Power distribution network risk situation judgment model training method and judgment method based on same

Abstract

The application relates to the technical field of power system operation and control, and discloses a power distribution network risk situation research and judgment model training method and a research and judgment method based on the same, wherein the method comprises the following steps: obtaining the historical operation state of each node of a power distribution network, and constructing a power grid data set; inputting the power grid data set into a preset power prediction model, obtaining prediction data, and obtaining risk sensitive data according to the prediction data and the power grid data set; inputting the power grid data set, the prediction data and the risk sensitive data into a pre-trained benchmark region model for region division, and obtaining the benchmark region model; and updating the benchmark distance of the benchmark safety region model based on the prediction data to obtain a power distribution network operation risk situation research and judgment model. The technical scheme provided by the application can realize a research and judgment method that is independent of full-node power measurement and has strict electrical physical support, and can solve the problems of poor engineering adaptability, low boundary precision, high missed judgment and false report rates and lack of early warning capability of related technologies.

Description

Technical Field

[0001] This application relates to the field of power system operation and control technology, and in particular to a training method for a risk situation assessment model of a distribution network and an assessment method based thereon. Background Technology

[0002] With the large-scale integration of distributed generation (DG) into medium- and low-voltage distribution networks, the operating characteristics of these networks have shifted from a traditional unidirectional radial pattern to a bidirectional interactive pattern. Random load fluctuations and intermittent fluctuations in DG output lead to frequent operational risks such as voltage exceeding limits and branch overloads, posing significant challenges to the safe and stable operation of the distribution network. Risk assessment of the distribution network is the core foundation for proactive grid control and predictive maintenance; its accuracy and foresight directly determine the scientific validity and reliability of distribution network dispatch decisions. Currently, medium- and low-voltage distribution networks generally suffer from comprehensive voltage measurement coverage but low power measurement coverage. Furthermore, the strong uncertainties brought about by DG integration place entirely new demands on the adaptability, robustness, and foresight of traditional risk assessment technologies.

[0003] Predicting the future operating status and development trend of active distribution networks helps in advance scheduling and control, thereby improving the economy and reliability of active distribution network operation. Currently, active distribution networks contain numerous uncertainties, such as load uncertainty and renewable energy output uncertainty. Coupled with the increasing complexity of distribution network systems, accurate distribution network situation prediction has become a key research focus. Data-driven active distribution network situation prediction mainly includes the following research directions: load forecasting, renewable energy output forecasting, and distribution network operating status prediction. For example, introducing attention mechanisms and combining CNN and GRU to achieve short-term power load forecasting. Using the inverse error method to weightedly combine LSTM and XGBoost reduces the error of a single prediction model, achieving ultra-short-term power load forecasting. A distribution network abnormal state early warning method based on fuzzy weighted SVDD collects electrical measurement data such as voltage, power, and current at all nodes of the distribution network, calculates fuzzy membership degrees based on sample space density, trains and improves the SVDD model using offline historical data to obtain a fixed safety boundary, and inputs real-time operating data into the model to achieve abnormal state identification and alarm.

[0004] In view of this, how to solve the problems of sparse power measurement, strong uncertainty of DG output leading to poor engineering adaptability of risk assessment technology, low boundary accuracy, high rate of missed judgment and false alarm and lack of early warning capability has become a technical problem that needs to be solved. Summary of the Invention

[0005] This application provides a training method for a distribution network risk situation assessment model and an assessment method based thereon, which solves the technical problems of poor engineering adaptability, low boundary accuracy, high false alarm rate and lack of early warning capability caused by sparse power measurement and strong uncertainty of DG output.

[0006] To achieve the above objectives, the main technical solutions adopted in this application include: Firstly, embodiments of this application provide a training method for a distribution network operation risk situation assessment model.

[0007] Obtain the historical operating status of each node in the power distribution network and construct a power grid dataset; Input the power grid dataset into a preset power prediction model to obtain prediction data, and based on the prediction data and the power grid dataset, obtain risk-sensitive data to characterize the dynamic confidence features of the prediction data; The power grid dataset, the predicted data, and the risk-sensitive data are input into a pre-trained benchmark region model for region division. Through iterative optimization, a benchmark region model for outputting benchmark safe regions is obtained. Based on the reliability of the predicted data, the baseline distance of the baseline safety area model is updated to obtain the distribution network operation risk situation assessment model. The distribution network operation risk situation assessment model is used to output the assessment area and the predicted data.

[0008] This embodiment provides a training method for a distribution network operation risk assessment model. It uses clustering to divide typical operating scenarios, leverages voltage data to drive predictions to fill in missing power and extract risk features, and then constructs an assessment region that can adaptively scale with prediction reliability and scenario requirements, achieving accurate assessment of distribution network operation risks. The technical solution provided in this embodiment can effectively perceive risks even when power measurements are significantly missing, maintaining high-sensitivity early warning for high-risk scenarios such as severe photovoltaic fluctuations and concentrated electric vehicle charging, while reducing false alarm rates in stable operating scenarios.

[0009] In one implementation, obtaining a reference region model for outputting a reference security region includes: The power grid dataset is divided into safe samples according to a preset threshold, and a comprehensive weight is constructed based on the risk-sensitive data. Using the comprehensive weights and preset update coefficients to characterize iterative stability, a boundary region for enclosing the safe sample is constructed, and the region center of the boundary region and the region distance from the region center to the boundary are determined. Iterate over the regional distance and the regional center until a preset condition is met to obtain the baseline regional model.

[0010] This embodiment filters safe samples using thresholds such as node voltage and branch load rate, and obtains comprehensive weights by combining risk-sensitive data. Then, an adaptive kernel function is introduced to construct an optimization problem, finding a boundary region that minimizes the size of all safe samples, and the original problem is transformed into a dual problem using the Lagrange multiplier method. Subsequently, based on the obtained support vectors, the region center is determined by weighting, and the mean distance from the support vectors to the center is taken as the benchmark distance, resulting in a benchmark safe region model defined by the region center and the benchmark distance.

[0011] In one implementation, the risk-sensitive data includes power fit confidence, load volatility, and load coupling coefficient, and the comprehensive weighting is obtained through the following methods: The power completion confidence weight is determined based on the ratio of the power fitting confidence to the maximum value of the power fitting confidence. Based on the load volatility and the load coupling coefficient, risk-sensitive weights are constructed to characterize the degree of influence of different samples. The comprehensive weight is determined by multiplying the power completion confidence weight and the risk sensitivity weight.

[0012] This embodiment determines the comprehensive weight by multiplying the power completion confidence weight and the risk-sensitive weight. The power completion confidence weight is constructed from the ratio of the power fitting confidence score to its maximum value, used to suppress the interference of low-confidence samples on boundary accuracy. The risk-sensitive weight is constructed based on load volatility and load coupling coefficient, and is adjusted by a balancing coefficient to adaptively scale the weight with the risk level, thereby enhancing the ability to capture rare high-risk samples. The comprehensive weight, resulting from the multiplication of the two, simultaneously considers the differences in power completion quality and the sample risk level, guiding subsequent training to both emphasize differences in completion confidence and strengthen the perception and penalty for high-risk situations.

[0013] In one implementation, the methods for obtaining the power fitting confidence level, load volatility, and load coupling coefficient include: The power fit confidence level is determined based on the numerical difference between the predicted data and the power grid dataset at the same node; Based on the power grid dataset within the preset time window, determine the power standard deviation and power average value within the time window; The load fluctuation rate is determined based on the ratio of the power standard deviation to the power average. The coordinated fluctuation rate is determined based on the degree of coordinated fluctuation between the output rate of distributed power sources and the power of nodes in the power grid dataset. The fitting standard deviation is determined by multiplying the standard deviation of the distributed power output rate by the standard deviation of the power. The load coupling coefficient is determined based on the ratio of the coordinated volatility to the fitted standard deviation.

[0014] This embodiment effectively quantifies the reliability of full-node power completion data by determining the power fitting confidence level, providing a reliable basis for subsequently suppressing the interference of low-confidence samples on boundary accuracy. Load volatility eliminates the influence of absolute levels through a ratio, objectively reflecting the severity of short-term fluctuations, while the load coupling coefficient accurately characterizes the coordinated changes in power output and load. Both provide key features for constructing risk-sensitive weights. This design significantly enhances the ability to perceive and distinguish high-risk operating states, ensuring the adaptive adjustment of comprehensive weights to differences in completion quality and risk levels, thereby improving the accuracy of the baseline safe zone model and the reliability of early warnings in rare extreme scenarios.

[0015] In one implementation, updating the baseline distance of the baseline safe zone model based on the confidence level of the predicted data includes: Obtain multiple prediction data at the current time step, and truncate the multiple prediction data based on a preset judgment interval to obtain the width of the confidence interval corresponding to the current time step; The baseline distance of the baseline safety zone model is updated according to the width of the confidence interval using a preset adjustment rule to obtain the distribution network operation risk situation assessment model.

[0016] This embodiment obtains multiple independent prediction data points simultaneously, arranges the prediction data in ascending order, and then extracts a portion using a preset judgment interval to obtain the normalized confidence interval width for the current time step, quantifying the uncertainty of power prediction. Then, based on preset adjustment rules, the benchmark distance of the benchmark safe zone model is dynamically updated online using this confidence interval width, thereby constructing a distribution network operation risk assessment model. This embodiment introduces the uncertainty of power prediction into the adaptive adjustment of the benchmark distance, ensuring that the benchmark distance remains unchanged when prediction uncertainty is minimal, and appropriately expands the benchmark boundary when uncertainty is maximum. This achieves adaptive scaling of the safe zone to random fluctuations of distributed power sources, effectively improving the accuracy and robustness of distribution network operation risk assessment.

[0017] Secondly, embodiments of this application provide a method for assessing the operational risk situation of a power distribution network, the method comprising: Obtain power grid operation data on the operating status of each node in the distribution network under test; The power grid operation data is input into the distribution network operation risk situation assessment model obtained by the training method described in any one of the claims above, the assessment area and prediction data are obtained, and the graded early warning level is output according to the changing trend of the prediction data and the assessment area.

[0018] In one implementation, the step of outputting a graded early warning level based on the changing trend of the predicted data and the assessed area includes: The predicted data are used to characterize the first-order difference of the risk change rate and the second-order difference of the risk change acceleration. Based on the predicted data, obtain the probability of breaking through the assessment area to determine the risk breakthrough probability; Based on the first-order difference, the second-order difference, the risk breakthrough probability, and the judgment area, a graded early warning level is output through preset output rules.

[0019] Thirdly, embodiments of this application provide a training device for a distribution network operation risk situation assessment model, the device comprising: The data acquisition unit is used to acquire the historical operating status of each node in the power distribution network and construct a power grid dataset. The data prediction unit is used to input the power grid dataset into a preset power prediction model, obtain prediction data, and obtain risk-sensitive data to characterize the dynamic confidence features of the prediction data based on the prediction data and the power grid dataset. The region generation unit is used to input the power grid dataset, the predicted data, and the risk-sensitive data into a pre-trained benchmark region model for region division, and obtain a benchmark region model for outputting a benchmark safe region through iterative optimization. The dynamic update unit is used to update the baseline distance of the baseline safety area model based on the reliability of the predicted data, so as to obtain the distribution network operation risk situation assessment model. The distribution network operation risk situation assessment model is used to output the assessment area and the predicted data.

[0020] Fourthly, embodiments of this application provide a computer device, including: The system includes a memory and a processor, which are interconnected. The memory stores computer instructions, and the processor executes the computer instructions to perform the training method for the distribution network operation risk situation assessment model or the assessment method described above.

[0021] Fifthly, embodiments of this application provide a computer-readable storage medium storing computer instructions, which are used to cause a computer to execute the power distribution network operation risk situation assessment model training method or the assessment method described above. Attached Figure Description

[0022] To more clearly illustrate the technical solutions in the specific embodiments of this application or the prior art, the drawings used in the description of the specific embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are some embodiments of this application. For those skilled in the art, other drawings can be obtained from these drawings without creative effort.

[0023] Figure 1 A flowchart illustrating a method for training a power distribution network operation risk assessment model, provided in this application embodiment; Figure 2 A flowchart of step S51 provided in an embodiment of this application; Figure 3 A flowchart of step S511 provided in an embodiment of this application; Figure 4 This application provides a training device for a power distribution network operation risk situation assessment model. Figure 5 This is a schematic diagram of the structure of a computer device provided in an embodiment of this application. Detailed Implementation

[0024] To make the objectives, technical solutions, and advantages of the embodiments of this application clearer, the technical solutions of the embodiments of this application will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of this application, not all embodiments. Based on the embodiments of this application, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of this application.

[0025] The relevant technologies rely on full-node power measurement data, which has limited practical applicability to engineering projects where power measurement is lacking in medium and low voltage distribution networks, and the boundary accuracy is generally poor in scenarios with missing measurements. The use of pure mathematical fuzzy weighting based on sample density cannot solve the problem of low-confidence samples interfering with safety boundaries under conditions of missing measurements, leading to boundary deviations. The use of a fixed-width Gaussian kernel function and offline fixed boundaries cannot adapt to the dynamic changes in distribution network conditions under high-proportion distributed generation (DG) access, resulting in serious problems of missed detections in high-fluctuation scenarios and false alarms in stable scenarios.

[0026] To address the aforementioned technical problems, an embodiment of the present application provides a training method for a distribution network risk situation assessment model. It should be noted that the steps shown in the flowchart in the accompanying drawings can be executed in a computer system such as a set of computer-executable instructions. Furthermore, although a logical order is shown in the flowchart, in some cases, the steps shown or described may be executed in a different order than that shown here.

[0027] This embodiment provides a training method for a distribution network operation risk assessment model. Figure 1 The flowchart of the training method for the distribution network risk situation assessment model provided in the embodiments of this application is as follows: Figure 1 As shown, the process includes the following steps: Step S1: Obtain the historical operating status of each node in the distribution network and construct a power grid dataset.

[0028] Specifically, the historical operating status of each node in the distribution network can be obtained through the Supervisory Control and Data Acquisition (SCADA) system, which collects real-time voltage data of all nodes, output data of distributed generation (DG) sources, and fixed topology parameters of the distribution network via feeder terminal units (FTUs). After processing, standardized historical operating data (which can be represented by feature vectors or basic feature matrices) and current scene labels are output to construct a power grid dataset.

[0029] Taking the implementation scenario as an example: An IEEE 33-node 12.66kV medium-voltage distribution network (nodes 10, 18, 25, and 33 are connected to 0.4MW, 0.6MW, 0.5MW, and 0.7MW of distributed photovoltaic (i.e., distributed power generation) respectively) is used. The data acquisition rules are set as follows: a 1-hour sliding window, a 15-minute step size, and each window contains four consecutive acquisition times (adjustable 15min, 30min, 45min, and 60min). The current time t is 12:00. The data acquisition objects are the voltage data of all 33 nodes, the power data of measurable nodes, the output data of the four distributed generators, and the fixed topology parameters of the distribution network. Subsequently, outlier removal and missing value supplementation are performed (for nodes without power measurement devices, their power data is not filled in during the data preprocessing stage, but is predicted and completed by the power prediction model in subsequent steps). The 3σ criterion is used to identify outliers, which are defined as: in, For the first in the sliding window Voltage data or DG output data This represents the arithmetic mean of the data within the sliding window. The standard deviation of the data within the sliding window. This is the total number of data points. It will exceed... Data within the range is identified as outliers caused by equipment failure and filled using the sliding median of the previous three acquisition cycles for the same node. For missing voltage measurements, if the missing percentage is less than 5% and the missing values ​​are random, linear interpolation is used (e.g., using the average of the preceding and following data); if the missing values ​​are consecutive for more than three cycles, the average voltage for the same period in the previous three days for the same node is used. For missing power data, no pre-filling is performed, and the missing values ​​are left for subsequent processing to reflect the actual engineering situation of missing measurements in medium and low voltage distribution networks. The basic feature matrix is ​​defined as a two-dimensional matrix composed of historical operating data of all nodes at all acquisition times within the sliding window. It serves as the basic data carrier for subsequent data standardization, operating scenario division, and subsequent modeling. The basic feature matrix dimension corresponding to a single 1-hour sliding window is [4 acquisition times × 37-dimensional features], where the 37-dimensional features include real-time voltage data of 33 nodes and power output data of 4 DG units.

[0030] Next, the min-max normalization method is used to map all continuous features to the [0,1] interval to eliminate dimensional differences and ensure the stability of subsequent training. in, These are the standardized eigenvalues. The original eigenvalues ​​are (a row of 37-dimensional data taken from the basic feature matrix at the current time). and These are the minimum and maximum values ​​of the corresponding features within the historical training data (or can be preset based on the maximum or minimum values ​​in other offline datasets), and they remain fixed throughout the process. This step strictly uses the extreme values ​​of the offline training set for standardization to prevent online data leakage and ensure that the distribution of online data is consistent with that of offline training data.

[0031] The real-time feature vector is taken from the row corresponding to the basic feature matrix at the current time, and is a one-dimensional vector formed after the above min-max standardization. All data are mapped to the interval [0,1] to eliminate the difference in dimensions. The dimension of the real-time feature vector is [1×37], which is aligned with the feature dimension of the basic feature matrix.

[0032] Scene labels are output by an offline-trained clustering model based on real-time feature vectors to characterize the current operating conditions of the distribution network. In this embodiment, the optimal number of clusters is determined in advance to be 5 through dual verification using the elbow rule and the profile coefficient, corresponding to 5 typical operating scenarios. The label determination is based on the distributed generation output rate within the sliding window (used to characterize the degree of influence of distributed generation on the current power balance) and the load coupling coefficient (DG-voltage coupling coefficient) (used to characterize the correlation between distributed generation output rate (DG output data) and node voltage data), and high and low fluctuation intervals and strong and weak load reverse coupling intervals are set respectively.

[0033] Subsequently, based on voltage data and distributed generation (DG) output data, four basic characteristics are calculated in real time, including: voltage fluctuation rate (the ratio of the standard deviation of voltage data to the average voltage data within a sliding window, used to characterize the degree of fluctuation of node voltage within the sliding window), and load coupling coefficient (within a preset sliding window, the Pearson correlation coefficient between the collected DG output data and corresponding node voltage data is calculated as the load coupling coefficient (DG-voltage coupling coefficient). The DG-voltage coupling coefficient (load coupling coefficient) can be used to determine the distributed generation output rate (DG output data) and the corresponding node voltage data. The strength and direction of coupling between power generation output (DG output data) and node voltage. When this coefficient is negative and has a large absolute value, it indicates a strong reverse coupling relationship between the DG output data and node voltage data changes; when the absolute value of this coefficient is small, it indicates a weak coupling effect of the DG output data on node voltage. This also includes time-coding characteristics (hourly sine / cosine coding, peak-valley segment labels (used to characterize the current operating segment), weekday / weekend labels) and topological characteristics (node ​​degree, branch impedance per unit value).

[0034] Finally, the offline-trained K-Means++ clustering model was used to classify the real-time operating data into scenarios based on four clustering features: load coupling coefficient, distributed generation output rate (e.g., the ratio of the sum of the real-time total active power output of the four DGs within the window (e.g., 0.32+0.58+0.45+0.63=1.98MW) to the current total load power of all 33 nodes), node voltage fluctuation rate, and peak-valley segment labels. The optimal number of clusters was determined to be 5 through dual verification using the elbow rule and silhouette coefficient. The scenario labels for the current moment were output, and the five typical scenarios are as follows: Scenario 1: High volatility + strong coupling of DG The load coupling coefficient is ≤-0.5 (strong reverse coupling), the distributed generation output rate is ≥0.5 (DG dominates power balance), and the voltage fluctuation rate is ≥0.05 (high fluctuation). This scenario corresponds to the midday peak photovoltaic power generation period, where the DG output accounts for a high proportion and experiences severe short-term fluctuations (such as due to cloud cover). The DG output and node load are strongly reverse coupled, and the voltage and power distribution of the distribution network are completely dominated by changes in DG output. This is a high-risk and sensitive scenario, and the warning boundaries need to be tightened.

[0035] Scenario 2: High volatility + weak coupling of DG The load coupling coefficient is >-0.5 (weak reverse coupling), the distributed generation output rate is <0.5 (weak DG impact), and the voltage fluctuation rate is ≥0.05 (high fluctuation). This scenario corresponds to the evening peak electricity consumption period for residents, when DG is basically out of service. The distribution network operation characteristics are dominated by residential load fluctuations, which are severe and prone to branch overload and voltage drop risks. This is a typical high-fluctuation operation scenario.

[0036] Scenario 3: Low volatility + strong coupling of DG The load coupling coefficient is ≤-0.5 (strong reverse coupling), the distributed generation output rate is ≥0.5 (DG dominates power balance), and the voltage fluctuation rate is <0.05 (low fluctuation). This scenario corresponds to a period of stable daytime sunlight, where DG output is stable and accounts for a high proportion, with strong reverse coupling to the node load, minimal load fluctuation, and stable distribution network operation. However, there is a potential risk of voltage drop caused by a sudden decrease in DG output, requiring close monitoring of DG output changes.

[0037] Scenario 4: Low volatility + weak coupling of DG The load coupling coefficient is >-0.5 (weak reverse coupling), the distributed generation output rate is <0.05 (no DG output), and the voltage fluctuation rate is <0.05 (low fluctuation). This scenario corresponds to the off-peak period of residential electricity consumption at night, when the DG is completely shut down, the load is stable with minimal fluctuation, and the distribution network is operating stably. This is a normal and safe operating scenario, and the warning boundary can be appropriately relaxed to reduce the false alarm rate.

[0038] Scenario 5: Extreme scenario of centralized charging of EVs The load coupling coefficient is >-0.1 (no coupling), the distributed generation output rate is <0.1 (no or weak output from DG), and the voltage fluctuation rate is >0.15 (extremely high fluctuation). This scenario represents an extreme load fluctuation caused by centralized charging of electric vehicles. The load increases suddenly in a short period of time, which can easily lead to serious operational risks such as line overload and node voltage exceeding the lower limit. It is a rare high-risk extreme scenario and requires the highest level of risk monitoring and early warning.

[0039] Step S3: Input the power grid dataset into the preset power prediction model to obtain prediction data, and based on the prediction data and the power grid dataset, obtain risk-sensitive data to characterize the dynamic confidence features of the prediction data.

[0040] Specifically, the power grid dataset is input to a pre-defined power prediction model (which can be a scenario-specific time-series model trained offline under different scenario labels), obtains full-node prediction data (including the power completion values ​​corresponding to the incomplete power data mentioned above), and outputs the real-time power verification residuals and real-time power flow calculation results of measurable nodes. Its core lies in fitting the node injection power based solely on voltage time-series data through the power prediction model, without relying on traditional power flow mechanism equations, thus adapting to scenarios with missing power measurements. The training method for the power prediction model is as follows: First, starting from the historical continuous one-year operating data (or IEEE 33-node full-scenario simulation data) collected from the target distribution network SCADA / FTU system, outlier removal, missing value processing, data standardization, and time-series sample construction are performed to form a basic training set covering different DG penetration rates, load levels, and extreme operating conditions. Then, the sample set is split for five typical operating scenarios, dividing the training set, validation set, and test set in a 7:2:1 ratio according to time sequence, strictly prohibiting random shuffling to ensure time-series continuity, and including rare high-frequency scenarios such as concentrated EV charging and sudden drops in DG output. In high-risk scenarios, Generative Adversarial Networks (GANs) are used for sample augmentation, expanding the sample size for extreme scenarios to 15% of the total sample size. Based on this, a Transformer time-series model adapted to non-mechanistic modeling of distribution networks is designed. The input layer consists of voltage data from 33 nodes at four 15-minute intervals, with learnable sinusoidal position encoding added to preserve temporal information. Then, two encoder layers are stacked (each layer contains an 8-head masked multi-head self-attention mechanism and a 2-layer feedforward fully connected network, with attention focusing only on historical information). Features are compressed using global average pooling, ultimately outputting predicted values ​​for active and reactive power at all nodes. During the training phase, five scenarios are trained independently using the AdamW optimizer (initial learning rate 1e-4, batch size 64, maximum training 200 epochs) and an early stopping mechanism (termination occurs if the verification loss does not decrease for 10 consecutive epochs) to prevent overfitting. Furthermore, the Sparrow Search Algorithm (SSA) is used to adaptively optimize the hyperparameters of the sub-scenario time-series model under each scenario label, ensuring optimal power inversion accuracy. Accuracy verification employs three metrics: mean absolute error, mean absolute percentage error, and goodness of fit. Models meeting the accuracy requirements are archived as offline .pth files and stored in the model library. In online applications, the corresponding scenario-specific time-series model is directly matched and invoked based on the current-time scenario label in the power grid dataset. Power completion for all nodes can be achieved simply by performing forward inference, eliminating the need for online training and fully meeting the computing power requirements for real-time distribution network operation.

[0041] Step S5: Input the power grid dataset, prediction data, and risk-sensitive data into the pre-trained benchmark region model for region division, and obtain the benchmark region model for outputting the benchmark safe region through iterative optimization.

[0042] Specifically, firstly, based on a comparison of data such as node voltage, branch load rate, and load coupling coefficient with preset thresholds, safe samples are divided from the power grid dataset. Data that does not meet the conditions or has historical fault periods are classified as risk samples, and rare high-risk samples are extracted from these risk samples to construct a comprehensive weighting. Next, an adaptive kernel function is introduced to construct a minimum hypersphere that encloses all safe samples. This optimization problem is transformed into a dual problem using the Lagrange multiplier method and solved iteratively using a sequential minimum optimization algorithm to obtain the optimal Lagrange multipliers. Based on this, support vectors are selected, and the weighted sum of the support vectors is used to determine the region center. Then, the average distance from all support vectors to the center is calculated as the benchmark distance, finally obtaining a benchmark safe region model defined by the region center and the benchmark distance.

[0043] Step S7: Based on the reliability of the prediction data, update the benchmark distance of the benchmark safe area model to obtain the distribution network operation risk situation assessment model. The distribution network operation risk situation assessment model is used to output the assessment area and prediction data.

[0044] Specifically, the baseline safety region output by the baseline safety region model defines three operating regions: Safety Region: the distance from the operating point to the center. ,in The distance from the operating point to the center of the baseline safety zone. Baseline distance; Warning and alert zone: Risk zone exceeding limits: Based on scene tags, dynamic adjustment rules for the warning zone are defined to balance missed detections and false alarms: For high-fluctuation + strongly coupled DG scenarios and extreme scenarios with concentrated EV charging, the warning zone is tightened; for low-fluctuation and stable scenarios, the warning zone is relaxed; for normal scenarios, the baseline warning zone is maintained. The baseline distance... The system can be adjusted based on the reliability of the predicted data to adapt to the random fluctuations of distributed power sources. This allows the boundary to maintain a baseline distance when the prediction uncertainty is minimized, and to appropriately expand the boundary when the uncertainty is maximized, achieving adaptive scaling of the safe zone to power fluctuations. This results in an assessment area consisting of a dynamic safe zone, a dynamic early warning zone, and a dynamic over-limit risk zone. The assessment area and the predicted data obtained by the power prediction model constitute the final output of the distribution network operation risk situation assessment model.

[0045] This embodiment provides a training method for a distribution network operation risk assessment model. It uses clustering to divide typical operating scenarios, leverages voltage data to drive predictions to fill in missing power and extract risk features, and then constructs an assessment region that can adaptively scale with prediction reliability and scenario requirements, achieving accurate assessment of distribution network operation risks. The technical solution provided in this embodiment can effectively perceive risks even when power measurements are significantly missing, maintaining high-sensitivity early warning for high-risk scenarios such as severe photovoltaic fluctuations and concentrated electric vehicle charging, while reducing false alarm rates in stable operating scenarios.

[0046] Figure 2 The flowchart provided in this application for obtaining a reference region model for outputting a reference security region may include the following steps: Step S51: Divide the power grid dataset into safe samples according to a preset threshold, and construct a comprehensive weight based on risk-sensitive data.

[0047] Specifically, samples that simultaneously meet the following conditions are defined as safe samples: the per-unit value of the node voltage is within the range of [0.93, 1.07], the branch load rate is ≤80%, and the load coupling coefficient (DG-voltage coupling coefficient) λ > -0.5 (no risk of strong reverse coupling of DG). These samples are used only for training the benchmark regional model. Data that do not meet any of the conditions or are within one hour before or after a historical fault are defined as risk samples. These samples are used only for verifying the benchmark regional model. Furthermore, samples of scenarios such as sudden changes in DG output, concentrated charging of EVs, and extreme load fluctuations are extracted from the risk samples to form a rare high-risk sample set, which is specifically used to verify the regional model's ability to capture low-probability extreme risks.

[0048] Step S53: Using the comprehensive weights and preset update coefficients used to characterize the stability of the iteration, construct the boundary region for wrapping the safe sample, and determine the region center of the boundary region and the region distance from the region center to the boundary.

[0049] Specifically, to adapt to the differentiated security boundary requirements under different operating scenarios, an adaptive kernel function is constructed as follows: In the formula: and Given the feature vectors of two samples, The Euclidean distance between the two is given by the following example: For instance, for the DG access node No. 10, during the midday peak photovoltaic period, corresponding to scenario label 1 (high volatility + DG strong coupling scenario), its feature vector is [0.9920, 0.3920, 0.0120, 0.4160, 0.0680, 0.7200, 0.0050, 0, 0.1250, 0.0812, 1, 1]. The Euclidean distance between the two 12-dimensional feature vectors is the square root of the sum of the squares of the differences in the corresponding dimensions of the two vectors. For the first The adaptive kernel width for each sample is expressed as follows: in, The baseline kernel width is determined by optimization using a 5-fold cross-validation grid. For example, the candidate value range is [0.01, 0.1, 1, 10, 100]. Each candidate value is tried in turn, its performance is evaluated, and finally the one with the best performance is selected. The optimization goal is to maximize the value of the validation set. , The weighting coefficients are all set to 0.3 to ensure... This design allows the kernel width to be automatically scaled according to the fluctuation intensity or coupling degree of the samples, thereby flexibly adjusting the similarity measurement scale in the feature space.

[0050] The original optimization objective of constructing the baseline region model is to find a boundary region in the high-dimensional feature space that can encompass all safe samples. The original problem of the boundary region is formalized as follows: in: The radius (the distance from the center of the region to the boundary of the region); As the regional center; For high-dimensional feature mapping induced by adaptive kernel functions, it can be achieved through... Obtain; The penalty coefficient is determined through grid optimization. The overall weight of the sample; It is a relaxation factor; This represents the total number of samples in the training set.

[0051] To solve this convex optimization problem, Lagrange multipliers are introduced. and The original constrained problem is transformed into an unconstrained Lagrangian function, which is expressed in the following form: If the sample does not meet the constraints, then For positive, to If we take it to infinity, the Lagrange function will become infinite.

[0052] If the sample satisfies the constraint that the term in parentheses is less than or equal to 0, then... After maximization, the Lagrange function will be equal to the original objective function.

[0053] Therefore, the problem of minimizing the original objective function is completely equivalent to the problem of minimizing and maximizing the Lagrange function.

[0054] right , , By taking the partial derivatives and setting them to zero, we obtain the boundary region, thus transforming the original problem into a dual optimization problem. The specific method for obtaining the boundary region is as follows: To minimize the original variables in the Lagrange function, we need to take the partial derivatives with respect to each of the three variables and set the partial derivatives to zero.

[0055] Condition 1: Take the partial derivative with respect to radius R and set the partial derivative to 0.

[0056] Simplifying, we get: Condition 2: Take the partial derivative with respect to the center of the sphere a, and set the partial derivative to 0.

[0057] Simplifying, we get: That is, the region center 'a' is directly represented by Lagrange multipliers.

[0058] Condition 3: Take the partial derivative with respect to the relaxation factor and set the partial derivative to 0. get: Combination ≥0 can be obtained Substitute these conditions back into the Lagrange function and use the adaptive kernel function. By replacing the inner product, all variables can be eliminated, resulting in a dual optimization problem containing only Lagrange multipliers: Eliminate sub-items 2, 7, and 8 according to condition 3; eliminate sub-items 1 and 6 according to condition 1; and eliminate sub-items 2 according to condition 2. use and Substitution, and according to Mercer's theorem: a symmetric, continuous function It can be used as an adaptive kernel function (i.e., there exists a mapping in a certain feature space such that...). .

[0059] The final representation of the boundary region is as follows: Step S55: Iterate over the regional distance and regional center until the preset conditions are met to obtain the baseline regional model.

[0060] Specifically, the distance of a safe sample to the current boundary region determines its potential to become a support vector—the closer a safe sample is to the boundary, the higher its corresponding support vector potential. The easier it is for the weight of a sample to be greater than zero; while the overall weight of a safe sample... Then by raising upper limit This gives secure samples a stronger ability to influence boundary regions. While meeting the requirements... and Under the constraints, continuously select the two security samples with the most severe boundary violations and update their corresponding... The process continues until all safe samples meet the conditions, at which point the optimal Lagrange multiplier is finally output. Based on this, it will satisfy The sample was determined to be a support vector. These samples happen to be located on or near the boundary of the boundary region.

[0061] The center of the baseline safe region is obtained by a weighted combination of all support vectors, i.e. For the baseline distance, arbitrarily select a support vector. Calculate the distance from the support vector to the center of the baseline safe region, and take the average of the distances of all support vectors as the final baseline distance. : This embodiment filters safe samples using thresholds such as node voltage and branch load rate, and obtains comprehensive weights by combining risk-sensitive data. Then, an adaptive kernel function is introduced to construct an optimization problem, finding a boundary region that minimizes the size of all safe samples, and the original problem is transformed into a dual problem using the Lagrange multiplier method. Subsequently, based on the obtained support vectors, the region center is determined by weighting, and the mean distance from the support vectors to the center is taken as the benchmark distance, resulting in a benchmark safe region model defined by the region center and the benchmark distance.

[0062] Figure 3 A flowchart illustrating the comprehensive weight acquisition method provided in this application embodiment, the process may include the following steps: Step S511: Determine the power completion confidence weight based on the ratio of the power fitting confidence to the maximum power fitting confidence.

[0063] Specifically, power completion confidence weights To address the issue of boundary accuracy interference in low-confidence samples, its representation is as follows: in, For the first The confidence level of the power fit for each sample; The maximum absolute confidence score for fitting the power of all samples; weights The smaller the residual, the closer the sample weight is to 1, which indicates a higher confidence level in power completion, thus effectively suppressing the interference of low-confidence samples on boundary accuracy.

[0064] Step S513: Based on load volatility and load coupling coefficient, construct risk-sensitive weights to characterize the degree of influence of different samples.

[0065] Specifically, in order to effectively characterize the impact of different samples on model training and address the problem of easily missing rare high-risk samples, this paper constructs risk-sensitive weights based on load volatility and the DG-load coupling coefficient. This weight directly affects the loss function or sample importance weighting, causing the model to pay more attention to high-risk samples with drastic changes in operating state and tense coupling relationships. Its calculation formula is: In the formula: For the first The load volatility of each sample reflects the degree of drastic change in load over time. For the first The absolute value of the load coupling coefficient of each sample measures the strength of the coupling relationship between distributed power sources and loads. The larger the load coupling coefficient, the higher the risk of power flow reversal or strong coupling. As the balance coefficient, when It can guarantee that... , , That is, the baseline weight is 1, and the higher the risk, the closer the weight is to 2. Through this construction, samples with more extreme load fluctuations and stronger DG coupling will receive greater weights, thereby significantly enhancing the model's ability to capture rare high-risk samples under imbalanced data distribution and reducing the risk of missing detection in low-probability extreme operating scenarios.

[0066] Step S515: Determine the comprehensive weight based on the product of the power completion confidence weight and the risk sensitivity weight.

[0067] Specifically, the overall sample weight is determined by multiplying the power completion confidence weight and the risk sensitivity weight. Its representation is as follows: in, The confidence weights for power completion are used to address the issue of sample interference boundary accuracy in low-power fitted confidence scores. The aforementioned risk-sensitive weights are used to address the issue of easily overlooking rare, high-risk samples. Multiplying these two factors allows the final comprehensive weights to adaptively adjust based on both the power fit confidence and the sample risk level. This guides subsequent training to both emphasize filling quality differences and strengthen the perception and punishment of high-risk situations, thereby improving the reliability and security of the overall decision-making process.

[0068] This embodiment determines the comprehensive weight by multiplying the power completion confidence weight and the risk-sensitive weight. The power completion confidence weight is constructed from the ratio of the power fitting confidence score to its maximum value, used to suppress the interference of low-confidence samples on boundary accuracy. The risk-sensitive weight is constructed based on load volatility and load coupling coefficient, and is adjusted by a balancing coefficient to adaptively scale the weight with the risk level, thereby enhancing the ability to capture rare high-risk samples. The comprehensive weight, resulting from the multiplication of the two, simultaneously considers the differences in power completion quality and the sample risk level, guiding subsequent training to both emphasize differences in completion confidence and strengthen the perception and penalty for high-risk situations.

[0069] In an optional embodiment, the power fitting confidence, load volatility, and load coupling coefficient are obtained in the following ways: The power fit confidence level is determined based on the numerical difference between the predicted data at the same node and the power grid dataset.

[0070] Specifically, for nodes in the power grid dataset with measured power, the numerical difference between the predicted data and the corresponding measured value in the power grid dataset, i.e., the power fitting confidence score, is used to quantify the confidence score of the node's completed data. Its representation is as follows: in, For nodes The power fit confidence score is the numerical difference between the predicted data and the power grid dataset. For nodes Measured power values ​​in the power grid dataset; For nodes The predicted data is used. For nodes without actual power measurements, the average value of the numerical differences between nodes of the same type and in the same scenario is used as an estimate of their fitting residuals, thereby determining their power fitting confidence. After obtaining the total power of all nodes, real-time power flow calculations are performed using the forward-backward substitution method, combined with fixed topology parameters, to obtain the per-unit voltage values ​​of all nodes, the transmission power of all branches, and the load factor. These values ​​are then verified against the measured voltages to ensure the electrical rationality of the total power.

[0071] Based on the power grid dataset within a preset time window, determine the power standard deviation and power average within that time window.

[0072] The load fluctuation rate is determined by the ratio of the power standard deviation to the power average.

[0073] Specifically, to quantify the severity of load fluctuations over a short period, a 1-hour sliding window is used to continuously scan the load power data in the power grid dataset. For each sampling time, the active power data of the load within the preceding 1-hour window is extracted, and the standard deviation of the power within the window is calculated. and average power The ratio of the two is taken as the load volatility at that moment. , This indicator eliminates the influence of absolute load levels by using a ratio, thus objectively reflecting the relative fluctuation range of load over short time scales. By iterating through all time points using a sliding window, the corresponding load volatility can be determined for each sample. This provides a foundation for the subsequent construction of risk-sensitive weights.

[0074] The coordinated fluctuation rate is determined based on the degree of coordinated fluctuation between the power output rate of each power source and the power of each node in the power grid data.

[0075] The fitting standard deviation is determined by multiplying the standard deviation of the power output by the standard deviation of the power.

[0076] The load coupling coefficient is determined based on the ratio of the coordinated volatility to the fitted standard deviation.

[0077] Specifically, the coordinated fluctuation rate is determined based on the degree of coordinated fluctuation between the power output rate (DG output) and node power in the power grid data. The fitting standard deviation is determined by multiplying the standard deviation of power output by the standard deviation of power. Based on this, the load coupling coefficient is... Used to characterize the coupling characteristics between the output rate of each power source and the nodal load, expressed as the ratio of the modulation fluctuation rate to the standard deviation of the fit: in: The covariance of the power output rate of each power source and the load of each node within the sliding window is called the coordinated volatility, which reflects the degree of coordinated volatility between the two. The standard deviation of the output rate of each power source within the sliding window; is the standard deviation of the node load within the sliding window; the product of the two is the fitted standard deviation, used to eliminate dimensions and normalize the cooperative fluctuation amplitude. The larger the absolute value of 1, the stronger the coupling between the power output rate of each power source and the load, and vice versa.

[0078] This embodiment effectively quantifies the reliability of full-node power completion data by determining the power fitting confidence level, providing a reliable basis for subsequently suppressing the interference of low-confidence samples on boundary accuracy. Load volatility eliminates the influence of absolute levels through a ratio, objectively reflecting the severity of short-term fluctuations, while the load coupling coefficient accurately characterizes the coordinated changes in power output and load. Both provide key features for constructing risk-sensitive weights. This design significantly enhances the ability to perceive and distinguish high-risk operating states, ensuring the adaptive adjustment of comprehensive weights to differences in completion quality and risk levels, thereby improving the accuracy of the baseline safe zone model and the reliability of early warnings in rare extreme scenarios.

[0079] Figure 4 The flowchart provided in this application embodiment updates the baseline distance of the baseline safe region model based on the confidence level of the predicted data. The flowchart may include the following steps: Step S71: Obtain multiple prediction data at the current time step, and truncate the multiple prediction data based on a preset judgment interval to obtain the width of the confidence interval corresponding to the current time step.

[0080] Specifically, power prediction values ​​are obtained for a power prediction model with a 15-minute timescale. The Dropout layer of the power prediction model is kept enabled, and 100 independent forward inferences are performed, randomly discarding 20% ​​of the neurons each time, resulting in 100 independent prediction data points. These prediction data points are then truncated based on a preset decision interval, i.e., the prediction data are sorted in ascending order, and the 2.5th percentile is used as the lower limit. Using the 97.5th percentile as the upper limit This gives us the width of the confidence interval corresponding to the current time step, calculated using the following formula: in: The confidence interval width is denoted as , and normalized to . interval; The larger the value, the greater the uncertainty of future power.

[0081] Step S73: Update the baseline distance of the baseline safe zone model according to the width of the confidence interval through a preset adjustment rule to obtain the distribution network operation risk situation assessment model.

[0082] Specifically, based on the uncertainty of future power forecasts, the baseline radius is dynamically adjusted online to adapt to the random fluctuations of distributed power sources. The specific update formula is defined as follows: in: for The updated baseline distance at any given time; for The width of the confidence interval at each time point is normalized to... interval; To adjust the coefficient, we set it to 0.8, thus ensuring that the dynamic radius always meets the requirements. This rule ensures that when the prediction uncertainty is minimized ( The updated baseline distance maintains the baseline distance. When uncertainty is at its maximum ( The boundary is appropriately expanded to 1.8 times the base distance to achieve adaptive scaling of the safe zone model to power fluctuations. The updated safe zone model is used as the model for assessing the operational risk situation of the distribution network.

[0083] This embodiment obtains multiple independent prediction data points simultaneously, arranges the prediction data in ascending order, and then extracts a portion using a preset judgment interval to obtain the normalized confidence interval width for the current time step, quantifying the uncertainty of power prediction. Then, based on preset adjustment rules, the benchmark distance of the benchmark safe zone model is dynamically updated online using this confidence interval width, thereby constructing a distribution network operation risk assessment model. This embodiment introduces the uncertainty of power prediction into the adaptive adjustment of the benchmark distance, ensuring that the benchmark distance remains unchanged when prediction uncertainty is minimal, and appropriately expands the benchmark boundary when uncertainty is maximum. This achieves adaptive scaling of the safe zone to random fluctuations of distributed power sources, effectively improving the accuracy and robustness of distribution network operation risk assessment.

[0084] In one optional embodiment, power grid operation data of the operating status of each node in the distribution network under test is obtained.

[0085] The power grid operation data is input into the training method from S1 to S5 to obtain the distribution network operation risk situation assessment model, obtain the assessment area and prediction data, and output the graded early warning level based on the changing trend of the prediction data and the assessment area.

[0086] Specifically, in practical use, an adaptive kernel function identical to that used in offline training is employed to calculate the distance from future running points in the predicted data to the judgment center of the judgment region. The standardized feature vector for future moments is... , To predict the time scale, prediction data for three different time spans—15 min, 30 min, and 60 min—were collected. This prediction data can be obtained simply by extending the power prediction model within the distribution network operation risk assessment model to different time dimensions. Therefore, the squared distance from this data to the assessment center of the assessment area is: For predictive data; For the prediction time scale, values ​​of 15 min / 30 min / 60 min are used; This refers to the optimal Lagrange multiplier obtained during the training process.

[0087] Risk levels are determined using assessment rules that conform to the conservative principle of distribution network dispatching. A veto rule is applied: if the risk exceeds the probability threshold at any time scale... If it does not, it is directly judged as extremely high risk and enters the red alert preparedness state. Trend reinforcement rule: when the first difference And second-order difference When this occurs, it indicates that the running point is accelerating towards the boundary, and even if the current risk breakthrough probability is less than 50%, the risk level will be increased by one level. Scenario adaptation rule: For scenario labels with high volatility and strong DG coupling, all risk levels are increased by half a level. False alarm avoidance rule: If the risk breakthrough probability is less than 30% and the first-order difference... If the trend is positive, it is considered a safe trend and no warning is triggered.

[0088] The specific triggering conditions and corresponding response recommendations for tiered early warnings are defined as follows: Blue Alert (Attention Level): Triggered by a risk spike probability within the next 60 minutes. And first-order difference This could indicate that the current operating point has entered a relaxed alert zone. The prediction timescale is 60 minutes in advance, and the recommended action is to strengthen operational monitoring and prepare contingency plans.

[0089] Yellow Alert (Alert Level): Triggered by the probability of a risk breakthrough within the next 30 minutes. And first-order difference Second-order difference The current operating point may enter the baseline warning zone. The prediction time scale is 30 minutes in advance, and the recommended action is to activate lightweight pre-control measures and optimize reactive power compensation and DG output.

[0090] Orange alert (emergency level): Triggered by the probability of a risk breach within the next 15 minutes. Furthermore, the distance from the first-order difference continues to increase, or the operating point enters the tightening warning zone. The prediction time scale is 15 minutes in advance, and the recommended action is to immediately implement the pre-control plan, limit interruptible loads, and optimize power flow distribution.

[0091] Red Alert (Crisis Level): Triggered by a risk exceeding the threshold within the next 15 minutes. And it exhibits the characteristic of accelerating towards the boundary. This could indicate that the operating point is about to exceed the safety boundary. The prediction timescale is 15 minutes in advance, and the recommended response is to implement the emergency prevention and control plan, including load shedding, blocking DG output, and isolating the risk section.

[0092] Finally, based on the above calculation results, a risk heat map of the distribution network nodes and a risk trend curve for the next hour are generated and pushed to the distribution network dispatch and monitoring system in real time. At the same time, a handling work order corresponding to the warning level is automatically generated and pushed to the operation and maintenance management system to achieve closed-loop management of the entire process of risk.

[0093] In an optional embodiment, based on the changing trend of the predicted data and the assessed area, a graded early warning level is output, including: Determine the first-order difference of the forecast data to characterize the rate of risk change and the second-order difference to characterize the acceleration of risk change.

[0094] Based on the predicted data, the probability of a breakthrough is obtained to characterize the area to be judged, and the probability of a risk breakthrough is determined.

[0095] Based on the first-order difference, second-order difference, risk breakthrough probability, and assessment area, a graded early warning level is output through preset output rules.

[0096] Specifically, risk trend indicators are constructed from three dimensions to capture the direction and speed of risk evolution. First, the first-order difference is used to characterize the velocity and distance of future points in the forecast data as they approach / move from the boundary. Its representation is as follows: in, The distance from the current operating point to the analysis center. The distance from the forecast data 15 minutes later to the analysis center.

[0097] Second, the second-order difference is used to depict the acceleration trend near the boundary, and its representation is as follows: in, The first difference is thirty minutes later. This is the first-order difference 15 minutes later.

[0098] Third, to quantify the risk of boundary breaches under random fluctuations, Monte Carlo simulations were performed based on the 95% confidence interval of the predictions. Assuming the power prediction error follows a normal distribution, 1000 independent samples were taken within the confidence interval to generate 1000 sets of prediction data scenarios. For each scenario, power flow calculation, feature construction, and distance calculation were performed sequentially to obtain the distance values ​​from 1000 future operating points to the assessment center. Combined with the predicted dynamic boundary radius at the corresponding time scale Count the number of scenarios that breach the boundary and calculate the probability of risk breach: in: =1000 represents the number of Monte Carlo samplings; This is an indicator function that takes the value 1 if the condition is met, and 0 otherwise. This represents the probability of a risk breakthrough at a future moment, with a value range of [0,1].

[0099] Accordingly, please refer to Figure 5 A block diagram of a power distribution network operation risk situation assessment model training device provided in this application embodiment, the device comprising: Data acquisition unit 101 is used to acquire the historical operating status of each node in the distribution network and construct a power grid dataset; The data prediction unit 103 is used to input the power grid dataset into a preset power prediction model, obtain prediction data, and obtain risk-sensitive data to characterize the dynamic confidence features of the prediction data based on the prediction data and the power grid dataset. The region generation unit 105 is used to input the power grid dataset, prediction data and risk-sensitive data into a pre-trained benchmark region model for region division, and obtain a benchmark region model for outputting the benchmark safe region through iterative optimization. The dynamic update unit 107 is used to update the baseline distance of the baseline safety area model based on the credibility of the prediction data, so as to obtain the distribution network operation risk situation assessment model. The distribution network operation risk situation assessment model is used to output the assessment area and prediction data.

[0100] In some optional implementations, the data prediction unit 103 includes: The power grid dataset is divided into safe samples according to a preset threshold, and a comprehensive weight is constructed based on risk-sensitive data. Using comprehensive weights and preset update coefficients to characterize iterative stability, a boundary region for wrapping safe samples is constructed, and the region center and the region distance from the region center to the boundary are determined. Iterate through the regional distance and regional center until the preset conditions are met to obtain the baseline regional model.

[0101] In some optional implementations, the data prediction unit 103 includes: The power completion confidence weight is determined based on the ratio of the power fit confidence score to the maximum power fit confidence score. Based on load volatility and load coupling coefficient, risk-sensitive weights are constructed to characterize the degree of impact of different samples. The overall weight is determined by multiplying the power completion confidence weight and the risk sensitivity weight.

[0102] In some optional implementations, the data prediction unit 103 includes: The power fit confidence level is determined based on the numerical difference between the predicted data and the power grid dataset at the same node. Based on the power grid dataset within the preset time window, determine the power standard deviation and power average within the time window; The load fluctuation rate is determined by the ratio of the power standard deviation to the power average. The coordinated fluctuation rate is determined based on the degree of coordinated fluctuation between the output rate of centralized distributed power sources and the power of nodes in the power grid data. The fitting standard deviation is determined by multiplying the standard deviation of the output rate of the distributed power source by the standard deviation of the power. The load coupling coefficient is determined based on the ratio of the coordinated volatility to the fitted standard deviation.

[0103] In some optional implementations, the region generation unit 105 includes: Obtain multiple prediction data at the current time step, and truncate the multiple prediction data based on a preset judgment interval to obtain the width of the confidence interval corresponding to the current time step; The baseline distance of the baseline safe zone model is updated according to the width of the confidence interval through preset adjustment rules to obtain the distribution network operation risk situation assessment model.

[0104] In this embodiment, a power distribution network operation risk situation assessment model training device is presented in the form of functional units. Here, a unit refers to an ASIC (Application Specific Integrated Circuit) circuit, a processor and memory that execute one or more software or fixed programs, and / or other devices that can provide the above functions.

[0105] Please see Figure 5 , Figure 5 This application provides a schematic diagram of the structure of a computer device, as shown in the embodiment of the present application. Figure 5 As shown, the computer device includes one or more processors 10, memory 20, and interfaces for connecting the components, including high-speed interfaces and low-speed interfaces. The components communicate with each other via different buses and can be mounted on a common motherboard or otherwise installed as needed. The processors can process instructions executed within the computer device, including instructions stored in or on memory to display graphical information of a GUI on external input / output devices (such as display devices coupled to the interfaces). In some alternative implementations, multiple processors and / or multiple buses can be used with multiple memories and multiple memory modules, if desired. Similarly, multiple computer devices can be connected, each providing some of the necessary operations (e.g., as a server array, a group of blade servers, or a multiprocessor system). Figure 5 Take a processor 10 as an example.

[0106] Processor 10 may be a central processing unit, a network processor, or a combination thereof. Processor 10 may further include a hardware chip. The hardware chip may be an application-specific integrated circuit (ASIC), a programmable logic device (PLD), or a combination thereof. The programmable logic device may be a complex programmable logic device (CAMP), a field-programmable gate array (FPGA), a general-purpose array logic (GDA), or any combination thereof.

[0107] The memory 20 stores instructions executable by at least one processor 10 to cause at least one processor 10 to perform the method shown in the above embodiments.

[0108] The memory 20 may include a program storage area and a data storage area. The program storage area may store the operating system and applications required for at least one function; the data storage area may store data created based on the use of the computer device. Furthermore, the memory 20 may include high-speed random access memory and may also include non-transitory memory, such as at least one disk storage device, flash memory device, or other non-transitory solid-state storage device. In some alternative embodiments, the memory 20 may optionally include memory remotely located relative to the processor 10, and these remote memories may be connected to the computer device via a network. Examples of such networks include, but are not limited to, the Internet, intranets, local area networks, mobile communication networks, and combinations thereof.

[0109] The memory 20 may include volatile memory, such as random access memory; the memory may also include non-volatile memory, such as flash memory, hard disk or solid-state drive; the memory 20 may also include a combination of the above types of memory.

[0110] The computer device also includes a communication interface 30 for communicating with other devices or communication networks.

[0111] This application also provides a computer-readable storage medium. The methods described in this application can be implemented in hardware or firmware, or implemented as recordable on a storage medium, or implemented as computer code downloaded over a network and originally stored on a remote storage medium or a non-transitory machine-readable storage medium and subsequently stored on a local storage medium. Thus, the methods described herein can be processed by software stored on a storage medium using a general-purpose computer, a dedicated processor, or programmable or dedicated hardware. The storage medium can be a magnetic disk, optical disk, read-only memory, random access memory, flash memory, hard disk, or solid-state drive, etc.; further, the storage medium can also include combinations of the above types of memory. It can be understood that the computer, processor, microprocessor controller, or programmable hardware includes a storage component capable of storing or receiving software or computer code, which, when accessed and executed by the computer, processor, or hardware, implements the methods shown in the above embodiments.

[0112] The apparatus and units described in the above embodiments can be implemented by a computer chip or physical entity, or by a product with a certain function. A typical implementation device is a computer. Specifically, a computer can be, for example, a personal computer, a laptop computer, a cellular phone, a camera phone, a smartphone, a personal digital assistant, a media player, a navigation device, an email device, a game console, a tablet computer, a wearable device, or any combination of these devices.

[0113] For ease of description, the above devices are described separately by function as various units. Of course, in implementing this application, the functions of each unit can be implemented in one or more software and / or hardware.

[0114] Those skilled in the art will understand that the embodiments of this application can be provided as methods or apparatus. Therefore, this application can take the form of a completely hardware embodiment, a completely software embodiment, or an embodiment combining software and hardware aspects. Furthermore, this application can take the form of a computer program product embodied on one or more computer-usable storage media (including but not limited to disk storage, CD-ROM, optical storage, etc.) containing computer-usable program code.

[0115] This application is described with reference to flowchart illustrations and / or block diagrams of methods, apparatuses, and devices according to embodiments of this application. It will be understood that each block of the flowchart illustrations and / or block diagrams, and combinations of blocks in the flowchart illustrations and / or block diagrams, can be implemented by computer program instructions. These computer program instructions can be provided to a processor of a general-purpose computer, special-purpose computer, embedded processor, or other programmable data processing apparatus to produce a machine, such that the instructions, which execute via the processor of the computer or other programmable data processing apparatus, generate instructions for implementing the process. Figure 1 One or more processes and / or boxes Figure 1 A device that provides the functions specified in one or more boxes.

[0116] These computer program instructions may also be stored in a computer-readable storage medium that can direct a computer or other programmable data processing device to function in a particular manner, such that the instructions stored in the computer-readable storage medium produce an article of manufacture including instruction means, which are implemented in a process Figure 1 One or more processes and / or boxes Figure 1 The function specified in one or more boxes.

[0117] These computer program instructions may also be loaded onto a computer or other programmable data processing equipment to cause a series of operational steps to be performed on the computer or other programmable equipment to produce a computer-implemented process, thereby providing instructions that execute on the computer or other programmable equipment for implementing the process. Figure 1 One or more processes and / or boxes Figure 1 The steps of the function specified in one or more boxes.

[0118] It should also be noted that 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. Unless otherwise specified, an element defined by the phrase "comprising one..." does not exclude the presence of other identical elements in the process, method, article, or apparatus that includes that element.

[0119] The various embodiments in this specification are described in a progressive manner. Similar or identical parts between embodiments can be referred to mutually. Each embodiment focuses on describing the differences from other embodiments. In particular, the apparatus embodiments are basically similar to the method embodiments, so the description is relatively simple; relevant parts can be referred to the descriptions of the method embodiments.

[0120] The above are merely embodiments of this application and are not intended to limit the scope of this application. Various modifications and variations can be made to this application by those skilled in the art. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of this application should be included within the scope of the claims of this application.

[0121] Although embodiments of this application have been described in conjunction with the accompanying drawings, those skilled in the art can make various modifications and variations without departing from the spirit and scope of this application, and such modifications and variations all fall within the scope defined by the appended claims.

Claims

1. A training method for a distribution network operation risk situation assessment model, characterized in that, The method includes: Obtain the historical operating status of each node in the power distribution network and construct a power grid dataset; Input the power grid dataset into a preset power prediction model to obtain prediction data, and based on the prediction data and the power grid dataset, obtain risk-sensitive data to characterize the dynamic confidence features of the prediction data; The model is trained using the power grid dataset, the predicted data, and the risk-sensitive data, and iteratively optimized to obtain a benchmark safe region model for outputting the benchmark safe region. Obtaining the benchmark safe region model includes: dividing the power grid dataset into safe samples according to a preset threshold, and constructing comprehensive weights based on the risk-sensitive data; using the comprehensive weights and preset update coefficients to characterize iterative stability, constructing a boundary region to enclose the safe samples, and determining the region center and the distance from the region center to the boundary; iterating over the region distance and the region center until a preset condition is met to obtain the benchmark safe region model. The risk-sensitive data includes power fitting confidence, load volatility, and load coupling coefficient. The comprehensive weights are obtained by: determining the power completion confidence based on the ratio of the power fitting confidence to the maximum value of the power fitting confidence. The weighting is as follows: Based on the load volatility and the load coupling coefficient, a risk-sensitive weight is constructed to characterize the degree of influence of different samples; a comprehensive weight is determined by multiplying the power completion confidence weight and the risk-sensitive weight; the acquisition methods of the power fitting confidence, load volatility, and load coupling coefficient include: determining the power fitting confidence based on the numerical difference between the predicted data and the power grid dataset under the same node; determining the power standard deviation and power average under the power grid dataset within a preset time window; determining the load volatility based on the ratio of the power standard deviation to the power average; determining the coordinated volatility based on the degree of coordinated volatility between the distributed generation output rate and the node power in the power grid dataset; determining the fitting standard deviation based on the product of the standard deviation of the distributed generation output rate and the power standard deviation; and determining the load coupling coefficient based on the ratio of the coordinated volatility to the fitting standard deviation. Based on the reliability of the predicted data, the baseline distance of the baseline safety area model is updated to obtain the distribution network operation risk situation assessment model. The distribution network operation risk situation assessment model is used to output the assessment area and the predicted data.

2. The method according to claim 1, characterized in that, The step of updating the baseline distance of the baseline safe zone model based on the confidence level of the predicted data includes: Obtain multiple prediction data at the current time step, and truncate the multiple prediction data based on a preset judgment interval to obtain the width of the confidence interval corresponding to the current time step; The baseline distance of the baseline safety zone model is updated according to the width of the confidence interval using a preset adjustment rule to obtain the distribution network operation risk situation assessment model.

3. A method for assessing the operational risk situation of a power distribution network, characterized in that, The assessment methods include: Obtain power grid operation data on the operating status of each node in the distribution network under test; The power grid operation data is input into the distribution network operation risk situation assessment model obtained by the training method of any one of claims 1-2 to obtain the assessment area and prediction data. Based on the changing trend of the prediction data and the assessment area, a graded early warning level is output.

4. The method according to claim 3, characterized in that, The step of outputting a graded early warning level based on the changing trend of the predicted data and the assessed area includes: The predicted data are used to characterize the first-order difference of the risk change rate and the second-order difference of the risk change acceleration. Based on the predicted data, obtain the probability of breaking through the assessment area to determine the risk breakthrough probability; Based on the first-order difference, the second-order difference, the risk breakthrough probability, and the judgment area, a graded early warning level is output through preset output rules.

5. An apparatus for implementing the training method for the distribution network operation risk situation assessment model according to any one of claims 1-2, characterized in that, The device includes: The data acquisition unit is used to acquire the historical operating status of each node in the power distribution network and construct a power grid dataset. The data prediction unit is used to input the power grid dataset into a preset power prediction model, obtain prediction data, and obtain risk-sensitive data to characterize the dynamic confidence features of the prediction data based on the prediction data and the power grid dataset. The region generation unit is used to train the model using the power grid dataset, the prediction data, and the risk-sensitive data, and to obtain a benchmark safe region model for outputting the benchmark safe region through iterative optimization. The dynamic update unit is used to update the baseline distance of the baseline safety area model based on the reliability of the predicted data, so as to obtain the distribution network operation risk situation assessment model. The distribution network operation risk situation assessment model is used to output the assessment area and the predicted data.

6. A computer device, characterized in that, include: The system includes a memory and a processor, which are interconnected. The memory stores computer instructions, and the processor executes the computer instructions to perform the power distribution network operation risk situation assessment model training method as described in any one of claims 1 to 2.

7. A computer-readable storage medium, characterized in that, The computer-readable storage medium stores computer instructions, which are used to cause the computer to execute the power distribution network operation risk situation assessment model training method according to any one of claims 1 to 2.

Images