A method and system for identifying the islanded state of a microgrid with a grid-connected energy storage converter

Abstract

This invention discloses a method and system for isolating microgrids using a grid-connected energy storage converter, belonging to the field of microgrid state identification technology. The method includes: collecting and cleaning multi-source monitoring data of the microgrid; using a disturbance baseline separation algorithm based on dynamic time window integration and threshold filtering to separate the data into steady-state and disturbance components; inputting the two components into a pre-trained state identification deep learning model, which is processed sequentially through a disturbance feature decoupling module, a dual-channel spatiotemporal feature fusion module, and a spatiotemporal feature collaborative extraction module to extract spatial feature vectors containing spatiotemporal correlations; and finally, calculating the state probability distribution using an adaptive state classification head based on Mahalanobis distance to determine the operating state. This invention significantly improves the identification accuracy and stability under noisy, transient, and transitional states by effectively separating disturbances, decoupling features, and collaboratively modeling spatiotemporal information.

Description

Technical Field

[0001] This invention belongs to the field of microgrid status identification technology, specifically relating to a method and system for identifying the isolated grid status of a microgrid with a grid-type energy storage converter. Background Technology

[0002] With the rapid evolution of distributed energy technologies, grid-connected energy storage converters play a core role in power quality regulation and islanded grid support in microgrids. In actual operation, microgrids face complex switching conditions such as main grid disconnections, sudden load changes, and various fault disturbances, leading to frequent and drastic changes in system operating states. To ensure stable operation of the microgrid and achieve smooth switching of control strategies, accurate and real-time identification of the islanded grid state is essential. Existing identification technologies mainly include judgment methods based on single electrical quantity thresholds, machine learning methods based on statistical features, and the recently emerging deep learning methods. These methods can achieve certain identification results under ideal conditions, and data preprocessing techniques improve the applicability of the models to some extent.

[0003] However, existing technologies still have significant limitations when facing complex operating environments containing strong noise interference, transient disturbances, and transitional states. First, conventional data preprocessing methods often employ simple normalization or filtering, which fails to effectively separate power frequency fluctuations and instantaneous disturbances. This results in key, subtle disturbance features being masked or obscured by background noise, leading to feature distortion. Second, existing single-channel feature extraction models often neglect the nonlinear coupling between steady-state and disturbance features, failing to dynamically adjust the contribution of each physical quantity according to operating conditions. Furthermore, they struggle to simultaneously capture the deep temporal and spatial correlations of multiple source features such as voltage, current, and frequency, resulting in insufficient sensitivity for identifying transitional states and transient events. In addition, conventional classifiers and single loss functions lack optimization of the feature space structure, limiting the model's generalization ability when processing boundary samples. Therefore, improvements to existing technologies are urgently needed to address the problems of low accuracy and poor robustness in microgrid isolated state identification caused by incomplete feature separation and insufficient spatiotemporal correlation mining under complex operating conditions. Summary of the Invention

[0004] The purpose of this invention is to overcome the shortcomings of the prior art and provide a method and system for identifying the isolated state of a microgrid using a grid-type energy storage converter, so as to solve the technical problems of low accuracy and poor robustness in identifying the isolated state of a microgrid under noise interference, transient processes and transition states.

[0005] To achieve the above objectives, the present invention adopts the following technical solution: In a first aspect, the present invention provides a method for identifying the islanded state of a microgrid using a grid-connected energy storage converter, comprising the following steps: S100: By deploying a multi-source sensor system at key nodes of the grid-type energy storage converter, multi-source monitoring data during the operation of the microgrid is collected, and the multi-source monitoring data is initially cleaned by timestamp alignment and outlier removal. S200: The multi-source monitoring data after preliminary cleaning is preprocessed using a disturbance baseline separation algorithm. Through dynamic time window integration and threshold filtering, the multi-source monitoring data is separated into steady-state components and disturbance components. S300: Input the steady-state component and the disturbance component into the pre-trained state recognition deep learning model, and control the model to sequentially perform disturbance feature decoupling, dual-channel spatiotemporal feature fusion and spatiotemporal feature collaborative extraction operations, and output a spatial feature vector containing spatiotemporal correlation information; S400: Input the spatial feature vector into the adaptive state classification head of the model, calculate the state probability distribution, and determine the current operating state of the microgrid based on the state probability distribution. The operating state includes normal grid-connected state, islanded operation state, transition state, and fault state.

[0006] Preferably, the multi-source sensor system in S100 includes a voltage transformer, a current transformer, a power meter, and a frequency meter, and the collected multi-source monitoring data includes voltage, current, active power, reactive power, frequency, and load change rate.

[0007] Furthermore, the specific process of preprocessing the multi-source monitoring data using the perturbation baseline separation algorithm in S200 includes: Based on the multi-source monitoring data, an original feature vector is constructed. The original feature vector is then subjected to a sliding integral using a dynamic time window. The average value within the window is calculated as the steady-state component. The length of the dynamic time window is set to be the product of the grid fundamental period and the scaling factor. The steady-state component and a preset perturbation threshold are subtracted from the original feature vector, and the ReLU activation function is applied to the result of the subtraction to obtain the non-negative perturbation component.

[0008] Furthermore, in S300, the process of decoupling the perturbation features of the model includes: The disturbance component is processed by an adaptive gating function based on the Sigmoid function, and a gating value is output to dynamically adjust the activation level of the disturbance component. By using the trainable filter weight matrix, the elements of the steady-state component and the corresponding perturbation component elements processed by the adaptive gating function are weighted and combined to generate a decoupled perturbation feature vector.

[0009] Furthermore, in S300, the process of dual-channel spatiotemporal feature fusion in the model includes: The steady-state components are input into a long short-term memory network to extract steady-state temporal feature vectors. The decoupled perturbation feature vector is input into a temporal convolutional network to extract the perturbation temporal feature vector; The steady-state time-series feature vector and the perturbation time-series feature vector are weighted and summed using learnable adaptive fusion coefficients, and the Hadamard product interaction term between the two is added to obtain the fused feature vector.

[0010] Furthermore, in step S300, the process of the model performing spatiotemporal feature collaborative extraction includes: An extended one-dimensional convolution operation is applied to the fused feature vector to capture long temporal dependencies and output a convolutional feature vector. Based on the convolutional feature vector, the elements in the vector are regarded as graph node features. The node features are projected onto the similarity space using a learnable projection matrix. The Euclidean distance between node feature vectors is calculated, and the Euclidean distance is converted into the connection strength between nodes using an exponential kernel function to construct an adaptive adjacency matrix. The convolutional feature vector and the adaptive adjacency matrix are input into a graph convolutional network. Feature propagation and aggregation are performed on the graph structure defined by the adaptive adjacency matrix to output the spatial feature vector.

[0011] Furthermore, in S400, the working process of the adaptive state classification head includes: The Mahalanobis distance between the spatial feature vector and the prototype vector of each state category is calculated, and the Mahalanobis distance is calculated using the invertible covariance matrix of the corresponding state category. The Mahalanobis distance is converted into probability values ​​using an exponential function and then normalized to obtain the probability distribution of the spatial feature vector belonging to each state category.

[0012] Furthermore, the state recognition deep learning model is trained based on a hybrid loss function, which includes: Cross-entropy loss is used to measure the difference between the predicted probability distribution and the true label; The perturbation separation and reconstruction loss is calculated based on the squared Frobenius norm between the original feature vector and the reconstructed feature vector, and is used to constrain the completeness of the perturbation separation. The metric learning loss is calculated based on the Euclidean distance between the spatial feature vector and the class prototype vector, and is used to compress the distance between samples of the same class and expand the distance between classes.

[0013] Secondly, the present invention also provides a microgrid islanding status identification system for grid-connected energy storage converter microgrids, comprising: The data acquisition module is configured to collect multi-source monitoring data during the operation of the microgrid through a multi-source sensor system deployed at key nodes of the grid-type energy storage converter, and to perform preliminary cleaning of the multi-source monitoring data by timestamp alignment and outlier removal. The data preprocessing module is configured to preprocess the pre-cleaned multi-source monitoring data using a disturbance baseline separation algorithm, and to separate the multi-source monitoring data into steady-state components and disturbance components through dynamic time window integration and threshold filtering. The feature extraction module is configured to input the steady-state component and the disturbance component into a pre-trained state recognition deep learning model, and control the model to sequentially perform disturbance feature decoupling, dual-channel spatiotemporal feature fusion and spatiotemporal feature collaborative extraction operations, and output a spatial feature vector containing spatiotemporal correlation information. The state recognition module is configured to input the spatial feature vector into the adaptive state classification head of the model, calculate the state probability distribution, and determine the current operating state of the microgrid based on the state probability distribution.

[0014] This invention discloses a method and system for identifying the islanded state of a microgrid using a grid-connected energy storage converter. Its core lies in purifying data at the source through an innovative disturbance baseline separation technique and utilizing a deep learning model integrating feature decoupling, dual-channel spatiotemporal fusion, and graph convolutional collaboration to perform refined modeling and joint analysis of steady-state and disturbance features. Compared to existing technologies, this invention, through the disturbance baseline separation algorithm, solves the problem of strong noise and power frequency fluctuations masking key disturbance features at the source, providing the model with high signal-to-noise ratio steady-state and disturbance component inputs. The adaptive gated feature decoupling module effectively breaks the nonlinear coupling between steady-state and disturbance features, achieving dynamic enhancement of sensitivity to sudden disturbances. The designed dual-channel spatiotemporal feature fusion and graph convolutional collaboration structure overcomes the shortcomings of single-channel model spatiotemporal correlation mining, achieving comprehensive capture of the multi-scale spatiotemporal evolution of the microgrid's operating state. Finally, the adaptive classifier based on Mahalanobis distance and the hybrid loss function are jointly optimized, which not only improves classification accuracy but also significantly enhances the model's robustness and generalization ability in distinguishing transitional states, boundary samples, and noisy environments. The entire scheme works synergistically to fundamentally improve the accuracy and reliability of isolated network state identification under complex working conditions. Attached Figure Description

[0015] To more clearly illustrate the technical solutions in the embodiments of the present invention, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0016] Figure 1 This is a flowchart illustrating the method for identifying the isolated state of a microgrid using a grid-type energy storage converter, as provided in an embodiment of the present invention.

[0017] Figure 2 This is a structural block diagram of the microgrid islanding status identification system for grid-type energy storage converters provided in an embodiment of the present invention.

[0018] Figure 3 This is a schematic diagram of signal comparison used in an embodiment of the present invention to verify the effect of the perturbation baseline separation algorithm.

[0019] Figure 4 This is a box plot illustrating the frequency characteristic distribution under different operating conditions in an embodiment of the present invention.

[0020] Figure 5 This is a scatter plot (before decoupling) used in an embodiment of the present invention to demonstrate the effect of the feature decoupling module.

[0021] Figure 6 This is a scatter plot (after decoupling) used in an embodiment of the present invention to demonstrate the effect of the feature decoupling module.

[0022] Figure 7 This is a bar chart comparing the identification performance (F1 score) of different methods under four types of power grid conditions in the embodiments of the present invention.

[0023] Figure 8 This is a comparison curve of the recognition accuracy of different methods under different noise levels in the embodiments of the present invention.

[0024] The following are the markings in the attached diagram: 100. Data acquisition module; 200. Data preprocessing module; 300. Feature extraction module; 400. Status recognition module. Detailed Implementation

[0025] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. All other embodiments obtained by those skilled in the art based on the embodiments of the present invention without creative effort are within the scope of protection of the present invention.

[0026] In the description of this invention, it should be noted that the terms "center," "upper," "lower," "left," "right," "vertical," "horizontal," "inner," and "outer," etc., indicate the orientation or positional relationship based on the orientation or positional relationship shown in the accompanying drawings. They are only for the convenience of describing this invention and simplifying the description, and do not indicate or imply that the device or element referred to must have a specific orientation, or be constructed and operated in a specific orientation. Therefore, they should not be construed as limiting this invention.

[0027] Traditional microgrid islanded state identification technologies struggle to effectively isolate power frequency fluctuations and transient disturbances in complex operating environments containing strong noise interference, transient disturbances, and transitional states. This leads to distortion of key features. Existing single-channel feature extraction models neglect the nonlinear coupling between steady-state and disturbance features, failing to capture the deep temporal and spatial correlations of multi-source features, resulting in insufficient sensitivity for identifying transitional states and transient events. Furthermore, conventional classifiers and single loss functions lack optimization for the feature space structure, limiting the model's generalization ability when processing boundary samples, leading to low identification accuracy and poor robustness.

[0028] In response, this application proposes a method for identifying the islanded state of a microgrid with a grid-connected energy storage converter. This method includes the following steps: Step 100: Collect multi-source monitoring data during the operation of the microgrid by using a multi-source sensor system deployed at key nodes of the grid-type energy storage converter, and perform preliminary cleaning of the multi-source monitoring data by timestamp alignment and outlier removal; Step 200: The multi-source monitoring data after initial cleaning is preprocessed using the perturbation baseline separation algorithm. Through dynamic time window integration and threshold filtering, the multi-source monitoring data is separated into steady-state components and perturbation components. Step 300: Input the steady-state component and the disturbance component into the pre-trained state recognition deep learning model, and control the model to perform disturbance feature decoupling, dual-channel spatiotemporal feature fusion and spatiotemporal feature collaborative extraction operations in sequence, and output a spatial feature vector containing spatiotemporal correlation information; Step 400: Input the spatial feature vector into the adaptive state classification head of the model, calculate the state probability distribution, and determine the current operating state of the microgrid based on the state probability distribution. The operating state includes normal grid-connected state, islanded operation state, transition state, and fault state.

[0029] For ease of understanding, the following explains some key terms in this embodiment: A grid-connected energy storage converter microgrid refers to a small power system that uses a grid-connected energy storage converter as its core and can operate independently or in parallel with the main power grid. This system has voltage and frequency support capabilities, and can maintain stable operation within the microgrid and supply power to critical loads when the main power grid fails or disconnects.

[0030] A multi-source sensor system refers to an integrated sensing device deployed at key electrical nodes of a microgrid to collect various electrical quantities and operating parameters in real time. This system typically includes voltage transformers, current transformers, power meters, and frequency meters, and can acquire multi-dimensional data such as voltage, current, active power, reactive power, frequency, and load change rate, providing comprehensive information input for microgrid status identification.

[0031] The disturbance baseline separation algorithm is a data preprocessing technique aimed at effectively separating the slowly changing steady-state components from the rapidly changing disturbance components in raw monitoring data. This algorithm eliminates the masking effect of background noise and power frequency fluctuations on key disturbance features, allowing subsequent feature extraction to focus more on disturbance information reflecting transient changes in the system.

[0032] Steady-state components refer to the relatively stable and slowly changing components in microgrid operating data, mainly reflecting the basic operating trend and power frequency characteristics of the system. Disturbance components refer to the rapidly changing and transient components superimposed on the steady-state components, caused by system events (such as load changes, faults, grid-connected / islanded switching), which contain key information for identifying abnormal system states.

[0033] A state recognition deep learning model refers to a pre-trained neural network structure designed to automatically learn and extract high-level spatiotemporal features from processed microgrid data, thereby classifying the operating state of the microgrid. This model can process multi-dimensional time-series data and has pattern recognition capabilities.

[0034] Perturbation feature decoupling refers to processing the steady-state component and the perturbation component within the deep learning model to eliminate the nonlinear coupling relationship that may exist between them, and dynamically adjusting the activation degree of the perturbation component to generate a more representative perturbation feature representation.

[0035] Dual-channel spatiotemporal feature fusion refers to the model extracting temporal features from the steady-state component and the decoupled perturbation features respectively, and effectively integrating these features from different channels through an adaptive mechanism, while considering their interactions, to form a fused feature containing comprehensive spatiotemporal information.

[0036] Spatiotemporal feature co-extraction refers to the process where, based on fused features, the model further utilizes a specific neural network structure to capture the deep correlations between data in the temporal and spatial dimensions, thereby generating a spatial feature vector that can characterize the current state of the microgrid.

[0037] An adaptive state classification head is a classification module at the end of a deep learning model. It can adaptively calculate the probability that the input spatial feature vector belongs to different predefined operating states (such as normal grid-connected state, islanded state, transition state, and fault state) based on the input spatial feature vector, and determine the current operating state of the microgrid accordingly.

[0038] This embodiment provides a method for identifying the isolated grid state of a microgrid using a grid-connected energy storage converter, specifically: In step 100, a multi-source sensor system deployed at key nodes of the grid-type energy storage converter is used to collect multi-source monitoring data during the operation of the microgrid, and the multi-source monitoring data is initially cleaned by timestamp alignment and outlier removal.

[0039] In step 200, the pre-cleaned multi-source monitoring data is preprocessed using a disturbance baseline separation algorithm. The core of this algorithm is to extract the steady-state trend (steady-state component) of the data by performing a moving average through a dynamic time window associated with the fundamental period of the power grid. Then, the steady-state component is subtracted from the original data and small fluctuations are filtered out to obtain the prominent disturbance component (disturbance component).

[0040] In step 300, the steady-state component and the perturbation component are input into a pre-trained state recognition deep learning model. The core design of this model is that it does not simply concatenate the two components, but first dynamically modulates the perturbation component through an adaptive gating mechanism and cross-combines it with the steady-state component to achieve feature decoupling. Subsequently, through a specially designed long short-term memory network and a temporal convolutional network, the deep temporal patterns of the steady-state and perturbation components are mined separately and adaptively fused. Finally, through the collaborative operation of dilated convolution and graph convolution, spatiotemporal correlation information is further extracted from the fused features.

[0041] In step 400, the spatial feature vector is input into the adaptive state classification head of the model. This classification head uses a probability calculation method based on Mahalanobis distance, which not only considers the distance between the feature vector and the center of each class, but also the distribution shape (covariance matrix) of each class in the feature space, thus having better robustness to boundary states and noise interference.

[0042] This embodiment purifies data at the source using a disturbance baseline separation algorithm, effectively separating steady-state and disturbance components. Subsequently, through feature decoupling, dual-channel spatiotemporal fusion, and spatiotemporal collaborative extraction operations in a deep learning model, it delves into the multi-scale spatiotemporal correlation information of the data. Finally, it utilizes an adaptive classification head based on statistical distribution for decision-making. This complete technological chain works synergistically to effectively improve the accuracy and robustness of microgrid islanded state identification under complex operating conditions such as strong noise interference, transient disturbances, and transitional states, providing a reliable basis for the stable operation of microgrids and the switching of control strategies.

[0043] This application further proposes that, in step 100 above, the multi-source sensor system includes a voltage transformer, a current transformer, a power meter, and a frequency meter, and the collected multi-source monitoring data includes voltage, current, active power, reactive power, frequency, and load change rate.

[0044] Specifically, by collecting these key electrical quantities, the steady-state operation level, power balance, and dynamic response characteristics of the microgrid can be comprehensively characterized. The load change rate, as a differential characteristic, is particularly sensitive to transient events such as load abrupt changes. This multi-dimensional and complementary data acquisition strategy lays a solid foundation for subsequent detailed analysis.

[0045] As a specific implementation method, the sampling frequency can be dynamically adjusted according to the fundamental frequency period of the power grid to capture complete event characteristics. After initial cleaning, the collected data is labeled with state categories based on microgrid operation event logs and expert knowledge. The labeled categories include normal grid connection, islanded operation, transitional state (such as frequency deviation or load change), and fault state (such as voltage collapse), for a total of four categories. After labeling, the data is divided into training set, validation set, and test set for subsequent model training and evaluation. By collecting and labeling data covering different operating conditions (such as normal grid connection, load switching, fault disturbance, etc.), the diversity of data and the effectiveness of the state identification model are ensured.

[0046] The above technical solutions clarify the specific composition of the multi-source sensor system. These rich and complementary data sources can more completely characterize the features of the microgrid under different operating conditions, especially when switching between grid-connected, islanded, transitional, and fault states, and can provide more sensitive and comprehensive state change signals.

[0047] This application further proposes a specific process for preprocessing multi-source monitoring data using the disturbance baseline separation algorithm in step 200, including: constructing an original feature vector based on the multi-source monitoring data; performing a sliding integral on the original feature vector using a dynamic time window; calculating the average value within the window as the steady-state component; wherein the length of the dynamic time window is set as the product of the grid fundamental period and the scaling factor; subtracting the steady-state component and a preset disturbance threshold from the original feature vector; and applying the ReLU activation function to the result after subtraction to obtain the non-negative disturbance component.

[0048] Specifically, an original feature vector is constructed based on multi-source monitoring data. This involves integrating collected data such as voltage, current, active power, reactive power, frequency, and load change rate at each sampling time to form a unified input vector containing all monitoring parameters, ensuring that all relevant data points can be included in the analysis. A moving integral is applied to the original feature vector using a dynamic time window to calculate the average value within the window as the steady-state component. Smoothing is then applied to extract the long-term trend and baseline of the data, effectively filtering out high-frequency noise and instantaneous fluctuations. This average value represents the operating characteristics of the microgrid under relatively stable conditions. The steady-state component and a preset disturbance threshold are subtracted from the original feature vector, and then the ReLU activation function is applied to obtain the non-negative disturbance component. Through this process, the deviation of the data from the baseline is first obtained, then small fluctuations and noise are filtered out, retaining only significant deviations as effective disturbances. The non-negative characteristic allows subsequent models to focus on analyzing actual abnormal events.

[0049] As a specific implementation method, the original data is processed using a perturbation baseline separation algorithm to eliminate noise and power frequency fluctuations, and the steady-state component and perturbation component are extracted respectively. The specific steps are as follows: 1) Steady-state component extraction By applying a moving integral to the original eigenvectors using a dynamic time window, the average value of the original eigenvectors within the window is calculated. Then, this average value is subtracted from the original eigenvectors to obtain the steady-state component. This process eliminates power frequency fluctuations and low-frequency noise interference contained in the original eigenvectors, preserving the characteristics representing stable system operation. This can be expressed as: In the formula, Represents the steady-state component, with dimension . Preserve the characteristics of stable system operation, dimensions Corresponding voltage, current, active power, reactive power, frequency, and load change rate; This represents the original feature vector, with dimension . High-dimensional input representing the raw monitoring data, including voltage Current Active power reactive power ,frequency and load change rate ; This represents the length of the dynamic time window, used to define the integration interval, and is calculated as follows: ,in For the fundamental frequency period of the power grid, This is the scaling factor; Indicates the current time point, representing the moment the data was collected; Indicates time interval Integrate the original feature vectors within the range, and eliminate power frequency fluctuations by using a moving average. This represents the normalization factor, used to calculate the average value of the feature vectors within the window.

[0050] 2) Extraction of disturbance components The steady-state component and perturbation threshold are subtracted from the original feature vector, and then the ReLU activation function is applied to obtain the perturbation component. This process filters out small fluctuations and ensures the non-negativity of the perturbation component, thereby extracting effective perturbation features, as shown below: In the formula, Represents the perturbation component, with dimension . Extract effective perturbation features; This represents the activation function of the rectified linear unit, ensuring the non-negativity of the output; This represents the disturbance threshold, used to filter out minor fluctuations; a value of 0.05 is preferred.

[0051] Through the above technical solution, this application can effectively solve the problem that traditional methods struggle to accurately separate steady-state and disturbance components in the complex and ever-changing operating environment of microgrids. The introduction of a dynamic time window, particularly its length's correlation with the grid's fundamental period, enables the extraction of steady-state components to adaptively respond to grid frequency changes, ensuring the accuracy of the steady-state baseline. Simultaneously, by precisely subtracting the steady-state components and a preset disturbance threshold from the original data, and combining this with the ReLU activation function, noise and unimportant fluctuations can be effectively filtered out, retaining only the non-negative disturbance information that truly represents system anomalies or state changes.

[0052] In one embodiment, to verify the actual effect of the disturbance baseline separation algorithm, a comparative experiment with conventional preprocessing methods was conducted. By simulating voltage signals during microgrid operation (containing a 50Hz fundamental frequency, random noise, and sudden disturbances), the separation capabilities of the proposed method and the conventional Min-Max normalization method for key signal components were compared. In the experimental data, the original voltage signal contained significant power frequency fluctuations, high-frequency noise, and two sudden disturbances (a positive pulse at time points 100-150 and a negative pulse at time points 300-350). Figure 3As shown, the conventional Min-Max normalization result (vertical axis: normalized value) reveals severe waveform distortion: power frequency fluctuations are compressed, noise components are amplified, and disturbance features are completely obscured by noise. However, the result processed using the dynamic time window integration and threshold filtering technique of this invention (vertical axis: volts) clearly shows that the steady-state component (green curve) fully preserves the smooth power frequency fluctuation characteristics; the disturbance component (red curve) accurately captures two sudden disturbances while filtering out minor fluctuations. This experiment demonstrates that the algorithm of this invention can effectively isolate noise and power frequency interference, providing high signal-to-noise ratio steady-state and disturbance features for subsequent state identification.

[0053] This application further proposes a process for decoupling perturbation features in step 300. This process includes processing the perturbation component using an adaptive gating function based on the Sigmoid function, outputting a gating value to dynamically adjust the activation level of the perturbation component; and using a trainable filter weight matrix to weight and combine the elements of the steady-state component and the corresponding perturbation component elements processed by the adaptive gating function to generate a decoupled perturbation feature vector.

[0054] Specifically, the adaptive gating function can dynamically adjust its behavior according to the characteristics of the input data. After receiving the disturbance component, it compresses the output value to the 0-1 interval through the Sigmoid function to form the gating value. It realizes the mapping of real values ​​to the (0,1) interval by relying on the S-shaped curve characteristics, and serves as a gating signal to represent the degree of "on" or "off".

[0055] The trainable filter weight matrix is ​​a learnable and adjustable parameter used during model training to define the combination of steady-state components and gated perturbation components. It achieves a weighted combination by summing the corresponding elements of the two components, rather than simply superimposing them. Instead, it intelligently allocates the relative importance of the two types of information using learned weights, generating a decoupled perturbation feature vector. This separates and highlights key features in the perturbation component, avoiding information masking and noise interference, and helping the model capture perturbation patterns related to state changes. This matrix can be a linear transformation matrix or a convolution kernel matrix.

[0056] As a specific implementation method, a learnable decoupling filter bank and an adaptive gating mechanism are used to nonlinearly decouple the steady-state and disturbance components to enhance the contribution of the disturbance characteristics. The specific steps are as follows: 1) Define the adaptive gating function An adaptive gating function based on the Sigmoid function is adopted. This function takes the element values ​​of the perturbation component as input, combines the slope parameter and the perturbation threshold parameter to output a gating value, thereby achieving dynamic and smooth control of the activation degree of the perturbation component. This ensures that only perturbations that significantly exceed the perturbation threshold participate in the subsequent feature decoupling, which can be expressed as: In the formula, This represents an adaptive gating function used to dynamically adjust the activation level of disturbance components; This represents the input scalar, that is, the value of a certain element of the disturbance component; The parameter representing the slope of the control gate function has the preferred value. This affects the steepness of the function's transition region; This represents the perturbation threshold parameter, used to determine the activation critical point, and is initialized using data quantiles; This represents the filter index, with a value range of 100. ; For feature dimensions.

[0057] 2) Calculate the decoupling disturbance characteristics Using a trainable filter weight matrix, each element of the steady-state component and the corresponding perturbation component element processed by the adaptive gating function are weighted and combined to obtain each element of the decoupled perturbation feature vector. This achieves feature crossover and nonlinear mapping between the steady-state component and the perturbation component, generating the decoupled perturbation feature vector, as follows: In the formula, The first eigenvector represents the decoupled perturbation eigenvector. One element; This represents the decoupled perturbation feature vector, with dimension . ; This represents the weight matrix of the trainable filter, with dimension . Feature crossover is achieved through linear combination, where To output feature dimensions, Input feature dimension; Represents the weight matrix of the first element. Line number The elements of the column represent the first... The input feature for the th input feature The weights of each output feature; The first component representing the steady-state component One element, As a steady-state component, the stable operating characteristics of the system are preserved; The first component representing the disturbance component One element, For perturbation components, extract the effective perturbation; This indicates that the adaptive gating function is applied to the first disturbance component. Given a set of elements, output the gate value; This represents the feature index, with a value range of 100. ; This represents the dimension of the input feature vector, with a preferred value of 6, corresponding to the voltage. Current Active power reactive power ,frequency and load change rate .

[0058] The above technical solution first utilizes an adaptive gating function based on the sigmoid function to process the perturbation component during feature extraction, dynamically adjusting its activation level to suppress noise and highlight key perturbation features. Subsequently, a trainable filter weight matrix is ​​used to weight and combine the gated perturbation component with the steady-state component to generate a decoupled perturbation feature vector. This method effectively separates and enhances perturbation patterns crucial for state recognition, while avoiding noise interference and information redundancy.

[0059] This application further proposes an operation for dual-channel spatiotemporal feature fusion in a state recognition deep learning model. The specific process includes: inputting the steady-state component into a long short-term memory network to extract the steady-state temporal feature vector; inputting the decoupled perturbation feature vector into a temporal convolutional network to extract the perturbation temporal feature vector; weighting and summing the steady-state temporal feature vector and the perturbation temporal feature vector using learnable adaptive fusion coefficients, and superimposing the Hadamard product interaction term between the two to obtain the fused feature vector.

[0060] Specifically, the steady-state component is input into a Long Short-Term Memory (LSTM) network to extract steady-state temporal feature vectors. This network, a special type of recurrent neural network, addresses the gradient problem in traditional recurrent neural networks when processing long sequences by selectively memorizing or forgetting information through gating mechanisms and cell states. It accurately captures the time-varying patterns and long-term dependencies of the steady-state component, forming a temporal feature vector reflecting the stable behavior of the system. Simultaneously, the decoupled perturbation feature vector is input into a temporal convolutional network. This network uses causal convolution to ensure temporal rationality and combines dilated convolution to expand the receptive field. While controlling network complexity, it captures the long-range dependencies of the perturbation component, efficiently extracting features and correlation patterns of perturbation events of different durations, generating perturbation temporal feature vectors. Based on this, the two types of temporal feature vectors are weighted and summed using adaptive fusion coefficients automatically learned during model training. Simultaneously, their Hadamard product is superimposed as an interaction term to form a fused feature vector. This fusion mechanism dynamically balances the contribution weights of steady-state and perturbation information, capturing the nonlinear synergistic effect between features through element-wise multiplication, highlighting the characterization ability of composite features for specific fault states, and strengthening feature complementarity.

[0061] As a specific implementation method, a dual-channel spatiotemporal encoder is constructed to extract the temporal features of steady state and disturbance respectively, and then fused through an adaptive mechanism to enhance the complementarity between features. The specific steps are as follows: 1) Steady-state time series feature extraction The steady-state components are input into a long short-term memory network for processing to capture the long-term temporal dependencies inherent in the steady-state components. The output is a steady-state temporal feature vector representing this dependency, expressed as: In the formula, This represents the steady-state temporal feature vector output by the Long Short-Term Memory network, with dimension 1. , characterizing the long-term time dependence of steady-state components; This represents a Long Short-Term Memory (LSTM) network, used to process sequential data and capture long-term dependencies. The steady-state components are represented and input into the long short-term memory network to preserve the stable operating characteristics of the system; This represents the set of trainable parameters for a Long Short-Term Memory (LSTM) network, including gating weights such as the weights and biases of the input gate, forget gate, and output gate.

[0062] 2) Extraction of temporal features of disturbance The decoupled perturbation feature vector is input into a temporal convolutional network for processing. The temporal convolutional network captures the long-range dependencies in the perturbation components, represented as follows: The output is a perturbation-based temporal feature vector representing long-range dependencies. The perturbation-based temporal feature vector output by the temporal convolutional network is defined as follows: , dimension , characterizing the long-range dependence of the perturbation component; among which, This represents a temporal convolutional network that uses a dilated convolutional structure to expand the receptive field and capture long-range dependencies. This represents the set of trainable parameters for a temporal convolutional network.

[0063] 3) Adaptive feature fusion The steady-state and perturbation time-series feature vectors are weighted and summed using adaptive fusion coefficients. The Hadamard product between these two vectors is used as an interaction term, which is then multiplied by an interaction strength parameter before being summed to output the final fused feature vector. This process dynamically adjusts the contribution ratio of steady-state and perturbation features and enhances the nonlinear interaction between features through the interaction term, as expressed below: In the formula, Represents the fused feature vector; Represents the adaptive fusion coefficient. The calculation method is expressed as ; For the Sigmoid function; It is a trainable weight matrix used to dynamically adjust the contribution ratio of steady-state and perturbation features; This represents the interaction strength parameter, with a preferred value of [value to be filled in]. Control the degree of influence of interactive items; This represents the Hadamard product, which is an element-wise multiplication used to enhance non-linear interactions between features.

[0064] The above technical solution utilizes a Long Short-Term Memory (LSTM) network to extract the temporal features of the steady-state components, capturing the long-term operating trend of the system. Simultaneously, a temporal convolutional network is used to extract the temporal patterns of the decoupled perturbation features, characterizing dynamic perturbation behavior. These two methods are weighted and combined using adaptive fusion coefficients, and a Hadamard product interaction term is introduced to capture the nonlinear synergistic relationship between steady-state and perturbation characteristics, achieving deep feature fusion and complementary enhancement.

[0065] This application further proposes an operation for collaborative extraction of spatiotemporal features, the specific process of which includes: applying an extended one-dimensional convolution operation to the fused feature vector to capture long temporal dependencies and outputting a convolutional feature vector; based on the convolutional feature vector, treating the elements in the vector as graph node features, projecting the node features to a similarity space using a learnable projection matrix, calculating the Euclidean distance between node feature vectors, and converting the Euclidean distance into the connection strength between nodes using an exponential kernel function to construct an adaptive adjacency matrix; inputting the convolutional feature vector and the adaptive adjacency matrix into a graph convolutional network, performing feature propagation and aggregation on the graph structure defined by the adaptive adjacency matrix, and outputting the spatial feature vector.

[0066] The process involves applying an extended one-dimensional convolution operation to the fused feature vector. By inserting holes between convolution kernel elements, the receptive field is expanded, efficiently capturing long-term temporal dependencies and dispersed correlation information along the time axis without increasing parameters or sacrificing resolution. The output is a convolutional feature vector containing both local and long-range temporal patterns, laying the foundation for subsequent spatial feature extraction. Secondly, the elements of the convolutional feature vector are treated as graph node features. A learnable projection matrix is ​​used to map these features to a similarity space, calculating the Euclidean distance between node feature vectors. This distance is then converted to the node connection strength in the 0-1 interval using an exponential kernel function, constructing an adaptive adjacency matrix that dynamically reflects the nonlinear correlation strength of different monitoring features of the microgrid over time. Finally, the convolutional feature vector and the adaptive adjacency matrix are input into a graph convolutional network. Feature propagation and aggregation are performed on this graph structure. By integrating the node's own and its neighbors' features to update representations, local and global structural information is captured, outputting a spatial feature vector that integrates temporal and microgrid topological and functional correlation information.

[0067] As a specific implementation method, temporal and spatial features are extracted sequentially through the joint operation of dilated convolution and graph convolution to enhance the capture of spatiotemporal correlations in transient events. Specifically, this includes: 1) Dilated Convolution Feature Extraction An extended one-dimensional convolution operation is applied to the fused feature vector, and then the ReLU activation function is applied to the convolution result to output a convolutional feature vector. By expanding the receptive field, long-term temporal dependencies in the fused feature vector are captured, as shown below: In the formula, Represents the convolutional feature vector; This indicates a dilated one-dimensional convolution operation, using the dilation rate. This is used to expand the receptive field and capture long-term temporal dependencies; This represents the number of convolutional layers.

[0068] 2) Adaptive Adjacency Matrix Construction The features of each node in the convolutional feature vector are projected onto the similarity space using a learnable projection matrix. Then, the Euclidean distance between the projected node feature vectors is calculated. Finally, the distance is converted into the connection strength between nodes using an exponential kernel function to construct an adaptive adjacency matrix, represented as follows: In the formula, Represents the th in the adaptive adjacency matrix Line number The elements of the column represent the first... The node and the first The connection strength between nodes; Represents an adaptive adjacency matrix with dimension . Used to define graph structure; The number of nodes; Denotes the Euclidean norm and calculates the distance between vectors; Represents the learnable projection matrix, with dimension . This is used to project features into a similarity space; Indicates the first The feature vector of each node comes from the convolution feature vector. ; Indicates the first The feature vector of each node comes from the convolution feature vector. ; This represents the similarity bandwidth parameter, which controls the rate of similarity decay; the optimal value is [value to be specified]. .

[0069] 3) Graph convolution feature extraction The convolutional feature vector and the adaptive adjacency matrix are input into the graph convolutional network. The graph convolution operation propagates features on the graph structure defined by the adaptive adjacency matrix. Then, the ReLU activation function is applied to the aggregation result, and the spatial feature vector is output, thereby capturing the spatial correlation of node features, as shown below: The specific calculation for graph convolution is as follows: In the formula, This represents the spatial feature vector output by the graph convolution, which is used for subsequent state classification. This represents the operation of a graph convolutional network; Indicates the first The output feature matrix of the layer; Indicates the first The output feature matrix of the layer; The adjacency matrix with added self-joins is represented as follows: ; It is the identity matrix; express The degree matrix, i.e. ; This represents the trainable weight matrix for graph convolution.

[0070] By applying the above technical solution to the fused feature vector using an expanded one-dimensional convolution operation, the receptive field can be enlarged, long-term temporal dependencies can be captured, and more comprehensive temporal features can be extracted. Based on this convolutional feature vector, an adaptive adjacency matrix is ​​constructed, transforming the inherent correlation between different monitoring data into graph-structured connection strength. This enables the graph convolutional network to perform feature propagation and aggregation on a dynamic graph structure, allowing the model to not only identify single time-series patterns but also mine complex nonlinear interactions and structured information between different monitoring variables.

[0071] This application further proposes that in step 400, the working process of the adaptive state classification head of the model includes: calculating the Mahalanobis distance between the spatial feature vector and the prototype vector of each state category, wherein the Mahalanobis distance is calculated using the invertible covariance matrix of the corresponding state category; converting the Mahalanobis distance into probability values ​​using an exponential function and performing normalization processing to obtain the probability distribution of the spatial feature vector belonging to each state category.

[0072] Specifically, spatial feature vectors are abstract high-dimensional data representing the current operating state of a microgrid, output by the model after multi-layer processing. The prototype vector for each state category is the representative center point of that category in the feature space, obtained by averaging or clustering similar spatial feature vectors in the training data. Mahalanobis distance, as a distance metric that considers data covariance, takes into account both feature dimensional correlation and scale differences. Compared to Euclidean distance, it is more accurate in assessing distances for non-spherical or skewed distribution data, and can truly reflect the similarity between samples and category prototypes. Mahalanobis distance calculation requires the use of the invertible covariance matrix of the corresponding state category. This matrix describes the distribution shape and direction of the category data in the feature space, capturing the feature dimensional variance and inter-dimensional covariance. It is estimated through sample statistical analysis or learning algorithms during the model training phase and can adaptively adjust the distance calculation weights to improve the accuracy of the statistical distance calculation between samples and category prototypes. After calculating the Mahalanobis distance, it is converted into a non-negative fraction proportional to the similarity through a negative exponential function, with the smaller the distance, the higher the score; then, after normalization, the sum of the probability values ​​of all state categories is made up to 1, forming an effective probability distribution, which provides a quantitative basis for determining the current operating state of the microgrid.

[0073] As a specific implementation method, a dynamic decision-making mechanism based on Mahalanobis distance is used to calculate the state probability distribution to enhance the ability to identify transition states and boundary conditions. Specifically, this includes: 1) Calculation of Mahalanobis distance The Mahalanobis distance between the spatial feature vector and the prototype vector of each state category is calculated. This distance calculation considers the inverse of the covariance matrix of the corresponding category and characterizes the degree of difference between the spatial feature vector and the center of each category in the statistical feature space. It can be expressed as: In the formula, The Mahalanobis distance between the spatial feature vector and the prototype vector of the c-th category represents the statistical difference in features. This represents the prototype vector of the c-th category, with dimension . It is updated during the training process and represents the central position of the class in the feature space; Let represent the invertible covariance matrix of the c-th category, with dimension . It is updated during the training process, representing the distribution shape of the class in the feature space; This represents the matrix transpose operation; This represents the inverse of the covariance matrix, used for Mahalanobis distance calculation; This represents the status category index, with a range of values ​​corresponding to different running states.

[0074] 2) Calculation of state probabilities The calculated Mahalanobis distance between the spatial feature vector and the prototype vectors of each category is transformed using an exponential function. Then, the transformation results for all categories are summed and normalized to obtain the probability distribution of the spatial feature vector belonging to each state category, thus realizing dynamic probabilistic decision-making for the state, expressed as: In the formula, This represents the probability that the spatial feature vector belongs to the c-th category, and the output is the state recognition result; This represents the scaling factor, used to adjust the degree to which distance affects probability; Represents the spatial eigenvectors and the first The Mahalanobis distance between the prototype vectors of each category is calculated as follows: ; This represents the summation of exponential distances across all categories, achieving probability normalization; The category index represents the summation, traversing all possible state categories; Indicates the first The prototype vectors of each category, with dimensions of... ; Indicates the first The invertible covariance matrix of each category has dimensions of . .

[0075] Through the above technical solution, the adaptive state classification head abandons simple geometric distance dependence. By calculating the Mahalanobis distance between the spatial feature vector and the prototype vector of each state category, and combining it with the invertible covariance matrix of the corresponding category, it accurately characterizes the distribution characteristics of each category in the feature space, effectively capturing the complexity of feature distribution, non-spherical features, and dimensional correlation, and optimizing the accuracy of classification boundaries. After exponential function transformation and normalization, the precise probability distribution of each state category is obtained, which not only improves the accuracy and robustness of isolated network state recognition, but also provides quantitative confidence for subsequent decisions.

[0076] This application proposes a method for training a deep learning model for state recognition, which is trained based on a hybrid loss function. The hybrid loss function includes cross-entropy loss, perturbation separation and reconstruction loss, and metric learning loss.

[0077] Specifically, cross-entropy loss is used to measure the difference between the predicted probability distribution and the true labels. This loss function quantifies the accuracy of classification by comparing the probability distribution of each state category output by the model with the actual microgrid operating state labels. During training, the model updates its parameters based on the gradient of the cross-entropy loss to reduce the gap between the predicted and true distributions, thereby improving the model's classification accuracy of microgrid operating states.

[0078] The disturbance separation and reconstruction loss is calculated based on the squared Frobenius norm between the original and reconstructed eigenvectors, and is used to constrain the completeness of disturbance separation. In microgrid operating state identification, accurately separating steady-state and disturbance components is crucial. This loss function forces the model to effectively reconstruct the separated components back to the original eigenvectors after disturbance separation, thus ensuring the integrity and information preservation of the disturbance separation process. The squared Frobenius norm, as a measure of matrix or vector differences, effectively quantifies the deviation between the original and reconstructed data, prompting the model to learn more accurate and representative steady-state and disturbance components.

[0079] The metric learning loss, calculated based on the Euclidean distance between spatial feature vectors and class prototype vectors, is used to compress the distance between samples of the same class and expand the distance between classes. After the deep learning model extracts spatial feature vectors containing spatiotemporal correlation information, the metric learning loss plays a crucial role in making these feature vectors more discriminative during classification. This loss function aims to optimize the feature space, minimizing the distance between samples belonging to the same microgrid operating state (i.e., samples of the same class) and maximizing the distance between samples belonging to different microgrid operating states (i.e., samples of different classes). By calculating the Euclidean distance between the spatial feature vectors and predefined class prototype vectors for each state, and optimizing based on this distance, the model can learn more compact intra-class clusters and more obvious inter-class separations, thereby significantly enhancing the model's ability to distinguish between different operating states.

[0080] As a specific implementation method, a hybrid loss function is constructed by jointly using classification loss and feature decoupling loss. This optimizes classification accuracy while ensuring the structure of the feature space. Specifically, it includes: 1) Calculate the disturbance separation and reconstruction loss The squared Frobenius norm between the original and reconstructed eigenvectors is calculated as the loss. Minimizing the reconstruction error ensures the integrity of information during the separation of steady-state and perturbation components, avoiding the loss of important information. This is expressed as: In the formula, This represents the perturbation separation and reconstruction loss, used to ensure the completeness of perturbation separation; This represents the Frobenius norm, which is calculated by taking the square root of the sum of the squares of all elements in a matrix and is used to measure the reconstruction error.

[0081] 2) Calculate the learning loss. Calculate the Euclidean distance between the feature vector and its true class prototype vector, and calculate the difference between this distance and the Euclidean distances between prototype vectors of different classes. Take the maximum value of this difference and zero as the loss, thereby optimizing the feature space structure, compressing the distance between features of the same type of samples and the class prototype, and expanding the distance between prototypes of different classes, as expressed as: In the formula, This represents the metric learning loss, used to optimize the feature space structure; This represents a function that maximizes the value, ensuring that the loss is non-negative. This represents the boundary spacing parameter, which controls the minimum distance between classes and indicates the preferred value. ; This represents the Euclidean distance function, which calculates the straight-line distance between two vectors.

[0082] 3) Calculate the total loss function The cross-entropy loss, perturbation separation and reconstruction loss, and metric learning loss are weighted and summed according to their respective weight coefficients to obtain the total loss function. This function optimizes classification accuracy while also considering the completeness of perturbation separation and the optimization of the feature space structure, thus balancing the multiple learning objectives of the model. It is expressed as: In the formula, This represents the total loss function, used for model training and optimization. Cross-entropy loss is used to calculate the difference between the predicted probability distribution and the true label, and is used to optimize classification accuracy. The weighting coefficient represents the perturbation separation and reconstruction loss, controlling the contribution of this loss term to the total loss, and its optimal value is selected. ; This represents the weight coefficient that measures the learning loss, controlling the degree to which this loss term contributes to the total loss, and its optimal value is determined by the weight coefficient. .

[0083] By combining the cross-entropy loss, perturbation separation and reconstruction loss, and metric learning loss into a hybrid loss function, the deep learning model for state recognition can be optimized and trained from multiple dimensions. Cross-entropy loss directly improves the model's classification accuracy; perturbation separation and reconstruction loss ensures the effectiveness and completeness of separating steady-state and perturbation components, providing high-quality input for subsequent feature extraction; and metric learning loss further optimizes the discriminative ability of the spatial feature vectors learned by the model, enabling features from different operating states to form clear boundaries in the feature space.

[0084] Furthermore, such as Figure 2 As shown in the embodiments of this application, an islanded state identification system for a microgrid with a grid-connected energy storage converter is also disclosed, including: The data acquisition module 100 is configured to collect multi-source monitoring data during the operation of the microgrid through a multi-source sensor system deployed at key nodes of the grid-type energy storage converter, and to perform preliminary cleaning of the multi-source monitoring data by timestamp alignment and outlier removal. The data preprocessing module 200 is configured to preprocess the pre-cleaned multi-source monitoring data using a disturbance baseline separation algorithm, and to separate the multi-source monitoring data into steady-state components and disturbance components through dynamic time window integration and threshold filtering. The feature extraction module 300 is configured to input the steady-state component and the disturbance component into a pre-trained state recognition deep learning model, and control the model to sequentially perform disturbance feature decoupling, dual-channel spatiotemporal feature fusion and spatiotemporal feature collaborative extraction operations, and output a spatial feature vector containing spatiotemporal correlation information. The state recognition module 400 is configured to input the spatial feature vector into the adaptive state classification head of the model, calculate the state probability distribution, and determine the current operating state of the microgrid based on the state probability distribution.

[0085] In this embodiment, the data acquisition module 100 acquires microgrid operation data through a multi-source sensor system, providing a comprehensive data foundation for state identification. The data preprocessing module 200 employs a disturbance baseline separation algorithm, using dynamic time window integration and threshold filtering to separate multi-source monitoring data into steady-state and disturbance components, effectively preventing power frequency fluctuations from masking key disturbance features. The feature extraction module 300 eliminates the nonlinear coupling between steady-state and disturbance features through disturbance feature decoupling operations, and integrates the temporal features of steady-state and disturbance components through a dual-channel spatiotemporal feature fusion mechanism. Furthermore, it utilizes spatiotemporal feature collaborative extraction operations to mine deep correlations in the time and spatial dimensions of the data, outputting a spatial feature vector with high representational capability. The adaptive state classification head of the state identification module 400 calculates the state probability distribution based on the spatial feature vector, achieving accurate determination of the microgrid's operating state. Through the above technical solution, this embodiment can effectively solve the problems of low recognition accuracy and poor robustness caused by incomplete feature separation and insufficient spatiotemporal correlation mining under complex working conditions, significantly improve the reliability of microgrid state recognition under strong noise interference and transient disturbance environment, and provide solid technical support for the stable operation of microgrid and control strategy switching.

[0086] Furthermore, in one embodiment, to verify the effectiveness of the multi-source monitoring data acquisition in step 100, particularly the ability to distinguish frequency characteristics under different operating conditions, a frequency characteristic distribution comparison experiment was designed. The experiment selected four typical operating conditions: normal grid-connected state, isolated grid operation state, transition state, and fault state. The frequency distribution characteristics were analyzed using box plots (the vertical axis is in Hertz, the boxes represent the interquartile range, the whiskers represent the normal fluctuation range, and the discrete points represent abnormal disturbances). The results are as follows: Figure 4 As shown, under normal grid-connected conditions, the frequency is highly concentrated around 50Hz (due to the narrow and high position of the enclosure); during islanded operation, the center of the enclosure slightly drops to around 49.8Hz and the range expands; under transitional conditions, the enclosure further shifts down to 49.5Hz and the distribution range significantly expands, with more discrete points; under fault conditions, the enclosure shifts significantly down to 48.5Hz and the distribution range is widest. These results verify the significant differences in key electrical quantity characteristics under different conditions and also demonstrate the necessity of dynamically adjusting the sampling frequency to capture complete event characteristics.

[0087] Furthermore, in one embodiment, to verify the effectiveness of the perturbation feature decoupling module in step 300 in improving feature independence, a comparative experiment before and after feature decoupling was designed. The experiment selected two key dimensions: steady-state feature components and perturbation feature components. A scatter plot was used to display the changes in feature distribution before and after decoupling, and an ellipse fitting was used to visually present the differences in correlation between features. For example... Figure 5 and Figure 6 As shown, before decoupling, the data points for the steady-state and perturbation features exhibit a distinct oblique distribution, and the major axis of the fitted ellipse is tilted, indicating a strong correlation between the features. After decoupling, the data point distribution tends to be circular, and the fitted ellipse is closer to a perfect circle, indicating a significant reduction in the correlation between the two feature dimensions. The improved feature independence helps the model extract complementary rather than redundant information from steady-state and perturbation data, thereby improving the accuracy of state recognition.

[0088] Furthermore, in one embodiment, to comprehensively evaluate the state recognition performance of the method of the present invention, a multi-method comparison experiment was designed. Support Vector Machine, Random Forest, and Long Short-Term Memory Network were selected as comparison objects. The F1 score (combined precision and recall) was used as the performance index to evaluate the recognition capability of four types of power grid states (normal grid connection, islanded operation, transitional state, and fault state). The results are as follows: Figure 7As shown in the figure, all methods perform best in the normal grid-connected state where features are most stable. In isolated network operation and transitional states, the performance of conventional methods declines significantly, with Support Vector Machines and Random Forests dropping their F1 scores by approximately 15%, while Long Short-Term Memory Networks perform slightly better but still fall short. The method of this invention maintains the highest performance across all four states, with a particularly significant advantage in isolated network operation (thanks to the enhanced sensitivity from perturbation feature decoupling) and transitional states (thanks to the effective capture of frequency deviations by the adaptive gating mechanism). In fault state identification, the spatiotemporal feature collaborative extraction capability of this invention significantly outperforms conventional single spatiotemporal models.

[0089] Furthermore, in one embodiment, to verify the robustness of the method of the present invention in noisy environments, a noise robustness comparison experiment was designed. The experiment simulated a typical interference scenario of a microgrid, with noise levels covering 0-0.4 times the standard deviation (simulating sensor acquisition errors and electromagnetic interference), and tested the recognition accuracy of each method under different noise levels. Figure 8 As shown, all methods exhibit similar performance (accuracy > 90%) in the absence of noise. When noise exceeds 0.15 standard deviations, the performance of traditional methods degrades sharply, with Support Vector Machines (SVMs) being most affected by noise, experiencing an accuracy drop of approximately 35%. Random Forests and Long Short-Term Memory (LSTM) networks achieve accuracy below 70% under strong noise. In contrast, the method proposed in this invention maintains an accuracy above 85% when noise is ≤ 0.3 standard deviations, and still outperforms traditional methods by more than 20 percentage points when noise is = 0.4 standard deviations. These experimental results comprehensively demonstrate that the perturbation separation algorithm, adaptive gating mechanism, and Mahalanobis distance classifier proposed in this invention work synergistically to effectively filter noise, suppress interference, and enhance the robustness of the model in noisy environments.

[0090] The above are merely embodiments of this application and are not intended to limit the scope of protection of this application. Various modifications and variations can be made to this application by those skilled in the art. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of this application should be included within the scope of protection of this application.

Claims

1. A method for identifying a state of an isolated network of a microgrid of a network-constructed energy storage converter, characterized in that, The method includes the following steps: Step 100: Collect multi-source monitoring data during the operation of the microgrid through a multi-source sensor system deployed at key nodes of the grid-type energy storage converter, and perform preliminary cleaning of the multi-source monitoring data by timestamp alignment and outlier removal; Step 200: The multi-source monitoring data after preliminary cleaning is preprocessed using the perturbation baseline separation algorithm. Through dynamic time window integration and threshold filtering, the multi-source monitoring data is separated into steady-state components and perturbation components. Step 300: Input the steady-state component and the disturbance component into the pre-trained state recognition deep learning model, and control the model to sequentially perform disturbance feature decoupling, dual-channel spatiotemporal feature fusion and spatiotemporal feature collaborative extraction operations, and output a spatial feature vector containing spatiotemporal correlation information; Step 400: Input the spatial feature vector into the adaptive state classification head of the model, calculate the state probability distribution, and determine the current operating state of the microgrid based on the state probability distribution. The operating state includes normal grid-connected state, islanded operation state, transition state, and fault state.

2. The method for identifying the islanded state of a microgrid with a grid-type energy storage converter according to claim 1, characterized in that, The multi-source sensor system in step 100 includes a voltage transformer, a current transformer, a power meter, and a frequency meter. The collected multi-source monitoring data includes voltage, current, active power, reactive power, frequency, and load change rate.

3. The method for identifying the islanded state of a microgrid with a grid-type energy storage converter according to claim 1, characterized in that, The specific process of preprocessing the multi-source monitoring data using the perturbation baseline separation algorithm in step 200 includes: Based on the multi-source monitoring data, an original feature vector is constructed. The original feature vector is then subjected to a sliding integral using a dynamic time window. The average value within the window is calculated as the steady-state component. The length of the dynamic time window is set to be the product of the grid fundamental period and the scaling factor. The steady-state component and a preset perturbation threshold are subtracted from the original feature vector, and the ReLU activation function is applied to the result of the subtraction to obtain the non-negative perturbation component.

4. The method for identifying the islanded state of a microgrid with a grid-type energy storage converter according to claim 1, characterized in that, In step 300, the process of decoupling the perturbation features of the model includes: The disturbance component is processed by an adaptive gating function based on the Sigmoid function, and a gating value is output to dynamically adjust the activation level of the disturbance component. By using the trainable filter weight matrix, the elements of the steady-state component and the corresponding perturbation component elements processed by the adaptive gating function are weighted and combined to generate a decoupled perturbation feature vector.

5. The method for identifying the islanded state of a microgrid with a grid-type energy storage converter according to claim 4, characterized in that, In step 300, the process of the model performing dual-channel spatiotemporal feature fusion includes: The steady-state components are input into a long short-term memory network to extract steady-state temporal feature vectors. The decoupled perturbation feature vector is input into a temporal convolutional network to extract the perturbation temporal feature vector; The steady-state time-series feature vector and the perturbation time-series feature vector are weighted and summed using learnable adaptive fusion coefficients, and the Hadamard product interaction term between the two is added to obtain the fused feature vector.

6. The method for identifying the islanded state of a microgrid with a grid-type energy storage converter according to claim 5, characterized in that, In step 300, the process of the model performing spatiotemporal feature collaborative extraction includes: An extended one-dimensional convolution operation is applied to the fused feature vector to capture long temporal dependencies and output a convolutional feature vector. Based on the convolutional feature vector, the elements in the vector are regarded as graph node features. The node features are projected onto the similarity space using a learnable projection matrix. The Euclidean distance between node feature vectors is calculated, and the Euclidean distance is converted into the connection strength between nodes using an exponential kernel function to construct an adaptive adjacency matrix. The convolutional feature vector and the adaptive adjacency matrix are input into a graph convolutional network. Feature propagation and aggregation are performed on the graph structure defined by the adaptive adjacency matrix to output the spatial feature vector.

7. The method for identifying the islanded state of a microgrid with a grid-type energy storage converter according to claim 1, characterized in that, In step 400, the operation of the adaptive state classification head includes: The Mahalanobis distance between the spatial feature vector and the prototype vector of each state category is calculated, and the Mahalanobis distance is calculated using the invertible covariance matrix of the corresponding state category. The Mahalanobis distance is converted into probability values ​​using an exponential function and then normalized to obtain the probability distribution of the spatial feature vector belonging to each state category.

8. The method for identifying the islanded state of a microgrid with a grid-type energy storage converter according to claim 1, characterized in that, The state recognition deep learning model is trained based on a hybrid loss function, which includes: Cross-entropy loss is used to measure the difference between the predicted probability distribution and the true label; The perturbation separation and reconstruction loss is calculated based on the squared Frobenius norm between the original feature vector and the reconstructed feature vector, and is used to constrain the completeness of the perturbation separation. The metric learning loss is calculated based on the Euclidean distance between the spatial feature vector and the class prototype vector, and is used to compress the distance between samples of the same class and expand the distance between classes.

9. A microgrid islanding status identification system for grid-connected energy storage converter microgrids, characterized in that, include: The data acquisition module is configured to collect multi-source monitoring data during the operation of the microgrid through a multi-source sensor system deployed at key nodes of the grid-type energy storage converter, and to perform preliminary cleaning of the multi-source monitoring data by timestamp alignment and outlier removal. The data preprocessing module is configured to preprocess the pre-cleaned multi-source monitoring data using a disturbance baseline separation algorithm, and to separate the multi-source monitoring data into steady-state components and disturbance components through dynamic time window integration and threshold filtering. The feature extraction module is configured to input the steady-state component and the disturbance component into a pre-trained state recognition deep learning model, and control the model to sequentially perform disturbance feature decoupling, dual-channel spatiotemporal feature fusion and spatiotemporal feature collaborative extraction operations, and output a spatial feature vector containing spatiotemporal correlation information. The state recognition module is configured to input the spatial feature vector into the adaptive state classification head of the model, calculate the state probability distribution, and determine the current operating state of the microgrid based on the state probability distribution.

Images