Physical data fusion driven new energy power grid transient trajectory calculation method and device
Patent Information
- Application Number
- CN202611291531.X
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2026-08-25
- Publication Date
- 2026-09-22
AI Technical Summary
[0006]为解决现有数据驱动方法可解释性不足、传统数值仿真计算耗时长及物理机理变化下适应性较弱的问题，本发明提供一种物理数据融合驱动的新能源电网暂态轨迹计算方法及装置，以提高暂态轨迹计算的准确性、物理一致性和在线推理效率
[0024]1、本发明通过建立数据驱动预测分支和物理机理计算分支，使暂态轨迹预测结果同时受到样本数据和物理规律约束，实现了数据驱动模型与物理机理模型的协同求解，提升了新能源电网暂态轨迹计算的准确性、可信性和可解释性。
Smart Images
Figure CN122797355A_ABST
Abstract
Description
Technical Field
[0001] This invention belongs to the field of power system transient calculation technology, specifically relating to a method and device for calculating transient trajectories of new energy power grids driven by physical data fusion. Background Technology
[0002] Transient trajectory calculation is a crucial foundation for power grid security and stability analysis, post-fault dynamic behavior characterization, and online safety early warning. Traditional power grids are primarily supported by the inertia of synchronous generators, and their transient processes can usually be solved using electromechanical transient models and numerical integration methods. However, with the large-scale integration of wind power, photovoltaics, energy storage, and various power electronic equipment, the dynamic behavior of new energy power grids exhibits characteristics such as low inertia, strong nonlinearity, multi-timescale coupling, and time-varying operating mechanisms, posing new challenges to traditional transient calculation methods.
[0003] Currently, the calculation of transient trajectories of power grids mainly relies on numerical integration methods. These methods have clear physical meanings, but in power grids with large-scale new energy integration, their differential-algebraic equations have high dimension, strong rigidity, many parameters, and limited computational step size, resulting in long computation time and making it difficult to meet the needs of online rapid calculation and batch evaluation in multiple scenarios.
[0004] In recent years, deep learning methods have been used for power grid dynamic trajectory prediction and transient stability assessment. These methods can learn the mapping relationship between inputs and outputs from historical simulation data or actual operating data, and have a relatively fast inference speed. However, purely data-driven models often lack physical constraints, making them prone to problems such as predicted trajectories not conforming to power grid dynamics, insufficient generalization ability to unseen operating conditions, sensitivity to noisy data, and poor adaptability to sudden changes in physical mechanisms.
[0005] Furthermore, existing physical information neural network methods typically incorporate the physical equation residuals as part of the loss function, improving the model's physical consistency through unidirectional constraints. However, in the transient solution of complex renewable energy power grids, simply embedding the physical residuals into the loss function fails to fully leverage the complementary advantages between the physical model and the data model, and also struggles to handle complex scenarios such as abrupt changes in physical mechanisms and parameter uncertainties. Therefore, a novel architecture for transient trajectory calculation is urgently needed that can simultaneously leverage the interpretability of the physical model and the rapid fitting capability of the data model, achieving bidirectional synergy and reliable fusion between data-driven prediction results and physical mechanism calculation results. Summary of the Invention
[0006] To address the issues of insufficient interpretability, long computation time, and weak adaptability to changes in physical mechanisms in existing data-driven methods, this invention provides a physical data fusion-driven method and apparatus for calculating transient trajectories of new energy power grids, thereby improving the accuracy, physical consistency, and online inference efficiency of transient trajectory calculation.
[0007] To achieve the above objectives, the present invention adopts the following technical solution:
[0008] A method for calculating the transient trajectory of a new energy power grid driven by physical data fusion includes:
[0009] Step 1: Construct a transient dynamic model of the new energy power grid, and express synchronous generators, new energy converters, network constraints and load dynamics in a unified form of differential-algebraic equations;
[0010] Step 2: Construct a transient trajectory dataset and a physical mechanism sample library;
[0011] Step 3: Establish a physics-data joint learning architecture; the physics-data joint learning architecture includes a data-driven prediction branch and a physical mechanism calculation branch; the data-driven prediction branch is used to predict the transient trajectory of the system based on the initial state, time series, operating conditions and disturbance information of the system; the physical mechanism calculation branch is used to generate transient responses that meet physical constraints based on the transient dynamics model of the new energy power grid, the physical parameters to be identified or the physical trajectory library;
[0012] Step 4: Construct the joint learning loss function; the joint learning loss function includes the data loss function, the physical residual loss function, the data-to-physical coupling loss function, the physical-to-data coupling loss function, and the total loss function;
[0013] Step 5: Based on the joint learning loss function, alternately train and update the data-driven prediction branch and the physical mechanism calculation branch with bidirectional error feedback to obtain the trained physical-data joint learning model.
[0014] Step 6: Calculate the transient trajectory of the new energy power grid under given initial state and operating conditions using the trained physical-data joint learning model, and output the transient stability evaluation index.
[0015] This invention also provides a physical data fusion-driven transient trajectory calculation device for new energy power grids, comprising:
[0016] The transient dynamics model construction module is used to construct transient dynamics models of new energy power grids and to uniformly represent synchronous generators, new energy converters, network constraints and load dynamics in the form of differential-algebraic equations.
[0017] The dataset building module is used to build transient trajectory datasets and physical mechanism sample libraries;
[0018] The data-driven prediction module is used to predict the transient trajectory of the system based on the initial state, time series, operating conditions and disturbance information of the system.
[0019] The physical mechanism calculation module is used to generate transient responses that meet physical constraints based on the transient dynamics model of the new energy power grid, the physical parameters to be identified, or the physical mechanism sample library.
[0020] The joint learning training module is used to construct the data loss function, physical residual loss function, data-to-physical coupling loss function, physical-to-data coupling loss function and total loss function, and to perform alternating training and bidirectional error feedback updates on the data-driven prediction module and the physical mechanism calculation module.
[0021] The transient trajectory output module is used to output the transient trajectory of the new energy power grid using the trained physical-data joint learning model;
[0022] The stability evaluation module is used to calculate transient stability evaluation indicators based on the output transient trajectory of the new energy power grid.
[0023] Beneficial effects:
[0024] 1. This invention establishes a data-driven prediction branch and a physical mechanism calculation branch, so that the transient trajectory prediction results are simultaneously constrained by sample data and physical laws, realizing the collaborative solution of the data-driven model and the physical mechanism model, and improving the accuracy, reliability and interpretability of transient trajectory calculation of new energy power grid.
[0025] 2. This invention constructs a data-to-physics coupling loss function and a physics-to-data coupling loss function, enabling data-driven prediction results to correct equivalent parameters or physical trajectory selection results in the physical mechanism branch in reverse. At the same time, it enables the physical mechanism calculation results to constrain the training direction of the data-driven model, overcoming the shortcomings of traditional physical information neural networks that unidirectionally embed physical residuals.
[0026] 3. After training, this invention can quickly output transient trajectories and stability evaluation indicators. Compared with traditional pointwise numerical integration methods, it has higher online inference efficiency and stronger physical consistency and generalization ability compared with pure data-driven methods. Attached Figure Description
[0027] Figure 1 This is a flowchart of the physical data fusion-driven transient trajectory calculation method for new energy power grids according to the present invention.
[0028] Figure 2 Phase angles of different methods in embodiments of the present invention Transient trajectory comparison diagram; where a) is the predicted phase angle of Long Short-Term Memory Network (LSTM), b) is the predicted phase angle of Gated Recurrent Unit (GRU), c) is the predicted phase angle of Ordinary Temporal Convolutional Network (TCN), d) is the predicted phase angle of Transformer Network, e) is the predicted phase angle of Physically Guided Temporal Convolutional Network (PINN), and f) is the predicted phase angle of the method proposed in this invention.
[0029] Figure 3 Frequency of different methods in the embodiments of the present invention Transient trajectory comparison diagram; where a) is the predicted frequency of Long Short-Term Memory Network (LSTM), b) is the predicted frequency of Gated Recurrent Unit (GRU), c) is the predicted frequency of Ordinary Temporal Convolutional Network (TCN), d) is the predicted frequency of Transformer Network, e) is the predicted frequency of Physically Guided Temporal Convolutional Network (PINN), and f) is the predicted frequency of the method proposed in this invention. Detailed Implementation
[0030] To make the objectives, technical solutions, and advantages of this invention clearer, the invention will be further described in detail below with reference to the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are merely illustrative and not intended to limit the invention. Furthermore, the technical features involved in the various embodiments of this invention described below can be combined with each other as long as they do not conflict with each other.
[0031] like Figure 1 As shown, this invention provides a method for calculating the transient trajectory of a new energy power grid driven by physical data fusion, including:
[0032] Step 1: Construct a transient dynamic model of the new energy power grid;
[0033] Step 2: Construct a transient trajectory dataset and a physical mechanism sample library;
[0034] Step 3: Establish a physics-data joint learning architecture;
[0035] Step 4: Construct the joint learning loss function;
[0036] Step 5: Conduct alternating training and two-way error feedback updates;
[0037] Step 6: Output transient trajectory and transient stability evaluation index.
[0038] Preferably, step 1 includes:
[0039] Step 11: Establish a differential-algebraic transient dynamic model for the new energy power grid:
[0040] (1)
[0041] In the formula, The system's differential state variables include at least one or more of the following: synchronous machine power angle, angular velocity, new energy converter control state, voltage control state, and frequency control state; t represents time. These are system algebraic variables, including at least one or more of the following: node voltage magnitude, node voltage phase angle, network injected current, and network power balance variables; Input variables for the system include at least one or more of the following: load disturbance, renewable energy output, fault type, fault location, and fault duration. The system physical parameters include at least one or more of the following: inertia, damping, line reactance, controller parameters, converter parameters, and load parameters. This represents the nonlinear dynamic function corresponding to the system's differential equation; For system algebraic constraint functions.
[0042] Step 12: Based on the sampling time series The transient trajectory of the system is represented as:
[0043] (2)
[0044] In the formula, Let N be the system transient trajectory matrix, where N is the number of sampling times, and the superscript T indicates the transpose of the matrix.
[0045] Step 13: Construct parameterized transient mapping relationships for different operating conditions, disturbance scenarios, and physical parameters:
[0046] (3)
[0047] In the formula, Let be the transient trajectory mapping function determined by the transient dynamics model of the new energy power grid. Let T be the initial state of the new energy power grid to be analyzed (i.e., the initial state of the system), and let T be the sampling time series. Input variables into the system.
[0048] Preferably, step 2 includes:
[0049] Step 21: Construct a transient trajectory dataset:
[0050] (4)
[0051] In the formula, For transient trajectory datasets, This represents the number of transient trajectory samples. Let be the initial state of the i-th transient trajectory sample. Let be the input variable for the i-th transient trajectory sample. Let be the physical parameters of the i-th transient trajectory sample. Let i be the sampling time series of the i-th transient trajectory sample. Let be the true transient trajectory of the i-th transient trajectory sample.
[0052] Step 22: Construct a physical mechanism sample library:
[0053] (5)
[0054] In the formula, As a physical mechanism sample library, The number of physical samples. The system parameters corresponding to the k-th physical sample are... This represents the physical transient trajectory corresponding to the k-th physical sample. This represents the physical residual information corresponding to the k-th physical sample.
[0055] Step 23: Normalize the transient trajectory dataset and the physical mechanism sample library:
[0056] (6)
[0057] In the formula, The normalized transient trajectory, The mean of the transient trajectory samples. This represents the standard deviation of the transient trajectory samples.
[0058] The physical samples are normalized in the same way.
[0059] Preferably, step 3 includes:
[0060] Step 31: Establish the data-driven prediction branch. The data-driven prediction branch outputs transient trajectory prediction values based on the system's initial state, input variables, sampling time series, and operating condition characteristics.
[0061] (7)
[0062] In the formula, To predict the transient trajectory of the branch output based on data-driven forecasting. The parameters for the data-driven prediction branch are: The data-driven prediction network is denoted by z, which represents the operating condition feature or disturbance feature.
[0063] Step 32: Establish a physical mechanism calculation branch. The physical mechanism calculation branch includes one or more of the following: explicit physical solution unit, parameterized physical identification unit, and physical trajectory database lookup unit.
[0064] Among them, the parameterized physical identification unit estimates the equivalent physical parameters based on the transient trajectory output of the data-driven prediction branch:
[0065] (8)
[0066] In the formula, To identify the obtained equivalent physical parameters, Identify network parameters for physical parameters The physical parameter identification network.
[0067] The physical mechanism calculation branch generates the physical transient trajectory based on the equivalent physical parameters:
[0068] (9)
[0069] In the formula, The transient trajectory of the branch output is calculated based on the physical mechanism. For physical solution operators;
[0070] When using the physical trajectory database lookup method, the index of the best-matching physical parameter is calculated according to the following formula:
[0071] (10)
[0072] In the formula, The index of the variable that minimizes the objective function. This represents the equivalent physical parameters obtained from the identification of the object to be matched. The physical sample index with the smallest distance. This represents the 2-norm operator.
[0073] And the physical transient trajectory is obtained based on the index of the best-matching physical parameters:
[0074] (11)
[0075] In the formula, For physical sample indexing The corresponding transient trajectory.
[0076] Preferably, step 4 includes:
[0077] Step 41: Calculate the data loss function:
[0078] (12)
[0079] In the formula, For data loss function, This represents the number of training samples in a batch. For the i-th transient trajectory sample, the data-driven trajectory prediction is performed. Let be the true trajectory of the i-th transient trajectory sample.
[0080] Step 42: Calculate the physical residual loss function:
[0081] (13)
[0082] In the formula, The physical residual loss function, Let be the data-driven state prediction value of the i-th transient trajectory sample at time t. For the corresponding algebraic variable estimates, These are the estimated values of the corresponding physical parameters. The nonlinear dynamic function corresponding to the system's differential equation. For system algebraic constraint functions.
[0083] Step 43: Calculate the data-to-physical coupling loss function:
[0084] (14)
[0085] In the formula, This is a data-to-physics coupling loss function used to constrain data-driven prediction results to approximate the results calculated based on physical mechanisms. For the first i The physical predicted trajectory of each sample.
[0086] Step 44: Calculate the physical-to-data coupling loss function:
[0087] (15)
[0088] In the formula, The physical-to-data coupling loss function, These are physical features generated by the physical mechanism calculation branch.
[0089] Step 45: Calculate the total loss function:
[0090] (16)
[0091] In the formula, For the total loss function, , , , and These are the weight coefficients for the data loss function, physical residual loss function, data-to-physical coupling loss function, physical-to-data coupling loss function, and regularization loss function, respectively. This is the regularization loss function.
[0092] Preferably, step 5 includes:
[0093] Step 51: Fix the physical mechanism calculation branch and update the data-driven prediction branch parameters. :
[0094] (17)
[0095] In the formula, The learning rate for data-driven prediction branches. For parameter update operations.
[0096] Step 52: Fix the data-driven prediction branch and update the physical parameters to identify network parameters. Or adjustable physical parameters in the physical mechanism calculation branch:
[0097] (18)
[0098] In the formula, The learning rate is calculated for the branch of physical mechanism.
[0099] Step 53: Repeat steps 51 and 52 until the maximum number of training rounds is reached or the total loss function meets the convergence condition.
[0100] Preferably, step 6 includes:
[0101] Step 61: Input the initial state of the new energy power grid to be analyzed. Operating conditions (i.e., system input variables), sampling time series and disturbance information .
[0102] Step 62: Output the transient trajectory using the trained physics-data joint learning model:
[0103] (19)
[0104] In the formula, For the final transient trajectory output, The fusion weights for the data-driven prediction branch and the physical mechanism calculation branch.
[0105] Step 63: Calculate the transient stability evaluation index based on the transient trajectory.
[0106] The maximum power angle deviation, used as an evaluation indicator, is calculated according to the following formula:
[0107] (20)
[0108] In the formula, For the maximum power angle deviation, To predict the trajectory of the work angle, This is the steady-state work angle.
[0109] The mean square error of the trajectory, used as an evaluation indicator, is calculated according to the following formula:
[0110] (twenty one)
[0111] In the formula, MSE is the mean square error of the trajectory. for Predicted state at any given time for The actual state at time N is the total number of transient trajectory sampling points, and j represents the sequence number of the transient trajectory sampling point, j=1,2,…,N.
[0112] This invention also provides a physical data fusion-driven transient trajectory calculation device for new energy power grids, comprising:
[0113] The transient dynamics model construction module is used to construct transient dynamics models of new energy power grids and to uniformly represent synchronous generators, new energy converters, network constraints and load dynamics in the form of differential-algebraic equations.
[0114] The dataset building module is used to build transient trajectory datasets and physical mechanism sample libraries;
[0115] The data-driven prediction module is used to predict the transient trajectory of the system based on the initial state, time series, operating conditions and disturbance information of the system.
[0116] The physical mechanism calculation module is used to generate transient responses that meet physical constraints based on the transient dynamics model of the new energy power grid, the physical parameters to be identified, or the physical mechanism sample library.
[0117] The joint learning training module is used to construct the data loss function, physical residual loss function, data-to-physical coupling loss function, physical-to-data coupling loss function and total loss function, and to perform alternating training and bidirectional error feedback updates on the data-driven prediction module and the physical mechanism calculation module.
[0118] The transient trajectory output module is used to output the transient trajectory of the new energy power grid using the trained physical-data joint learning model;
[0119] The stability evaluation module is used to calculate transient stability evaluation indicators based on the output transient trajectory of the new energy power grid.
[0120] Example:
[0121] To verify the effectiveness of the physical-data joint learning method proposed in this invention, this embodiment takes the modified 39-node renewable energy power grid as the research object and uses system transient simulation data to perform transient trajectory calculation. The phase angle of the target unit is selected. and frequency state As the state variable to be predicted, the system state vector is represented as:
[0122] ;
[0123] This embodiment uses a parameter scanning method to construct a physical mechanism sample library. The equivalent physical parameter to be identified is denoted as p, with values ranging from 1.0 to 100.0 and an interval of 0.1, resulting in 991 sets of transient trajectories corresponding to different physical parameters. The simulation time range for each set of trajectories is 0–10 s, and each trajectory contains 951 time sampling points; therefore, the time interval between adjacent sampling points is approximately 0.010526 s. In the constructed data file, the phase trajectory matrix and frequency trajectory matrix both have a dimension of 951×991, and the physical parameter vector has a dimension of 991×1.
[0124] In this embodiment, the transient trajectory corresponding to the physical parameter p=1.0 is taken as the true trajectory of the target system, denoted as:
[0125] ;
[0126] To simulate errors in actual measurement devices and communication noise, Gaussian noise with a mean of zero was superimposed on the real trajectory. The standard deviation of the phase measurement noise was set to 0.05, and the standard deviation of the frequency measurement noise was set to 0.002. The noisy trajectory was used as training data for the data-driven prediction branch, while the un-noisy transient trajectory served as the benchmark trajectory for evaluating the model's computational accuracy.
[0127] In this embodiment, the data-driven prediction branch employs a temporal convolutional network. The network input includes time t and the initial phase angle. and initial frequency Therefore, the input dimension is 3; the network output consists of the predicted phase angle and frequency at the corresponding time point, with an output dimension of 2. The neural network contains 4 causal convolutional layers, with the number of channels in each layer being 32, 64, 64, and 64 respectively. The kernel size is set to 5, and the dropout rate is set to 0.1. Each convolutional layer uses dilated causal convolution to extract multi-timescale variation features of the transient trajectory.
[0128] The physical mechanism calculation branch employs a combination of physical trajectory database lookup and adjacent trajectory interpolation. The data-driven prediction branch first outputs the transient trajectory prediction results, and then the physical parameter identification unit obtains the identified equivalent physical parameters to be matched. Then, the equivalent physical parameters obtained from the identification to be matched are calculated. Parameters in the physical mechanism sample library The L2 distance between them is used to select the physical sample with the smallest distance. When the equivalent physical parameters obtained from the identification are used for matching... When located between two adjacent discrete parameters, linear interpolation is performed on the two adjacent physical trajectories based on the parameter distance to obtain the transient trajectory output by the physical mechanism calculation branch.
[0129] The data-driven prediction branch and the physical mechanism calculation branch are jointly optimized using an alternating training approach. The total number of training epochs is set to 40,000. In each training epoch, the data-driven prediction branch is first fixed, and the physical mechanism calculation branch is updated for 20 steps; subsequently, the physical mechanism calculation branch is fixed, and the data-driven prediction branch is updated for 20 steps. The learning rate for both the data-driven prediction branch and the physical mechanism calculation branch is set to 0.01.
[0130] The loss weights during the learning process are set as follows:
[0131] ;
[0132] in, Used to adjust the error between data-driven predicted trajectories and actual trajectories. Used to adjust the coupling error between data-driven predicted trajectories and physical mechanism trajectories. Used to adjust the error between the physical trajectory and the actual trajectory. Used to adjust the feedback constraint strength of the physical mechanism branch on the data-driven branch.
[0133] During training, the data-driven predicted trajectory, the trajectory calculated based on the physical mechanism, the values of various loss functions, and the results of physical parameter identification are saved every 100 rounds. All evaluation metrics are calculated after the predicted trajectory is denormalized to the original physical quantity space.
[0134] To verify the technical effects of this invention, Long Short-Term Memory (LSTM), Gated Recurrent Unit (GRU), Conventional Temporal Convolutional Network (TCN), Transformer Network, and Physically Guided Temporal Convolutional Network (PINN) were used as comparison methods. Specifically, the Physically Guided Temporal Convolutional Network used a fixed physical parameter p=1.1, while the other comparison methods used the same training and evaluation data as this invention.
[0135] Figure 2This section presents a comparison of the transient phase angle trajectories obtained using different methods with the actual phase angle trajectories. Figure 2 a) represents the phase angle prediction result of the Long Short-Term Memory (LSTM) network. Figure 2 b) represents the predicted phase angle of the gated cyclic unit (GRU). Figure 2 c) represents the phase angle prediction result of a conventional temporal convolutional network (TCN). Figure 2 d) represents the predicted phase angle of the Transformer network. Figure 2 e) represents the predicted phase angle of the Physically Guided Temporal Convolutional Network (PINN). Figure 2 f) represents the phase angle prediction result of the method proposed in this invention. Figure 2 It is evident that Long Short-Term Memory (LSTM) networks, gated recurrent units (GRUs), and ordinary temporal convolutional networks exhibit varying degrees of amplitude deviation near the transient oscillation peak. While physically guided temporal convolutional networks with fixed physical parameters can maintain the basic oscillation trend, systematic errors still exist between their predicted and actual trajectories due to deviations in the physical parameters. The phase angle trajectory obtained by the method of this invention maintains a high degree of consistency with the actual trajectory in terms of oscillation amplitude, oscillation frequency, and decay trend.
[0136] Figure 3 This section presents a comparison of the frequency transient trajectories obtained using different methods with the actual frequency trajectories. Figure 3 a) represents the predicted frequency of the Long Short-Term Memory (LSTM) network. Figure 3 b) represents the predicted frequency of the gated cyclic unit (GRU). Figure 3 c) represents the predicted frequency of a conventional temporal convolutional network (TCN). Figure 3 d) represents the frequency prediction result of the Transformer network. Figure 3 e) represents the predicted frequency of the Physically Guided Temporal Convolutional Network (PINN). Figure 3 f) represents the frequency prediction result of the method proposed in this invention. Figure 3 It is evident that the method of the present invention can accurately characterize the rapid frequency change, peak position, and attenuation process after a fault, and no significant cumulative error appears in the later stage of the transient process.
[0137] The above description is merely an embodiment of the present invention and does not limit the scope of the present invention. Any equivalent structural or procedural transformations made based on the content of the present invention specification and drawings, or direct or indirect applications in other related system fields, are similarly included within the protection scope of the present invention.
[0138] The contents not described in detail in this specification are existing technologies known to those skilled in the art.
Claims
1. A method for calculating transient trajectories of new energy power grids driven by physical data fusion, characterized in that, Includes the following steps: Step 1: Construct a transient dynamic model of the new energy power grid, and express synchronous generators, new energy converters, network constraints and load dynamics in a unified form of differential-algebraic equations; Step 2: Construct a transient trajectory dataset and a physical mechanism sample library; Step 3: Establish a physics-data joint learning architecture; the physics-data joint learning architecture includes a data-driven prediction branch and a physical mechanism calculation branch; the data-driven prediction branch is used to predict the transient trajectory of the system based on the initial state, time series, operating conditions and disturbance information of the system; the physical mechanism calculation branch is used to generate transient responses that meet physical constraints based on the transient dynamics model of the new energy power grid, the physical parameters to be identified or the physical trajectory library; Step 4: Construct the joint learning loss function; the joint learning loss function includes the data loss function, the physical residual loss function, the data-to-physical coupling loss function, the physical-to-data coupling loss function, and the total loss function; Step 5: Based on the joint learning loss function, alternately train and update the data-driven prediction branch and the physical mechanism calculation branch with bidirectional error feedback to obtain the trained physical-data joint learning model. Step 6: Calculate the transient trajectory of the new energy power grid under given initial state and operating conditions using the trained physical-data joint learning model, and output the transient stability evaluation index.
2. The method for calculating transient trajectories of a new energy power grid driven by physical data fusion according to claim 1, characterized in that, Step 1 includes: Step 11: Establish a differential-algebraic transient dynamic model for the new energy power grid; Step 12: Based on the sampled time series, represent the system transient trajectory as a trajectory matrix; Step 13: Construct parameterized transient mapping relationships for different operating conditions, disturbance scenarios, and physical parameters.
3. The method for calculating transient trajectories of new energy power grids driven by physical data fusion according to claim 1, characterized in that, Step 2 includes: normalizing the transient trajectory dataset and the physical mechanism sample library.
4. The method for calculating transient trajectories of a new energy power grid driven by physical data fusion according to claim 1, characterized in that, In step 3, the data-driven prediction branch outputs transient trajectory prediction values based on the system's initial state, input variables, sampling time series, and operating condition characteristics; the physical mechanism calculation branch includes one or more of the following: explicit physical solution unit, parameterized physical identification unit, and physical trajectory database lookup unit.
5. The method for calculating transient trajectories of a new energy power grid driven by physical data fusion according to claim 4, characterized in that, The parameterized physical identification unit estimates the equivalent physical parameters based on the transient trajectory output by the data-driven prediction branch; the physical mechanism calculation branch generates the physical transient trajectory based on the equivalent physical parameters.
6. The method for calculating transient trajectories of a new energy power grid driven by physical data fusion according to claim 4, characterized in that, When using the physical trajectory library lookup unit, the physical sample index that has the smallest distance to the equivalent physical parameters obtained from the identification to be matched is calculated, and the physical transient trajectory is obtained based on the physical sample index.
7. The method for calculating transient trajectories of new energy power grids driven by physical data fusion according to claim 1, characterized in that, Step 4 includes: calculating the data loss function, the physical residual loss function, the data-to-physical coupling loss function, the physical-to-data coupling loss function, and the total loss function.
8. The method for calculating transient trajectories of a new energy power grid driven by physical data fusion according to claim 1, characterized in that, Step 5 includes: Step 51: Fix the physical mechanism calculation branch and update the data-driven prediction branch parameters; Step 52: Fix the data-driven prediction branch and update the adjustable physical parameters in the physical parameter identification network parameter or physical mechanism calculation branch; Step 53: Repeat steps 51 and 52 until the maximum number of training rounds is reached or the total loss function meets the convergence condition.
9. The method for calculating transient trajectories of a new energy power grid driven by physical data fusion according to claim 1, characterized in that, Step 6 includes: Step 61: Input the initial state of the new energy power grid to be analyzed as the system initial state, operating conditions, sampling time series and disturbance information; Step 62: Output the transient trajectory using the trained physics-data joint learning model; Step 63: Calculate transient stability evaluation index based on transient trajectory.
10. A physical data fusion-driven transient trajectory calculation device for a new energy power grid, used to implement the physical data fusion-driven transient trajectory calculation method for a new energy power grid as described in any one of claims 1-9, characterized in that, include: The transient dynamics model construction module is used to construct transient dynamics models of new energy power grids and to uniformly represent synchronous generators, new energy converters, network constraints and load dynamics in the form of differential-algebraic equations. The dataset building module is used to build transient trajectory datasets and physical mechanism sample libraries; The data-driven prediction module is used to predict the transient trajectory of the system based on the initial state, time series, operating conditions and disturbance information of the system. The physical mechanism calculation module is used to generate transient responses that meet physical constraints based on the transient dynamics model of the new energy power grid, the physical parameters to be identified, or the physical mechanism sample library. The joint learning training module is used to construct the data loss function, physical residual loss function, data-to-physical coupling loss function, physical-to-data coupling loss function and total loss function, and to perform alternating training and bidirectional error feedback updates on the data-driven prediction module and the physical mechanism calculation module. The transient trajectory output module is used to output the transient trajectory of the new energy power grid using the trained physical-data joint learning model; The stability evaluation module is used to calculate transient stability evaluation indicators based on the output transient trajectory of the new energy power grid.