical neural network, for any given system initial state variable, the transient process of the system state variable is calculated as follows:

[0089] (19)

[0090] wherein, a system dynamics function learned with gradient physical neural networks, a system state variable at time t.

[0091] Step 34, extract the time series features:

[0092] The self-attention network (Transformer) network is used to extract the time series features of the system state variable:

[0093] The stacked matrix of the state variable is calculated as follows:

[0094] (20)

[0095] In the formula, represents the stacked input matrix of the Transformer network, and d represents the dimension of the hidden layer of the Transformer network; is the value of the system state variable at each sampling time, is the time, is a real set.

[0096] Then, the multi-head attention layer is used to extract the time series features layer by layer:

[0097] (21)

[0098] (22)

[0099] (23)

[0100] is the number of layers of multi-head attention, is the total number of multi-head attention layers, is the output of the lth multi-head attention layer, represents the multi-head attention operator, and h represents the number of heads of multi-head attention, represents the i-th multi-head attention (i takes values from 1 to h), represents the softmax activation function, is the Q, K, V key-value pair corresponding to the i-th multi-head attention, is the Q, K, V projection matrix corresponding to the i-th multi-head attention, is the dimension of K / Q in a single attention.

[0101] ​​​​​​​​Step 35, calculating the aggregated features:

[0102] Based on the system initial state variables extracted by the full connection neural network in the foregoing step 32 and the time sequence features of the system state variables extracted by the Transformer network in step 34, the aggregated features are calculated according to the following formula:

[0103] (24)

[0104] In the formula, is the aggregated feature, is the final output of the Transformer network, is the final output of the full connection neural network, is the weight term of the linear mapping, is the bias term of the linear mapping.

[0105] Step 36, calculating the output probability:

[0106] The final transient stability classification probability is calculated according to the following formula:

[0107] (25)

[0108] In the formula, is the probability of the i-th classification label, is the exponential function, C is the total number of multi-target classification, is the i-th dimension of the aggregated feature, is the j-th dimension of the aggregated feature.

[0109] Finally, the transient stability performance of the system is determined according to the output probability distribution, such as judging instability / stability, judging the equilibrium point to which the system is stable, judging the type of instability to which the system belongs, etc. In addition, according to the output of the accompanying gradient physical neural network, the transient dynamic behavior of the system can be obtained, and further analysis can be made on the transient process of the system.

[0110] The present application also provides a new energy power system transient stability evaluation device, comprising:

[0111] a transient dynamics model construction module, which constructs a transient dynamics model of the new energy power system; including establishing a transient dynamics model of a synchronous machine, establishing a transient dynamics model of a virtual synchronous machine, and establishing a transient dynamics model of the whole system based on the transient dynamics model of the synchronous machine and the transient dynamics model of the virtual synchronous machine;

[0112] The transient characteristic prediction module predicts the transient characteristic of the system by using the adjoint gradient physical neural network, and comprises the following steps: constructing a transient dynamic data set, calculating a data loss function, a physical loss function and a total loss function, calculating an adjoint state and its gradient, and updating the parameters of the adjoint gradient physical neural network.

[0113] The evaluation module evaluates the transient stability performance of the new energy power system by using a self-attention network.

[0114] Those skilled in the art will appreciate that embodiments of the present application can be provided as methods, systems, or computer program products. Accordingly, the present application can take the form of an entirely hardware embodiment, an entirely software embodiment, or an embodiment combining software and hardware aspects. Furthermore, the present application can take the form of a computer program product implemented on one or more computer-usable storage media (including, but not limited to, disk storage, CD-ROMs, optical storage, etc.) containing computer usable program code. The solutions in the embodiments of the present application can be implemented in various computer languages.

[0115] The present application is described with reference to flowcharts and / or block diagrams of methods, devices (systems), and computer program products according to embodiments of the present application. It should be understood that each flow and / or block in the flowcharts and / or block diagrams, as well as combinations of flows and / or blocks in the flowcharts and / or block diagrams, can be implemented by computer program instructions. These computer program instructions can be provided to a processor of a general-purpose computer, a special-purpose computer, an embedded processor, or other programmable data processing apparatus to produce a machine, so that the instructions executed by the processor of the computer or other programmable data processing apparatus produce a device that implements the functions specified in the flowcharts and / or block diagrams. Figure 1 one or more flows and / or blocks Figure 1 an apparatus that performs the functions specified in the flowcharts and / or block diagrams.

[0116] These computer program instructions can also be stored in a computer-readable memory that can direct the computer or other programmable data processing apparatus to work in a specific manner, so that the instructions stored in the computer-readable memory produce a manufactured product including instruction apparatus, which implements the functions specified in the flowcharts and / or block diagrams. Figure 1 one or more flows and / or blocks Figure 1 an apparatus that performs the functions specified in the flowcharts and / or block diagrams.

[0117] These computer program instructions can also be loaded onto a computer or other programmable data processing apparatus, so that a series of operation steps are performed on the computer or other programmable data processing apparatus to produce a computer-implemented process, so that the instructions executed on the computer or other programmable data processing apparatus provide a process for implementing the functions specified in the flowcharts and / or block diagrams. Figure 1 one or more flows and / or blocks Figure 1 an apparatus that performs the functions specified in the flowcharts and / or block diagrams.

[0118] While the preferred embodiments of the application have been described, additional variations and modifications can be made to the preferred embodiments by those skilled in the art once they have the benefit of the present disclosure. Therefore, the present application intends all such additional variations and modifications to be included within the scope of the present application.

[0119] It is apparent that those skilled in the art can make modifications and variations to the present application without departing from the spirit and scope of the application. Accordingly, it is intended to include all such modifications and variations in the scope of the present application.

[0120] The above description is only the preferred embodiments of the present application, and is not intended to limit the scope of the present application, and any equivalent structure or equivalent process transformation made by using the content of the present application specification and drawings, or directly or indirectly applied in other related system fields, are also included in the protection scope of the present application.

[0121] The content not described in detail in the present application specification is the prior art known to those skilled in the art.

Claims

1. A method for transient stability assessment of a new energy power system, characterized in that, Comprising: Step 1, constructing a transient dynamics model of a new energy power system; including establishing a transient dynamics model of a synchronous machine, establishing a transient dynamics model of a virtual synchronous machine, and establishing a transient dynamics model of the whole system based on the transient dynamics model of the synchronous machine and the transient dynamics model of the virtual synchronous machine; Step 2, predicting the transient characteristics of the system using an adjoint gradient physical neural network; including: constructing a transient dynamics dataset, calculating a data loss function, a physical loss function and a total loss function, calculating an adjoint state and its gradient, and updating the parameters of the adjoint gradient physical neural network; Step 3, evaluating the transient stability performance of the new energy power system using a self-attention network, including: Step 31, preparing a transient stability discrimination dataset: Based on simulation and actual data, the following transient stability discrimination dataset is constructed: （15） In the formula, is a transient stability discrimination data set, is the total number of samples in the transient stability discrimination data set, is the system state variable sample of the i th initial time, is the label true value corresponding to the i th sample; Step 32, extracting the initial state variables of the system in the dataset using a fully connected neural network: （16） （17） （18） wherein, represents the initial value information of the state variable after feature extraction, represents the output of the fully connected neural network, is the output of the kth fully connected neural network, is the kth layer of the fully connected neural network, is the total number of layers of the fully connected neural network, is the weight term of the kth layer of the fully connected neural network, is the bias term of the kth layer of the fully connected neural network, represents the activation function operator; Step 33, calculating the transient process of the system state variables; Step 34, extracting the time sequence characteristics of the system state variables using a self-attention network; Step 35, calculating the aggregated features; Step 36, calculating the final transient stability classification probability.

2. The method of claim 1, wherein, Step 1 includes: Step 11, establishing a transient dynamics model of a synchronous machine: The following 4th order synchronous machine transient dynamics equation is established: （1） wherein is the phase of the ith synchronous machine, is the reference angular frequency, is the angular frequency of the ith synchronous machine, is the synchronous angular frequency, is the inertia constant of the ith synchronous machine, is the mechanical input power of the ith synchronous machine, is the electrical output power of the ith synchronous machine, is the damping coefficient of the ith synchronous machine, is the q-axis transient electromotive force of the ith synchronous machine, is the field voltage of the ith synchronous machine, is the d-axis open-circuit transient time constant of the ith synchronous machine, is the d-axis synchronous reactance of the ith synchronous machine, is the d-axis transient reactance of the ith synchronous machine, is the component of the stator current in the d-axis, is the d-axis transient electromotive force of the ith synchronous machine, is the q-axis open-circuit transient time constant of the ith synchronous machine, is the q-axis synchronous reactance of the ith synchronous machine, is the q-axis transient reactance of the ith synchronous machine, is the component of the stator current in the q-axis; Its algebraic equation for the stator is: （2） wherein is the stator d-axis voltage of the i-th synchronous machine, is the stator q-axis voltage of the i-th synchronous machine, is the stator winding resistance; Step 12, establishing a transient dynamics model of a virtual synchronous machine: The following 2nd order transient dynamics equation of the virtual synchronous machine is established: （3） wherein, is the phase of the kth virtual synchronous machine, is the angular frequency of the kth virtual synchronous machine, is the virtual inertia constant of the kth virtual synchronous machine, is the input power of the ith virtual synchronous machine, is the output power of the ith virtual synchronous machine, is the virtual damping coefficient of the ith virtual synchronous machine; Step 13, establishing a transient dynamics model of the whole system: Based on the transient dynamics model of the synchronous machine in step 11 and the transient dynamics model of the virtual synchronous machine in step 12, the following transient dynamics model of the whole system is established: （4） wherein is the number of synchronous machines, is the number of virtual synchronous machines, is the number of synchronous machines, is the identity matrix of dimension n, is the diagonal matrix of inertia constants of all synchronous machines, is the diagonal matrix of d-axis open-circuit transient time constants of all synchronous machines, is the diagonal matrix of q-axis open-circuit transient time constants of all synchronous machines, is the identity matrix of dimension n, is the identity matrix of dimension n, is the diagonal matrix of virtual inertia constants of all virtual synchronous machines; ω for the angular frequency of all synchronous machines, P in for the mechanical input power of all synchronous machines, P out for the output power of all synchronous machines, J for the inertia coefficient of all synchronous machines, E for the excitation voltage of all synchronous machines, E q for the q-axis transient electromotive force of all synchronous machines, X d for the d-axis synchronous reactance of all synchronous machines, X d for the d-axis transient reactance of all synchronous machines, H for the Hadamard product, I d for the d-axis current of all synchronous machines, E d for the d-axis transient electromotive force of all synchronous machines, X q for the q-axis synchronous reactance of all synchronous machines, X q for the q-axis transient reactance of all synchronous machines, I q for the q-axis current of all synchronous machines, ω for the angular frequency deviation of all virtual synchronous machines, P in for the input power of all virtual synchronous machines, P out for the output power of all virtual synchronous machines, J for the virtual damping coefficient of all virtual synchronous machines; State variables of the system, in particular including the phase angle of all synchronous machines the angular frequency deviation of all synchronous machines the q-axis transient electromotive force of all synchronous machines the d-axis transient electromotive force of all synchronous machines the phase angle of all virtual synchronous machines the angular frequency deviation of all virtual synchronous machines .

3. The method of claim 2, wherein, Step 2 includes: Step 21, constructing a transient dynamics dataset: Based on simulation and actual data, the following dataset is constructed: （6） （7） In the formula, For the labeled dataset, For the accompanying dataset, For the sampling time in the labeled dataset, For the label data to be concentrated The values ​​of the state variables at each time step are: For the sampling time in the accompanying dataset, For accompanying data to be concentrated The value of the system's nonlinear dynamic function at time t is... The total number of samples in the labeled dataset. This represents the total number of samples in the accompanying dataset; Step 22, calculating a data loss function: Sampling from the label dataset and calculating the data loss function: （8） In the formula, is a data loss function, is the number of samples extracted from the label data set, is is the true value of the state variable at time t, is is the predicted value of the state variable at time t; Step 23, calculating a physical loss function: The physical loss function is calculated according to all the data in the adjoint dataset: （9） wherein, is a physical loss function, is a system dynamics function accompanying gradient physical neural network learning, is a system dynamics function accompanying gradient physical neural network learning, is a system dynamics function accompanying gradient physical neural network learning, Step 24, calculating a total loss function: The total loss function is calculated: （10） wherein is a total loss function, is a weight coefficient of the data loss function, is a weight coefficient of the physical loss function; Step 25, calculating an adjoint state and its gradient: Step 26, updating the parameters of the adjoint gradient physical neural network; Steps 22 to 26 are repeated until the maximum number of training rounds is reached.

4. The method of claim 3, wherein, In step 25, for all samples in the label dataset, the adjoint state is calculated: （11） wherein is a state variable In the accompanying state; The gradient of the adjoint state is also calculated: （12） wherein is the gradient of the state, represents the partial differential operator, is the system dynamics function learned by the adjoint gradient physical neural network.

5. The method of claim 4, wherein, In step 26, the change in the loss function gradient is calculated as follows: （13） In the formula, is the change amount of the loss function gradient; is the system dynamics function learned with the gradient physical neural network Then the parameters of the adjoint gradient physical neural network are updated as follows: （14） In the formula, is a network parameter update weight.

6. The method of claim 5, wherein, Step 33 includes: Based on the trained adjoint gradient physical neural network, for any given system initial state variable, the transient process of the system state variable is calculated as follows: （19） wherein, is a system dynamics function learned with gradient physical neural networks, is is a system state variable at time instant 7. The method of claim 6, wherein, Step 35 includes: Based on the system initial state variables extracted by the fully connected neural network in the aforementioned step 32 and the time sequence characteristics of the system state variables extracted by the Transformer network in step 34, the aggregated features are calculated as follows: （24） wherein is a characteristic of the polymerization, is the final output of the Transformer network, is the final output of the fully connected neural network, is a weight term of the linear mapping, is a bias term of the linear mapping.

8. The method of claim 7, wherein, Step 36 includes: The final transient stability classification probability is calculated as follows: （25） wherein, is the probability of the i-th classification label, is the exponential function, C is the total number of multi-objective classification, is the i-th dimension of the aggregated feature, is the j-th dimension of the aggregated feature.

9. A device for transient stability assessment of a new energy power system, characterized in that, Comprising: A transient dynamics model construction module constructs a transient dynamics model of the new energy power system, including establishing a transient dynamics model of a synchronous machine, establishing a transient dynamics model of a virtual synchronous machine, and establishing a transient dynamics model of the whole system based on the transient dynamics model of the synchronous machine and the transient dynamics model of the virtual synchronous machine; A transient characteristic prediction module predicts the transient characteristics of the system by using an adjoint gradient physical neural network, including: constructing a transient dynamics dataset, calculating a data loss function, a physical loss function and a total loss function, calculating an adjoint state and its gradient, and updating the parameters of the adjoint gradient physical neural network; An evaluation module evaluates the transient stability performance of the new energy power system by using a self-attention network, including: Preparing a transient stability discrimination dataset: In combination with simulation and actual data, the following transient stability discrimination dataset is constructed: （15） In the formula, is a transient stability discrimination data set, is the total number of samples in the transient stability discrimination data set, is the system state variable sample of the i-th initial time, is the label true value corresponding to the i-th sample; Using a fully connected neural network to extract the initial state variables of the system in the dataset: （16） （17） （18） In the formula, represent the initial value information of the state variable after feature extraction, represent the output of the fully connected neural network, is the output of the kth fully connected neural network, is the kth layer of the fully connected neural network, is the total number of fully connected neural networks, is the weight term of the kth layer of the fully connected neural network, is the bias term of the kth layer of the fully connected neural network, represent the activation function operator; Calculating the transient process of the system state variables; Using a self-attention network to extract the time sequence characteristics of the system state variables; Calculating the aggregated features; Calculating the final transient stability classification probability.

Images