Regional adaptive protection setting method and system considering new energy access

Abstract

The application discloses a regional adaptive protection setting method and system considering new energy access, and relates to the technical field of relay protection of power systems.The method comprises the following steps: obtaining wide-area measurement data of a distribution network and identifying topological characteristics; mapping nonlinear transient quantities to a high-dimensional linear latent space by using a deep contrast Kupmann network to obtain linearized state variables; matching a Kupmann evolution matrix according to the topological characteristics and solving a linear algebraic equation to calculate a transient short-circuit current waveform; and calculating and issuing adaptive protection setting values based on the waveform.The application is used to solve the problems that the analysis of nonlinear transient current is inaccurate, the calculation is time-consuming, and the setting scheme is difficult to cover complex working conditions under topological mutation caused by new energy access, and realizes the rapid analysis and accurate setting of protection setting values.

Description

Technical Field

[0001] This invention relates to the field of power system relay protection technology, and more specifically, to a regional adaptive protection setting method and system that takes into account the access of new energy sources. Background Technology

[0002] In new distribution networks, the large-scale integration of inverter-type renewable energy sources alters the system's short-circuit transient characteristics. Traditional synchronous generator-dominated grids exhibit large short-circuit current amplitudes and stable attenuation characteristics during short-circuit faults. However, inverter-type distributed power sources, limited by the thermal capacity of power electronic devices and low-voltage ride-through control strategies, typically exhibit short-circuit transient currents with limited amplitude, phase shifts, and highly nonlinear distortion. This multi-source, bidirectional fault condition with weak feeder characteristics poses a severe challenge to the sensitivity and selectivity of distribution network relay protection settings.

[0003] Currently, for protection setting calculations in distribution networks containing new energy sources, mainstream solutions typically rely on time-domain simulation models or simplified equivalent analytical models of controlled current sources. The setting system collects steady-state operating data and node switch states of the system, matches them with an offline preset operating mode library, and then calls nonlinear differential equations to solve for fault characteristics step by step through integration, or directly uses simplified steady-state phasor equations to estimate short-circuit current indices. Finally, based on the obtained data, it performs setting calculations and issues the settings.

[0004] The aforementioned existing technologies have significant limitations in engineering practice. On the one hand, because the inverter control strategies of various new energy manufacturers are not publicly disclosed, traditional simplified analytical models cannot accurately characterize the nonlinear constraints in the initial transient phase; while nonlinear time-domain simulation is accompanied by a huge amount of iterative calculation time, which cannot meet the timeliness requirements of adaptive tuning for analytical speed. On the other hand, the distribution network topology is frequently reconfigured, and conventional nonlinear analysis models are prone to mathematical operator divergence or basis function failure when faced with topological changes. This makes it difficult for existing calculation schemes to maintain the stability of the evolution path under topological changes, and the calculated transient current has a large deviation. Static or conventional dynamic tuning schemes cannot cover complex grid operating conditions, which greatly increases the risk of protection failure and cascading tripping in scenarios with high penetration of new energy. Summary of the Invention

[0005] To overcome the aforementioned deficiencies of the prior art, embodiments of the present invention provide a regional adaptive protection setting method that considers the access of new energy sources. By utilizing a deep contrastive Koopman network to map nonlinear transient electrical quantities to a high-dimensional linear latent space, and combining the topological characteristics of the power grid operation to perform linear algebra analysis to rapidly obtain short-circuit waveforms, the protection settings are then dynamically calculated and issued. This addresses the technical problems of long transient current analysis time, easy divergence of traditional analytical models, and inability of static protection settings to adapt to dynamic operation modes in high-proportion new energy distribution networks under topology reconstruction and nonlinear distortion conditions.

[0006] To achieve the above objectives, the present invention provides the following technical solution: An adaptive protection setting method for a region considering renewable energy access includes the following steps: acquiring wide-area measurement data of a distribution network containing renewable energy inverters; identifying the operating topology of the distribution network based on the wide-area measurement data to obtain topological features; using an encoder in a pre-trained deep contrastive Koopman network to map the nonlinear transient electrical quantities in the wide-area measurement data to a high-dimensional linear latent space to obtain linearized state variables; matching a linear Koopman evolution matrix according to the topological features, solving linear algebraic equations in the high-dimensional linear latent space, and calculating the transient short-circuit current waveform of the renewable energy inverter based on the linearized state variables; and calculating adaptive protection settings based on the transient short-circuit current waveform and sending them to the protection equipment.

[0007] In addition, this invention also provides a regional adaptive protection setting system considering the access of new energy sources, comprising the following modules: a topology feature identification module, used to acquire wide-area measurement data of a distribution network containing new energy inverter power sources, identify the operating topology of the distribution network based on the wide-area measurement data, and obtain topology features; a latent space linearization mapping module, used to map nonlinear transient electrical quantities in the wide-area measurement data to a high-dimensional linear latent space using an encoder in a pre-trained deep contrastive Koopman network, and obtain linearized state variables; a transient waveform analytical calculation module, used to match a linear Koopman evolution matrix according to the topology features, solve linear algebraic equations in the high-dimensional linear latent space, and calculate the transient short-circuit current waveform of the new energy inverter power source in combination with the linearized state variables; and a setting value distribution module, used to calculate adaptive protection settings based on the transient short-circuit current waveform and distribute them to the protection equipment.

[0008] The technical effects and advantages of the regional adaptive protection setting method for new energy access in this invention are as follows: This invention identifies distribution network topology features in real time by acquiring wide-area measurement data. Using a pre-trained deep contrastive Koopman network, it maps complex nonlinear transient electrical quantities affected by renewable energy inverters to a high-dimensional linear latent space. Then, based on the topology features, it matches the corresponding linear evolution matrix and directly solves the linear algebraic equations within this high-dimensional space. This mechanism cleverly transforms the traditional time-consuming and easily divergent nonlinear time-domain solution into pure linear algebraic operations, greatly improving the analytical speed and model accuracy of transient short-circuit current waveforms. Finally, based on these transient short-circuit current waveforms, it rapidly calculates and issues adaptive protection settings, effectively overcoming the setting difficulties caused by the nonlinear characteristics of weak renewable energy feedstocks and dynamic reconfiguration of the power grid topology. This achieves online, dynamic, and precise adaptive protection for the distribution network protection system. Attached Figure Description

[0009] Figure 1 This is a schematic diagram of the regional adaptive protection setting method considering new energy access provided in an embodiment of the present invention; Figure 2 A flowchart of offline training and online mapping of a deep contrastive Koopman network provided in an embodiment of the present invention; Figure 3 This is a three-dimensional visualization diagram of the evolution of high-dimensional latent space features provided in an embodiment of the present invention; Figure 4 This is a bar chart comparing computation time at different system scales provided in embodiments of the present invention; Figure 5 A schematic diagram showing the accuracy comparison between the multi-transient short-circuit current waveform and the traditional electromagnetic transient simulation waveform provided in the embodiments of the present invention; Figure 6 This is a schematic diagram of a three-dimensional frequency response pre-distortion compensation surface provided in an embodiment of the present invention; Figure 7 A schematic diagram of the actual waveform tracking error distribution on the primary side provided in an embodiment of the present invention; Figure 8 This is a block diagram of a regional adaptive protection setting system that takes into account the access of new energy sources, provided in an embodiment of the present invention. Detailed Implementation

[0010] 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, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those of ordinary skill in the art without creative effort are within the scope of protection of the present invention.

[0011] Example 1, Figure 1 The present invention provides a regional adaptive protection setting method considering new energy access, comprising the following steps: S1. Obtain wide-area measurement data of the distribution network including new energy inverter power sources, identify the operating topology of the distribution network based on the wide-area measurement data, and obtain topology characteristics.

[0012] In this embodiment, S1 includes: S101, Obtain the node switch status information and distributed power output data of the power distribution network; Furthermore, to obtain synchronous real-time data that meets the requirements of adaptive protection actions, the acquisition of measurement data relies on the coordination of the underlying hardware and communication links at the field. Specifically, the operating data of primary side equipment is collected in real time through intelligent distribution terminals (DTU / FTU) deployed at major nodes and branch boxes of the distribution network and wide-area measurement systems (PMUs) on the substation side. At the data transmission level, the IEC 61850 substation communication network and system standard is adopted, and the remote signaling data of the switch positions of each node is transmitted in high real-time through GOOSE (General Object-Oriented Substation Event) messages. This status data uses 0 or 1 Boolean values ​​to represent the opening and closing status of circuit breakers or load switches. At the same time, the three-phase voltage phasors, three-phase current phasors, and active / reactive power output telemetry data of the grid connection point (PCC) of the renewable energy inverter power supply (IBRs) are continuously collected through SV (Sample Value) messages or SCADA system. The aforementioned discrete switching quantities and continuous analog quantities are aggregated via a high-speed communication private network to the data access gateway of the regional protection and control center, serving as the basic input source for subsequent topology identification and evolution calculation algorithm modules.

[0013] S102, Generate a node admittance matrix based on the switch state information; In this embodiment, the process of converting discrete physical switching quantities into continuous mathematical matrices for control logic transformation is as follows: The static branch impedance parameter table of the entire distribution network is pre-stored in the control center's memory database. Upon receiving data pushed by the gateway, the control center's algorithm module dynamically adjusts the connectivity of the corresponding network graph branches based on the real-time switch state variables extracted in S101, thereby reconstructing the node admittance matrix of the distribution network online. For example, if the distribution network contains... Nodes, Nodes With nodes The static branch admittance between them is denoted as The switch state variable corresponding to this branch is denoted as , among which when When the switch is closed, the branch circuit is connected. The time indicates that the switch is open; at the same time, the node is recorded. Earth admittance The off-diagonal elements of the node admittance matrix. and diagonal elements The calculation formulas are as follows: (1) (2) In the formula, The first node admittance matrix after reconstruction Line number Column elements; These are the self-admittance elements of the reconstructed nodal admittance matrix; These are real-time remote signaling variables dynamically captured by the system; and These are the constant physical and electrical parameters retrieved by the system from the database. The execution of this sub-step accurately and with low latency maps the physical topology changes in the field distribution network caused by fault isolation or load transfer into a sparse matrix structure that can be processed by computer algorithms.

[0014] S103, the topological features are obtained by extracting features from the node admittance matrix using a graph convolutional neural network.

[0015] As a preferred implementation, in order to fully explore the spatiotemporal correlation between the physical structure and operating state of the distribution network, the algorithm module first converts the node admittance matrix reconstructed in S102 into an undirected graph adjacency matrix. The logical operation involves extracting the non-zero features of the off-diagonal elements of the nodal admittance matrix using a binarization mapping function, ignoring specific electrical impedance magnitudes, and retaining only the connectivity of the physical topology. For [the following]... The adjacency matrix of a distribution network with nodes. elements in The logic for determining the value is as follows: (3) In the formula, Representation Nodes With nodes There are physical branch connections between them. This indicates that there is no direct connection or that it is a self-loop.

[0016] Furthermore, execute the node feature matrix. Construction logic: Time-align the discrete-time electrical quantities acquired by S101 from the PMU or SCADA system. For any node in the distribution network... Extract the three-phase voltage amplitude at its grid connection point (PCC). Contributing merit and effort and no effort put in (If no distributed power source is connected to the node, the corresponding output term is padded with zeros), and these are combined to form the local feature vector of the node. Subsequently, the entire network The feature vectors of each node are stacked and concatenated in row-wise order according to their node numbers to construct a structure of size [size missing]. Dense node feature matrix .

[0017] After constructing the underlying graph data, a Graph Convolutional Neural Network (GCN) is used to extract local and global topological connectivity features by aggregating neighbor node information. The core tensor operation formula for the forward propagation of a single hidden layer in the GCN is as follows: (4) In the above formula, This represents the network adjacency matrix after introducing self-loops. It is the identity matrix; To and The corresponding degree matrix has diagonal elements that satisfy ; The graph convolutional network represents the first... The node feature matrix of the layer, for the initial input layer, satisfies ; For the first The trainable weight parameter matrix of a layer network; For non-linear activation functions, the ReLU function is used in this embodiment.

[0018] Finally, after undergoing multiple layers of graph convolution operations, the network is described by a global average pooling layer, which incorporates information on topology and electrical distribution. The 3D node feature map is compressed into a dense vector of fixed dimension, which yields the final topological feature vector. .

[0019] This step effectively solves the problem of delayed operational status perception faced by high-penetration new energy distribution networks during islanded operation or fault reconfiguration by combining high-frequency acquisition of physical terminal data with deep computation of graph neural networks. It reduces the dimension of discrete and complex non-Euclidean power grid topology relationships into continuous feature vectors, providing a high-precision structural benchmark for stable alignment of subsequent linearization mapping.

[0020] S2 utilizes a pre-trained deep contrastive Koppman network to map nonlinear transient electrical quantities in wide-area measurement data to a high-dimensional linear latent space, thereby obtaining linearized state variables.

[0021] In this embodiment, the deep contrastive Koopman network needs to be trained offline to achieve physical binding in space, such as... Figure 2 As shown, the specific execution and training steps of S2 include: S201, Obtain historical transient electrical quantity samples and corresponding historical topological features of the distribution network; Furthermore, addressing the engineering challenge of the highly "black box" nature of the internal control logic of renewable energy inverters (IBRs), which is often encapsulated by manufacturers and thus difficult for traditional analytical models to accurately describe their short-circuit transient characteristics, this step relies on a hardware platform to construct an offline training set. Specifically, the algorithm module utilizes a real-time digital simulation system (RTDS) or high-precision electromagnetic transient simulation software (such as PSCAD / EMTDC) to construct a hardware-in-the-loop simulation platform for the distribution network. Within this physical simulation platform, various topology change conditions are set in batches (e.g., critical line commissioning / decommissioning, tie switch state switching, ring main unit opening, etc.), and different types of short-circuit faults are injected at different spatial locations. Through simulation operation, massive amounts of multi-node three-phase voltage and current discrete-time series data are collected and constructed as input tensors. The dimension of this historical transient electrical quantity sample tensor is defined as... (in This refers to the batch size for network training. The sampling time step span for a single transient waveform record. (This refers to the number of electrical characteristic channels, such as three-phase voltage and current). While acquiring the above waveforms, the algorithm module synchronously generates and records the corresponding historical topological feature vectors under this simulation condition, according to the graph convolutional network (GCN) feature extraction logic in step S1, as supervision labels for subsequent training.

[0022] S202, The historical transient electrical quantity samples are mapped into high-dimensional features through the encoder of the deep contrastive Koopman network; In this embodiment, to address the difficulty in solving nonlinear dynamic systems, an encoder is introduced to perform nonlinear upscalar computation. The encoder's network structure consists of alternating stacks of a one-dimensional residual convolutional network (1D-ResNet) and a multilayer perceptron (MLP). The original, low-dimensional transient electrical input vector, exhibiting highly nonlinear characteristics, is projected into a high-dimensional latent space with extremely high degrees of freedom by passing through multiple nonlinear activation layers (such as GELU or ReLU layers) of the encoder. The formula for the encoder's forward mapping is as follows: (5) In the formula, This represents the original low-dimensional nonlinear transient electrical input extracted from the batch tensor, with a dimension of... ; The high-dimensional feature vector representing the network output (i.e., the state variables within a high-dimensional linear latent space) has a dimension of . ; The parameter is The nonlinear mapping function of the deep neural network encoder. This mapping process requires... This ensures that the high-dimensional space has sufficient analytical degrees of freedom, enabling the distortion and nonlinear dynamic trajectory caused by the inverter's low-voltage ride-through current limiting characteristics in the low-dimensional space to be completely "flattened" into a linear evolution trajectory that can be expressed by a unified matrix.

[0023] S203, calculate the first feature similarity of transient electrical quantity samples with the same historical topological characteristics in the high-dimensional linear latent space; calculate the second feature similarity of transient electrical quantity samples with different historical topological characteristics in the high-dimensional linear latent space; As a preferred implementation, to establish the correlation constraints between waveforms and topology in the latent space, the algorithm module dynamically constructs batch sample pairs in memory. Based on the tag information synchronously acquired in S201, if two historical transient electrical quantity samples belong to the same physical power grid topology, they are classified as positive sample pairs; conversely, if they come from different operating topologies, they are classified as negative sample pairs. Subsequently, the algorithm module quantifies the spatial distance between state variables by calculating the cosine similarity of the upgraded high-dimensional feature vectors in the latent space. The formula for calculating the feature similarity is as follows: (6) In the formula, For any two eigenvectors in a high-dimensional space and Spatial cosine similarity; This represents the L2 norm of the vector. The output range of this function is [-1, 1]. The closer the calculated value is to 1, the smaller the geometric angle between the two transient eigenvectors in the high-dimensional latent space, and the closer their spatial distance. In other words, the time evolution of the short-circuit current waveforms represented by the two vectors in the latent space is highly consistent.

[0024] S204, construct the contrast loss function based on the first feature similarity and the second feature similarity, so that transient evolution features corresponding to the same running topology are clustered, and transient evolution features corresponding to different running topologies are mutually exclusive; In this embodiment, to address the fundamental challenges faced by the Koopman operator in traditional mathematics when encountering abrupt changes in the physical topology of a distribution network, such as mismatched basis function dictionaries and operator divergence, the algorithm module introduces a supervised comparison loss mechanism. The underlying mathematical logic of this mechanism is expressed as the following loss function formula with a temperature coefficient: (7) In the formula, This is the set of indices for all electrical quantity samples within the current calculation batch; For anchor point samples Positive samples with the same topological feature labels Indexed collection, Let be the cardinality of the set; To include all positive and negative samples within the same batch The complete set of indexes; This is a scalar temperature hyperparameter used to control the smoothness of the spatial probability distribution. From a physical verification perspective, this loss function is essentially a gradient penalty mechanism: it forces the neural network to adjust its weights through backpropagation. This allows all short-circuit transient waveforms under the same distribution network topology to have their characteristic states in the high-dimensional latent space "pulled closer (clustered)" to each other, regardless of where the actual fault point is located on the feeder or the size of the transition resistance. Conversely, once the power grid topology is reconfigured, the characteristic states corresponding to different topological features are strongly "pushed apart (mutually exclusive)" in space.

[0025] like Figure 3 The image shows a 3D visualization of the high-dimensional latent space feature evolution obtained after comparing the Koopman network mapping. Principal Component Analysis (PCA) dimensionality reduction algorithm is used to reduce the high-dimensional state variables... The manifold is visualized in a 3D space (PC1-PC2-PC3). Trajectory bundles of different colors and spatial locations represent short-circuit transient characteristic time series under different operating topologies. It can be seen that after introducing a contrast loss mechanism with a temperature coefficient, under the same operating topology (e.g., the red trajectory bundle region), even with different fault locations or transition resistances, the transient evolution paths are highly clustered in the latent space, converging into a compact linear subspace. However, when the power grid undergoes topology reconfiguration due to circuit breaker opening (e.g., a jump to the blue and green trajectory bundle regions), the corresponding characteristic bundles are significantly pushed apart in the 3D latent space, exhibiting strong mutual exclusivity and orthogonal separability at their boundaries. This 3D manifold data distribution intuitively verifies that this method successfully overcomes the basis function divergence problem of the traditional Koopman operator during topological abrupt changes, achieving deep alignment between the physical structure and linear evolution characteristics.

[0026] S205, with the goal of constraining the spatial alignment between the high-dimensional features and the corresponding historical topological features, a total loss function including a contrastive loss function is constructed, and the parameters of the deep contrastive Koopman network are updated based on the total loss function.

[0027] Furthermore, construct a system that includes the aforementioned contrast loss. Reconstruction loss and linear evolution prediction loss The total loss function. The reconstruction loss is used to constrain the decoder network. The formula for the accuracy of restoring the original transient waveform is: (8) Linear evolutionary prediction loss is used to constrain state variables to follow a linear Koopman evolution matrix in a high-dimensional latent space. The dynamic law of is expressed by the following formula: (9) After clarifying the individual losses mentioned above, they are linearly combined to generate the total loss function, as shown in the following formula: (10) In the formula, Let this be the total loss function for network training; The input samples are the original low-dimensional nonlinear transient electrical quantities. These are high-dimensional state variables; For parameters decoder function; and High-dimensional features are respectively and The state variable at any given time; The linear Koopman evolution matrix; and Hyperparameters are used to balance the weights of each loss term.

[0028] During the offline training phase, the AdamW optimizer is invoked, based on... The gradient is calculated, and the network weight parameters are backpropagated and iteratively updated until the loss function converges. After offline training, the network parameters are fixed, and the nonlinear transient electrical quantities in the wide-area measurement data acquired online are directly input into this deep contrastive Koopman network, which can then be mapped to obtain linearized state variables in a high-dimensional linear latent space.

[0029] This step introduces a contrastive learning mechanism into the dynamic analysis of relay protection, solving the operator divergence problem that occurs in traditional analytical models during distribution network topology reconfiguration. By establishing spatial correspondence constraints between the physical topology and transient evolution paths in the loss function, the high-dimensional state variables possess stable separability, providing a reliable characteristic foundation for the subsequent linear algebraic solution of the short-circuit current waveform.

[0030] S3. Based on the topological features, a linear Koopman evolution matrix is ​​matched, and the linear algebraic equation is solved in the high-dimensional linear latent space. Combined with the linearized state variables, the transient short-circuit current waveform of the new energy inverter power supply is calculated.

[0031] In this embodiment, S3 includes: S301, In the preset topology-matrix mapping library, query the target linear Koopman evolution matrix corresponding to the topological feature; Furthermore, a topology-matrix mapping table containing multiple sets of typical network operating structures is pre-constructed in the static storage area of ​​the algorithm module. After the system receives real-time data, the execution logic will use the topology feature vector generated in step S1. The index is entered into the mapping table as the retrieval key. By calling the nearest neighbor search algorithm (KNN) or calculating the cosine distance metric between high-dimensional features, the target index with the highest spatial similarity is quickly matched within the mapping database, and the target linear Koopman evolution matrix corresponding to the real-time topological state is retrieved from there. The objective is a linear Koopman evolution matrix. The matrix elements have been trained and solidified in the previous offline stage based on contrastive learning. In a physical and mathematical sense, they represent the global linear dynamic evolution operator of the system state variables in the high-dimensional latent space under the current specific distribution network topology.

[0032] S302, using the target linear Koopman evolution matrix, construct a first-order linear differential equation for the linearized state variables; In this embodiment, the initial time-linearized state variables obtained by the encoder in step S2 are linearized. As the initial mathematical value for system evolution, based on the Koopman operator theory, the evolution of complex power systems, including nonlinear characteristics such as inverter dead time and limiting elements, is rigorously transformed into the time-domain motion of a purely linear system in a high-dimensional latent space. Based on this theory, the algorithm module uses the matched target linear Koopman evolution matrix to construct a first-order linear ordinary differential equation (ODE) for the state variables in this latent space, the expression of which is as follows: (11) In the formula, express State variables that are always in a high-dimensional linear latent space; This is the target linear Koopman evolution matrix retrieved in step S301. The construction of this step simplifies the complex physical nonlinear partial differential equations that were previously required in the time-domain simulation of power distribution networks into an initial value solution logic that can be handled using standard linear algebra tools.

[0033] S303, Solve the matrix exponential function corresponding to the first-order linear differential equation to obtain the time evolution trajectory of the linearized state variable in the high-dimensional linear latent space; As a preferred implementation, to avoid the computational delay caused by iterative integration in traditional electromagnetic transient nonlinear simulations, the execution logic directly obtains the analytical solution of the aforementioned first-order ordinary differential equation using matrix exponential operations in the processor. The analytical expression of the time evolution trajectory of the linearized state variables in the high-dimensional linear latent space is as follows: (12) To implement this analytical formula in digital real-time tuning software, the algorithm module employs Taylor series expansion or Padé approximation algorithm for the matrix exponent in the continuous time domain. Perform fast discretization approximation calculations. This is achieved by setting the digital simulation step size at the software layer. (e.g., preset) This allows us to calculate the linearized state sequence matrix within the future observation time window after the fault occurs. This sequence matrix constitutes the complete temporal evolution trajectory of the linearized state variables in the high-dimensional linear latent space. The aforementioned purely algebraic analytical process based on matrix exponential approximation significantly reduces the time complexity of equation solving compared to conventional nonlinear numerical simulations.

[0034] like Figure 4 As shown, traditional numerical simulation methods, when faced with an expanding distribution network, require solving a massive nonlinear Jacobian matrix, resulting in an exponential increase in computation time. With 118 nodes, the computation time approaches 6 seconds, completely failing to meet the millisecond-level online setting requirements of relay protection. In contrast, this invention pre-defines the nonlinear topological characteristics as a target linear evolution matrix. Online computation only requires algebraic multiplication and addition operations on the constant coefficients of the matrix, and its time consumption only shows a slow linear increase with the node size. At the scale of hundreds of nodes, the computation time of this method is still strictly controlled within 20 milliseconds, which proves the ability of the method in this embodiment to support the online real-time tuning of adaptive protection in terms of timeliness.

[0035] S304, the time evolution trajectory is inversely transformed by the decoder of the deep contrast Koopman network to obtain the transient short-circuit current waveform; Furthermore, after calculating the temporal evolution trajectory of the linearized state variable in the high-dimensional linear latent space, the sequence matrix corresponding to the temporal evolution trajectory calculated in S303 is... As input, the data is passed to the decoder module of the deep contrastive Koopman network. This decoder consists of multiple linear layers stacked with nonlinear activation functions. Through an inverse nonlinear mapping algorithm, it maps the high-dimensional abstract state features point-by-point to electrical waveform data in the original physical coordinate system. The mathematical formula corresponding to this inverse transformation is as follows: (13) In the formula, This is the instantaneous value vector of transient short-circuit current at the grid connection point of the new energy inverter power supply obtained by inverse transformation of the time evolution trajectory through the decoder. This is a decoder function with fixed parameters. The restored transient short-circuit current waveform accurately reproduces the impact of the internal control strategy of the new energy inverter in the early stage of a fault, so that the waveform signal contains the true current-limiting chamfer distortion and amplitude-limiting characteristics.

[0036] like Figure 5 The figure shows a comparison of the accuracy of the transient short-circuit current waveform calculated by the method of this embodiment with that of the traditional electromagnetic transient simulation waveform. The horizontal axis represents time (ms), and the vertical axis represents the normalized current amplitude (pu). The black solid line represents the real reference waveform acquired through RTDS physical simulation; the red dashed line represents the reconstructed waveform output by the inverse transformation after linear analysis in high-dimensional latent space according to this invention; and the blue dotted line represents the waveform calculated by the traditional equivalent controlled current source model that ignores nonlinear characteristics. The figure demonstrates the data reconstruction capability of this method: the red dashed line accurately follows the "flat-top chamfer" distortion at the peak of the reference waveform caused by the inverter's low-voltage ride-through (LVRT) phase-locked loop offset and limiting logic, while the traditional model (blue dotted line), lacking nonlinear mapping capability, still exhibits standard sinusoidal attenuation, resulting in an amplitude prediction error as high as 18% at the peak. This experimental result verifies that this invention maintains the extremely fast solution of algebraic equations without losing any key high-frequency and nonlinear short-circuit characteristics.

[0037] This step replaces the time-consuming nonlinear time-domain simulation in the conventional setting process by performing linear algebraic analytical operations in a high-dimensional latent space. Given the known characteristics of the power grid topology, this method enables rapid numerical prediction of transient fault characteristics in new energy sources, providing a physically reliable waveform data basis for the instantaneous updating of adaptive protection settings.

[0038] S4 calculates the adaptive protection setting based on the transient short-circuit current waveform and sends it to the protection device.

[0039] In this embodiment, S4 includes: S401, extract the fault characteristic quantities in the transient short-circuit current waveform, the fault characteristic quantities include the limited current amplitude and phase deflection angle affected by the new energy inverter power supply control strategy; Furthermore, regarding the discrete-time series output by step S3... The algorithm module performs a full-wave discrete Fourier transform (DFT) on the time-series waveform within a preset short data window (e.g., half a power frequency cycle or one complete power frequency cycle) to achieve feature analysis from the time domain to the frequency domain. The formula for feature analysis is as follows: (14) (15) In the formula, The total number of discrete sampling points within a single power frequency cycle; The data sampling time interval; For the first The instantaneous value of the short-circuit current at each sampling moment; The complex phasor of the fault fundamental current is obtained through discrete transformation; The extracted restricted current amplitude; This is the extracted phase deflection angle.

[0040] In practical power distribution network applications, traditional synchronous generators can provide short-circuit currents up to 5 to 10 times their rated current in the initial stage of a short circuit; however, due to the thermal capacity limitations of power electronic devices, the output current of renewable energy inverters (IBRs) is limited. Typically, the current is hardware-limited to between 1.1 and 1.5 times the rated current. Simultaneously, during a deep voltage dip in the grid, the inverter triggers low-voltage ride-through (LVRT) control logic and forcibly injects reactive current, leading to… A significant nonlinear shift occurs. This step involves precise calculation. and The weak feed characteristics and phase distortion of the inverter are established as the physical boundary conditions for subsequent adaptive tuning calculations.

[0041] S402, Calculate the optimal protection level difference that satisfies the sensitivity and selectivity requirements based on the fault characteristic quantity and the topology characteristics; In this embodiment, the step difference of traditional relay protection is usually a fixed empirical value (such as 0.2s or 0.3s). However, in distribution networks with high penetration of new energy sources, fault current is limited and injected from multiple sources in reverse, making a fixed step difference prone to selective coordination failure. To address this, the algorithm module establishes a dynamic adaptive optimal step difference calculation logic.

[0042] First, the algorithm module processes the obtained topological feature vectors. Analysis: Based on the preset topology-matrix mapping library in the algorithm module (shared with step S301), using... To retrieve the physical directed graph adjacency list corresponding to the current operating state, the search key is used to match the current operating state. Then, using the grid-connected substation nodes and new energy inverter nodes of the distribution network as root nodes, a breadth-first search (BFS) algorithm is used to traverse the directed graph adjacency list, extracting all node pairs with upstream and downstream power flow adjacency relationships. This maps and generates the upstream and downstream node cooperation constraint set for the current operating state. Elements in the set This indicates that under the current topology, the node The upstream main protection node is located near the power supply side. It is its corresponding downstream neighboring node.

[0043] Subsequently, for each pair of cooperating nodes in the constraint set, the algorithm module calculates the optimal protection level difference by comprehensively considering the physical hardware delay, the error margin of the weak power supply characteristics of new energy sources, and the complexity of the network structure. The specific calculation formula is as follows: (16) In the formula, For the calculated nodes With nodes The optimal protection level difference between them; It is the sum of the physical constants of the inherent opening time of the circuit breaker and the operation time of the protection device; This is the rated load current of the feeder; The limited short-circuit current amplitude extracted from S401. This is the current limiting compensation coefficient. The physical meaning is: when the short-circuit current The smaller the value (i.e., the more severe the current restriction on new energy sources and the weaker the fault characteristics), the higher the risk of measurement error amplification, and the system automatically increases the time margin. The predefined topological complexity penalty weight vector. The term represents a dynamically added safety margin compensation based on the complexity of the ring network structure or the number of branches in the power grid topology.

[0044] After calculating the optimal protection level difference, the setting calculation module further constructs a set of mathematical constraint equations to solve the protection start-up setting and action time, in order to meet the dual engineering requirements of sensitivity and selectivity. The set of equations is as follows: (17) (18) (19) In the formula, The setting value for the starting current of this level of overcurrent protection is to be determined. This represents the real-time sensitivity coefficient. This is the minimum sensitivity threshold required by power industry regulations; and Upstream nodes With downstream nodes The protection action time. This is determined by solving the above solution, which includes topological constraints. With dynamic level difference The system of equations and the algorithm realize the calculation of the optimal combination of constant values ​​under the current topology and new energy output state.

[0045] S403, Generate the adaptive protection setting based on the optimal protection level difference.

[0046] Furthermore, after completing the differential coordination calculation, the execution logic of the setting control system, based on the calculation results of each node obtained in step S402, finally summarizes and generates the adaptive protection setting. It should be clarified that the adaptive protection setting is mainly determined by the starting current threshold directly involved in the device tripping judgment. And the action time of this level of protection calculated based on the optimal protection level cascade. composition.

[0047] After generating the settings, they are packaged and converted into setting files conforming to the Substation Configuration Description Language (SCL) standard or standard XML format. During the data transfer and physical distribution phase, the system remotely distributes the updated setting files to the corresponding topology nodes via the dedicated power communication data network using the IEC 61850 Manufacturing Message Specification (MMS) protocol. The setting cache register of the intelligent electronic device (IED), distribution automation terminal unit (DTU / FTU), or microprocessor-based protection device is then used. Subsequently, the control center issues a "setting zone switching" or "online refresh" control command to the protection device. Upon receiving the command, the protection device replaces the old setting value in its logic judgment circuit with the newly generated starting current threshold. With protective action time This allows the adaptive protection setting to take effect immediately during the operation protection of the primary equipment.

[0048] This step integrates the microscopic transient waveform analysis results with the macroscopic relay protection setting principles. By dynamically calculating and remotely issuing settings online, it avoids the risk of protection failure that is prone to occur in scenarios with low short-circuit capacity of new energy sources under traditional fixed settings, and builds a dynamic defense mechanism for high-proportion new energy distribution networks.

[0049] S5, after sending the adaptive protection setting to the protection device, further includes: generating a test waveform signal based on the transient short-circuit current waveform; controlling the primary and secondary complete set of testing equipment to inject a large current corresponding to the test waveform signal into the primary side current transformer circuit of the protection device; monitoring the operation behavior of the protection device under the injected large current to verify the adaptive protection setting.

[0050] In this embodiment, S5 includes: S501, extract the non-standard sinusoidal waveform features from the transient short-circuit current waveform, the non-standard sinusoidal waveform features include transient DC bias, low-frequency harmonic components and current limiting chamfer; Furthermore, to achieve accurate separation of complex electromagnetic transient signals, the digital signal processing module processes the discrete transient waveform sequence obtained through inverse transformation in step S3. The system extracts real physical characteristics. Specifically, the control logic uses a low-pass filter (LPF) and a fast Fourier transform (FFT) to extract the DC attenuation term and the 2nd to 7th low-frequency harmonic components from the transient waveform. Simultaneously, by calculating the first-order difference (i.e., waveform derivative) of the time-domain waveform, it detects the flat-top phenomenon (i.e., current-limiting chamfer) in the current waveform caused by the closed-loop control of the renewable energy inverter reaching saturation. After the above feature separation, a comprehensive digital test waveform containing the real distortion characteristics of the renewable energy system is constructed, and its mathematical expression is as follows: (20) In the formula, The reference signal for the constructed digital test waveform; and These represent the initial amplitude and decay time constant of the transient DC bias component, respectively. For the first The amplitude of the second harmonic component, The initial phase angle corresponding to the harmonic. For power frequency angular velocity, The highest harmonic order for analysis; This is a nonlinear piecewise compensation function characterizing the inverter's saturation current-limiting chamfer characteristics. The execution of this sub-step accurately reproduces, at the digital level, the complex electrical characteristics of the new energy power supply's output under real short-circuit conditions—characterized by non-sinusoidal and amplitude-limited behavior.

[0051] S502, based on the power amplifier response characteristics of the primary and secondary complete set of testing equipment, the non-standard sinusoidal waveform characteristics and the power frequency fundamental wave are digital-to-analog conversion and power amplification are performed to synthesize the test waveform signal; As a preferred implementation, to convert digital domain signals into high-power electromagnetic excitation in the physical world without distortion, the execution logic employs a hardware-based pre-distortion compensation technique. When the digital test waveform... Before being fed into the digital-to-analog converter (DAC), the control logic performs digital pre-distortion processing on the signal by combining the frequency response characteristic curves of the power amplifier inside the primary and secondary integrated testing equipment. Because the leakage inductance of the current transformer and the capacitive load of the power amplifier can cause severe signal distortion and phase shift when a large current of 500A to 1000A is injected into the primary side of the physical circuit, the algorithm module incorporates the inverse model of the power amplifier's transfer function into the control link. The control domain formula for the pre-distortion compensation is as follows: (twenty one) In the formula, Laplace transform of the reference signal for the digital test waveform; To detect the inverse transfer function of the power amplifier inside the testing equipment; This represents the inverse Laplace transform operation; This is the digital control voltage signal used to drive the DAC and after compensation. By combining this inverse model compensation algorithm with high-power linear / switching power amplifier hardware, the system can faithfully reproduce the real high-current test signal, including current-limiting chamfer and transient DC bias, on the primary physical circuit.

[0052] like Figure 6 The figure shows a schematic diagram of the three-dimensional frequency response pre-distortion compensation surface of a primary and secondary complete testing equipment when injecting a non-standard large current. This figure illustrates the control module's response to different frequency components (…). (shaft, 0-1000Hz) and different load impedances of the primary side current transformer ( Axis, 0-5 The dynamically generated inverse model gain compensation surface ( (shaft, unit dB). For example... Figure 7 The figure shows a schematic diagram of the measured waveform tracking error distribution on the primary side. This figure illustrates the total tracking error (THDe) of the current waveform before and after compensation under an 800A peak current injection. It can be seen that without applying... During pre-distortion compensation (gray scatter group), due to the superposition of the power amplifier capacitive load effect and the DC bias saturation of the transformer, the high-order harmonics and chamfers of the waveform are severely distorted, with the maximum tracking error exceeding 15%. However, after using the three-dimensional pre-distortion inverse model compensation of this embodiment (green solid scatter group), the test waveform errors in all frequency bands are converged and suppressed to within the tolerance band of 2.5%.

[0053] S503, determine whether the actual action delay and tripping logic of the protection device are consistent with the expected action target of the adaptive protection setting; In this embodiment, the high-current injection phase relies on hardware-in-the-loop timing alignment logic to achieve closed-loop verification. The high-frequency current output terminal of the primary and secondary integrated testing equipment is clamped and directly connected to the through-bus of the primary side current transformer (CT) of the protection equipment, directly injecting the synthesized high current into the primary system. Simultaneously, the high-speed switching input contacts of the testing equipment are hardwired to the corresponding node's intelligent distribution terminal unit (DTU / FTU) or the trip output (BO) relay of the microprocessor-based protection device. During the test, a high-precision microsecond-level timer inside the testing equipment records the start time of the high-current injection. Simultaneously capture the actual moment of physical closure of the trip contact. The system verifies the expected target based on the following time deviation determination inequality: (twenty two) In the formula, For the actual moment of action captured; This is the start time of current injection; The algorithm in step S4 is calculated and distributed to this node. The set value for the action time of this level of protection; To account for the allowable time tolerance boundary (typically set to 20ms to 30ms) of the transmission error of the primary transformer and the inherent dispersion of the circuit breaker mechanism, this step effectively verifies the actual response performance of the protection hardware device when faced with non-standard characteristic currents from new energy sources and transient saturation of the transformer.

[0054] S504 If there is a discrepancy, a setting correction feedback signal is generated, and the adaptive protection setting of the distribution network is updated according to the feedback signal.

[0055] Furthermore, when the system determines that the actual action delay exceeds the tolerance range... When this manifests as protection failure to operate or severely delayed operation, it indicates that the transient transmission characteristics of the primary-side current transformers in this batch are poor, or that the Fourier filter algorithm inside the protection device cannot effectively identify the distorted current-limiting chamfer waveform. At this time, the control logic immediately extracts the time error signal and generates a setting correction feedback command, which is then transmitted to the adaptive protection calculation module. After receiving the command, the calculation module executes the setting margin compensation update formula: (twenty three) In the formula, The updated sensitivity verification coefficient threshold; The threshold value for the initial sensitivity verification coefficient used in the previous round of calculations; The learning step size is adaptively adjusted by the algorithm; The penalty function is constructed based on the measured time deviation, wherein the time deviation According to the above formula, the system re-triggers the S4 setting calculation module by adjusting the sensitivity threshold parameter of the setting verification, thereby generating a new adaptive protection setting with greater safety tolerance, and completing the iterative reconstruction from "calculation and distribution" to "physical verification" and then to "feedback correction".

[0056] This step integrates the logical deduction of the software algorithm with the actual physical response of the primary transformer and the secondary transmission circuit, forming a feedback control closed loop, which effectively ensures the reliability and effectiveness of the adaptive protection action of the distribution network area under the high proportion of new energy access.

[0057] Example 2, Figure 8 A regional adaptive protection setting system considering the access of new energy sources is presented, including: The topology feature recognition module is used to acquire wide-area measurement data of the distribution network that includes new energy inverter power sources, and to identify the operating topology of the distribution network based on the wide-area measurement data to obtain topology features. The latent space linearization mapping module is used to map nonlinear transient electrical quantities in wide-area measurement data to a high-dimensional linear latent space using an encoder in a pre-trained deep contrastive Koppman network, thereby obtaining linearized state variables. The transient waveform analysis and calculation module is used to match the linear Koopman evolution matrix according to the topological features, solve the linear algebraic equation in the high-dimensional linear latent space, and calculate the transient short-circuit current waveform of the new energy inverter power supply by combining the linearized state variables. The setting and distribution module is used to calculate adaptive protection settings based on transient short-circuit current waveforms and distribute them to protection devices.

[0058] The above formulas are all dimensionless calculations. The formulas are derived from software simulations based on a large amount of collected data to obtain the most recent real-world results. The preset parameters in the formulas are set by those skilled in the art according to the actual situation.

[0059] The above embodiments can be implemented, in whole or in part, by software, hardware, firmware, or any other combination thereof. When implemented using software, the above embodiments can be implemented, in whole or in part, in the form of a computer program product.

[0060] Those skilled in the art will recognize that the modules and algorithm steps of the various examples described in conjunction with the embodiments disclosed herein can be implemented in electronic hardware, or a combination of computer software and electronic hardware. Whether these functions are implemented in hardware or software depends on the specific application and design constraints of the technical solution. Those skilled in the art can use different methods to implement the described functions for each specific application, but such implementation should not be considered beyond the scope of this application.

[0061] In addition, the functional modules in the various embodiments of this application can be integrated into one processing module, or each module can exist physically separately, or two or more modules can be integrated into one module.

[0062] The above description is merely a specific embodiment of this application, but the scope of protection of this application is not limited thereto. Any variations or substitutions that can be easily conceived by those skilled in the art within the scope of the technology disclosed in this application should be included within the scope of protection of this application. Therefore, the scope of protection of this application should be determined by the scope of the claims.

[0063] In conclusion, the above description is only a preferred embodiment of the present invention and is not intended to limit the present invention. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the protection scope of the present invention.

Claims

1. A regional adaptive protection setting method considering the access of new energy sources, characterized in that, include: Acquire wide-area measurement data of the distribution network that includes new energy inverter power supply, identify the operating topology of the distribution network based on the wide-area measurement data, and obtain topology characteristics; By using the encoder in a pre-trained deep contrastive Koppmann network, nonlinear transient electrical quantities in wide-area measurement data are mapped to a high-dimensional linear latent space to obtain linearized state variables. Based on the topological features, a linear Koopman evolution matrix is ​​matched, and combined with linearized state variables, linear algebraic equations are solved in a high-dimensional linear latent space to calculate the transient short-circuit current waveform of the new energy inverter power supply. Based on the transient short-circuit current waveform, the adaptive protection setting is calculated and sent to the protection device.

2. The method according to claim 1, characterized in that, The acquisition of the topological features includes: Obtain the node switch status information of the distribution network; Generate a node admittance matrix based on the switch state information; The topological features are obtained by extracting features from the node admittance matrix using a graph convolutional neural network.

3. The method according to claim 1, characterized in that, The pre-trained deep contrastive Koopman network is trained through the following steps: Obtain historical transient electrical quantity samples and corresponding historical topological features of the distribution network; The encoder of the deep contrastive Koopman network maps the historical transient electrical quantity samples into high-dimensional features. With the goal of constraining the spatial alignment between the high-dimensional features and the corresponding historical topological features, a total loss function including a contrastive loss function is constructed, and the parameters of the deep contrastive Koopman network are updated based on the total loss function.

4. The method according to claim 3, characterized in that, The construction of the total loss function, which includes a contrastive loss function, includes: Calculate the first feature similarity of transient electrical quantity samples with the same historical topological characteristics in the high-dimensional linear latent space; Calculate the second feature similarity of transient electrical quantity samples with different historical topological characteristics in the high-dimensional linear latent space; The contrastive loss function is constructed based on the first feature similarity and the second feature similarity, so that transient evolution features corresponding to the same running topology are clustered, and transient evolution features corresponding to different running topologies are mutually exclusive.

5. The method according to claim 1, characterized in that, The calculated transient short-circuit current waveform of the new energy inverter power supply includes: In the preset topology-matrix mapping library, query the target linear Koopman evolution matrix corresponding to the topological feature; Using the target linear Koopman evolution matrix, a first-order linear differential equation is constructed for the linearized state variables; Solving the matrix exponential function corresponding to the first-order linear differential equation yields the time evolution trajectory of the linearized state variable in the high-dimensional linear latent space; The transient short-circuit current waveform is obtained by inversely transforming the time evolution trajectory using the decoder of the deep contrastive Koppman network.

6. The method according to claim 1, characterized in that, The calculation of the adaptive protection setting includes: The fault characteristic quantities are extracted from the transient short-circuit current waveform, including the limited current amplitude and phase deflection angle affected by the new energy inverter power supply control strategy. Based on the fault characteristic quantities and the topological characteristics, calculate the optimal protection level difference that satisfies the sensitivity and selectivity requirements; The adaptive protection setting is generated based on the optimal protection level difference.

7. The method according to any one of claims 1 to 6, characterized in that, After the adaptive protection setting is sent to the protection device, the following is also included: A test waveform signal is generated based on the transient short-circuit current waveform; The control system injects a large current corresponding to the test waveform signal into the primary side current transformer circuit of the protection device using the primary and secondary complete set of testing equipment. The behavior of the protection device under a large current injection is monitored to verify the adaptive protection setting.

8. The method according to claim 7, characterized in that, The step of generating a test waveform signal based on the transient short-circuit current waveform includes: Extract the non-standard sinusoidal waveform features from the transient short-circuit current waveform. The non-standard sinusoidal waveform features include transient DC bias, low-frequency harmonic components, and current-limiting chamfer. Based on the power amplifier response characteristics of the primary and secondary complete testing equipment, the non-standard sine waveform characteristics and the power frequency fundamental wave are subjected to digital-to-analog conversion and power amplification to synthesize the test waveform signal.

9. The method according to claim 7, characterized in that, The monitoring of the protective device's behavior under the injected large current, in order to verify the adaptive protection setting, includes: Determine whether the actual action delay and tripping logic of the protection device are consistent with the expected action target of the adaptive protection setting; If there is a discrepancy, a setpoint correction feedback signal is generated, and the adaptive protection setpoint of the distribution network is updated based on the feedback signal.

10. A regional adaptive protection setting system considering the access of new energy sources, characterized in that, include: The topology feature recognition module is used to acquire wide-area measurement data of the distribution network that includes new energy inverter power sources, and to identify the operating topology of the distribution network based on the wide-area measurement data to obtain topology features. The latent space linearization mapping module is used to map nonlinear transient electrical quantities in wide-area measurement data to a high-dimensional linear latent space using an encoder in a pre-trained deep contrastive Koppman network, thereby obtaining linearized state variables. The transient waveform analysis and calculation module is used to match the linear Koopman evolution matrix based on topological features, solve the linear algebraic equation in the high-dimensional linear latent space, and calculate the transient short-circuit current waveform of the new energy inverter power supply by combining the linearized state variables. The setting and distribution module is used to calculate adaptive protection settings based on transient short-circuit current waveforms and distribute them to protection devices.

Images