A full-quantity user voltage out-of-limit automatic identification method, system, device and medium

Abstract

The application provides a full-user voltage overrun automatic identification method, system, device and medium, fuses multiple power grid data and generates a typical scene library through clustering, solves the problem of incomplete scene coverage under high proportion of new energy, uses expert rules and particle swarm optimization variational mode decomposition-Stacking integrated learning double-path identification to greatly improve efficiency, and uses a combination weighting method to construct a voltage overrun sensitivity model and quantify the risk level, realizes fine diagnosis of user overrun, and completely overcomes the defects of incomplete scene coverage, low identification efficiency and poor pertinence of traditional artificial experience methods; the method comprises the following steps: acquiring multiple power grid data and preprocessing; clustering to generate a typical voltage overrun scene library; extracting key factors and weighting to construct a sensitivity model; double identification through expert rules and PSO-VMD+Stacking integrated learning; calculating a risk value and outputting a report.

Description

Technical Field

[0001] This invention relates to the field of power engineering technology, specifically to a method, system, device, and medium for automatic identification of voltage over-limit for all users. Background Technology

[0002] Voltage quality is a key indicator for evaluating the operational level of a power distribution network, directly impacting the normal operation of user equipment and electricity safety. According to statistics from the State Grid Corporation of China, the national power distribution network voltage qualification rate was 99.82% in 2022, but voltage exceedances still occurred for 0.18% of the time, equivalent to approximately 15.8 hours of voltage exceedance throughout the year. Voltage exceedances not only affect the safe operation of user equipment but can also lead to damage to power grid equipment and even system collapse.

[0003] With the large-scale integration of distributed energy resources into the distribution network, the traditional radial distribution network structure is transforming into a multi-source structure, which makes voltage control more complex. The grid connection of intermittent power sources such as photovoltaic and wind power leads to more frequent voltage fluctuations in the distribution network, significantly increasing the risk of voltage exceeding limits.

[0004] Meanwhile, with the deepening of the construction of new power systems, the large-scale grid connection of distributed new energy sources and the diversified development of user loads (such as the widespread access of electric vehicles and energy storage devices) have made the voltage operation characteristics of the power grid increasingly complex. Data from the National Energy Administration shows that in 2023, complaints about voltage exceeding limits by users of 10kV and below distribution networks accounted for 41.2% of all power supply quality complaints nationwide, and the delays in handling these complaints due to the untimely identification of the causes of the voltage exceeding limits accounted for more than 60%.

[0005] Traditional voltage management relies on manual inspection, using experience-based identification methods such as manual patrols, judgment, and qualitative analysis. These methods are passive, lagging, and inefficient, suffering from incomplete scenario coverage, low identification efficiency, and poor targeting, and can no longer meet the requirements of "reliable, high-quality, and efficient" power supply services. Furthermore, with the large-scale integration of renewable energy into the grid, the operating characteristics of the power grid have fundamentally changed, manifesting as: spatiotemporal mismatch between source and load and bidirectional power flow, strong volatility and uncertainty, the interactive complexity brought about by power electronics, and an increase in the dimensions of power system operation. This makes traditional manual experience-based voltage limit identification methods inadequate for handling the complex scenarios under high-proportion renewable energy integration. Therefore, constructing typical voltage limit scenarios and achieving automatic identification is of significant importance and value. Summary of the Invention

[0006] This invention aims to address the technical deficiencies of existing technologies by providing a method, system, device, and medium for automatic identification of voltage overruns for all users. By fusing multi-source power grid data and clustering to generate a typical scenario library, it solves the problem of incomplete scenario coverage under high-proportion renewable energy sources. It utilizes a dual-path identification approach combining expert rules and particle swarm optimization variational mode decomposition-Stacking ensemble learning to significantly improve efficiency. Furthermore, it constructs a voltage overrun sensitivity model and quantifies risk levels using a combined weighting method, achieving refined diagnosis of user overruns. This comprehensively overcomes the shortcomings of traditional manual experience methods, such as incomplete scenario coverage, low identification efficiency, and poor targeting.

[0007] This invention provides the following technical solution: The first aspect discloses a method for automatic identification of voltage over-limit for all users, comprising the following steps:

[0008] S1. Acquire multi-source power grid data from the dispatching system, electricity consumption information collection system, power distribution automation system, new energy monitoring system and power grid GIS platform, and sequentially perform data cleaning, time-series alignment and feature extraction on the multi-source power grid data to obtain processed data;

[0009] S2. Based on the processed data, perform cluster analysis to generate a scenario library containing multiple typical voltage over-limit scenarios;

[0010] S3. Based on the processed data, extract multiple key factors affecting voltage over-limit, use the combined weighting method to determine the weight of each key factor, and construct a voltage over-limit sensitivity quantitative model based on the weight of each key factor.

[0011] S4. Match the processed data with the scene library, and perform reasoning through the preset expert rule library to obtain the first identification result;

[0012] The variational mode decomposition method based on particle swarm optimization algorithm is used to decompose multiple key factors to obtain decomposition results. The decomposition results are then input into the Stacking ensemble learning model for identification to obtain a second identification result.

[0013] S5. Based on the first identification result and the second identification result, calculate the voltage over-limit risk value of the current state based on the voltage over-limit sensitivity quantification model, determine the risk level based on the voltage over-limit risk value, and output the final identification report based on the risk level.

[0014] Furthermore,

[0015] The clustering analysis in S2 uses the K-means++ algorithm and uses the silhouette coefficient to evaluate the clustering effect, so as to generate a multi-level typical voltage over-limit scenario library including substation layer, line layer, distribution transformer layer, user layer and new energy layer.

[0016] Furthermore,

[0017] The key factors in S3 include:

[0018] The main transformer capacity, tap position, short-circuit impedance, compensation capacity, and switching frequency at the substation level;

[0019] Line length, resistance, reactance, line load factor, and current of the line layer;

[0020] Distribution transformer capacity, tap level, load factor, and three-phase current imbalance of the distribution transformer layer;

[0021] Maximum load and load volatility at the user level;

[0022] The access capacity, output fluctuation rate, and power factor of the new energy layer.

[0023] Furthermore,

[0024] The combined weighting method in S3 is as follows:

[0025] The subjective weights of each key factor are determined using the eigenvector method:

[0026] A pairwise comparison judgment matrix, denoted as A, is constructed using a 1-9 scale among the key influencing factors for voltage limit exceedance. In the formula: n is the number of key influencing factor categories involved in subjective weighting, a ij F represents the category of the i-th key influence factor. i Relative to the j-th key impact factor category F j The importance comparison of the degree of impact on the risk of voltage exceeding limits;

[0027] A consistency check is then performed, including calculating the consistency index CI and the consistency ratio CR, using the following formulas:

[0028] ,

[0029] ,

[0030] In the formula: λmax is the largest eigenvalue of the judgment matrix, RI is the average random consistency index, and when CR < 0.1, the judgment matrix A satisfies the consistency requirement;

[0031] Solve the system of equations AW=λmaxW to obtain the eigenvector W. After normalization, the subjective weights of each factor are:

[0032] ,

[0033] In the formula: W j W represents the weight of the j-th evaluation factor. i Let be the weight of the i-th evaluation factor;

[0034] The objective weights of each key factor are determined using the information entropy method:

[0035] Construct a standardized decision matrix by standardizing the factor data to obtain the matrix. , where r ij Let m be the element in the i-th row and j-th column of matrix R, m be the number of samples, n be the number of factors, and the information entropy of the j-th factor be... The calculation formula is as follows:

[0036] ,

[0037] In the formula: ,like ,but ,

[0038] The objective weight of the j-th factor is:

[0039] ,

[0040] The combined weights are calculated using a linear weighting method. :

[0041] ,

[0042] The coefficient for subjective weight is 0.4, and the coefficient for objective weight is 0.6.

[0043] Furthermore,

[0044] The voltage over-limit sensitivity quantification model in S3 is constructed based on combined weights, and its expression is as follows:

[0045] ,

[0046] In the formula: S is the voltage over-limit sensitivity value; x j x is the actual value of the j-th key factor; j0 Let j be the standard value of the j-th key factor; Let be the combined weight of the j-th key factor.

[0047] Furthermore,

[0048] The expert rule base preset in S4 includes: rules for unbalanced three-phase load of distribution transformers, rules for unreasonable distribution transformer taps, rules for lack of coordination between main and distribution networks, rules for line overload, rules for line light load, and rules for photovoltaic backfeeding.

[0049] Furthermore,

[0050] In step S4, the decomposition result is input into the Stacking ensemble learning model for identification to obtain the second identification result. The steps include:

[0051] S4-1. Parameter optimization of variational mode decomposition algorithm based on particle swarm optimization algorithm;

[0052] The particle swarm optimization algorithm is used to search for variational mode decomposition algorithms globally and perform optimization combinations. The optimal variational mode decomposition algorithm is obtained by iteratively finding the most reasonable penalty factor, the number of decomposition layers, and the optimal number of modes.

[0053] S4-2, Key Factor Extraction and Signal Decomposition;

[0054] The optimal variational mode decomposition algorithm is used to decompose the original voltage signal, and the key factors affecting voltage over-limit are accurately identified and extracted from the complex original data.

[0055] S4-3, Constructing a primary meta-model group;

[0056] Three heterogeneous neural networks were constructed and trained in parallel as meta-models for the Stacking framework, including a multilayer perceptron (MLP) for capturing nonlinear static features of data, a long short-term memory (LSTM) network for mining long-range time series correlations of voltage waveforms, and a gated recurrent unit (GRU) for optimizing computational efficiency.

[0057] S4-4, Model fusion based on stacking strategy;

[0058] The Stacking ensemble learning algorithm is adopted. The prediction results of the three meta-models in S4-3 are used as input and fed into the secondary learner. The advantages of each meta-model are integrated through the fusion strategy to correct the prediction bias of the single model and construct the Stacking ensemble learning model.

[0059] S4-5. Identification of voltage over-limit results;

[0060] The trained Stacking ensemble learning model is used to identify whether the output voltage exceeds the limit in real-time or test data.

[0061] The second aspect discloses a full-user voltage over-limit automatic identification system, including:

[0062] The acquisition module is used to acquire multi-source power grid data from the dispatching system, electricity consumption information collection system, distribution automation system, new energy monitoring system, and power grid GIS platform; the multi-source power grid data is sequentially cleaned, time-series aligned, and feature extracted to obtain the processed data;

[0063] The analysis module is used to perform cluster analysis based on the processed data using clustering algorithms to generate a scenario library containing multiple typical voltage over-limit scenarios.

[0064] A module is constructed to extract multiple key factors affecting voltage over-limit based on processed data; a combined weighting method is used to determine the weight of each key factor, and a voltage over-limit sensitivity quantitative model is constructed based on the weight of each key factor.

[0065] The identification module is used to match the processed data with the scene library and perform reasoning through a preset expert rule base to obtain the first identification result;

[0066] The variational mode decomposition method based on particle swarm optimization algorithm is used to decompose multiple key factors to obtain decomposition results. The decomposition results are then input into the Stacking ensemble learning model for identification to obtain a second identification result.

[0067] The output module is used to calculate the voltage over-limit risk value of the current state based on the voltage over-limit sensitivity quantification model according to the first identification result and the second identification result; determine the risk level according to the voltage over-limit risk value; and output the final identification report according to the risk level.

[0068] The third aspect discloses an automatic identification device for voltage over-limit of all users, including a processor and a memory, wherein the processor implements the automatic identification method for voltage over-limit of all users when executing a computer program stored in the memory.

[0069] The fourth aspect discloses a computer-readable storage medium for storing a computer program, wherein the computer program, when executed by a processor, implements a method for automatic identification of voltage over-limit for all users.

[0070] Compared with the prior art, the present invention has the following technical effects:

[0071] This invention systematically solves three major pain points of traditional manual experience methods in the context of high proportion of new energy sources by adopting a technical chain of full data fusion → full scenario coverage → dual-path efficient identification → factor quantification and rating:

[0072] (1) Incomplete scenario coverage: Automatic clustering generates a typical scenario library that encompasses the complex patterns brought about by new energy fluctuations; This invention integrates multi-source data such as scheduling system, electricity information collection system, distribution automation system, new energy monitoring system and power grid GIS platform, covering the complex operating characteristics such as power fluctuations and bidirectional power flow generated after the access of high proportion of new energy (such as photovoltaic and wind power). Through cluster analysis, a scenario library containing multiple typical voltage over-limit scenarios is automatically generated from massive historical data, which can cover intermittent and random voltage over-limit modes that are difficult to enumerate with traditional manual experience (such as transient over-limit caused by sudden changes in new energy output, and the interactive influence of multiple feeders, etc.), completely changing the limitation of manual experience relying on limited typical operating conditions, and realizing full coverage of new energy high penetration scenarios;

[0073] (2) Low identification efficiency: Parallel rule reasoning and optimization ensemble learning model to achieve fast and automatic identification of all users; Two parallel identification paths are designed: ① Fast reasoning based on expert rule base to achieve millisecond-level response to common over-limit patterns; ② Adaptive decomposition of key factors by variational mode decomposition (PSO-VMD) improved by particle swarm optimization algorithm, and then input into Stacking ensemble learning model for deep identification. This method automatically decomposes complex nonlinear and non-stationary voltage sequences and integrates multiple base learners. No manual trial and error and parameter tuning are required, which greatly shortens the analysis time. The risk value is automatically calculated and the final report is output. No manual investigation of each household is required. Compared with manual experience identification, the efficiency is improved by several times to tens of times, especially in the scenario of massive users (all users).

[0074] (3) Poor targeting: The combined weighted sensitivity model and risk classification output differentiated and operable identification reports, ultimately providing grid operators with a comprehensive, fast and accurate auxiliary decision-making tool for voltage overruns, which strongly supports the safe and stable operation of a high proportion of new energy grids; The combined weighting method (integrating subjective and objective weights) is used to determine the weights of multiple key factors affecting voltage overruns, and a voltage overrun sensitivity quantitative model is constructed. This model can distinguish the dominance of each factor under different users, different time periods and different operating modes, and achieve "one feature per scenario". Based on the first and second identification results, the voltage overrun risk value of the current state is calculated in combination with the sensitivity model, and the risk level (such as high, medium and low) is divided. The final identification report can accurately point to high-risk users, overrun dominant factors and suggested disposal measures, which changes the rough judgment of manual experience "one-size-fits-all" and realizes refined and personalized diagnosis and early warning for each user / each overrun event. Attached Figure Description

[0075] Figure 1 This is a flowchart of the method of the present invention;

[0076] Figure 2This is a schematic diagram of the framework for the Stacking ensemble learning model. Detailed Implementation

[0077] The technical solutions of the present invention will be clearly and completely described below with reference to the accompanying drawings in the embodiments of the present invention. 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.

[0078] See Figure 1 This embodiment provides a method for automatic identification of voltage over-limit for all users, including the following steps:

[0079] S1. Acquire multi-source power grid data from the dispatching system, electricity consumption information collection system, power distribution automation system, new energy monitoring system and power grid GIS platform, and sequentially perform data cleaning, time-series alignment and feature extraction on the multi-source power grid data to obtain processed data;

[0080] S2. Based on the processed data, perform cluster analysis to generate a scenario library containing multiple typical voltage over-limit scenarios;

[0081] S3. Based on the processed data, extract multiple key factors affecting voltage over-limit, use the combined weighting method to determine the weight of each key factor, and construct a voltage over-limit sensitivity quantitative model based on the weight of each key factor.

[0082] S4. Match the processed data with the scene library, and perform reasoning through the preset expert rule library to obtain the first identification result;

[0083] The variational mode decomposition method based on particle swarm optimization algorithm is used to decompose multiple key factors to obtain decomposition results. The decomposition results are then input into the Stacking ensemble learning model for identification to obtain a second identification result.

[0084] S5. Based on the first identification result and the second identification result, calculate the voltage over-limit risk value of the current state based on the voltage over-limit sensitivity quantification model, determine the risk level based on the voltage over-limit risk value, and output the final identification report based on the risk level.

[0085] The present invention will be further explained below with reference to the accompanying drawings and specific embodiments:

[0086] Step 1: In this embodiment, the following multi-source power grid data is obtained by acquiring data from the dispatching system, electricity consumption information collection system, distribution automation system, new energy monitoring system, and power grid GIS platform:

[0087] Dispatch system: Provides data such as substation bus voltage, line power flow, and transformer tap position, with a time resolution typically of 1-5 minutes;

[0088] Electricity Information Acquisition System: Provides users with measurement data such as voltage, current, and power, with a time resolution of 15 minutes;

[0089] Power distribution automation system: provides information such as the status of power distribution equipment and switch positions;

[0090] New energy monitoring system: provides power output data for distributed photovoltaic, wind power and other new energy sources;

[0091] Power Grid GIS Platform: Provides static data such as power grid geographic information and equipment parameters.

[0092] Since the aforementioned multi-source power grid data is heterogeneous, it is necessary to handle missing values ​​and outliers separately.

[0093] This embodiment uses linear interpolation to fill in data with fewer than 5 consecutive missing points:

[0094] (1)

[0095] In the formula: x t x represents the voltage value at the current moment. t-1 x represents the voltage value at the previous moment. t+1 This indicates the voltage value at the next moment.

[0096] For data with 5-20 consecutive missing points, the Autoregressive Integrated Moving Average (ARIMA) is used, where the autoregressive model AR can be expressed as:

[0097] (2)

[0098] In the formula: c is a constant term, φi is the model regression coefficient, and ε t AR(P) represents white noise error, indicating that the current value can be represented by a linear combination of the past p voltage historical values ​​plus a random error.

[0099] The moving average (MA) model indicates that the current value can be represented by a linear combination of voltage random error terms over the past q periods:

[0100] (3)

[0101] In the formula: μ is the average value of the voltage sequence, θ i ε represents the coefficients of the moving average model. t This represents the white noise error term for the current and past time points.

[0102] Difference is used to transform a non-stationary voltage sequence into a stationary voltage sequence. A d-order difference indicates performing d-fold difference operations.

[0103] (4)

[0104] By combining the AR, MA, and difference methods above, we can obtain the ARIMA model.

[0105] That is, an ARMA(p,q) model is established for the differentially differentiated stationary voltage sequence Yt, and the specific expression is as follows:

[0106] (5)

[0107] In the formula: Yt = △dxt.

[0108] In this embodiment, abnormal voltage values ​​are first detected using the 3σ criterion.

[0109] Suppose a voltage dataset is as follows:

[0110] (6)

[0111] Its calculated standard deviation is:

[0112] (7)

[0113] (8)

[0114] If |x i If -μ|>3σ, then the voltage value can be determined to be an abnormal value.

[0115] In this embodiment, since the time resolution of data from different systems and devices cannot be kept consistent, time-series alignment processing is also required. The processing strategy is as follows:

[0116] (a) Downsampling high-resolution data to low resolution;

[0117] (b) Upsample low-resolution data to high resolution using interpolation methods;

[0118] (c) Use unified timestamps to ensure data synchronization from heterogeneous sources.

[0119] Considering the influence of spatiotemporal, load, and renewable energy characteristics in distribution network voltage data, the feature extraction in this embodiment includes extracting the following features to improve the accuracy of the distribution network voltage over-limit model:

[0120] (a) Time characteristics, including hours, weekdays / weekends, seasons, etc.;

[0121] (b) Load characteristics, including active power, reactive power, power factor, load factor, etc.;

[0122] (c) Power grid characteristics, including line length, cross-sectional area, impedance parameters, etc.;

[0123] (d) Characteristics of new energy power generation, including photovoltaic power output rate, wind power output rate, penetration rate, etc.;

[0124] (e) Voltage-related characteristics of the distribution network, including voltage deviation rate, voltage fluctuation rate, etc.

[0125] Step 2: In this embodiment, the K-means++ algorithm is used to cluster 1.26 million historical out-of-limit cases. The algorithm steps are as follows:

[0126] (a) Randomly select a sample as the first cluster center;

[0127] (b) Calculate the shortest distance D(x) between each sample and the existing cluster centers;

[0128] (c) According to probability Select the next cluster center;

[0129] (d) Repeat steps (b)-(c) until k cluster centers are selected;

[0130] (e) Perform the standard K-means algorithm.

[0131] The silhouette coefficient was used to evaluate the clustering effect.

[0132] (9)

[0133] In the formula: a(i) represents the average distance from the voltage sample to all other samples in its cluster. The smaller the value, the more the sample belongs to the cluster; b(i) represents the distance from the voltage sample to its nearest neighboring cluster. The larger the value, the better the cluster of the sample is separated from other clusters; s(i) represents the silhouette coefficient, which ranges from [-1, 1]. The closer the value is to 1, the better the clustering effect.

[0134] As shown in Table 1, a cluster analysis was performed on 1.26 million kV voltage over-limit events in a certain region in a certain year. The results generated 6 major categories and 18 subcategories of scenarios, including substation level, line level, distribution transformer level, user level, new energy grid connection and composite factor over-limit scenarios.

[0135] As shown in Table 1, the distribution transformer level (accounting for 28.02%) and the line level (accounting for 27.3%) are the main levels at which voltage over-limit occurs. Among them, "main and distribution network not coordinated", "three-phase load imbalance of distribution transformer", and "unreasonable distribution transformer tap" are the three most frequent typical scenarios.

[0136] Table 1:

[0137]

[0138] Step 3: This embodiment is based on the full path analysis of "station-line-transformer-customer". Combining physical mechanisms and data mining results, it identifies 26 key influencing factors in 8 categories for voltage exceeding limits. These factors are classified according to their impact level, as shown in Table 2. They include: main transformer capacity, tap position, short-circuit impedance, compensation capacity, and switching frequency at the substation level; line length, resistance, reactance, line load rate, and current at the line level; distribution transformer capacity, tap position, load rate, and three-phase current imbalance at the distribution transformer level; maximum load and load fluctuation rate at the user level; and access capacity, output fluctuation rate, and power factor at the renewable energy level.

[0139] Table 2:

[0140]

[0141] This embodiment uses a combined weighting method to calculate the weights:

[0142] AHP Subjective Weights: Ten experts were invited to construct a pairwise comparison judgment matrix among the key influencing factors of each voltage limit exceeding the limit using a 1-9 scale, denoted as A.

[0143] In the formula: n is the number of key influencing factor categories involved in subjective weighting, a ij F represents the category of the i-th key influence factor. i Relative to the j-th key impact factor category F j The importance comparison of the degree of impact on the risk of voltage exceeding limits;

[0144] Consistency check: Calculate the consistency index CI and the consistency ratio CR, using the following formulas:

[0145] (10)

[0146] (11)

[0147] In the formula: λmax is the largest eigenvalue of the judgment matrix, and RI is the average random consistency index (RI=1.41 when n=8).

[0148] When CR < 0.1, the judgment matrix satisfies the consistency requirement.

[0149] Constructing a standardized decision matrix: Standardize the factor data to obtain the matrix. , where r ij Let m represent the element in the i-th row and j-th column of matrix R, where m is the number of samples and n is the number of factors.

[0150] Calculate the subjective weights using the eigenvector method for each factor. Solve the system of equations AW=λ. max W, where λ max It is the largest eigenvalue of the matrix. After normalizing the eigenvector W, the subjective weights of each factor are:

[0151] (12)

[0152] In the formula W j W represents the weight of the j-th evaluation factor. i This represents the weight of the i-th evaluation factor.

[0153] Information entropy of the j-th factor The calculation formula is as follows:

[0154] (13)

[0155] In the formula: ,like ,but .

[0156] The objective weight calculation formula for the j-th factor is as follows:

[0157] (14)

[0158] The combined weights are calculated using a linear weighting method. The weighting coefficients for subjective and objective weights are 0.4 and 0.6 respectively, determined based on expert experience and data reliability analysis.

[0159] (15)

[0160] Constructing a sensitivity model based on combined weights:

[0161] (16)

[0162] In the formula: S is the voltage over-limit sensitivity value; x j x is the actual value of the j-th key factor; j0 Let j be the standard value of the j-th key factor; Let be the combined weight of the j-th key factor.

[0163] The larger the S value, the higher the risk of voltage over-limit in this scenario.

[0164] Table 3 shows the validation results of the sensitivity model. The results in Table 3 show that the overall goodness of fit of the model is 0.82 and the risk identification accuracy is 88.6%. The model performs best in the transformer layer and line layer scenarios, indicating that the model can effectively quantify the influence of each factor on voltage over-limit and can be used for early warning of over-limit risks.

[0165] Table 3:

[0166]

[0167] Table 4 shows the sensitivity contribution ranking of key factors. From Table 4, it can be analyzed that the transformer tap size (weight 0.21), line impedance (0.18), and photovoltaic capacity (0.15) are the three core factors. Among them, the model performance of the transformer layer and line layer scenarios is the best, indicating that the model can effectively quantify the influence of each factor on voltage over-limit and can be used for early warning of over-limit risk.

[0168] Table 4:

[0169]

[0170] Step 4: Based on the integration of expert experience and ensemble learning, this embodiment designs a three-layer architecture automatic identification system comprising a data input layer, a core processing layer, and a result output layer.

[0171] Data input layer: The processed data received from step one is accessed in real time through the data interface;

[0172] Core processing layer: Based on the typical voltage over-limit scenario constructed in step two, the expert decision system matches the new input voltage over-limit scenario with the typical scenario. At the same time, based on Stacking and machine learning methods, the key influencing factors extracted in step three are identified to determine the specific form of the input system voltage over-limit and the cause of the voltage over-limit.

[0173] Results output layer: Outputs identification results such as the type of violation, cause, risk level, and handling suggestions, automatically generates reports and pushes them out.

[0174] In this embodiment, the core processing layer consists of two identification layers. One identification layer is based on the experience of power experts, which establishes a rule base for identifying voltage over-limit conditions. The detailed rules are as follows:

[0175] (a) Three-phase load imbalance rules of distribution transformers

[0176] If the three-phase voltage imbalance is greater than 3%, the load rate is greater than 30%, and the three-phase current imbalance is greater than 25%, then the three-phase load imbalance of the distribution transformer exceeds the limit, with a confidence level of 0.9.

[0177] (b) Unreasonable rules for distribution transformer gear positions

[0178] (If the proportion of users exceeding the lower limit is >50% and the transformer outlet voltage is <0.9 times the rated voltage and the transformer's current tap is not at the highest tap) or (if the proportion of users exceeding the upper limit is >50% and the transformer outlet voltage is >1.1 times the rated voltage and the transformer's current tap is not at the lowest tap), then the distribution tap is unreasonable and exceeds the limit, with a confidence level of 0.9.

[0179] (c) Rules for non-coordination between primary and distribution networks

[0180] If (the distance between the transformer and the line is ≤20% of the power supply radius and the transformer tap is not in the low position) or (20% of the power supply radius < the distance between the transformer and the line is ≤60% of the power supply radius and the transformer tap is not in the middle position) or (the distance between the transformer and the line is >60% of the power supply radius and the transformer tap is not in the high position), then the main distribution network does not coordinate and exceeds the limit, with a confidence level of 0.9.

[0181] (d) Line overload / overload rules

[0182] If the load factor is >85%, the power factor is <0.85, and the terminal voltage is <0.93 times the rated voltage, then the heavy load low voltage exceeds the limit with a confidence level of 0.9.

[0183] (e) Light Load Rules for Lines

[0184] If the load factor is <30% and the photovoltaic output is >80% of the load and the first-end voltage is >1.05 times the rated voltage, then the light load high voltage exceeds the limit with a confidence level of 0.85.

[0185] (f) Solar power feedback rules

[0186] If the photovoltaic output is greater than the load and the voltage is greater than 1.04 times the rated voltage, then the photovoltaic backfeed exceeds the limit with a confidence level of 0.95.

[0187] It adopts a forward reasoning mechanism, based on actual data, and draws conclusions through rule matching. That is, it reads the current running data from the actual database, matches the facts with the conditions in the rule base, selects the rule according to priority when multiple rules match at the same time, and finally outputs the over-limit type and confidence level.

[0188] Another method is Stacking ensemble learning, a strategy that integrates multiple prediction algorithms to improve generalization ability by combining the strengths of each model. The steps of this invention for identifying grid voltage exceedances based on Stacking ensemble learning are as follows:

[0189] Step 1: PSO-based VMD parameter optimization (preprocessing optimization layer)

[0190] The Particle Swarm Optimization (PSO) algorithm is used to search for the Variational Mode Decomposition (VMD) algorithm globally for optimal combination. The most reasonable penalty factor, number of decomposition layers, and optimal number of modes are found through iteration to overcome the blindness of the VMD algorithm when manually tuning parameters.

[0191] Step 2: Key Factor Extraction and Signal Decomposition (Feature Identification Layer)

[0192] The optimal VMD algorithm obtained in step 1 is used to decompose the original voltage signal, accurately identify and extract the key factors affecting voltage over-limit from the complex original data, and provide high-quality input features for subsequent models.

[0193] Step 3: Construct the initial meta-model group (base learner layer)

[0194] Three heterogeneous neural networks were constructed and trained in parallel as meta-models for the Stacking framework: a multilayer perceptron (MLP) for capturing nonlinear static features of data, a long short-term memory (LSTM) network for mining long-range time series correlations of voltage waveforms, and a gated recurrent unit (GRU) for optimizing computational efficiency.

[0195] Step 4: Model fusion based on stacking strategy (ensemble learning layer)

[0196] The Stacking ensemble learning algorithm is used to take the prediction results of the three meta-models (MLP, LSTM, GRU) in step 4 as input and feed them into the meta-learner. By integrating the advantages of each meta-model through the fusion strategy, the prediction bias of the single model is corrected, thereby constructing a high-performance dynamic identification model.

[0197] Step 5: Voltage Exceedance Result Identification (Output Prediction Layer)

[0198] The trained Stacking ensemble model is used to identify real-time or test data, and the output is the judgment result of whether the voltage exceeds the limit. The strong generalization ability of ensemble learning is used to achieve fast and accurate identification of voltage status.

[0199] PSO obtains the optimal solution through cooperation and information sharing among individuals in the population, and has advantages such as fast convergence speed, few tuning parameters, and ease of implementation. Given training samples , where x i y represents a high-dimensional fault feature vector. i This indicates the sample's label, specifically the voltage over-limit type. A non-linear function is used. Its function is to project linearly inseparable data from the original low-dimensional space onto a high-dimensional feature space, making the data linearly separable or easier to process in that space. This involves mapping samples to a k-dimensional space (k>d) and constructing a recursive function:

[0200] (17)

[0201] In the formula: x is the sample vector, ω is the weight vector, and b is the classification threshold. The selection of the number of decomposition layers, penalty factor, and modality is crucial to the VMD decomposition effect. If these three parameters are not set properly, it will directly lead to "over-decomposition" or "under-decomposition," thus resulting in the inability to effectively extract voltage over-limit features. Specifically, it includes the following three parts:

[0202] Modular function monolithic frequency decomposition:

[0203] (18)

[0204] In the formula: F1 represents the spectrum of the modulus function, δ(t) is the impulse function, j represents the imaginary unit, and u k (t) represents the intrinsic mode function component, and t represents the time domain variable.

[0205] Solving for the correction function:

[0206] (19)

[0207] In the formula: F2 represents the correction function, ω k This represents the center frequency of the intrinsic modulus function. This represents the complex exponential modulation term.

[0208] Solving variational constraint models:

[0209] (20)

[0210] In the formula: f is the input signal, This represents the k-th modal component obtained from the decomposition. This represents the center frequency of the k-th modal component. Represents the unit impact function. This represents the partial derivative with respect to time.

[0211] like Figure 2 The diagram shown is a framework diagram of the Stacking ensemble learning model in this embodiment. Figure 2 The MLP algorithm in [the text is incomplete and likely refers to a specific algorithm or algorithm] is a type of artificial neural network with a tendency structure. The calculation formula is as follows:

[0212] (a) Hidden layer input formula:

[0213] (twenty one)

[0214] In the formula: S j W represents the input of the j-th node. jiLet X represent the weight between the j-th nodes. i The input to the input layer is h, where h is the number of neurons, and θ is the input to the input layer. j This represents the threshold of the j-th neuron node.

[0215] (b) Output the solution formula

[0216] The activation function is the sigmoid function. The output formulas for each node in the hidden layer are then calculated.

[0217] (twenty two)

[0218] In the formula: H j It is the output value of the j-th hidden layer neuron node.

[0219] (c) Hidden layer output layer formula

[0220] (twenty three)

[0221] In the formula: O k It is the input of the k-th neuron node, γ jk β is the weight between the j-th and k-th neurons. k The threshold value is the threshold value for the k-th neuron node.

[0222] LSTM is an improved recurrent neural network, consisting of an input gate, an output gate, and a forget gate. The input gate can be represented as:

[0223] (twenty four)

[0224] In the formula: W i Let h be the input gate weight matrix. t-1 Indicates the output of the hidden state in the previous moment, x t Indicates the current input, b i This represents the bias vector.

[0225] The forget gate can be represented as:

[0226] (25)

[0227] In the formula: W f Let h be the forget gate weight matrix. t-1 Indicates the output of the hidden state in the previous moment, x t Indicates the current input, b f This represents the bias vector.

[0228] The output gate can be represented as:

[0229] (26)

[0230] In the formula: W o Let b be the output gate weight matrix. o This represents the bias vector.

[0231] GRU neural networks are a mainstream neural network algorithm in the field of deep learning. The core calculation formula is:

[0232] (27)

[0233] (28)

[0234] (29)

[0235] (30)

[0236] In the formula: a is the Sigmoid activation function, W z W r And W are the corresponding parameter weighting coefficients, x t For GRU input, y t-1 Indicates the hidden state in the previous time step, y t For GRU output, r t and z t These are the GRU reset gate and update gate, respectively. This is the output of the hidden layer.

[0237] This embodiment verifies the method of the present invention, using the following experimental platform: hardware environment includes Intel Xeon Gold 6248R CPU, NVIDIA Tesla V100 GPU, 256GB RAM; software includes Python 3.8, TensorFlow 2.5, Scikit-learn 0.24, and PyTorch 1.8.

[0238] The data used is the annual data of a certain provincial and municipal power grid, including: 526,000 10kV bus voltage data, 38,000 distribution transformer data, 1,283,000 user voltage data, and 245,000 new energy power generation data. The data is divided into training set, validation set and test set in a ratio of 7:2:1.

[0239] The model performance is evaluated using four metrics: accuracy, precision, recall, and decision value. The calculation formulas for these metrics are as follows:

[0240] (31)

[0241] (32)

[0242] (33)

[0243] (34)

[0244] In the formula: TP is a true positive (correctly identified as exceeding the limit), TN is a true negative (correctly identified as normal), FP is a false positive (normal but misjudged as exceeding the limit), and FN is a false negative (exceeding the limit but misjudged as normal).

[0245] Different models were tested on the test set, and their performance is shown in Table 5. The test results show that the Stacking ensemble model has the best performance, with an accuracy of 92.7%, proving that the present invention can accurately identify the type and cause of voltage over-limit faults in the power grid, significantly improving the level of automation and shortening the time for voltage over-limit fault diagnosis.

[0246] Table 5:

[0247]

[0248] Based on the same inventive concept, this invention also provides a full-user voltage over-limit automatic identification system. The principle behind this system is similar to the aforementioned full-user voltage over-limit automatic identification method. In specific implementation, the full-user voltage over-limit automatic identification system provided by this invention specifically includes:

[0249] The acquisition module is used to acquire multi-source power grid data from the dispatching system, electricity consumption information collection system, distribution automation system, new energy monitoring system, and power grid GIS platform; the multi-source power grid data is sequentially cleaned, time-series aligned, and feature extracted to obtain the processed data;

[0250] The analysis module is used to perform cluster analysis based on the processed data using clustering algorithms to generate a scenario library containing multiple typical voltage over-limit scenarios.

[0251] A module is constructed to extract multiple key factors affecting voltage over-limit based on processed data; a combined weighting method is used to determine the weight of each key factor, and a voltage over-limit sensitivity quantitative model is constructed based on the weight of each key factor.

[0252] The identification module is used to match the processed data with the scene library and perform reasoning through a preset expert rule base to obtain the first identification result;

[0253] The variational mode decomposition method based on particle swarm optimization algorithm is used to decompose multiple key factors to obtain decomposition results. The decomposition results are then input into the Stacking ensemble learning model for identification to obtain a second identification result.

[0254] The output module is used to calculate the voltage over-limit risk value of the current state based on the voltage over-limit sensitivity quantification model according to the first identification result and the second identification result; determine the risk level according to the voltage over-limit risk value; and output the final identification report according to the risk level.

[0255] Accordingly, this embodiment of the invention also provides an automatic identification device for full user voltage over-limit, including a processor and a memory, wherein the processor executes a computer program stored in the memory to implement the automatic identification method for full user voltage over-limit provided in this embodiment of the invention.

[0256] Accordingly, embodiments of the present invention also provide a computer-readable storage medium for storing a computer program, wherein the computer program, when executed by a processor, implements the above-described automatic identification method for full-scale user voltage over-limit as provided in embodiments of the present invention.

[0257] The preferred embodiments of the present invention have been described in detail above with reference to the accompanying drawings. However, the present invention is not limited to the above embodiments. Within the scope of knowledge possessed by those skilled in the art, various changes can be made without departing from the spirit of the present invention. These changes involve related technologies well known to those skilled in the art, and all of them fall within the protection scope of the present invention.

[0258] Many other changes and modifications can be made without departing from the concept and scope of this invention. It should be understood that this invention is not limited to the specific embodiments, and the scope of this invention is defined by the appended claims.

Claims

1. A method for automatic identification of voltage over-limit for all users, characterized in that, Includes the following steps: S1. Acquire multi-source power grid data from the dispatching system, electricity consumption information collection system, power distribution automation system, new energy monitoring system and power grid GIS platform, and sequentially perform data cleaning, time-series alignment and feature extraction on the multi-source power grid data to obtain processed data; S2. Based on the processed data, perform cluster analysis to generate a scenario library containing multiple typical voltage over-limit scenarios; S3. Based on the processed data, extract multiple key factors affecting voltage over-limit, use the combined weighting method to determine the weight of each key factor, and construct a voltage over-limit sensitivity quantitative model based on the weight of each key factor. S4. Match the processed data with the scene library, and perform reasoning through the preset expert rule library to obtain the first identification result; The variational mode decomposition method based on particle swarm optimization algorithm is used to decompose multiple key factors to obtain decomposition results. The decomposition results are then input into the Stacking ensemble learning model for identification to obtain a second identification result. S5. Based on the first identification result and the second identification result, calculate the voltage over-limit risk value of the current state based on the voltage over-limit sensitivity quantification model, determine the risk level based on the voltage over-limit risk value, and output the final identification report based on the risk level.

2. The automatic identification method for voltage over-limit of all users according to claim 1, characterized in that, The clustering analysis in S2 uses the K-means++ algorithm and uses the silhouette coefficient to evaluate the clustering effect, so as to generate a multi-level typical voltage over-limit scenario library including substation layer, line layer, distribution transformer layer, user layer and new energy layer.

3. The automatic identification method for voltage over-limit of all users according to claim 2, characterized in that, The key factors in S3 include: The main transformer capacity, tap position, short-circuit impedance, compensation capacity, and switching frequency at the substation level; Line length, resistance, reactance, line load factor, and current of the line layer; Distribution transformer capacity, tap level, load factor, and three-phase current imbalance of the distribution transformer layer; Maximum load and load volatility at the user level; The access capacity, output fluctuation rate, and power factor of the new energy layer.

4. The automatic identification method for voltage over-limit of all users according to claim 1, characterized in that, The combined weighting method in S3 is as follows: The subjective weights of each key factor are determined using the eigenvector method: A pairwise comparison judgment matrix, denoted as A, is constructed using a 1-9 scale among the key influencing factors for voltage limit exceedance. In the formula: n is the number of key influencing factor categories involved in subjective weighting, a ij F represents the category of the i-th key influence factor. i Relative to the j-th key impact factor category F j The importance comparison of the degree of impact on the risk of voltage exceeding limits; A consistency check is then performed, including calculating the consistency index CI and the consistency ratio CR, using the following formulas: ， ， In the formula: λmax is the largest eigenvalue of the judgment matrix, RI is the average random consistency index, and when CR < 0.1, the judgment matrix A satisfies the consistency requirement; Solve the system of equations AW=λmaxW to obtain the eigenvector W. After normalization, the subjective weights of each factor are: ， In the formula: W j W represents the weight of the j-th evaluation factor. i Let be the weight of the i-th evaluation factor; The objective weights of each key factor are determined using the information entropy method: Construct a standardized decision matrix by standardizing the factor data to obtain the matrix. , where r ij Let m be the element in the i-th row and j-th column of matrix R, m be the number of samples, n be the number of factors, and the information entropy of the j-th factor be... The calculation formula is as follows: ， In the formula: ,like ,but , The objective weight of the j-th factor is: ， The combined weights are calculated using a linear weighting method. : ， The coefficient for subjective weight is 0.4, and the coefficient for objective weight is 0.

6.

5. The automatic identification method for voltage over-limit of all users according to claim 4, characterized in that, The voltage over-limit sensitivity quantification model in S3 is constructed based on combined weights, and its expression is as follows: ， In the formula: S is the voltage over-limit sensitivity value; x j x is the actual value of the j-th key factor; j0 Let j be the standard value of the j-th key factor; Let be the combined weight of the j-th key factor.

6. The automatic identification method for voltage over-limit of all users according to claim 1, characterized in that, The expert rule base preset in S4 includes: rules for unbalanced three-phase load of distribution transformers, rules for unreasonable distribution transformer taps, rules for lack of coordination between main and distribution networks, rules for line overload, rules for line light load, and rules for photovoltaic backfeeding.

7. The automatic identification method for voltage over-limit of all users according to claim 1, characterized in that, In step S4, the decomposition result is input into the Stacking ensemble learning model for identification to obtain the second identification result. The steps include: S4-1. Parameter optimization of variational mode decomposition algorithm based on particle swarm optimization algorithm; The particle swarm optimization algorithm is used to search for variational mode decomposition algorithms globally and perform optimization combinations. The optimal variational mode decomposition algorithm is obtained by iteratively finding the most reasonable penalty factor, the number of decomposition layers, and the optimal number of modes. S4-2, Key Factor Extraction and Signal Decomposition; The optimal variational mode decomposition algorithm is used to decompose the original voltage signal, and the key factors affecting voltage over-limit are accurately identified and extracted from the complex original data. S4-3, Constructing a primary meta-model group; Three heterogeneous neural networks were constructed and trained in parallel as meta-models for the Stacking framework, including a multilayer perceptron (MLP) for capturing nonlinear static features of data, a long short-term memory (LSTM) network for mining long-range time series correlations of voltage waveforms, and a gated recurrent unit (GRU) for optimizing computational efficiency. S4-4, Model fusion based on stacking strategy; The Stacking ensemble learning algorithm is adopted. The prediction results of the three meta-models in S4-3 are used as input and fed into the secondary learner. The advantages of each meta-model are integrated through the fusion strategy to correct the prediction bias of the single model and construct the Stacking ensemble learning model. S4-5. Identification of voltage over-limit results; The trained Stacking ensemble learning model is used to identify whether the output voltage exceeds the limit in real-time or test data.

8. A full-user voltage over-limit automatic identification system, characterized in that, include: The acquisition module is used to acquire multi-source power grid data from the dispatching system, electricity consumption information collection system, distribution automation system, new energy monitoring system, and power grid GIS platform; The multi-source power grid data is sequentially cleaned, time-series aligned, and feature-extracted to obtain the processed data. The analysis module is used to perform cluster analysis based on the processed data using clustering algorithms to generate a scenario library containing multiple typical voltage over-limit scenarios. A module is constructed to extract multiple key factors affecting voltage over-limit based on processed data; a combined weighting method is used to determine the weight of each key factor, and a voltage over-limit sensitivity quantitative model is constructed based on the weight of each key factor. The identification module is used to match the processed data with the scene library and perform reasoning through a preset expert rule base to obtain the first identification result; The variational mode decomposition method based on particle swarm optimization algorithm is used to decompose multiple key factors to obtain decomposition results. The decomposition results are then input into the Stacking ensemble learning model for identification to obtain a second identification result. The output module is used to calculate the voltage over-limit risk value of the current state based on the voltage over-limit sensitivity quantification model according to the first identification result and the second identification result. The risk level is determined based on the voltage over-limit risk value, and a final identification report is output based on the risk level.

9. An automatic voltage over-limit identification device for all users, characterized in that, It includes a processor and a memory, wherein the processor executes a computer program stored in the memory to implement the full-user voltage over-limit automatic identification method as described in any one of claims 1 to 7.

10. A computer-readable storage medium, characterized in that, Used to store a computer program, wherein the computer program, when executed by a processor, implements a full-user voltage over-limit automatic identification method as described in any one of claims 1 to 7.

Images