A method for generating a space-time hidden variable topology simulation dataset for a flexible interconnected power distribution network

Abstract

The application discloses a kind of space-time hidden variable topological simulation dataset generation methods for flexible interconnected distribution network, comprising: constructing the double-layer topological structure including discrete switch layer and continuous modulation parameter layer, obtain the extended basic simulation model;Based on the extended basic simulation model, the probability double-track data generation mechanism of implicit node is established, the full observation topological model and the partial observation topological model are maintained synchronously, and the initialized double-track output structure is obtained;The initialized double-track output structure is input into the event scheduling engine with causal chain label, and the event direct layer update, power flow calculation layer update and feature evolution layer update are sequentially executed at each time step, to obtain node electrical characteristics, topological state and causal label array;The above data is aligned and packaged according to time step index, and the simulation dataset is output.The application provides high-quality simulation data basis for power distribution network topology identification and other artificial intelligence tasks.

Description

Technical Field

[0001] This invention relates to the field of power system distribution network simulation technology, and in particular to a method for generating spatiotemporal latent variable topology simulation datasets for flexible interconnected distribution networks. Background Technology

[0002] With the construction and development of new power systems, the deployment scale of flexible interconnection equipment (such as smart soft open nodes (SOPs), flexible multi-state switches (FDSs), and smart soft normally open nodes (SNOPs) in low-voltage distribution networks is continuously expanding. These devices enable continuous and rapid regulation of active and reactive power, gradually transforming traditional radial distribution networks into flexible interconnected, multi-terminal power supply systems. Simultaneously, the widespread integration of distributed photovoltaic systems, user-side energy storage, and adjustable loads further exacerbates the dynamics and uncertainties of distribution network topology. Accurately sensing the real-time topology status of the distribution network is a crucial foundation for realizing line loss analysis, fault location, voltage control, and digital twin construction.

[0003] However, existing methods for generating distribution network topology identification datasets have the following three technical limitations.

[0004] First, traditional topology models cannot represent the continuously adjustable topology states brought about by flexible interconnect devices. Existing simulation methods typically treat the topology as a binary discrete event of "switch closing / opening," using branch open / close tables to simulate topology changes. However, the operating states of flexible devices such as SOPs and FDSs far exceed the binary scope; the electrical connections between their ports have controllable power transmission characteristics, and the equivalent impedance changes dynamically with the control strategy. Taking SOPs as an example, their power support capabilities between feeders vary under different modes such as constant voltage control, constant power control, and droop control. This topology drift in a continuous parameter space cannot be described by discrete switch models. Therefore, existing datasets can only cover discrete transition scenarios and lack data representation of the continuous topology states of flexible interconnect devices.

[0005] Second, the existence of implicit nodes leads to incomplete observation problems in topology identification. Low-voltage distribution networks contain a large number of user nodes that cannot upload operational data in real time (such as old meters and nodes with communication failures). The actual connection relationships and operational status of these implicit nodes cannot be directly observed, resulting in incomplete training labels. Existing data generation methods have never considered providing both "observable representations" and "complete labels" in the dataset, making it impossible to effectively evaluate the true performance of the trained model on the implicit node inference task.

[0006] Third, existing datasets lack causal chain annotations between topological events and electrical features. Traditional methods only record the time of event occurrence and compare voltage amplitude changes before and after the event, failing to establish fine-grained causal alignment from "event type" to "feature change pattern" at the data level. This causes topology recognition models trained on such datasets to primarily learn statistical correlations, resulting in insufficient generalization ability when faced with unseen changes in causal mechanisms.

[0007] In summary, how to construct a simulation dataset generation method that can express the continuous topological state of flexible interconnected devices, handle the problem of incomplete observation of implicit nodes, and have the ability to label causal structures is a technical problem that urgently needs to be solved in this field. Summary of the Invention

[0008] The purpose of this invention is to provide a method for generating spatiotemporal latent variable topology simulation datasets for flexible interconnected distribution networks, in order to solve at least some of the problems existing in the prior art.

[0009] Technical solution: A method for generating spatiotemporal latent variable topology simulation datasets for flexible interconnected distribution networks, comprising the following steps:

[0010] The basic distribution network topology model is extended to construct a two-layer topology structure containing a discrete switching layer and a continuous modulation parameter layer, resulting in an extended basic simulation model.

[0011] Based on the extended basic simulation model, a probabilistic dual-track data generation mechanism for implicit nodes is established, and the full observation topology model and the partial observation topology model are maintained simultaneously to obtain the initialized dual-track output structure.

[0012] The initialized dual-track output structure is input into an event scheduling engine with causal chain annotations. Dynamic topology events are injected in consecutive time steps, an event propagation state queue is maintained, and in each time step, the first-layer event direct layer update, the second-layer power flow calculation layer update, and the third-layer feature evolution layer update are executed sequentially to obtain the node electrical characteristics, topology state, and causal chain of the current time step. label Array;

[0013] The node's electrical characteristics, topology state, and causal data at the current time step. label The array is aligned and encapsulated by time step index, and the output is a simulation dataset containing node electrical characteristics, topology labels, and causal labels.

[0014] Beneficial effects: This invention generates a power distribution network simulation dataset that can simultaneously express the continuous topology drift of flexible interconnected devices, support incomplete observation-supervised learning of implicit nodes, and possess explicit causal structure information by constructing a two-layer topology structure including a discrete switching layer and a continuous modulation parameter layer, establishing a probabilistic dual-track data generation mechanism for implicit nodes, and introducing an event scheduling engine with causal chain annotation. Attached Figure Description

[0015] Figure 1 This is a flowchart of the overall solution of the present invention. Detailed Implementation

[0016] To enable those skilled in the art to better understand the present invention, the technical solutions of the present invention will be clearly and completely described below with reference to the accompanying drawings of the embodiments of the present invention. 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 skilled in the art without creative effort should fall within the scope of protection of the present invention.

[0017] like Figure 1 As shown in the figure, this embodiment describes in detail the data processing flow of a method for generating spatiotemporal latent variable topology simulation datasets for flexible interconnected distribution networks, specifically including:

[0018] S1: Extend the basic distribution network topology model to construct a two-layer topology structure containing a discrete switching layer and a continuous modulation parameter layer, thus obtaining the extended basic simulation model.

[0019] In this application, the basic distribution network topology model is first extended to construct a two-layer topology structure containing a discrete switching layer and a continuous modulation parameter layer. The discrete switching layer is used to represent the physical on / off states of traditional branches in the distribution network, such as the connecting wires between branch boxes and meter boxes, and the incoming lines between meter boxes and users. The on / off states of these branches are usually controlled by physical switching devices such as circuit breakers, disconnectors, or relays, and their state changes are discrete transition events. The continuous modulation parameter layer is used to represent the continuously adjustable operating parameters of flexible interconnection devices, including smart soft open nodes (SOP), flexible multi-state switches (FDS), and smart soft normally open nodes (SNOP). Their operating states cannot be described by a simple "closed / open" binary state, but require multiple continuous-dimensional parameters to characterize them.

[0020] Specifically, when constructing the two-layer topology, the physical structure data of the target low-voltage distribution network is first collected, including transformer location coordinates, branch box hierarchy, meter box distribution locations, user node affiliation information, and distributed photovoltaic access point locations. After structured parsing of this data, a set of basic electrical nodes and a table of branch connection relationships are obtained. Next, a basic topology model is established in the PandaPower simulation framework, and the branch connection relationships are written into the discrete switch layer. Each branch is assigned a switch object by default, and its state is initialized to closed.

[0021] Furthermore, in the continuous modulation parameter layer, a corresponding continuous parameter space is defined for each flexible interconnect device connected to the distribution network. For example, for an SOP device, its continuous parameter space is defined as (P... ref Q ref control mode ) triples, where P ref Q represents the active power reference value. ref Indicates the reactive power reference value, control mode This indicates the current control mode, which can be selected from constant voltage control mode, constant power control mode, droop control mode, or reactive power compensation control mode. For FDS equipment, its continuous parameter space is defined as (α... abc , θ abc Six-dimensional parameters, α abc θ represents the three-phase voltage amplitude modulation coefficient, with a value ranging from 0 to 1. abc This represents the three-phase phase adjustment angle, with a value ranging from -π to π. For SNOP devices, its continuous parameter space is defined as a port power allocation matrix W∈R. n×n Where n represents the number of ports in the SNOP, and each element W in the matrix ij This represents the power weight allocated from port i to port j, and satisfies row sum and normalization constraints.

[0022] According to a further improvement of this embodiment, after the parameter space definition is completed, the continuous modulation parameter layer and the discrete switching layer are associated with the same electrical node model to form a unified simulation model. Specifically, each flexible interconnect device corresponds to an equivalent two-port or multi-port element in the model, and its electrical characteristics are determined by the continuous modulation parameters at the current moment. At the same time, the electrical connection relationship between its two ends or between each end is also constrained by the corresponding switching state in the discrete switching layer. Through the above processing, the extended basic simulation model is finally obtained. This model retains the expressive ability of discrete switching events in traditional distribution networks and adds the expressive ability of continuous parameter adjustment of flexible interconnect devices, laying the foundation for the subsequent generation of simulation datasets containing continuous topology drift characteristics.

[0023] S2: Based on the extended basic simulation model, a probabilistic dual-track data generation mechanism for implicit nodes is established, and the full observation topology model and the partial observation topology model are maintained simultaneously to obtain the initialized dual-track output structure.

[0024] After obtaining the extended basic simulation model, this scheme further establishes a probabilistic dual-track data generation mechanism for implicit nodes. Implicit nodes refer to electrical nodes in the distribution network that exist but cannot upload operational status data in real time. Examples include user meter nodes in some older distribution areas that are not connected to the data acquisition system, monitoring nodes with missing data due to communication failures, or end nodes that are intentionally not configured with telemetry functions. Since the actual topological connections and operational status of these nodes cannot be directly observed, a data generation method that can simultaneously provide "observable representations" and "complete labels" is needed when training the topology identification model. To this end, this application designs a probabilistic dual-track data generation mechanism that simultaneously maintains a fully observed topology model and a partially observed topology model.

[0025] Specifically, implicit nodes are first randomly inserted into the distribution network structure of the expanded basic simulation model. The insertion method involves randomly selecting a parent node from existing branch nodes or meter box nodes, adding one or more child nodes as implicit nodes to this parent node, and randomly assigning each implicit node a real electrical connection relationship and operating state, including either operational or deactivated status. This step constructs a fully observable topology model where the real information of the implicit nodes is fully recorded, but it is not exposed to the model during subsequent training.

[0026] According to one aspect of this application, conditional probability distribution parameters are established for each implicit node and its associated observable nodes in a fully observable topology model, based on a Bayesian hidden tree model. Specifically, for each implicit node L and the observable nodes in its parent node set Pa(L), a conditional probability P(L | Pa(L)) is defined, which represents the probability that the implicit node is in a certain connection state given the state of its parent node. Similarly, the operational state transition probability of the implicit node is also defined to simulate the behavior of the implicit node randomly switching between operation and shutdown.

[0027] In a preferred embodiment, at each time step, missing observation patterns are sampled according to conditional probability distribution parameters. The sampling process for missing patterns is as follows: For each implicit node, its state is determined by a certain probability to determine whether it can be "observed" at the current time step. This probability is determined by the conditional probability distribution parameters, resulting in some implicit nodes remaining hidden for a long time, while others are occasionally exposed. Subsequently, the states of implicit nodes in the fully observed topology model are hidden according to the sampled missing patterns, i.e., the existence information of the node in the observable topology is deleted, and the connection relationship between its parent and child nodes is adjusted so that the observable topology still maintains connectivity. This yields a partially observed topology model and corresponding observable topology labels.

[0028] Furthermore, the observable topology labels corresponding to the partially observed topology model are aligned and bound to the complete topology labels corresponding to the fully observed topology model on the time axis. The alignment and binding are as follows: for each time step t, two sets of label data are saved simultaneously: one set contains only the topology labels with observable node information (for model training), and the other set contains the complete topology labels with the true states of all implicit nodes (for verification). Both sets are associated with the electrical measurement features of the same time step. After the above processing, the initialized dual-track output structure is obtained.

[0029] S3: Input the initialized dual-track output structure into the event scheduling engine with causal chain annotations, inject dynamic topology events in consecutive time steps, maintain the event propagation state queue, and sequentially execute the first-layer event direct layer update, the second-layer power flow calculation layer update, and the third-layer feature evolution layer update at each time step to obtain the node electrical characteristics, topology state, and causal chain of the current time step. label Array.

[0030] After obtaining the initialized dual-track output structure, this scheme inputs it into an event scheduling engine with causal chain annotations. The core function of this engine is to inject dynamic topology events into consecutive time steps, maintain an event propagation state queue, and sequentially execute the first-layer event direct layer update, the second-layer power flow calculation layer update, and the third-layer feature evolution layer update at each time step. Finally, it outputs the node electrical characteristics, topology state, and causal chain for the current time step. label Array.

[0031] In this embodiment, event-cause tuples are first predefined. Each event-cause tuple contains four fields: cause type root node_set ,propagation path and effect timesteps Among them, cause typeUsed to identify the type of topology event, with possible values ​​including user disconnection, user reconnection, PV outgoing connection, PV reconnection, PV switching, SOP mode switching, FDS parameter adjustment, SNOP power redistribution, etc.; root node_set The set of nodes directly affected by an event, such as the root node of a user disconnection event. node_set The electrical node corresponding to this user; propagation path This is a directed list used to record the propagation order of electrical quantity changes in the power grid topology. For example, the propagation path of a SOP power regulation event can be represented as "SOP port → connecting feeder → branch box → meter box → user"; timesteps The number of time steps that the change propagates, typically ranging from 2 to 15.

[0032] Further, an event propagation state queue is initialized. A list of event causal tuples sorted by timeline is sequentially injected into the queue. Each event causal tuple carries its current propagation progress index in the queue, initially set to 0. The queue is then popped sequentially from the queue according to the time step order, showing the active events that need to be processed at the current time step.

[0033] According to one aspect of this application, the following three levels of operations are performed at each time step:

[0034] The first-level event direct-layer update: Retrieves the active events of the current time step from the event propagation state queue. For each active event, it locates its root. node_set The nodes in the system are updated with their load parameters (active power P and reactive power Q) or generation parameters (photovoltaic output). During the update process, a linear or exponential transition curve is used to adjust the effect. timesteps The parameter smoothly changes from its initial value to the target value within the specified number of time steps, avoiding physical transients caused by parameter jumps. After completing this layer update, a set of updated node injection power values ​​is obtained.

[0035] The second-layer power flow calculation update: Using the node injection power values ​​updated in the first layer as input, the pandapower power flow calculation tool is called to perform a Newton-Raphson iterative solution for the entire network. After the power flow calculation converges, the updated network voltage amplitude, phase angle, active power distribution, and reactive power distribution are obtained. These data reflect the steady-state operating point of the entire distribution network under the current topology and injected power conditions.

[0036] Third-layer feature evolution layer update: based on the propagation path. pathThe characteristic response delay parameters of each node are used to adjust the electrical quantities of the corresponding nodes. For example, for a propagation path "SOP port → feeder → branch box → meter box → user", the response delays of different node types are different: the response time of the SOP port itself is in the millisecond range, which can be regarded as an immediate response at the current time step; the response delay of the feeder node is about 1 time step; the branch box node is about 2 time steps; the meter box node is about 3 time steps; and the user node is about 4 time steps. According to the stage index of the current event in the propagation path, the corresponding set of nodes is selected, and their electrical quantities are adjusted by weighting according to a preset attenuation coefficient to simulate the process of the event's impact propagating step by step along the power grid.

[0037] After completing the above three levels of operations, a causal value for the current time step is generated based on all active events and their contribution weights at the current time step. label Array. causal label The array generates a list for each time step. Each element in the list contains three parts of information: a unique identifier for the referenced event causal tuple, the event's contribution weight at the current time step (ranging from 0 to 1, indicating the degree of influence of the event on the current electrical state), and the event's propagation... path The index of the current propagation stage. For example, if a SOP mode switching event propagates to the second stage at the current time step, with a contribution weight of 0.6, then causal label The array contains a variable of the form (event) id The tuple of , 0.6, 2).

[0038] Through the above processing, the node electrical characteristics (including node voltage amplitude, node current amplitude, active power P, reactive power Q, and power factor), topology state (including discrete switch state table and continuous modulation parameter values), and causal parameters of the current time step are obtained. label Array.

[0039] S4: Assign the node's electrical characteristics, topology state, and causal information at the current time step. label The array is aligned and encapsulated by time step index, and the output is a simulation dataset containing node electrical characteristics, topology labels, and causal labels.

[0040] After obtaining the node electrical characteristics, topology state, and causal data at each time step... label After arraying, this approach aligns and encapsulates these data according to time step indices, ultimately outputting a simulation dataset containing node electrical features, topology labels, and causal labels. This step ensures that each sample in the dataset corresponds to a unique time step, and that all information dimensions are strictly aligned on the time axis, facilitating the subsequent training and validation of the topology recognition model.

[0041] Specifically, the multidimensional data at each time step is first organized into three types of vectors. The first type is the feature vector: the voltage amplitude of all nodes at the current time step is arranged into a sequence according to the node number. Similarly, the node current amplitude, active power P, reactive power Q, and power factor are processed. Then, these five sequences are concatenated end to end to form a one-dimensional first feature vector. For example, if the distribution network has N nodes, the length of the first feature vector is 5N.

[0042] The second type is the topology state vector: This involves arranging the state of each switch (0 for open, 1 for closed) in the discrete switch state table of the current time step into a sequence according to switch number. Simultaneously, it includes all continuous modulation parameter values ​​of the current time step (including P of SOP). ref and Q ref α of FDS abc and θ abc The elements of the power allocation matrix W of SNOP are arranged into another sequence according to the device number and parameter type. Then, the two sequences are concatenated to form a one-dimensional second topology state vector.

[0043] The third category is causal label vectors: these vectors represent the causal label vectors at the current time step. label Each element in the array is converted into a fixed-length numeric vector. For example, the event unique identifier is mapped to an integer code, the contribution weight is directly treated as a floating-point number, and the propagation stage index is treated as an integer. Then, the encoded vectors of all events are concatenated in order to form the third causal label vector. If there are no active events at the current time step, the third causal label vector is a zero vector.

[0044] In a preferred embodiment, the simulation dataset has a time resolution of 1 minute, covering one month of continuous time-series data for each of the four seasons: spring, summer, autumn, and winter. Specifically, March is selected for spring, July for summer, September for autumn, and December for winter, with each season simulating for 30 consecutive days. Therefore, the entire dataset contains 120 days and 172,800 time steps. For each season, the load scaling factor and photovoltaic output curve are adjusted according to the typical solar radiation patterns and temperature characteristics of that season, so that the generated electrical characteristics conform to the seasonal variation patterns.

[0045] Furthermore, the first feature vector, the second topological state vector, and the third causal label vector are indexed and aligned in ascending order of time steps. Index alignment is as follows: the three vectors at time step t are combined into a single sample, and the sample's index is the time step number t. All samples are arranged in the order t = 1, 2, ..., 172800, forming a time step-sample pair mapping table.

[0046] Before outputting the simulation dataset, the following verification steps are performed according to further improvements in this embodiment:

[0047] First, the power balance rationality of the generated simulation dataset is verified. It checks whether the active and reactive power of each node in each time step satisfies Kirchhoff's laws, i.e., whether the difference between the sum of the injected power of all nodes and the sum of the losses of each branch is within the allowable error range. If the error of a certain time step exceeds a preset threshold, the data for that time step is marked as abnormal, and the power flow calculation is re-executed in S3. After successful verification, a power relationship verification report is output.

[0048] Second, perform topology-feature consistency verification. Check whether the changes in node electrical characteristics before and after dynamic topology events conform to the laws of electrical physics. For example, after a user's disconnection event, the voltage amplitude of that user should drop significantly, while the voltage change of the node in the meter box where that user is located should be relatively small; after a photovoltaic outgoing event, the voltage at the grid connection point should increase. Compare and analyze the feature data for five time steps before and after each event. If the direction of feature change is consistent with the physical expectation, the verification is successful.

[0049] Third, the feasibility of implicit node inference is verified. Using the observable topological labels in the dataset as input, a Bayesian inference method is attempted to recover the true state of the implicit nodes. The inference results are compared with the complete topological labels, and a consistency index (such as KL divergence) between the posterior probability and the preset prior distribution is calculated. If the consistency index is less than a preset threshold, the dual-track generation mechanism is considered reasonable and effective.

[0050] Fourth, the generated simulation dataset is statistically compared with the historical operating data of the actual transformer substations. A Kolmogorov-Smirnov test is performed to compare whether the cumulative distribution functions of key indicators such as voltage and power originate from the same distribution; simultaneously, the time-series autocorrelation function is compared to check whether the temporal correlation of the data conforms to the actual load variation pattern. After all the above comparisons pass, the statistical consistency certification result is obtained.

[0051] Finally, the fully validated simulation dataset will be output in a standard format (such as HDF5 or Parquet). This dataset contains a node electrical feature matrix (shape: time steps × feature dimension), a topology label matrix (shape: time steps × topology state dimension), and a causal label matrix (shape: time steps × causal coding dimension). It can be used for multiple distribution network artificial intelligence tasks such as low-voltage distribution network topology identification, user disconnection detection, photovoltaic access identification, and implicit node inference.

[0052] This invention constructs a two-layer topology structure combining a discrete switching layer and a continuous modulation parameter layer. This incorporates the continuously adjustable operating parameters of flexible interconnected devices (SOP, FDS, SNOP) into the distribution network topology representation framework, enabling the generated simulation dataset to reflect the complete spectrum of topology states from discrete jumps to continuous drifts. This fills the gap in existing methods that cannot simulate continuous topology behaviors such as power modulation and mode switching of flexible interconnected devices. Furthermore, a probabilistic dual-track data generation mechanism based on a Bayesian hidden tree model simultaneously maintains both a fully observed topology model and a partially observed topology model. It outputs strictly aligned observable topology labels and complete labels containing the true states of implicit nodes on the same time axis, providing a supervised learning validation benchmark for implicit node inference tasks and effectively solving the problem of missing training labels caused by incomplete observations in actual distribution networks. Furthermore, a causal chain labeling engine with event-causal tuples and a propagation queue is introduced. At each time step, the contribution weight of each topology event to changes in electrical characteristics and propagation stage information are explicitly recorded, enabling the dataset to distinguish between correlation and causality, thus improving the generalization performance of the training model in unseen operating scenarios. Furthermore, by covering continuous time-series data for all four seasons at a one-minute resolution and performing power balancing, topology-feature consistency, and statistical distribution cross-validation, the physical rationality, temporal integrity, and consistency with real-world operating characteristics of the dataset are ensured. In summary, this invention provides a high-quality, multi-dimensional, and interpretable simulation data foundation for artificial intelligence tasks such as flexible interconnected distribution network topology identification, user disconnection detection, photovoltaic access identification, and implicit node inference.

[0053] The preferred embodiments of the present invention have been described in detail above. However, the present invention is not limited to the specific details in the above embodiments. Within the scope of the technical concept of the present invention, various equivalent transformations can be made to the technical solutions of the present invention, and these equivalent transformations all fall within the protection scope of the present invention.

Claims

1. A method for generating a spatiotemporal latent variable topology simulation dataset for flexible interconnected distribution networks, characterized in that, The method includes: The basic distribution network topology model is extended to construct a two-layer topology structure containing a discrete switching layer and a continuous modulation parameter layer, resulting in an extended basic simulation model. Based on the extended basic simulation model, a probabilistic dual-track data generation mechanism for implicit nodes is established, and the full observation topology model and the partial observation topology model are maintained simultaneously to obtain the initialized dual-track output structure. The initialized dual-track output structure is input into an event scheduling engine with causal chain annotations. Dynamic topology events are injected in consecutive time steps, an event propagation state queue is maintained, and in each time step, the first-layer event direct layer update, the second-layer power flow calculation layer update, and the third-layer feature evolution layer update are executed sequentially to obtain the node electrical characteristics, topology state, and causal chain of the current time step. label Array; The node's electrical characteristics, topology state, and causal data at the current time step. label The array is aligned and encapsulated by time step index, and the output is a simulation dataset containing node electrical characteristics, topology labels, and causal labels.

2. The method according to claim 1, characterized in that, The continuous modulation parameter layer defines different continuous parameter spaces for the three types of flexible interconnect devices: For the intelligent soft-open node SOP device, the continuous modulation parameter layer is represented as (P ref Q ref control mode ) triples, where P ref Q is the active power reference value. ref This is the reactive power reference value, control mode Select from constant voltage control mode, constant power control mode, droop control mode or reactive power compensation control mode; For flexible multi-state switch (FDS) devices, the continuous modulation parameter layer is represented as (α) abc , θ abc Six-dimensional parameters, where α abc θ is the three-phase voltage amplitude modulation coefficient. abc For phase adjustment angle; For intelligent soft always-on (SNOP) nodes, the continuous modulation parameter layer is represented as a port power allocation matrix W∈R. n×n , where n is the number of ports, and the elements in W represent the power distribution weights among the ports.

3. The method according to claim 1, characterized in that, The initialized dual-track output structure includes: Implicit nodes are randomly inserted into the distribution network structure of the extended basic simulation model. Each implicit node is given a real electrical connection relationship and operating status to construct a fully observed topology model, where the operating status includes commissioning status or decommissioning status. Based on the Bayesian hidden tree model, conditional probability distribution parameters are established for each implicit node and its associated observable nodes in the fully observed topology model. At each time step, the missing observation information pattern is sampled according to the conditional probability distribution parameters. The implicit node states in the full observation topology model are hidden according to the missing pattern to obtain the partial observation topology model and the corresponding observable topology labels. Align and bind the observable topology labels corresponding to the partial observation topology model with the complete topology labels corresponding to the full observation topology model on the time axis to obtain the initialized dual-track output structure.

4. The method according to claim 1, characterized in that, An event scheduling engine with causal chain annotations is built and runs as follows: Predefine event causal tuples; Initialize an event propagation state queue by sequentially injecting a list of event causal tuples sorted by timeline into the queue. Each event causal tuple carries the current propagation progress index in the queue. In each time step, active events are extracted from the event propagation state queue, load parameters or generation parameters are updated, and the first-level event direct layer update is completed. The power flow calculation tool is called to perform Newton-Raphson iterative solution on the entire network, and the second power flow calculation layer is updated to obtain the updated voltage amplitude, phase angle, active power and reactive power distribution of the entire network. Based on the characteristic response delay parameter, adjust the electrical quantities of the corresponding nodes to complete the update of the third characteristic evolution layer; Generate the causal value for the current time step based on all active events and their contribution weights. label Array.

5. The method according to claim 2 or 4, characterized in that, Dynamic topology events include discrete switch action events and continuous parameter change trajectory events, where continuous parameter change trajectory events include: P of SOP devices ref Or Q ref Continuous variation within a specified time window; α of the FDS device abc or θ abc The port power allocation matrix W of the SNOP device changes continuously within a specified time window.

6. The method according to claim 1, characterized in that, The electrical characteristics of a node include at least the node voltage amplitude, node current amplitude, active power P, reactive power Q, and power factor; the topology labels include discrete switch status tables and continuous modulation parameter values. The simulation dataset has a time resolution of 1 minute and covers one month of continuous time-series data for each of the four seasons: spring, summer, autumn, and winter, forming a total of 120 days and 172,800 time steps.

7. The method according to claim 1, characterized in that, After outputting the simulation dataset, the following verification steps are also performed: The power balance rationality of the generated simulation dataset is verified by checking whether the active and reactive power of each node satisfies Kirchhoff's laws, and a power relationship verification report is obtained. Perform topology-feature consistency verification on the generated simulation dataset to check whether the changes in the electrical characteristics of nodes before and after dynamic topology events conform to the laws of electrical physics, and obtain a consistency verification report. The observable topological labels in the generated simulation dataset are compared with the complete topological labels. The consistency between the posterior probability of inferring the true state of the implicit node from the observable topological labels and the preset Bayesian prior distribution is evaluated, and an implicit node modeling accuracy report is obtained. The generated simulation dataset is statistically compared with the historical operation data of the real transformer substation. Kolmogorov-Smirnov tests and time series autocorrelation function comparisons are performed to obtain statistical consistency certification results.

8. The method according to claim 1, characterized in that, In the first-level event direct layer update, when updating load parameters or generation parameters, a linear transition curve or an exponential transition curve is used to complete the smooth change of the parameters from the initial value to the target value within a specified number of time steps.

9. The method according to claim 4, characterized in that, causal label The array generates a list for each time step, and each element in the list contains three parts of information: a unique identifier for the event causal tuple, the contribution weight of the current time step, and the index of the current propagation stage.

10. The method according to claim 1, characterized in that, The node's electrical characteristics, topology state, and causal data at the current time step. label The array is aligned and encapsulated according to the time step index, specifically including: The node voltage amplitude sequence, node current amplitude sequence, active power P sequence, reactive power Q sequence, and power factor sequence at each time step are concatenated into the first feature vector. The discrete switch state table and continuous modulation parameter values ​​at each time step are concatenated to form the second topological state vector; causal at each time step label The array serves as a third causal label vector; The first feature vector, the second topological state vector, and the third causal label vector are indexed and aligned in ascending order of time steps to form a time step-sample pair mapping table, and the simulation dataset is output.

Images