Distributed microgrid grid-connected protection and power balance intelligent regulation and control method

Abstract

The invention relates to the technical field of distributed micro-grids, and discloses a distributed micro-grid grid-connected protection and power balance intelligent regulation and control method, which comprises the following steps of: firstly, acquiring multi-source operation data of a micro-grid, extracting characteristics by using an intelligent analysis model to obtain an operation characteristic sequence, and calculating a grid correlation coefficient; dimensionality reduction is carried out on correlation matrixes of correlation coefficients of different regional power grids, key feature vectors are extracted, and regulation and control weights of the regional power grids are obtained through Sigmoid function conversion. And dynamically adjusting the reference regulation and control points of the operation feature sequence according to the weights, generating regulation and control parameters and updating regulation and control records. In addition, a multi-dimensional power grid regulation and control space is constructed to monitor abnormity, and a joint optimization model is established to solve an optimal regulation and control scheme. According to the method, the operation state can be accurately analyzed, and the grid-connected protection capability and the power balance regulation and control level of the distributed micro-grid are effectively improved.

Description

Technical Field

[0001] This invention relates to the field of distributed microgrid technology, specifically to a method for grid-connected protection and intelligent power balance control of distributed microgrids. Background Technology

[0002] With the continuous growth of global demand for clean energy, distributed microgrids, as a small-scale power supply and utilization system that integrates distributed power sources, energy storage devices, loads, and monitoring and protection devices, have received widespread attention and development. Distributed microgrids can effectively utilize renewable energy sources such as solar and wind power, promote local energy consumption, and are of great significance for driving energy transformation and building new power systems.

[0003] However, distributed microgrids face numerous severe challenges in actual operation. On the one hand, distributed power sources such as wind and solar power exhibit significant dispersion and intermittency. Weather changes cause substantial fluctuations in the power output of solar photovoltaic panels, and unstable wind speeds lead to inconsistent wind turbine output. This unstable power generation characteristic has a significant negative impact on the power quality of the grid, such as easily triggering voltage fluctuations and harmonic pollution, which in turn affect the normal operation of other equipment in the grid. Simultaneously, its randomness also poses significant difficulties for power balance control in the grid, making it difficult to accurately predict the output power of power sources, resulting in numerous uncertainties in grid dispatch and operation.

[0004] On the other hand, the integration of numerous distributed power sources has adversely affected the control, protection, and operational reliability of the power grid. When distributed power sources are integrated into the distribution network, the grid's topology and power flow distribution become more complex. Traditional protection devices may be unable to accurately identify the location and type of faults, increasing the risk of malfunctions or failures to operate. Furthermore, how distributed microgrids can coordinate with the main grid and ensure their own stable operation during grid faults or anomalies are also critical issues that urgently need to be addressed. For example, when the main grid experiences a power outage, distributed microgrids need to be able to quickly switch to islanded operation mode and maintain their internal power balance to ensure continuous power supply to critical loads; and after the fault is restored, they must be able to smoothly reconnect to the grid to avoid impacting the main grid.

[0005] Existing microgrid control methods have revealed significant limitations in addressing these complex issues. Some traditional control methods rely on simple empirical rules or fixed control strategies, failing to adequately consider the dynamic changes and uncertainties inherent in distributed microgrid operation, resulting in poor control effectiveness. For example, some methods lack efficient data processing and analysis capabilities when dealing with multi-source operational data, making it difficult to extract key information from massive datasets. Consequently, they cannot accurately grasp the microgrid's operating status, hindering the development of precise and effective control strategies. Traditional methods also suffer from insufficient accuracy and slow response speed in calculating grid correlation coefficients and performing power balance control, failing to meet the demands of distributed microgrids for rapid and precise control. Therefore, there is an urgent need for an innovative and intelligent distributed microgrid grid-connected protection and power balance intelligent control method to effectively address these problems, improve the operational stability, reliability, and power quality of distributed microgrids, and promote the efficient utilization of distributed energy resources. Summary of the Invention

[0006] The purpose of this invention is to provide a method for grid-connected protection and intelligent power balance control of distributed microgrids to solve the problems mentioned in the background art.

[0007] To achieve the above objectives, the present invention provides the following technical solution: a method for intelligent regulation and control of grid-connected protection and power balance in distributed microgrids, the method comprising:

[0008] Collect multi-source operation data of distributed microgrids, wherein the operation data includes grid voltage data, current data and power data;

[0009] An intelligent analysis model is used to extract features from the operating data to obtain an operating feature sequence, and the power grid correlation coefficient is calculated based on the fluctuations of the operating feature sequence.

[0010] The correlation matrix of the power grid correlation coefficients in different regions is reduced in dimension, key feature vectors are extracted, and the key feature vectors are transformed by probability distribution using the Sigmoid function to obtain the power grid regulation weights for each region.

[0011] The baseline control points of the operating characteristic sequence are dynamically adjusted according to the grid control weights to generate control parameters, and the intelligent control records of the distributed microgrid are updated based on the control parameters.

[0012] Preferably, the feature extraction includes:

[0013] The running data is extracted according to the preset encoding rules, a feature tensor is constructed, and feature embedding processing is performed.

[0014] A multi-head attention layer is used to perform global feature association on the feature tensor, and the output attention score is a weighted fusion of the feature vector.

[0015] The fusion result is input into a long short-term memory network, multi-scale temporal features are extracted through time step convolution, and the feature dimensions are filtered through an attention mechanism to generate a running feature sequence.

[0016] Preferably, the calculation of the power grid correlation coefficient includes:

[0017] Select the operating characteristic sequence of any region as the benchmark control point, and calculate the fluctuation difference of the characteristic vector of its n adjacent regions;

[0018] The ratio of the mean of the fluctuation difference to the cosine similarity of the benchmark control point is used as the correlation quantity within the region;

[0019] Calculate the mutual information between the benchmark control point and cross-regional characteristics, and take its normalized value as the inter-regional correlation quantity.

[0020] The harmonic mean of the correlation quantities within the region and the correlation quantities between regions is used as the power grid correlation coefficient.

[0021] Preferably, the extraction of key feature vectors includes:

[0022] Construct a correlation matrix of the power grid correlation coefficients in different regions, and perform principal component analysis to obtain orthogonal eigenvectors.

[0023] Select eigenvectors whose variance exceeds a set threshold to form a key feature subspace;

[0024] The correlation matrix is ​​mapped to the key feature subspace to obtain the dimensionality-reduced key feature vector.

[0025] Preferably, the process of obtaining the power grid regulation weights for each region includes:

[0026] The key feature vectors are standardized, and their dot product similarity with the preset control vector is calculated.

[0027] The similarity values ​​are input into a fully connected neural network, and the initial weights are generated after transformation by the hidden layer and the Tanh activation function.

[0028] The initial weights are exponentially smoothed using a time-sliding window to output the power grid control weights for each region.

[0029] Preferably, the generated control parameters include:

[0030] The weight adjustment parameters are obtained by multiplying the power grid control weights and the benchmark control points by matrix position.

[0031] The deviation between the weight adjustment parameter and the benchmark control point is calculated, and the deviation is corrected by state estimation using a Kalman filter.

[0032] The corrected deviation is superimposed on the benchmark control point to generate the control parameters.

[0033] Preferred options also include:

[0034] A multidimensional power grid control space is constructed based on control parameters, and outliers and fluctuation change regions are extracted within the space.

[0035] When the density of outliers exceeds the set threshold or the range of fluctuating abrupt changes is greater than the limit, it is determined to be an abnormal control state, and multi-level early warning instructions are generated.

[0036] Preferably, the construction of the multi-dimensional power grid control space includes:

[0037] The control parameters are mapped to a multidimensional Euclidean space according to the timestamp to generate a control distribution point set;

[0038] The kernel density estimation method is used to estimate the spatial density of the control distribution point set, and the variance gradient and kurtosis coefficient of the estimated space are calculated.

[0039] The gradient distribution is filtered out of noise using a median filtering algorithm to eliminate the effects of random disturbances.

[0040] Preferred options also include:

[0041] A joint optimization model of distributed microgrid operation data and control strategies is established, and the optimal control scheme is solved using Newton's method.

[0042] The optimal control scheme is coupled with the control parameters in real time to generate the optimal control strategy for power grid operation.

[0043] The solution to the joint optimization model includes:

[0044] The objective function is defined as the weighted sum of the absolute values ​​of the control deviation and the control overhead, and the constraint condition is the power grid operation safety boundary.

[0045] The objective function is decomposed into a control strategy subproblem and a running data subproblem by performing a Lagrange dual transformation.

[0046] The two subproblems are solved iteratively and alternately until convergence, and the optimal control scheme that satisfies the constraints is output.

[0047] Preferably, the method further includes:

[0048] Perform data validation on the running feature sequence and extract the difference between the validation feature and the preset standard feature;

[0049] Input the discrepancy values ​​into a support vector machine classifier to determine the type of data anomalies;

[0050] Data correction instructions are generated based on the anomaly type, and the running characteristic sequence is adjusted based on the correction instructions.

[0051] Compared with the prior art, the beneficial effects of the present invention are:

[0052] The intelligent control method for grid-connected protection and power balance of distributed microgrids proposed in this invention has many significant beneficial effects.

[0053] In the data processing and feature extraction stages, a meticulously designed intelligent analysis model can accurately extract multi-source operational data of distributed microgrids using preset coding rules, covering key information such as grid voltage, current, and power data, and cleverly constructing feature tensors. Subsequently, feature embedding processing maps the data to an appropriate space, laying a solid foundation for subsequent analysis. A multi-head attention layer is used to globally correlate features in the feature tensor, comprehensively capturing the complex relationships between different features and outputting a weighted fusion result of attention scores and feature vectors. This result is then input into a Long Short-Term Memory (LSTM) network, where time-step convolution successfully extracts multi-scale temporal features. An attention mechanism is then used to finely filter feature dimensions, ultimately generating an operational feature sequence that highly reflects the actual operating state of the microgrid. This series of operations significantly improves the efficiency and accuracy of data processing, providing solid and reliable data support for subsequent regulation and control analysis.

[0054] In calculating the grid correlation coefficient, this method innovatively selects the operating characteristic sequence of any region as the benchmark control point and meticulously calculates the fluctuation difference of the characteristic vectors of its n adjacent regions. The correlation quantity within the region is determined by the ratio of the mean of the fluctuation difference to the cosine similarity of the benchmark control point. Furthermore, the mutual information between the benchmark control point and cross-regional characteristics is calculated and normalized to represent the inter-regional correlation quantity. Finally, the grid correlation coefficient is obtained through the harmonic mean. This calculation method can comprehensively and accurately reflect the degree of grid correlation between different regions, providing a crucial basis for a deeper understanding of the microgrid's operating structure and the formulation of reasonable control strategies.

[0055] For key feature vector extraction and power grid control weight determination, a correlation matrix of the correlation coefficients of power grids in different regions is first constructed, and then principal component analysis is performed to obtain orthogonal feature vectors. Feature vectors whose eigenvalue explained variance exceeds a set threshold are selected to form a key feature subspace, and the correlation matrix is ​​mapped to this subspace to obtain the dimensionality-reduced key feature vectors. After standardizing the key feature vectors, their dot product similarity with the preset control vectors is calculated. This similarity value is input into a fully connected neural network, and initial weights are generated through hidden layers and Tanh activation functions. Then, exponential smoothing filtering is performed through a time sliding window to output the power grid control weights for each region. This process effectively reduces the dimensionality of the data, highlights key features, and, through neural networks and filtering, makes the determination of power grid control weights more scientific and reasonable, better adapting to the complex and ever-changing operating conditions of microgrids.

[0056] In the stages of generating control parameters and updating control records, the grid control weights are multiplied bitwise with the benchmark control point to obtain the weight adjustment parameters. The Euclidean distance deviation between these parameters and the benchmark control point is calculated, and the deviation is corrected for state estimation using a Kalman filter. Finally, the corrected deviation is superimposed on the benchmark control point to generate the control parameters. Updating the intelligent control records of the distributed microgrid based on these control parameters enables dynamic and precise adjustment of the benchmark control point, thereby better optimizing the microgrid's operating state and improving the accuracy and effectiveness of power balance control.

[0057] Furthermore, this invention effectively extracts outliers and fluctuating change regions by constructing a multi-dimensional power grid control space. When the outlier density exceeds a set threshold or the fluctuating change region exceeds a limit, it can promptly identify an abnormal control state and generate multi-level early warning commands. This helps staff to promptly identify and address potential problems in microgrid operation, ensuring the safe and stable operation of the microgrid. The established joint optimization model of distributed microgrid operation data and control strategies uses Newton's method to solve for the optimal control scheme and couples it with control parameters in real time to generate the optimal control strategy for power grid operation, further improving the overall operating efficiency and economy of the microgrid. Data verification of the operating feature sequence is performed. A support vector machine classifier is used to determine the data anomaly type and generate data correction commands. Based on this, the operating feature sequence is adjusted, ensuring the accuracy and reliability of the data and providing high-quality data support for the entire control process. Attached Figure Description

[0058] Figure 1 This is a schematic diagram illustrating the working principle of the distributed microgrid grid-connected protection and power balance intelligent control method described in this invention.

[0059] Figure 2 Design drawings for calculating power grid correlation coefficients;

[0060] Figure 3 Design diagrams generated for adjusting parameters;

[0061] Figure 4 Design diagram for joint optimization model and data validation. Detailed Implementation

[0062] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0063] Please see Figures 1-4 The present invention relates to a distributed microgrid grid-connected protection and power balance intelligent control method, the specific implementation steps of which are as follows:

[0064] Multi-source operational data of the distributed microgrid is collected, including grid voltage, current, and power data. This data acquisition is achieved by deploying high-precision voltage sensors, current sensors, and power measurement devices at key nodes of the distributed microgrid, such as the output terminals of distributed power sources, connection points to the main grid, and points where important loads are connected. The sensors convert the real-time analog signals into digital signals and transmit them to the data acquisition unit via high-speed communication lines.

[0065] An intelligent analysis model is used to extract features from operational data. Operational data is extracted according to preset encoding rules, a feature tensor is constructed, and feature embedding is performed. Then, a multi-head attention layer is used to globally correlate the feature tensor, outputting a weighted fusion of attention scores and feature vectors. The fusion result is then input into a Long Short-Term Memory (LSTM) network, where multi-scale temporal features are extracted through time-step convolution, and the feature dimensions are filtered using an attention mechanism to obtain the operational feature sequence. Simultaneously, the grid correlation coefficient is calculated based on the fluctuations of the operational feature sequence. Specifically, the operational feature sequence of any region is selected as the benchmark control point, and the fluctuation difference of the feature vectors of its n neighboring regions is calculated. The ratio of the mean of the fluctuation difference to the cosine similarity of the benchmark control point is used as the regional correlation quantity. The mutual information between the benchmark control point and cross-regional features is calculated, and its normalized value is used as the inter-regional correlation quantity. Finally, the harmonic mean of the regional and inter-regional correlation quantities is used as the grid correlation coefficient.

[0066] Subsequently, the correlation matrices of power grid correlation coefficients in different regions are reduced in dimension. First, correlation matrices of power grid correlation coefficients in different regions are constructed, and principal component analysis is performed to obtain orthogonal eigenvectors. Eigenvectors whose eigenvalues ​​explain variance exceeding a set threshold are selected to form a key feature subspace. The correlation matrices are mapped to this key feature subspace to obtain the dimensionality-reduced key feature vectors. Then, the key feature vectors are transformed by a sigmoid function to obtain the power grid control weights for each region. Specifically, the key feature vectors are standardized, and their dot product similarity with the preset control vectors is calculated. The similarity value is input into a fully connected neural network, and after transformation by hidden layers and a Tanh activation function, initial weights are generated. An exponential smoothing filter is applied to the initial weights through a time sliding window to output the power grid control weights for each region.

[0067] The benchmark control point of the operating characteristic sequence is dynamically adjusted based on the grid control weights to generate control parameters. The grid control weights and the benchmark control point are then multiplied bitwise to obtain the weight adjustment parameters. The Euclidean distance deviation between the weight adjustment parameters and the benchmark control point is calculated, and the deviation is corrected using a Kalman filter for state estimation. The corrected deviation is then superimposed onto the benchmark control point to generate the control parameters. Finally, the intelligent control record of the distributed microgrid is updated based on the control parameters, and the control parameters, along with relevant time information and grid state information before and after control, are stored in the intelligent control database of the distributed microgrid for subsequent querying, analysis, and optimization.

[0068] Example 1:

[0069] The operational data comes from various devices and nodes in the distributed microgrid, covering grid voltage, current, and power data. To ensure the comprehensiveness and accuracy of the collected data, sensors need to be deployed at multiple key locations. At the output end of distributed power sources, such as solar photovoltaic panels, voltage and current sensors are installed. These sensors can detect changes in the voltage and current output of the photovoltaic panels in real time, as factors such as changes in sunlight intensity and temperature fluctuations affect the output characteristics of the photovoltaic panels, and these changes can be captured by the sensors in a timely manner. At the connection point with the main grid, high-precision power measurement devices are also deployed. These devices can accurately measure the bidirectional flow of power between the microgrid and the main grid, that is, the power delivered by the microgrid to the main grid and the power obtained from the main grid. This is crucial for understanding the energy interaction relationship between the microgrid and the main grid. For important load connection points, such as factories and hospitals, where the stability of power supply is critical, multiple sensors are installed. These sensors can not only acquire voltage, current, and power data, but also monitor load change trends. For example, the power demand of a factory varies greatly at different production periods, and sensor data can clearly show this. These sensors convert the collected analog signals into digital signals, which are then rapidly transmitted to the data acquisition unit via high-speed communication lines, such as fiber optics or 5G communication technology. The data acquisition unit performs preliminary processing and buffering of the data, awaiting further processing.

[0070] Once the running data is transmitted to the data processing stage, it is truncated according to preset encoding rules. The formulation of these preset encoding rules requires comprehensive consideration of multiple factors. Firstly, it must align with the physical meaning of the data, such as grouping data representing the same electrical parameter (e.g., voltage) at different times together. Secondly, it must consider the time series characteristics, grouping data at appropriate time intervals, such as every 10 seconds. Following these rules, the running data is divided into data segments, which are then arranged and combined in a specific order to construct a three-dimensional feature tensor. When constructing the feature tensor, it is crucial to ensure that the correlation between different types of data is reflected in the tensor structure. For example, voltage, current, and power data at the same time should be placed in the same "slice" of the tensor so that the mutual influence between these data can be comprehensively considered during subsequent processing. After constructing the feature tensor, feature embedding processing is performed. At this stage, pre-trained embedding models are used. These models are typically trained on large amounts of power data and possess powerful feature extraction capabilities. Their working principle is to map each element in the feature tensor, i.e., each data point, into a low-dimensional vector space. In this low-dimensional vector space, data points with similar physical meanings are mapped to nearby locations. For example, voltage data at different times but with similar values ​​will be relatively close in the vector space. This mapping not only reduces the dimensionality of the data and the amount of subsequent computation, but also uncovers the potential similarities and correlations between data points.

[0071] After feature embedding, the feature tensor undergoes global feature association in a multi-head attention layer. This layer employs multiple attention heads, each with its unique focus. Some attention heads focus on the correlation between voltage data at different times. Since grid voltage fluctuates over time due to various factors, these attention heads can capture patterns in voltage fluctuations, such as trends in voltage changes during certain seasons or time periods. Other attention heads focus on the relationship between current and power data, as the magnitude and variation of current directly affect power transmission and consumption. Analyzing their correlation reveals the efficiency of power transmission in the microgrid and the impact of load changes on power. Still other attention heads focus on the connections between data from different regions. Distributed microgrids may consist of multiple regions, and the power conditions in each region influence each other. These attention heads can detect the synergy or interference in power transmission between regions. Each attention head processes the feature tensor separately, outputting a weighted fusion result of its respective attention score and feature vector. By further weighting and fusing the outputs of these different attention heads, a comprehensive and integrated extraction of the global features of the feature tensor can be achieved, resulting in a more representative fusion result.

[0072] The fusion results are then input into a Long Short-Term Memory (LSTM) network. LSM networks have unique advantages in processing time-series data, as their internal memory units can effectively capture long-term dependencies within the data. In distributed microgrid operation data, many parameters exhibit long-term trends and dependencies. For example, power consumption may show periodic changes with seasons, workdays, or user electricity consumption habits. The memory units of the LSM network can remember these long-term patterns. By setting convolutional kernels with different time steps, LSM networks can extract temporal features at different time scales. Smaller time-step kernels, such as those measured in seconds, can capture rapid fluctuations in short periods, such as instantaneous fluctuations in current and voltage caused by sudden load changes. Larger time-step kernels, such as those measured in hours or even days, can extract trend features over longer periods, such as peak and off-peak electricity consumption at different times of the day. After feature extraction, an attention mechanism is used to filter the feature dimensions. This mechanism assigns different weights to each feature dimension based on its importance in describing the microgrid's operating state. For feature dimensions that have a relatively minor impact on the grid's operational status, their weights are reduced, weakening their role in subsequent processing. Conversely, for feature dimensions that significantly impact the grid's operational status, such as those closely related to voltage stability and power balance, their weights are increased to highlight their importance. This screening process ultimately generates a representative sequence of operational features, which accurately reflects the operational status of the distributed microgrid and provides crucial information for subsequent calculations of grid correlation coefficients and the formulation of control strategies.

[0073] Example 2:

[0074] When calculating grid correlation coefficients, selecting a benchmark control point requires comprehensive consideration of multiple factors. The complexity of the grid structure varies across different regions within a distributed microgrid. Some regions may contain multiple distributed power sources, such as photovoltaic and wind power, as well as a large number of different types of loads, resulting in a relatively complex grid structure. Conversely, some regions may be dominated by a single type of distributed power source with simpler load types, leading to a relatively simple grid structure. Generally, for regions with complex grid structures and concentrated distributed power sources and loads, the operating characteristic sequence at the center of the region is preferentially selected as the benchmark control point. This is because the operating characteristic sequence at the center of the region can, to a certain extent, represent the overall grid operating status of the region, and its data includes the combined influence of power sources and loads in all surrounding directions. Simultaneously, the distribution density of distributed power sources within the region must also be considered. In areas with higher distribution density, the mutual influence between power sources is more pronounced; selecting the operating characteristic sequence at the center as the benchmark control point can better capture this mutual influence relationship. In addition, the importance of the load is also a key factor. For some important load areas that have high requirements for power supply stability, using the operating characteristic sequence of the vicinity as a benchmark control point will help to more accurately analyze the grid connection between this area and other areas, thereby ensuring the stable power supply of important loads.

[0075] After determining the benchmark control point, the fluctuation difference of the characteristic vectors of n adjacent regions is calculated. For each adjacent region, the difference between its characteristic vector and the benchmark control point's characteristic vector in each dimension needs to be calculated. These dimensions include multiple electrical parameters such as voltage amplitude, voltage phase, current magnitude, current phase, active power, and reactive power. During the calculation, the fluctuation difference is arranged according to a time series, recording the fluctuation difference in each dimension at each time point. Since the power system is a dynamically changing system, various electrical parameters change continuously over time. Therefore, arranging the fluctuation difference in a time series clearly reflects the correlation between adjacent regions and the benchmark control point. For example, in some time periods, the voltage amplitude fluctuation of adjacent regions may be similar to that of the benchmark control point, while in other time periods, there may be significant differences. This change can be intuitively observed through the time series fluctuation difference.

[0076] When calculating the mean of the fluctuation difference, the average value is calculated for the fluctuation difference sequence in each dimension. This process needs to consider the length of the fluctuation difference sequence and the distribution of the data. For long fluctuation difference sequences, appropriate statistical methods are required for mean calculation to ensure the accuracy of the results. Furthermore, since there may be outliers in the fluctuation difference sequence, which could be caused by measurement errors, transient interference, etc., directly calculating the mean could significantly affect the results. Therefore, before calculating the mean, the fluctuation difference sequence needs to be preprocessed, for example, by using filtering algorithms to remove outliers or correcting them. After preprocessing, the average fluctuation difference for each dimension is calculated. This yields results that more accurately reflect the average correlation between adjacent regions and the benchmark control point in that dimension.

[0077] When calculating the regional correlation, the average fluctuation difference is divided by the cosine similarity of the reference control point in the corresponding dimension. Cosine similarity is an index that measures the directional similarity of two vectors. By calculating the cosine similarity between the average fluctuation difference vector and the reference control point vector in the corresponding dimension, their directional similarity can be obtained. Dividing the average fluctuation difference by this cosine similarity comprehensively considers the magnitude and direction of the fluctuation, yielding the regional correlation for that dimension. This calculation is performed for each dimension, and then the regional correlations for all dimensions are comprehensively processed. The comprehensive processing method can assign different weights to different dimensions based on their importance. For example, voltage amplitude and phase dimensions, which are crucial for voltage stability, can be given higher weights, while some less important dimensions can be given lower weights. Through weighted summation and other methods, the overall regional correlation is finally obtained, which comprehensively reflects the tightness of the power grid correlation between various parts of the region.

[0078] When calculating inter-regional correlation, the mutual information algorithm is used to calculate the mutual information between the benchmark control point and cross-regional characteristics. The mutual information algorithm is a method to measure the degree of dependence between two variables. In power systems, the operating states of different regions are dependent on each other; for example, power changes in one region may affect the voltage and current in adjacent regions. By calculating the mutual information between the benchmark control point and cross-regional characteristics using the mutual information algorithm, a quantitative indicator of their dependence can be obtained. Since the value range of the mutual information may be large, it needs to be normalized to a range between 0 and 1 for easier subsequent comparison and calculation. The normalized mutual information is the inter-regional correlation, which accurately reflects the degree of grid correlation between the region where the benchmark control point is located and other cross-regional regions.

[0079] Finally, the harmonic mean of the inter-regional and inter-regional correlation quantities is calculated using the harmonic mean formula, serving as the grid correlation coefficient between this region and other regions. The harmonic mean is a special type of average that is more sensitive to smaller values, highlighting the impact of smaller correlation quantities when comprehensively considering both intra-regional and inter-regional correlation quantities. This is because both intra-regional and inter-regional correlation quantities are important in grid analysis; even a small correlation quantity can have a critical impact on the overall grid operation. The grid correlation coefficient obtained by calculating the harmonic mean comprehensively and objectively reflects the degree of grid correlation between this region and other regions, providing important reference for subsequent grid regulation and management. Using this grid correlation coefficient, the mutual influence relationships between different regions can be further analyzed, identifying key areas and weak links in the grid, thereby enabling targeted regulation strategies to improve the operational stability and reliability of distributed microgrids.

[0080] Example 3:

[0081] In the process of extracting key feature vectors, constructing a correlation matrix of the power grid correlation coefficients between different regions is the first step. This correlation matrix uses the power grid correlation coefficients between each region as matrix elements, thus forming a two-dimensional matrix. Each row and column of the matrix corresponds to a specific region, and the value of the matrix element reflects the magnitude of the power grid correlation coefficient between the corresponding two regions. For example, if the value of the element in the i-th row and j-th column of the matrix is ​​large, it indicates that the power grid correlation between region i and region j is relatively strong; conversely, if the value of the element is small, it indicates that the power grid correlation between the two regions is relatively weak. This matrix form can intuitively show the overall situation of power grid correlation between various regions in a distributed microgrid, providing a basic data structure for subsequent in-depth analysis.

[0082] After constructing the correlation matrix, the next step is to perform principal component analysis (PCA). The core principle of PCA is to transform the original data from its original coordinate system to a completely new coordinate system through a linear transformation. In this new coordinate system, the variance of the data can be maximized. This process generates a series of mutually orthogonal eigenvectors, each corresponding to the trend of data variation in a specific direction. The orientation of these eigenvectors in the new coordinate system is crucial because they determine the main direction of data variation across different dimensions. For example, some eigenvectors may reflect the main direction of power transmission variation between different regions, while others may reflect the correlation direction of voltage fluctuations between different regions. These eigenvectors obtained through PCA help us to more clearly understand the inherent structure and changing patterns of grid interconnections between different regions in a distributed microgrid.

[0083] Among the numerous eigenvectors obtained through principal component analysis, eigenvectors whose explained variance exceeds a set threshold need to be selected to form a key feature subspace. Eigenvalues ​​are important indicators of the contribution of each eigenvector to the data variance. Generally, the larger the eigenvalue, the greater the variance explained by that eigenvector, meaning it plays a more crucial role in describing the overall changes in the data. In this embodiment, the threshold is set to 80%. That is, only eigenvectors whose explained variance accounts for more than 80% of the total variance are selected. These selected eigenvectors can retain the information contained in the original data to the greatest extent. For example, in a distributed microgrid containing multiple regions, there may be dozens or even hundreds of eigenvectors, but by setting this threshold, the most representative few eigenvectors can be selected. The key feature subspace formed by these eigenvectors can effectively summarize the main information of the original data regarding the correlation coefficients of the power grid in different regions, greatly reducing the complexity of data processing while ensuring the integrity of key data information.

[0084] Mapping the correlation matrix to the key feature subspace is achieved using a linear transformation. A linear transformation is a mathematical operation that projects each element of the original correlation matrix onto the key feature subspace. Specifically, let the original correlation matrix be A, and the key feature subspace consist of a set of eigenvectors {v1, v2, ..., v...}. k} constitute (where k is the number of eigenvectors that satisfy the threshold condition), for any element a in matrix A ij Its projection in the key feature subspace can be calculated using the following formula: in '*' represents the projected element value, and '·' represents the dot product operation of the vectors. Through this projection operation, each element in the original correlation matrix is ​​transformed into a key feature subspace, thus obtaining the dimensionality-reduced key feature vectors. These key feature vectors not only greatly reduce the dimensionality of the data and significantly lower the complexity of subsequent calculations, but also completely preserve the key information between the correlation coefficients of different regional power grids. This enables us to formulate reasonable strategies more efficiently based on this concise but crucial information when analyzing and controlling distributed microgrids, thereby improving the stability and reliability of microgrid operation.

[0085] Example 4:

[0086] The control parameters are derived from calculations based on grid operation data, covering multiple aspects such as voltage control parameters, current control parameters, and power control parameters. These parameters reflect the operational status adjustment needs of the distributed microgrid in different dimensions. The first step in constructing a multidimensional grid control space is to map the control parameters to a multidimensional Euclidean space by timestamp. The dimensions of the multidimensional Euclidean space are determined based on the type of control parameter, with each parameter corresponding to a coordinate axis in the space. For example, voltage control parameters correspond to one coordinate axis, current control parameters to another, power control parameters to yet another, and so on, forming a high-dimensional spatial structure. Within this space, the control parameter values ​​at different times are used as coordinate values, thereby generating a control distribution point set. These point sets represent the positions of the distributed microgrid's operational control status at different points in time within the multidimensional space. Through the distribution of these point sets, the changing trends of the control status over time and the interrelationships between different control parameters can be intuitively observed.

[0087] After generating the control distribution point set, the kernel density estimation method is used to estimate its spatial density. Kernel density estimation is a nonparametric statistical method for estimating the spatial distribution density of data. Its principle is to place a kernel function at each data point; the kernel function can be understood as a distribution model centered on the data point, such as the common Gaussian kernel function. Then, all kernel functions are weighted and summed to obtain the density estimate for any point in space. During the calculation, the variance gradient and kurtosis coefficient of the estimated space also need to be considered. The variance gradient directly reflects the rate of density change in space. A large variance gradient indicates rapid density change within that region, meaning the control state transitions drastically between different locations; conversely, a small variance gradient indicates relatively gentle density change and a relatively stable control state. The kurtosis coefficient measures the steepness of the density distribution. A large kurtosis coefficient means the density distribution is more concentrated, with most data points concentrated in a certain area; a small kurtosis coefficient indicates a more dispersed density distribution, with data points distributed relatively evenly in space. Analyzing these parameters allows for a deeper understanding of the distribution characteristics of the control distribution point set in multidimensional space.

[0088] In actual measurement and calculation processes, data may be affected by various random disturbances, leading to deviations in density estimation results. To eliminate the influence of these random disturbances, median filtering is needed to remove noise from the gradient distribution. The median filtering algorithm replaces the gradient value of each data point with the median of the gradient values ​​of its neighboring data points. For example, for a specific data point, other data points within a certain range are selected, their gradient values ​​are sorted by magnitude, and the median value is taken as the new gradient value for that specific data point. In this way, abnormal gradient values ​​caused by random disturbances can be effectively removed, making the gradient distribution smoother and more accurate, thereby improving the reliability of the density estimation results.

[0089] After completing the above preparations, the next step is to extract outliers and fluctuating abrupt regions. The extraction process is based on the previously calculated density estimation results. First, an outlier threshold is set, which needs to comprehensively consider the normal operating range of the microgrid and the distribution of historical data. When the density value of a point is significantly lower than that of the surrounding area, it is identified as an outlier. This is because outliers usually represent abnormal conditions in the microgrid's operation, possibly due to equipment failure, sudden interference, or other reasons causing a significant difference between the control state and normal conditions. Fluctuating abrupt regions are extracted by comparing the density change rates of adjacent areas. When the density change rate between adjacent areas exceeds the set abrupt threshold, the area is identified as a fluctuating abrupt region. The appearance of a fluctuating abrupt region may indicate a rapid and significant change in the microgrid's operating state within that area, such as due to the sudden connection or disconnection of distributed power sources, or a sudden change in large-scale load. Accurately identifying these outliers and fluctuating abrupt regions is crucial for timely detection of potential problems in microgrid operation.

[0090] Once the outlier density exceeds a set threshold or the range of fluctuating abrupt changes exceeds a limit, the system will determine it to be in an abnormal control state. At this point, the system will generate different levels of warning commands based on the severity of the anomaly. The classification of warning command levels is usually based on pre-defined rules. For example, a minor anomaly might correspond to a yellow warning command; at this point, although some abnormalities have occurred, they have not yet seriously affected the normal operation of the microgrid, and staff can inspect and handle them at an appropriate time. A moderate anomaly generates an orange warning command, indicating that the anomaly is more obvious and may have affected some functions of the microgrid, requiring staff to pay attention and take appropriate measures as soon as possible. A severe anomaly generates a red warning command, meaning that the anomaly has seriously threatened the safe and stable operation of the microgrid, and emergency measures must be taken immediately to avoid serious consequences such as large-scale power outages. Through this multi-level warning mechanism, staff can be promptly alerted to the operating status of the microgrid and take effective countermeasures to ensure the reliable operation of the microgrid.

[0091] Simultaneously, this invention also includes establishing a joint optimization model for distributed microgrid operation data and control strategies, using Newton's method to solve for the optimal control scheme, and coupling the optimal control scheme with control parameters in real time to generate the optimal control strategy for grid operation. When establishing the joint optimization model, the objective function and constraints must first be clearly defined. The objective function is defined as the weighted absolute value sum of the control deviation and control costs. The control deviation is measured by calculating the difference between the actual control parameters and the ideal control parameters; it reflects the gap between the current control scheme and the optimal control state. The control costs consider factors such as energy consumption and equipment losses during the control process, because in actual control, not only the control effect but also the control cost must be considered. The constraints are set as the grid operation safety boundary to ensure that the microgrid operation is always within a safe and reliable range during control, and that the grid does not experience safety problems such as voltage exceeding limits or overload due to excessive pursuit of control effects.

[0092] By applying a Lagrange dual transformation to the objective function, the original complex optimization problem is cleverly decomposed into a control strategy subproblem and an operational data subproblem. This decomposition makes large-scale optimization problems, which were originally difficult to solve directly, much easier to handle. In the iterative process of solving the two subproblems, the solution to the operational data subproblem is first fixed, and efforts are focused on solving the control strategy subproblem. During this process, based on the given operational data, the optimal control strategy is sought to minimize the objective function. After obtaining the optimal control strategy under the current operational data, the solution to the control strategy subproblem is fixed again, and the operational data subproblem is solved. At this point, based on the determined control strategy, the optimal operational data configuration is sought, again with the objective function minimized. This iterative process is repeated until the solutions to both subproblems converge, reaching a stable state. The output at this point is the optimal control scheme that satisfies the constraints. This optimal control scheme comprehensively considers the microgrid's operational data and control strategy, achieving the best balance between control effectiveness and control cost while ensuring the safe operation of the power grid.

[0093] Finally, the optimal control scheme is coupled with the control parameters in real time. Since the microgrid's operating state is dynamic, the control parameters are also updated in real time. Through real-time coupling, the optimal control scheme is dynamically adjusted based on the latest control parameters, enabling the control strategy to adapt promptly to changes in the microgrid's operating state. For example, when the power generation of a distributed power source in the microgrid suddenly decreases due to weather conditions, the control parameters will change accordingly. Through real-time coupling, the optimal control scheme can quickly adjust, potentially increasing the power generation of other distributed power sources or adjusting load distribution to maintain the microgrid's power balance and stable operation. The resulting control strategy truly optimizes grid operation, improves the microgrid's operating efficiency and reliability, and provides users with a more stable and reliable power supply.

[0094] Furthermore, this invention also includes data verification of the operating feature sequence, extracting the difference between the verification features and preset standard features; inputting the difference into a support vector machine classifier to determine the type of data anomaly; generating data correction instructions based on the anomaly type, and adjusting the operating feature sequence based on the correction instructions. In the data verification stage, the preset standard features are determined based on a large amount of historical data accumulated by the distributed microgrid under normal operating conditions, combined with theoretical knowledge and experience of power system operation. These preset standard features represent the characteristic values ​​and variation patterns of various parameters of the microgrid under ideal operating conditions. The differences between the verification features and preset standard features in various dimensions are extracted, such as a detailed comparison in dimensions like voltage amplitude, current phase, and power factor, and these differences are used as input to the support vector machine classifier. The support vector machine classifier is a powerful classification tool based on statistical learning theory. By finding an optimal classification hyperplane in a high-dimensional space, it can accurately distinguish different types of data anomalies. Based on the classification results, the specific type of data anomaly is determined; it may be a voltage anomaly, such as excessively high or low voltage; it may be a current anomaly, such as excessive current fluctuations or harmonics; or it may be a power anomaly, such as power imbalance. For different anomaly types, corresponding data correction instructions are generated. For example, for voltage anomalies, if the voltage is too high, instructions may be generated to reduce the output voltage of the distributed power source or to activate reactive power compensation devices to absorb excess reactive power and lower the voltage; for current anomalies, if harmonic issues exist, instructions may be generated to activate harmonic mitigation equipment for filtering. Finally, based on these correction instructions, the operating characteristic sequence is adjusted. By correcting and optimizing relevant parameters, the operating characteristic sequence can more accurately reflect the actual operating status of the distributed microgrid, providing a reliable data foundation for subsequent analysis and control, and further ensuring the stable operation of the microgrid.

[0095] It should be noted that, in this document, relational terms such as "first" and "second" are used only to distinguish one entity or operation from another, and do not necessarily require or imply any such actual relationship or order between these entities or operations. Furthermore, the terms "comprising," "including," or any other variations thereof are intended to cover non-exclusive inclusion, such that a process, method, article, or apparatus that comprises a list of elements includes not only those elements but also other elements not expressly listed, or elements inherent to such process, method, article, or apparatus.

[0096] Although embodiments of the invention have been shown and described, it will be understood by those skilled in the art that various changes, modifications, substitutions and alterations can be made to these embodiments without departing from the principles and spirit of the invention, the scope of which is defined by the appended claims and their equivalents.

Claims

1. A method for grid-connected protection and intelligent power balance control of distributed microgrids, characterized in that, include: Collect multi-source operation data of distributed microgrids, wherein the operation data includes grid voltage data, current data and power data; An intelligent analysis model is used to extract features from the operating data to obtain an operating feature sequence, and the power grid correlation coefficient is calculated based on the fluctuations of the operating feature sequence. The correlation matrix of the power grid correlation coefficients in different regions is reduced in dimension, key feature vectors are extracted, and the key feature vectors are transformed by probability distribution using the Sigmoid function to obtain the power grid regulation weights for each region. The baseline control points of the operating characteristic sequence are dynamically adjusted according to the grid control weights to generate control parameters, and the intelligent control records of the distributed microgrid are updated based on the control parameters.

2. The intelligent control method for grid-connected protection and power balance of a distributed microgrid according to claim 1, characterized in that, The feature extraction includes: The running data is extracted according to the preset encoding rules, a feature tensor is constructed, and feature embedding processing is performed. A multi-head attention layer is used to perform global feature association on the feature tensor, and the output attention score is a weighted fusion of the feature vector. The fusion result is input into a long short-term memory network, multi-scale temporal features are extracted through time step convolution, and the feature dimensions are filtered through an attention mechanism to generate a running feature sequence.

3. The intelligent control method for grid-connected protection and power balance of a distributed microgrid according to claim 1, characterized in that, The calculation of the power grid correlation coefficient includes: Select the operating characteristic sequence of any region as the benchmark control point, and calculate the fluctuation difference of the characteristic vector of its n adjacent regions; The ratio of the mean of the fluctuation difference to the cosine similarity of the benchmark control point is used as the correlation quantity within the region; Calculate the mutual information between the benchmark control point and cross-regional characteristics, and take its normalized value as the inter-regional correlation quantity. The harmonic mean of the correlation quantities within the region and the correlation quantities between regions is used as the power grid correlation coefficient.

4. The intelligent control method for grid-connected protection and power balance of a distributed microgrid according to claim 1, characterized in that, The extraction of key feature vectors includes: Construct a correlation matrix of the power grid correlation coefficients in different regions, and perform principal component analysis to obtain orthogonal eigenvectors. Select eigenvectors whose variance exceeds a set threshold to form a key feature subspace; The correlation matrix is ​​mapped to the key feature subspace to obtain the dimensionality-reduced key feature vector.

5. The intelligent control method for grid-connected protection and power balance of a distributed microgrid according to claim 1, characterized in that, The obtained power grid regulation weights for each region include: The key feature vectors are standardized, and their dot product similarity with the preset control vector is calculated. The similarity values ​​are input into a fully connected neural network, and the initial weights are generated after transformation by the hidden layer and the Tanh activation function. The initial weights are exponentially smoothed using a time-sliding window to output the power grid control weights for each region.

6. The intelligent control method for grid-connected protection and power balance of a distributed microgrid according to claim 1, characterized in that, The generated control parameters include: The weight adjustment parameters are obtained by multiplying the power grid control weights and the benchmark control points by matrix position. The deviation between the weight adjustment parameter and the benchmark control point is calculated, and the deviation is corrected by state estimation using a Kalman filter. The corrected deviation is superimposed on the benchmark control point to generate the control parameters.

7. The intelligent control method for grid-connected protection and power balance of a distributed microgrid according to claim 1, characterized in that, Also includes: A multidimensional power grid control space is constructed based on control parameters, and outliers and fluctuation change regions are extracted within the space. When the density of outliers exceeds the set threshold or the range of fluctuating abrupt changes is greater than the limit, it is determined to be an abnormal control state, and multi-level early warning instructions are generated.

8. The intelligent control method for grid-connected protection and power balance of a distributed microgrid according to claim 7, characterized in that, The construction of a multi-dimensional power grid control space includes: The control parameters are mapped to a multidimensional Euclidean space according to the timestamp to generate a control distribution point set; The kernel density estimation method is used to estimate the spatial density of the control distribution point set, and the variance gradient and kurtosis coefficient of the estimated space are calculated. The gradient distribution is filtered out of noise using a median filtering algorithm to eliminate the effects of random disturbances.

9. The intelligent control method for grid-connected protection and power balance of a distributed microgrid according to claim 1, characterized in that, Also includes: A joint optimization model of distributed microgrid operation data and control strategies is established, and the optimal control scheme is solved using Newton's method. The optimal control scheme is coupled with the control parameters in real time to generate the optimal control strategy for power grid operation. The solution to the joint optimization model includes: The objective function is defined as the weighted sum of the absolute values ​​of the control deviation and the control overhead, and the constraint condition is the power grid operation safety boundary. The objective function is decomposed into a control strategy subproblem and a running data subproblem by performing a Lagrange dual transformation. The two subproblems are solved iteratively and alternately until convergence, and the optimal control scheme that satisfies the constraints is output.

10. The intelligent control method for grid-connected protection and power balance of a distributed microgrid according to claim 1, characterized in that, Also includes: Perform data validation on the running feature sequence and extract the difference between the validation feature and the preset standard feature; Input the discrepancy values ​​into a support vector machine classifier to determine the type of data anomalies; Data correction instructions are generated based on the anomaly type, and the running characteristic sequence is adjusted based on the correction instructions.

Images