Data-model combined driven voltage sag state estimation method

Abstract

The invention discloses a data-model combined driving voltage sag state estimation method, and relates to the technical field of electric energy quality analysis and intelligent power distribution network state estimation, and the method comprises the steps: constructing a power distribution network model and setting a short-circuit fault working condition at a data driving level, probabilistic modeling is carried out on the voltage sag key disturbance parameters based on historical power quality monitoring data, a voltage sag event sample set is generated, and voltage sag waveform data of a basic monitoring node and a target node are synchronously collected; on a model driving layer, calculating a voltage sag process based on the voltage sag key disturbance parameters in combination with a short-circuit equivalent model of the power distribution network model, and converting voltage sag waveform data into two-dimensional characteristics containing time sequence evolution information; and according to the voltage sag waveform data and the two-dimensional features, a voltage sag state estimation model of data-model combined driving is constructed to estimate the voltage sag state, and the robustness is improved while the estimation precision of the voltage sag state is considered.

Description

Technical Field

[0001] This application relates to the field of power quality analysis and smart distribution network state estimation technology, and in particular to a data-model jointly driven method for voltage sag state estimation. Background Technology

[0002] In recent years, with the continuous expansion of the power grid and the increasing proportion of renewable energy integration, the operating environment of the power system has become increasingly complex, and power quality problems have gradually become prominent. Among these, voltage sags, due to their high frequency, wide impact, and severe economic losses, have become one of the most representative power quality disturbances in distribution network operation. Voltage sags are usually caused by short circuits or ground faults. They have numerous sources of interference, a wide distribution range, and a significant amplification effect, and have been considered a major safety hazard in smart grids. However, due to the limited number of monitoring devices, existing monitoring systems are mostly concentrated at the point of common coupling, making it difficult to achieve comprehensive perception of voltage sag status across all nodes of the power grid. This results in insufficient system analysis and response capabilities for sag events, causing sensitive users to frequently suffer economic losses due to voltage disturbances.

[0003] Extensive research has been conducted by scholars both domestically and internationally on the problem of voltage sag state estimation (VSSE). Early studies primarily relied on physical models of the power system, estimating the sag state of nodes without foundation monitoring by constructing state equations. Relevant literature first introduced the concept of observability in voltage sag state estimation and established corresponding mathematical models to analyze the observability of the system; however, this method requires the deployment of numerous monitoring devices at the feeder ends, resulting in high engineering implementation costs. Related literature further explores the construction of state equations... In this model, H represents the frequency observations of the measured busbars, M is the system observation matrix, and X is the state variable to be estimated. By statistically analyzing the voltage disturbance frequency under a specific sag threshold, a mathematical model of the distribution network VSSE is achieved. However, its estimation accuracy is significantly affected by the method of constructing the observation matrix, resulting in slow convergence. To address these shortcomings, related literature introduces genetic algorithms to improve computational efficiency during the solution process, but the improvement in estimation accuracy is limited. Related literature uses the fault point method to reconstruct the observation matrix and combines it with singular value decomposition to solve the state equations, which significantly increases computational overhead while improving estimation accuracy. Related literature proposes a stochastic prediction model based on the adaptive trust region principle, integrating power node types and load characteristics during the modeling process to achieve a certain forward-looking prediction capability. However, its algorithm template is fixed, and its adjustable space is limited, making it difficult to adapt to complex operating scenarios. Related literature further extends the stochastic prediction framework, but it still suffers from high funding requirements, long monitoring cycles, and large data scales, weakening its engineering applicability.

[0004] With the increasing number of monitoring devices and the development of data-driven concepts, data-based VSSE methods have gradually gained attention. The complex coupling relationships between power grid nodes are often hidden within operational data, making them difficult to accurately characterize using analytical models. Related literature proposes state estimation methods based on single-node reasoning, but their estimation accuracy is easily affected by changes in operating conditions, and their generalization ability is limited. Related literature constructs a voltage sag pattern library and performs pattern matching to reconstruct the sag state of nodes without basic monitoring, but the method's performance is highly dependent on the completeness of the pattern library. Related literature estimates the sag frequency of nodes without basic monitoring based on probabilistic models, but cannot obtain the specific sag amplitude range. Related literature introduces the random forest algorithm to estimate the voltage amplitude of nodes without basic monitoring, but it is highly dependent on auxiliary equipment. With the introduction of deep learning methods, related literature utilizes neural networks to achieve multi-node voltage sag amplitude estimation, and uses convolutional neural networks to further improve the model's generalization ability, but requires separate model training for different nodes, limiting scalability. To overcome the above shortcomings, relevant literature introduces CycleGAN to learn the correlation between multi-node transient data, achieving high-precision estimation without the need for system topology information; relevant literature also uses bidirectional WaveNet to mine the time dependency characteristics in monitoring data, achieving high estimation accuracy even with very few monitoring points.

[0005] In summary, existing VSSE methods still have shortcomings in terms of estimation accuracy and engineering applicability, especially under conditions of limited monitoring resources, where it is difficult to balance accuracy and robustness. Therefore, how to integrate data-driven methods with power system mechanism information to achieve high-precision voltage sag state estimation under limited monitoring conditions remains a key research issue. Summary of the Invention

[0006] The purpose of this application is to provide a data-model jointly driven voltage sag state estimation method to solve the problems of low accuracy and poor robustness in voltage sag state estimation.

[0007] To achieve the above objectives, this application provides the following solution: This application provides a data-model jointly driven method for estimating voltage sag states, including: At the data-driven level, a distribution network model is constructed and short-circuit fault conditions are set. Based on historical power quality monitoring data, probabilistic modeling of key voltage sag disturbance parameters is performed to generate a voltage sag event sample set. Voltage sag waveform data of basic monitoring nodes and target nodes are collected simultaneously. The key voltage sag disturbance parameters include voltage sag duration, fault location, and load parameters. At the model-driven level, the voltage sag process is calculated based on the key disturbance parameters of the voltage sag and the short-circuit equivalent model of the distribution network model. Based on the voltage sag process, the voltage sag waveform data is converted into two-dimensional features containing time-series evolution information; Based on the voltage sag waveform data and the two-dimensional features, a data-model jointly driven voltage sag state estimation model is constructed to estimate the voltage sag state.

[0008] According to the specific embodiments provided in this application, this application has the following technical effects: This application provides a data-model jointly driven method for estimating voltage sag states. At the data-driven level, historical power quality monitoring data is combined with a distribution network model simulating short-circuit fault conditions to probabilistically model key voltage sag disturbance parameters, generating a voltage sag event sample set. This effectively expands the sample coverage while ensuring physical consistency of the samples. At the model-driven level, based on the key voltage sag disturbance parameters and combined with the short-circuit equivalent model of the distribution network model, the voltage sag process is calculated. The voltage sag waveform data is converted into two-dimensional features containing time-series evolution information, fully exploring the implicit electrical correlation between monitored nodes and unmonitored nodes (including target nodes). Based on the voltage sag waveform data and the two-dimensional features, a data-model jointly driven voltage sag state estimation model is constructed, enabling accurate estimation of the voltage sag state of unmonitored nodes. This method improves robustness while maintaining the estimation accuracy of the voltage sag state. Attached Figure Description

[0009] To more clearly illustrate the technical solutions in the embodiments of this application or the prior art, the drawings used in the embodiments will be briefly introduced below. Obviously, the drawings described below are only some embodiments of this application. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0010] Figure 1 A schematic flowchart of a data-model jointly driven voltage sag state estimation method provided in an embodiment of this application; Figure 2 This is a schematic diagram of the voltage sag state estimation model provided in an embodiment of this application; Figure 3 A schematic diagram illustrating the relationship between training set accuracy and Epoch under different estimation methods provided in an embodiment of this application; Figure 4 This diagram illustrates the relationship between the training set loss function and Epoch under different estimation methods provided in an embodiment of this application. Detailed Implementation

[0011] The technical solutions of the embodiments of this application will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of this application, and not all embodiments. Based on the embodiments of this application, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of this application.

[0012] To make the objectives, features and advantages of this application more apparent and understandable, the application will be further described in detail below with reference to the accompanying drawings and specific embodiments.

[0013] like Figure 1 As shown in the embodiments of this application, a data-model jointly driven voltage sag state estimation method is provided, including: S1: At the data-driven level, a distribution network model is constructed and short-circuit fault conditions are set. Based on historical power quality monitoring data, probabilistic modeling of key voltage sag disturbance parameters is performed to generate a voltage sag event sample set. Voltage sag waveform data of basic monitoring nodes and target nodes are collected simultaneously. The key voltage sag disturbance parameters include voltage sag duration, fault location, and load parameters.

[0014] S2: At the model-driven level, based on the key disturbance parameters of the voltage sag, and combined with the short-circuit equivalent model of the distribution network model, the voltage sag process is calculated.

[0015] S3: Based on the voltage sag process, the voltage sag waveform data is converted into a two-dimensional feature containing time-series evolution information.

[0016] S4: Based on the voltage sag waveform data and the two-dimensional features, construct a data-model jointly driven voltage sag state estimation model to estimate the voltage sag state.

[0017] This application addresses the shortcomings of existing voltage sag estimation methods, such as limited monitoring device deployment, insufficient sample size, and strong dependence of estimation results on system parameters. It proposes a data-model jointly driven voltage sag estimation method.

[0018] At the data level, by combining the power system short-circuit mechanism model with power quality monitoring data statistical modeling, probabilistic modeling is performed on key disturbance factors such as voltage sag duration, fault location and load parameters, and a monitoring-simulation joint voltage sag sample enhancement model is constructed, which effectively expands the sample coverage while ensuring the physical consistency of the samples.

[0019] At the model level, by reconstructing the features of the voltage sag time-series waveform and introducing a deep learning network suitable for time-series signal modeling, the implicit electrical correlation between monitored and unmonitored nodes is fully explored, thereby achieving accurate estimation of the voltage sag state of unmonitored nodes.

[0020] In this application, data-model joint driving means that in the process of voltage sag state estimation, the simulation model built based on the physical mechanism of the power system is organically integrated with the statistical characteristics and data-driven modeling method derived from power quality monitoring data, and plays a collaborative role in multiple stages such as sample generation, feature construction and state estimation model training, so as to jointly drive the realization of the voltage sag state estimation process.

[0021] Among them, model-driven focuses on using the short-circuit mechanism of the power system, the equivalent circuit of the power grid and the parameter relationship to construct a physical calculation model of voltage sag, which is used to constrain and generate voltage sag events with physical consistency; data-driven focuses on using historical power quality monitoring data to statistically model disturbance parameters, and to mine the implicit electrical correlation between monitoring nodes and unmonitored nodes through data feature reconstruction and deep learning methods.

[0022] The joint driving force is reflected in the fact that the physical model provides mechanistic constraints for data generation and sample construction, and the monitoring data provides statistical basis for model parameter distribution and learning process. The two constrain and complement each other in multiple steps, rather than being independent or simply linked together.

[0023] By adopting the above-mentioned data-model jointly driven voltage sag estimation method, the accuracy and stability of voltage sag estimation can be significantly improved without the need for comprehensive deployment of monitoring devices or precise acquisition of all system parameters. This reduces engineering implementation and operation and maintenance costs, and provides a practical and feasible technical means for power quality assessment and operation analysis of distribution networks.

[0024] In an exemplary embodiment, an IEEE 14-node standard distribution network is built on the Matlab / Simulink platform. Based on the statistical characteristics of historical monitoring data, probabilistic modeling of key voltage sag disturbance parameters is performed to generate voltage sag events. Voltage sag waveform data from monitoring nodes and target nodes are synchronously collected to form a voltage sag sample dataset for subsequent model training and validation. Specifically, probabilistic modeling of disturbance parameters such as sag duration, fault location, and load impedance based on the statistical characteristics of historical monitoring data constitutes the data-driven part. The voltage sag process is calculated based on these disturbance parameters and combined with the short-circuit equivalent model of the distribution network, constituting the model-driven part. This achieves joint driving of data and physical models in the sample generation stage. S1 specifically includes: S11: Construct a distribution network model on the simulation platform and determine the target node and basic monitoring node for voltage sag estimation; S12: Set up a short-circuit fault condition in the power distribution network model, perform probabilistic modeling of key voltage sag disturbance parameters based on historical power quality monitoring data, and generate a voltage sag event sample set. S13: Based on the voltage sag event sample set, voltage sag waveform data of the basic monitoring node and the target node are collected synchronously according to the preset data acquisition rules to form a data sample set for model training and verification.

[0025] In one exemplary embodiment, the distribution network model includes 14 nodes and 20 branches; Node 6 is selected as the target node, and based on the electrical distance principle, nodes 5, 11, and 12, whose coupling relationship with the target node is higher than a set threshold, are selected as basic monitoring nodes. The basic monitoring nodes are used to collect voltage sag waveform data.

[0026] In practical applications, the number of nodes and branches can be other data. For ease of description, the number of nodes is 14 and the number of branches is 20 in this application.

[0027] In practical applications, a simulation model of the IEEE 14-node standard power distribution system is constructed on the Matlab / Simulink platform. This system comprises 14 nodes and 20 branches, with a system power base of 100 MVA, a rated voltage of 138.0 kV, and a rated frequency of 50.0 Hz. Node 6 is selected as the target node for voltage sag estimation. Based on the electrical distance principle, nodes 5, 11, and 12, which have strong coupling relationships with the target node, are selected as basic monitoring nodes, denoted as Node A, Node B, and Node C, respectively, for collecting voltage sag waveform data.

[0028] It is worth noting that the selection of the above-mentioned basic monitoring nodes is based on the electrical proximity of the target nodes, so as to ensure that the collected voltage sag information can fully reflect the sag characteristics of the target nodes.

[0029] In an exemplary embodiment, S12 specifically includes: S121: Set up a short-circuit fault condition in the power distribution network model and determine the equivalent short-circuit circuit; the short-circuit fault condition is a three-phase short-circuit fault condition, which is set on the lines between nodes 4 and 5, nodes 10 and 11, and nodes 12 and 13 respectively, to simulate the occurrence of voltage sag events.

[0030] S122: Based on the short-circuit equivalent circuit, determine the impedance of the fault branch according to the fault location and the impedance per unit length of the line; the fault location is the distance between the fault point and the common connection point, used to characterize the spatial randomness of short circuits occurring in different line sections.

[0031] S123: Determine the fault current based on the equivalent power source electromotive force, the system equivalent impedance, and the fault branch impedance.

[0032] S124: Determine the voltage dip at the point of common coupling during a fault based on the equivalent power source electromotive force, the system equivalent impedance, and the fault current.

[0033] S125: Based on historical power quality monitoring data, perform probabilistic modeling on the key disturbance parameters of the voltage sag and construct a probability density function.

[0034] S126: Based on the probability density function, determine the probability density function of the total length interval of the line.

[0035] S127: Based on historical power quality monitoring data, the line is divided into several sections, and a segmented probability model is constructed according to the segment weights based on the probability density function of the total length interval of the line; the probability of fault occurrence is different in different sections.

[0036] S128: Based on the segmented probability model, determine the segment probability density function according to the load impedance.

[0037] S129: Based on the segment probability density function, sample and obtain the key voltage sag perturbation parameters for any voltage sag event.

[0038] S130: Based on the fault location in the key perturbation parameters of any voltage sag event, determine the residual voltage amplitude of the sag, and simultaneously determine the fault application and clearing time based on the voltage sag duration in the key perturbation parameters of any voltage sag event, thus forming a complete voltage sag event.

[0039] S131: Generate a voltage sag event sample set based on multiple voltage sag events.

[0040] In practical applications, a three-phase short-circuit fault condition is set in the IEEE 14-node distribution system simulation model to simulate the occurrence of voltage sag events; the short-circuit faults are respectively set on the lines between nodes 4–5, 10–11, and 12–13.

[0041] Based on the short-circuit equivalent circuit, let the equivalent power source electromotive force be E, the system equivalent impedance be ZS, and the fault branch impedance be ZF. Then the fault current IF satisfies: (1) Voltage dip at the point of common coupling (PCC) during a fault It can be represented as: (2) Furthermore, assuming the impedance per unit length of the line is Z, and the distance between the fault point and the PCC is lF, the impedance of the fault branch can be written as: (3) Substituting equation (3) into equations (1) and (2), we get: (4) Based on the aforementioned transient sag mechanism model, to ensure that transient event samples reflect the uncertainties in actual power grid operation, this application constructs probability density functions for key disturbance parameters such as transient duration T, distance lF from the fault point to the PCC, and load impedance ZL, based on statistical results from historical power quality monitoring data. Corresponding random sources are then configured in the simulation platform to perform Monte Carlo sampling. The transient duration T exhibits a non-negative right-skewed characteristic and is fitted using a Gamma distribution; its probability density function is: (5) Where k is the shape parameter. For scale parameters, Gamma function Fault location lF is used to characterize the spatial randomness of short circuits occurring in different line sections, when divided by the total line length interval. When assuming equal probability, a uniform distribution can be used, and its probability density function is: (6) When historical fault statistics show that the probability of occurrence differs in different sections, the line is divided into several sections. and according to segment weight Construct a piecewise probability model such that the sampling probability in the j-th segment satisfies and The load impedance ZL is used to characterize the randomness of parameters under fluctuating operating conditions. It is fitted with a small variance normal distribution, and its probability density function is: (7) in and These are the mean and variance, respectively, obtained from historical monitoring / operational statistics.

[0042] During simulation, the perturbation parameter combination for the i-th transient event is obtained by independent sampling according to the above distribution. ,Will l F，iSubstitute into equation (3) to calculate ZF and further substitute into equations (1) – (4) to obtain the corresponding sag residual voltage amplitude. T i Determine the timing of fault application and clearing to form a complete voltage sag event; repeat the above process n times to obtain a sample set of voltage sag events, whose Monte Carlo sampling process satisfies: (8) Where X represents the disturbance parameter to be sampled or the sag derived from it. Let n represent the probability density function that is consistent with the above-mentioned perturbation parameter distribution, where n is the number of samplings.

[0043] A three-phase short-circuit fault condition was set in the IEEE 14-node distribution system simulation model to simulate the occurrence of voltage sag events; the short-circuit faults were set in different line sections. A sag mechanism calculation model was established based on the short-circuit equivalent circuit. According to the fault current calculation relationship, PCC sag voltage calculation relationship, and the mapping relationship between fault location parameters and fault branch impedance described by equations (1) to (4), the calculated expression of the voltage sag amplitude changing with the fault location was obtained.

[0044] To ensure that the event samples can reflect the randomness and statistical distribution characteristics in actual power grid operation, this application constructs probability density function models for key disturbance parameters such as sag duration T, distance lF from the fault point to the PCC, and load impedance ZL based on the statistical results of historical power quality monitoring data, and configures corresponding random sources in the simulation platform to perform sampling. Among them, the sag duration is fitted with a Gamma distribution, the fault location parameter is modeled with a uniform distribution under the assumption of equal probability, and a segmented weighted spatial probability model is further constructed when there are segment differences. The load impedance is fitted with a small variance normal distribution. The above probability modeling relationships are represented by equations (5) to (7). During simulation, the perturbation parameter combination of the i-th voltage sag event is independently sampled according to the above distribution. The sampled fault location parameters are then substituted into equations (1) to (4) to calculate the residual voltage amplitude. Simultaneously, the fault application and clearing times are determined by the sampled sag duration, thus forming a complete voltage sag event. The above sampling-calculation-simulation process is repeated n times to obtain the voltage sag event sample set, whose Monte Carlo sampling process is characterized by equation (8). The event samples generated by this mechanism are statistically representative and consistent with the short-circuit mechanism calculation, thus providing effective data support for subsequent state estimation model training.

[0045] In an exemplary embodiment, S13 specifically includes: Based on the voltage sag event sample set, when a voltage sag event occurs in the distribution network model, voltage sag waveform data is synchronously collected at the basic monitoring nodes and target nodes according to preset data acquisition rules; wherein, the voltage sag waveform data acquisition process satisfies the following conditions: 1) The distance between the basic monitoring node and the target node in the electrical topology is less than a set distance threshold; 2) Perform synchronous measurements on the basic monitoring node and the target node, and collect voltage sag waveform data for at least three consecutive complete power frequency cycles for each voltage sag event.

[0046] In practical applications, when a voltage sag event occurs in the power distribution system, voltage sag waveform data is synchronously collected at the basic monitoring nodes A, B, C and the target node M. The voltage waveform is collected and stored by the power quality monitoring device.

[0047] The following conditions must be met during voltage sag data acquisition: 1) The basic monitoring nodes A, B, and C are relatively close to the target node M in the electrical topology and have a strong electrical coupling relationship; 2) Synchronous measurement is performed on the basic monitoring nodes and the target nodes. For each voltage sag event, voltage sag data of the above four nodes for at least three consecutive complete power frequency cycles are collected.

[0048] The simulation experiment duration was set to 0.5s, the sampling interval was 0.2ms, and a grouped data acquisition method was adopted. A set of sample data was formed for each voltage sag event, and finally 1000 sets of effective voltage sag samples were collected for the training and verification of the subsequent voltage sag state estimation model.

[0049] When a voltage sag event occurs in the power distribution system, voltage sag waveform data is synchronously collected at both the basic monitoring node and the target node. The voltage waveform is collected and stored by a power quality monitoring device. During data acquisition, the basic monitoring node and the target node are kept relatively close in the electrical topology, and synchronous measurements are performed on each node. For each voltage sag event, at least several consecutive power frequency cycles of voltage sag data are collected.

[0050] Meanwhile, a group of samples is formed for each voltage sag event by using a grouped data collection method, thereby constructing a voltage sag sample dataset with a unified structure and consistent timing for subsequent model training and validation.

[0051] In an exemplary embodiment, S2 performs feature construction processing on the voltage sag waveform data, converting the original one-dimensional voltage sag time-series signal into two-dimensional features containing time-series evolution information, which are then used as input to the subsequent voltage sag state estimation model. In this step, the voltage sag waveform data serves as the data-driven source, and two-dimensional features are constructed through statistical state transition relationships. Simultaneously, the feature construction process is based on quantities with clear physical meaning, such as the effective value of the voltage sag, thereby achieving a joint characterization of the physical properties of the voltage sag and the statistical structure of the data during the feature representation stage.

[0052] In one exemplary embodiment, a voltage sag state estimation model is constructed based on voltage sag waveform data and two-dimensional features, such as... Figure 2 As shown; after completing model training, the estimation performance of the constructed model is verified and compared based on an independent validation dataset and a comparison model. In this step, voltage sag sample data and two-dimensional features are used as data-driven inputs to complete the training and inference of the state estimation model; the network structure of the state estimation model is designed based on the physical understanding of the spatial correlation and temporal evolution characteristics of voltage sags, thereby realizing the joint driving of data-driven learning and model structure constraints in the state estimation stage. S3 specifically includes: S31: Based on the voltage sag process, the effective value of the voltage sag waveform data is calculated, and the timing features of the effective value of the voltage sag are extracted; the timing features of the effective value of the voltage sag are the original one-dimensional voltage sag timing signal.

[0053] S32: Based on the time-series characteristics of the effective voltage sag, construct a Markov transfer field feature matrix to determine two-dimensional features containing time-series evolution information.

[0054] In an exemplary embodiment, S31 specifically includes: S311: During the voltage sag process, a discrete voltage signal sequence is generated based on the voltage sag waveform data collected by the basic monitoring node and the target node.

[0055] S312: The effective voltage value of the discrete voltage signal sequence is calculated using the sliding window root mean square algorithm, as shown in equation (9).

[0056] S313: Based on the voltage RMS value, generate a voltage sag RMS value time series that reflects the change of voltage sag amplitude over time; the voltage sag RMS value time series includes multiple voltage sag RMS value time series features.

[0057] In practical applications, the acquired discrete voltage signal sequence The effective value of voltage is calculated using the sliding window root mean square algorithm. The calculation formula is as follows: (9) Where N is the window length and k is the index of the current sampling time.

[0058] The above calculations yield a time series of effective values ​​reflecting the change of voltage sag amplitude over time, which serves as the basis for subsequent feature construction.

[0059] In an exemplary embodiment, S32 specifically includes: S321: Discretize the effective voltage sag time series into multiple state intervals, and construct a Markov transition probability matrix based on the state transition relationship between adjacent time intervals.

[0060] S322: Introduce a time index to extend the Markov transition probability matrix into a Markov transition field feature matrix.

[0061] S323: Based on the Markov transfer field feature matrix, the time-series feature of the voltage sag effective value is converted into a two-dimensional feature containing time-series evolution information.

[0062] In practical applications, the effective voltage sag time series is discretized and divided into K state intervals. A Markov transition probability matrix D is constructed based on the state transition relationship between adjacent time steps, and its elements are defined as follows: (10) Based on this, a time index is introduced, and the Markov transition probability matrix is ​​extended to a Markov transition field matrix M, the elements of which are defined as: (11) in, q t This indicates the state interval number corresponding to time t.

[0063] Through the above processing, the one-dimensional voltage sag effective value time series data is converted into a two-dimensional feature matrix, which is used as input for the subsequent voltage sag state estimation model.

[0064] The effective voltage sag time series is discretized into multiple state intervals, and a Markov transition probability matrix is ​​constructed based on the state transition relationship between adjacent time intervals as shown in Equation (10). On this basis, a time index is introduced to extend the transition probability matrix into a Markov transition field matrix as shown in Equation (11).

[0065] Through the above feature construction process, the one-dimensional voltage sag time series signal is transformed into two-dimensional features, providing a unified data input format for subsequent deep learning models to extract sag spatiotemporal features.

[0066] In one exemplary embodiment, S4 specifically includes: S41: Divide the voltage sag waveform data and the two-dimensional features into a training dataset, a validation dataset, and a test dataset.

[0067] S42: Construct and train a data-model jointly driven voltage sag estimation model based on the training dataset; the voltage sag estimation model is a network structure combining a two-dimensional convolutional neural network and a gated recurrent unit.

[0068] S43: Based on the verification dataset, perform inference calculations on the voltage sag state estimation model to determine the interval to which the voltage sag amplitude of the unmonitored node belongs, so as to estimate the voltage sag state.

[0069] In one exemplary embodiment, S43 is followed by: S44: Compare the voltage sag state estimated by the voltage sag state model with the corresponding real voltage sag state in the validation dataset to determine the estimation performance of the voltage sag state model in the voltage sag state estimation task; the estimation performance includes classification accuracy, confusion matrix distribution and stability index.

[0070] S45: Under the same training dataset, validation dataset, and test dataset, construct multiple comparative voltage sag state estimation models using different estimation methods; the different estimation methods include constructing a comparative voltage sag state estimation model using a two-dimensional convolutional neural network alone, and constructing a comparative voltage sag state estimation model using a gated recurrent unit alone.

[0071] S46: Compare and analyze the estimation performance of the voltage sag estimation model and the comparative voltage sag estimation model.

[0072] In practical applications, S42 is as follows: A two-dimensional convolutional neural network is constructed using the two-dimensional feature matrix of the Markov transition field as input to extract spatial correlation information from voltage sag features. Its convolution operation expression is as follows: (12) in, The ReLU function is used as the activation function. (13) The feature sequence output by the convolutional network is input into the gated recurrent unit for temporal modeling, and its state update relationship is as follows: Update Gate: (14) Reset Door: (15) Candidate hidden state: (16) Hidden status update: (17) A two-dimensional convolutional neural network with the Markov transition field feature matrix as input is constructed to extract spatial correlation information in voltage sag features. Its convolution operation expression is Equation (12), and the activation function is Equation (13). The feature sequence output by the convolutional network is input into the gated recurrent unit for time-series modeling. Its state update relationship is as follows: update gate is as shown in Equation (14), reset gate is as shown in Equation (15), candidate hidden state is as shown in Equation (16), and hidden state update is as shown in Equation (17).

[0073] This enables joint modeling of the spatial and temporal evolution characteristics of voltage sags, and outputs the corresponding classification results for the voltage sag amplitude range.

[0074] In practical applications, multiple comparison scenarios are set up. Under the same data partitioning method and parameter training strategy, the performance of voltage sag state estimation methods under different model structures and sample augmentation strategies is compared and analyzed to verify the effectiveness and applicability of this application in the task of estimating voltage sag amplitude range. S45 specifically includes: In scenario 1, under the same data partitioning method and parameter training strategy, and with the introduction of sample augmentation methods, the performance differences of using a 2D Convolutional Neural Network (2D CNN), using a gated recurrent unit (GRU), and the 2D CNN-GRU hybrid neural network model proposed in this application in the task of estimating the voltage sag amplitude range are compared. This analysis examines the impact of different model structures on the ability to model the spatiotemporal features of voltage sag and the classification accuracy, and verifies the effectiveness and advancement of the model structure proposed in this paper.

[0075] Scenario 2: Under the premise of the same data partitioning method and parameter training strategy, the 2D CNN-GRU hybrid neural network proposed in this paper is used as the benchmark model. The performance of the model in the voltage sag amplitude range estimation task is compared and analyzed under the two conditions of introducing sample augmentation method and not introducing sample augmentation method, so as to evaluate the impact of sample augmentation strategy on the model classification accuracy, loss function convergence characteristics and generalization ability.

[0076] Two scenarios are set up. Scenario 1, under the premise of the same data partitioning method and parameter training strategy, and with the introduction of sample augmentation methods, compares the proposed voltage sag state estimation method that integrates data and model with methods using 2D CNN alone and methods using GRU alone. Table 1 shows that the classification accuracy of this application on the test set is 97.65%, which is significantly better than the 93.59% of the 2D CNN method and the 94.81% of the GRU method. Meanwhile, the cross-entropy loss of this application is 0.579, significantly lower than the 0.842 of the 2D CNN method and the 0.631 of the GRU method. This indicates that this application can output a predicted probability closer to the true label distribution in multi-class tasks within the voltage sag amplitude range (V-1 to V-6), reducing the probability of cross-class misclassification, thereby improving the reliability and consistency of sag amplitude range estimation. Furthermore, from... Figure 3 It can be seen that this application achieves rapid accuracy improvement in the early stages of training and maintains relatively small fluctuations in the later stages of training, demonstrating strong convergence stability; Figure 4 It can be seen that the loss function of this application decreases more smoothly and remains at a low level during the training process, indicating that its parameter optimization process is more stable and robust, and can achieve better fitting effect and stronger generalization ability with limited training rounds.

[0077] Table 1. Comparison of the estimation performance of the transient amplitude range under different estimation methods in Scenario 1

[0078] Scenario 2: Based on the same data partitioning method and parameter training strategy, the model training phase is further modified by whether or not sample augmentation is introduced. The application using sample augmentation is compared with that without. Table 2 shows that without sample augmentation, the model's classification accuracy drops to 87.63%, and the cross-entropy loss increases to 0.812, indicating that when the number of samples is limited and the class distribution is unbalanced, the model's ability to distinguish different voltage sag amplitude ranges is significantly constrained. In contrast, after introducing sample augmentation, the application shows significant improvements in both accuracy and loss function metrics. Combined with... Figure 3 and Figure 4 It can be seen that after introducing sample augmentation, the model exhibits a smoother accuracy improvement process and a more stable loss convergence trend during training. This indicates that sample augmentation can effectively enrich the training sample distribution, improve the model's learning sufficiency for different stagnation patterns, and thus enhance the model's generalization ability and estimation stability.

[0079] Table 2 Comparison of the estimation performance of the transient amplitude range under different sample strategies in Scenario 2

[0080] This application achieves the systematic generation of voltage sag samples by constructing a standard distribution network model on a simulation platform and combining probabilistic modeling of voltage sag disturbance parameters with multi-node synchronous data acquisition. On this basis, by calculating the effective value of the voltage sag time-series signal and constructing Markov transfer field features, the one-dimensional time-series data is converted into two-dimensional features suitable for deep learning processing. Furthermore, by using a voltage sag state estimation model combining a two-dimensional convolutional neural network and a gated recurrent unit, the joint modeling of the spatiotemporal characteristics of voltage sag and the inference of the voltage sag amplitude range of unmonitored nodes are realized, providing a complete and feasible technical solution for voltage sag state estimation of distribution systems.

[0081] Compared with the prior art, this application has the following advantages: 1. This application proposes a data-model jointly driven voltage sag state estimation method. By integrating simulation modeling and monitoring data processing, a complete state estimation process is constructed, which includes data acquisition, disturbance modeling, feature extraction, VSSE model design and performance evaluation. It can effectively estimate the voltage sag amplitude range without relying on the measured data of unmonitored nodes, and provides a systematic and feasible new method for state perception under typical sag disturbance scenarios.

[0082] 2. This application introduces a Monte Carlo random voltage sag event generation mechanism based on the statistical characteristics of measured data to probabilistically model key disturbance parameters such as sag duration, fault location, and load impedance. This replaces the traditional sample construction method that relies on empirical assumptions. The generated sag samples are closer to actual operating conditions in terms of both statistical distribution and physical mechanism. This provides a diverse and representative data foundation for subsequent state estimation model training, effectively improving the model's generalization ability and engineering applicability.

[0083] 3. This application uses the Markov transfer field method to preprocess the voltage sag waveform data, mapping the original one-dimensional time series signal into two-dimensional features. While maintaining the integrity of the time series evolution information, it enhances the stability and expressive power of the feature structure, and improves the deep learning model's perception ability and input adaptability to complex sag features.

[0084] 4. This application constructs a hybrid neural network model based on 2D CNN-GRU to achieve collaborative modeling of the spatial correlation characteristics and temporal evolution characteristics of voltage sag. Without relying on complete power grid topology parameters, it directly uses the information of monitoring nodes to complete the inversion estimation of the voltage sag state of unmonitored nodes, and shows high estimation accuracy and model robustness in various disturbance scenarios.

[0085] The technical features of the above embodiments can be combined in any way. For the sake of brevity, not all possible combinations of the technical features in the above embodiments are described. However, as long as there is no contradiction in the combination of these technical features, they should be considered to be within the scope of this specification.

[0086] This document uses specific examples to illustrate the principles and implementation methods of this application. The descriptions of the above embodiments are only for the purpose of helping to understand the methods and core ideas of this application. Furthermore, those skilled in the art will recognize that, based on the ideas of this application, there will be changes in the specific implementation methods and application scope. Therefore, the content of this specification should not be construed as a limitation of this application.

Claims

1. A data-model jointly driven method for estimating voltage sag states, characterized in that, include: At the data-driven level, a distribution network model is constructed and short-circuit fault conditions are set. Based on historical power quality monitoring data, probabilistic modeling of key voltage sag disturbance parameters is performed to generate a voltage sag event sample set. Voltage sag waveform data of basic monitoring nodes and target nodes are collected simultaneously. The key voltage sag disturbance parameters include voltage sag duration, fault location, and load parameters. At the model-driven level, the voltage sag process is calculated based on the key disturbance parameters of the voltage sag and the short-circuit equivalent model of the distribution network model. Based on the voltage sag process, the voltage sag waveform data is converted into two-dimensional features containing time-series evolution information; Based on the voltage sag waveform data and the two-dimensional features, a data-model jointly driven voltage sag state estimation model is constructed to estimate the voltage sag state.

2. The data-model jointly driven voltage sag state estimation method according to claim 1, characterized in that, Based on historical power quality monitoring data, probabilistic modeling of key voltage sag disturbance parameters is performed to generate a voltage sag event sample set. Simultaneously, voltage sag waveform data from both the basic monitoring nodes and the target nodes are collected, specifically including: A power distribution network model is constructed on the simulation platform, and the target node and basic monitoring node for voltage sag estimation are determined. In the power distribution network model, a short-circuit fault condition is set up, and the key disturbance parameters of voltage sag are probabilistically modeled based on historical power quality monitoring data to generate a voltage sag event sample set. Based on the voltage sag event sample set, voltage sag waveform data of the basic monitoring node and the target node are collected synchronously according to the preset data acquisition rules.

3. The data-model jointly driven voltage sag state estimation method according to claim 2, characterized in that, The distribution network model includes 14 nodes and 20 branches; Node 6 is selected as the target node, and based on the electrical distance principle, nodes 5, 11, and 12, whose coupling relationship with the target node is higher than a set threshold, are selected as basic monitoring nodes. The basic monitoring nodes are used to collect voltage sag waveform data.

4. The data-model jointly driven voltage sag state estimation method according to claim 2, characterized in that, In the power distribution network model, a short-circuit fault condition is set up. Based on historical power quality monitoring data, probabilistic modeling of key voltage sag disturbance parameters is performed to generate a voltage sag event sample set, specifically including: In the power distribution network model, a short-circuit fault condition is set up, and the equivalent circuit of the short circuit is determined. The short-circuit fault condition is a three-phase short-circuit fault condition, which is set on the lines between nodes 4 and 5, nodes 10 and 11, and nodes 12 and 13, respectively, to simulate the occurrence of voltage sag events. Based on the short-circuit equivalent circuit, the impedance of the fault branch is determined according to the fault location and the impedance per unit length of the line; the fault location is the distance between the fault point and the common connection point, which is used to characterize the spatial randomness of short circuits occurring in different line sections. The fault current is determined based on the equivalent power source electromotive force, the system equivalent impedance, and the fault branch impedance. The voltage dip at the point of common coupling during a fault is determined based on the equivalent power source electromotive force, the system equivalent impedance, and the fault current. Based on historical power quality monitoring data, a probability model is performed on the key disturbance parameters of the voltage sag, and a probability density function is constructed. Based on the probability density function, determine the probability density function for the total length interval of the line; Based on historical power quality monitoring data, the line is divided into several sections, and a segmented probability model is constructed according to the segment weights based on the probability density function of the total length interval of the line; the probability of fault occurrence is different in different sections. Based on the segmented probability model, the segment probability density function is determined according to the load impedance; Based on the segment probability density function, the key voltage sag perturbation parameters for any voltage sag event are sampled and obtained. Based on the fault location in the key perturbation parameters of any voltage sag event, the residual voltage amplitude of the sag is determined. At the same time, the fault application and clearing time is determined by the voltage sag duration in the key perturbation parameters of any voltage sag event, thus forming a complete voltage sag event. A voltage sag event sample set is generated based on multiple voltage sag events.

5. The data-model jointly driven voltage sag estimation method according to claim 2, characterized in that, Based on the voltage sag event sample set, and according to preset data acquisition rules, voltage sag waveform data of the basic monitoring node and the target node are collected synchronously, specifically including: Based on the voltage sag event sample set, when a voltage sag event occurs in the distribution network model, voltage sag waveform data is synchronously collected at the basic monitoring nodes and target nodes according to preset data acquisition rules; wherein, the voltage sag waveform data acquisition process satisfies the following conditions: 1) The distance between the basic monitoring node and the target node in the electrical topology is less than a set distance threshold; 2) Perform synchronous measurements on the basic monitoring node and the target node, and collect voltage sag waveform data for at least three consecutive complete power frequency cycles for each voltage sag event.

6. The data-model jointly driven voltage sag state estimation method according to claim 1, characterized in that, Based on the voltage sag process, the voltage sag waveform data is converted into two-dimensional features containing time-series evolution information, specifically including: Based on the voltage sag process, the effective value of the voltage sag waveform data is calculated, and the timing features of the effective value of the voltage sag are extracted; the timing features of the effective value of the voltage sag are the original one-dimensional voltage sag timing signal; Based on the time-series characteristics of the effective voltage sag, a Markov transfer field feature matrix is ​​constructed to determine two-dimensional features containing time-series evolution information.

7. The data-model jointly driven voltage sag state estimation method according to claim 6, characterized in that, Based on the voltage sag process, the effective value of the voltage sag waveform data is calculated, and the timing features of the effective value of the voltage sag are extracted, specifically including: During the voltage sag process, a discrete voltage signal sequence is generated based on the voltage sag waveform data collected by the basic monitoring node and the target node; The effective voltage value of the discrete voltage signal sequence is calculated using the sliding window root mean square algorithm; Based on the effective voltage value, a voltage sag effective value time series is generated, which reflects the change of voltage sag amplitude over time; the voltage sag effective value time series includes multiple voltage sag effective value time series features.

8. The data-model jointly driven voltage sag state estimation method according to claim 7, characterized in that, Based on the time-series characteristics of the effective voltage sag, a Markov transfer field feature matrix is ​​constructed to determine two-dimensional features containing time-series evolution information, specifically including: The effective voltage sag time series is discretized into multiple state intervals, and a Markov transition probability matrix is ​​constructed based on the state transition relationship between adjacent time intervals. By introducing a time index, the Markov transition probability matrix is ​​expanded into a Markov transition field feature matrix; Based on the Markov transfer field feature matrix, the time-series features of the effective voltage sag are converted into two-dimensional features containing time-series evolution information.

9. The data-model jointly driven voltage sag state estimation method according to claim 1, characterized in that, Based on the voltage sag waveform data and the two-dimensional features, a data-model jointly driven voltage sag state estimation model is constructed to estimate the voltage sag state, specifically including: The voltage sag waveform data and the two-dimensional features are divided into a training dataset, a validation dataset, and a test dataset. A data-model jointly driven voltage sag estimation model is constructed and trained based on the training dataset; the voltage sag estimation model is a network structure that combines a two-dimensional convolutional neural network with a gated recurrent unit. Based on the validation dataset, the voltage sag estimation model is inferred and calculated to determine the range to which the voltage sag amplitude of unmonitored nodes belongs, so as to estimate the voltage sag state.

10. The data-model jointly driven voltage sag state estimation method according to claim 9, characterized in that, Based on the validation dataset, the voltage sag estimation model is inferred and calculated to determine the interval to which the voltage sag amplitude of unmonitored nodes belongs, in order to estimate the voltage sag state. This process then includes: By comparing the voltage sag state estimated by the voltage sag state model with the corresponding real voltage sag state in the validation dataset, the estimation performance of the voltage sag state model in the voltage sag state estimation task is determined; the estimation performance includes classification accuracy, confusion matrix distribution, and stability index. Under the same training, validation, and test datasets, various comparative voltage sag estimation models were constructed using different estimation methods. These different estimation methods included constructing a comparative voltage sag estimation model using a two-dimensional convolutional neural network alone, and constructing a comparative voltage sag estimation model using a gated recurrent unit alone. The estimation performance of the voltage sag estimation model and the comparative voltage sag estimation model are compared and analyzed.

Images