A method, system, equipment, and medium for risk assessment of distribution network scheduling operations based on grey relational feature screening and adaptive NODE model.

Abstract

This invention relates to the field of distribution network dispatching and operation safety technology, and in particular to a method, system, equipment, and medium for risk assessment of distribution network dispatching operations based on grey relational analysis feature screening and an adaptive NODE model. The method involves acquiring multi-source monitoring data of distribution network dispatching operations, using grey relational analysis to perform quantitative analysis of correlations between indicators such as node voltage, line power flow, and load shedding, and screening to form a joint feature dataset. Based on this dataset, a NODE risk assessment model with a continuous state evolution equation as its core is constructed. A neural network is used to fit the differential change law of the distribution network state with dispatching operations. A population optimization strategy incorporating adaptive t-distribution perturbations is used to optimize the model's hyperparameter combination to obtain the optimal hyperparameter combination. The optimal hyperparameters are then substituted into the model to complete training, enabling continuous dynamic risk assessment of distribution network dispatching operations and outputting the results.

Description

Technical Field

[0001] This invention relates to the field of distribution network dispatching and operation safety technology, and in particular to a method, system, equipment and medium for risk assessment of distribution network dispatching operations based on grey relational feature screening and adaptive NODE model. Background Technology

[0002] Distribution network dispatching operations cover a variety of scenarios, including switching operations, load transfer, fault handling, and the access and disconnection of distributed power sources. These operations are frequent and affected by various factors such as fluctuations in the output of distributed power sources, load changes, and grid topology adjustments. They also carry multiple risks, such as voltage exceeding limits, line overload, power outages for users, and cascading failures. Therefore, accurate and real-time risk assessment of distribution network dispatching operations is crucial to ensuring the safe and stable operation of the distribution network.

[0003] Currently, the mainstream methods for risk assessment in distribution network dispatching operations mainly include static analytical methods, fuzzy comprehensive evaluation methods, and Monte Carlo simulation methods. Static analytical methods rely on human experience to build assessment models, making it difficult to adapt to the dynamic characteristics of distribution networks; fuzzy comprehensive evaluation methods are highly subjective, and their assessment accuracy is greatly affected by the quality of prior knowledge; Monte Carlo simulation methods have high computational complexity and poor real-time performance, making them unsuitable for online dispatching decision-making. With the large-scale integration of high-proportion distributed power sources, distribution networks exhibit characteristics such as nonlinearity, time-varying nature, and dynamic topology adjustments. Traditional methods cannot accurately capture the continuous evolution of risks throughout the entire dispatching operation process, and their limitations are becoming increasingly prominent.

[0004] Neural-Ordinary-Differential Equation (NODE), as a novel time-series modeling method, uses continuous differential equations to characterize the evolution of system states. It has advantages in the state deduction and prediction of nonlinear dynamic systems, providing a new technical approach for the dynamic assessment of distribution network scheduling operational risks. However, existing NODE models still have significant shortcomings in this scenario: the optimizer parameters are fixed, making it unable to adapt to dynamic changes such as distribution network topology adjustments; the solution step size for differential equations is rigid, making it difficult to balance real-time performance and accuracy; and the lack of correlation screening of multi-dimensional monitoring indicators leads to slow model convergence and weak generalization ability, making it difficult to support intelligent scheduling decisions in complex distribution network scenarios. Summary of the Invention

[0005] In view of the problems existing in the prior art, the present invention is proposed.

[0006] Therefore, the problem to be solved by this invention is how to address the shortcomings of existing distribution network scheduling operation risk assessment methods in complex scenarios such as high proportion of distributed power source access and frequent topology adjustments, which are unable to accurately capture the continuous evolution of risks throughout the scheduling operation process. Existing NODE models have fixed hyperparameters, lack adaptive optimization, and lack correlation quantification screening methods for multi-dimensional monitoring indicators of the distribution network, resulting in slow model convergence, weak generalization, and insufficient real-time assessment, making it difficult to support online risk warning and dynamic scheduling decisions.

[0007] To solve the above-mentioned technical problems, the present invention provides the following technical solution: In a first aspect, embodiments of the present invention provide a method for risk assessment of distribution network scheduling operations based on grey relational feature screening and adaptive NODE model, which includes: acquiring multi-source monitoring data of distribution network scheduling operations; performing quantitative analysis of the correlation between various monitoring indicators under multiple operation scenarios; and screening to form a joint feature dataset. Based on the joint feature dataset, a scheduling operation risk assessment model with continuous state evolution equation as its core is constructed. A population optimization strategy based on probability distribution adaptive perturbation is adopted to optimize the hyperparameter combination of the risk assessment model and obtain the optimal hyperparameter combination. Substitute the optimal hyperparameter combination into the risk assessment model to perform continuous dynamic risk assessment of distribution network scheduling operations and output the risk assessment results.

[0008] As a preferred embodiment of the distribution network dispatching operation risk assessment method based on grey relational feature screening and adaptive NODE model described in this invention, the method of screening and forming a joint feature dataset based on the correlation quantification analysis between various monitoring indicators under multiple operation scenarios includes determining the monitoring indicator sets before and after the operation for different dispatching operation scenarios, wherein the monitoring indicator sets include node voltage, line power flow and load shedding. Grey relational analysis was used to calculate the pairwise correlation between each indicator in the monitoring indicator set. Based on the degree of correlation, select the index combinations with strong correlation, and construct the selected index combinations into a joint feature dataset.

[0009] As a preferred embodiment of the distribution network scheduling operation risk assessment method based on grey relational feature screening and adaptive NODE model described in this invention, the construction of the scheduling operation risk assessment model with continuous state evolution equation as the core includes: constructing a shared preprocessing layer to map joint features to a unified hidden vector space. Construct multiple symmetric decision subnets. Each subnet performs multi-level differentiable threshold splitting on the hidden vector and outputs a feature representation vector with the same dimension as the hidden vector. After averaging the outputs of each subnet, the predicted probabilities of each risk category are obtained by mapping the output layer to the fully connected layer. A continuous state evolution equation driven by the distribution network state vector and the scheduling operation input vector is established, and the differential change law of the distribution network state with the scheduling operation is fitted by a neural network.

[0010] As a preferred embodiment of the distribution network scheduling operation risk assessment method based on grey relational feature screening and adaptive NODE model described in this invention, the continuous state evolution equation is fitted with the differential change law of the distribution network state using a three-layer fully connected neural network, and the current state vector of the distribution network, the scheduling operation input vector and the model learnable parameters are jointly mapped to the state change rate. By numerically integrating the continuous state evolution equation, the continuous evolution trajectory of the distribution network state during the entire scheduling operation is obtained, thus realizing the time-series dynamic deduction of the risk state.

[0011] As a preferred embodiment of the distribution network scheduling operation risk assessment method based on grey relational feature screening and adaptive NODE model described in this invention, the population optimization strategy further includes initializing the population, with each individual in the population corresponding to a set of hyperparameter combinations of the distribution network scheduling operation risk assessment model. Based on the fitness assessment results of individuals in various populations, individuals in the population are divided into three categories: discoverers, followers, and scouts. A t-distribution is constructed using the current iteration number as the degree of freedom parameter, and random perturbations of the t-distribution are introduced into the position update process of the three types of individuals respectively; The position update is iteratively executed until convergence. The hyperparameter combination corresponding to the optimal individual is substituted into the risk assessment model to complete the training. The risk assessment of the distribution network scheduling operation is performed, and the risk assessment result is output.

[0012] As a preferred embodiment of the distribution network scheduling operation risk assessment method based on grey relational feature screening and adaptive NODE model described in this invention, the location update strategies for the three types of individuals are as follows: the discoverer updates its own position according to the current global optimal position direction; Followers update their location based on the discoverer's location; When the overall fitness of the population is lower than the preset warning value, the scout jumps to the vicinity of the global optimal position to perform a random perturbation search.

[0013] As a preferred embodiment of the risk assessment method for distribution network dispatching operations based on grey relational feature screening and adaptive NODE model described in this invention, after outputting the risk assessment results, precision, recall and F1 score are selected as evaluation indicators to evaluate the performance of the optimized risk assessment model, and the evaluation results are compared and analyzed with the control model to verify the improvement effect of distribution network dispatching operation risk classification performance.

[0014] Secondly, embodiments of the present invention provide a risk assessment system for distribution network scheduling operations based on grey relational feature screening and adaptive NODE model, which includes a feature screening module for acquiring multi-source monitoring data of distribution network scheduling operations, and screening to form a joint feature dataset based on quantitative analysis of the correlation between various monitoring indicators under multiple operation scenarios. The model building module is used to construct a scheduling operation risk assessment model based on the joint feature dataset and with the continuous state evolution equation as its core. The hyperparameter optimization module is used to optimize the hyperparameter combination of the risk assessment model by adopting a population optimization strategy based on probability distribution adaptive perturbation, so as to obtain the optimal hyperparameter combination. The risk assessment module is used to substitute the optimal hyperparameter combination into the risk assessment model to perform continuous dynamic risk assessment of the distribution network scheduling operation and output the risk assessment results.

[0015] Thirdly, embodiments of the present invention provide a computer device, including a memory and a processor, wherein the memory stores a computer program, wherein: when the computer program instructions are executed by the processor, they implement the steps of the distribution network scheduling operation risk assessment method based on grey relational feature screening and adaptive NODE model as described in the first aspect of the present invention.

[0016] Fourthly, embodiments of the present invention provide a computer-readable storage medium having a computer program stored thereon, wherein: when the computer program instructions are executed by a processor, they implement the steps of the distribution network scheduling operation risk assessment method based on grey relational feature screening and adaptive NODE model as described in the first aspect of the present invention.

[0017] The beneficial effects of this invention are as follows: This invention uses grey relational analysis to perform pairwise correlation quantification calculations on three types of monitoring indicators—node voltage, line power flow, and load shedding—before and after distribution network dispatching operations. Based on the degree of correlation, a joint feature dataset is formed. Indicator combinations with coordinated response characteristics before and after the operation are used as model input, eliminating redundant features from the source and allowing the model to focus on feature information that truly discriminates distribution network dispatching risks. This avoids the feature noise problem caused by the indiscriminate input of all indicators in traditional methods. In terms of model construction, a NODE structure based on continuous state evolution equations is introduced, using differential dynamics to fit the continuous evolution of the distribution network state throughout the entire dispatching operation process. The model transforms the trajectory, breaking through the limitations of traditional static classification models that can only provide discrete risk judgments at a single point in time. This allows the assessment results to reflect the complete dynamic change process of risks such as voltage overruns, line overloads, and load shedding from their inception to accumulation, providing dispatchers with risk trend information with temporal continuity. In terms of hyperparameter optimization, a t-distribution is constructed with the current iteration number as the degree of freedom parameter. Adaptive perturbations are introduced into the position updates of individuals in the three population types: discoverers, followers, and scouts. This allows the optimization process to maintain a large exploration step size in the early stages of iteration and automatically narrow to a local fine search in the later stages. This solves the problem that the fixed hyperparameter configuration of the NODE model is difficult to adapt to the dynamic adjustment scenario of the distribution network topology, and obtains a better hyperparameter combination within the same number of iterations. Attached Figure Description

[0018] To more clearly illustrate the technical solutions of the embodiments of the present invention, the accompanying drawings used in the description of the embodiments will be briefly introduced below. Obviously, the accompanying drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0019] Figure 1 The flowchart shows a method for risk assessment of distribution network scheduling operations based on grey relational feature screening and adaptive NODE model. Figure 2 A computer equipment diagram for a risk assessment method for distribution network scheduling operations based on grey relational feature screening and adaptive NODE model; Figure 3 Another flowchart for the risk assessment method of distribution network scheduling operation based on grey relational feature screening and adaptive NODE model; Figure 4 This is a comparison chart of various models for risk assessment of distribution network scheduling operations based on grey relational feature screening and adaptive NODE model. Detailed Implementation

[0020] To make the above-mentioned objects, features and advantages of the present invention more apparent and understandable, the specific embodiments of the present invention will be described in detail below with reference to the accompanying drawings.

[0021] Many specific details are set forth in the following description in order to provide a full understanding of the invention. However, the invention may also be practiced in other ways different from those described herein, and those skilled in the art can make similar extensions without departing from the spirit of the invention. Therefore, the invention is not limited to the specific embodiments disclosed below.

[0022] Secondly, the term "one embodiment" or "embodiment" as used herein refers to a specific feature, structure, or characteristic that may be included in at least one implementation of the present invention. The phrase "in one embodiment" appearing in different places in this specification does not necessarily refer to the same embodiment, nor is it a single or selective embodiment that is mutually exclusive with other embodiments.

[0023] Example 1 Reference Figure 1 This is the first embodiment of the present invention, which provides a method for risk assessment of distribution network scheduling operations based on grey relational feature screening and an adaptive NODE model, including: S100: Acquire multi-source monitoring data of distribution network scheduling operations, and perform quantitative analysis of the correlation between various monitoring indicators under multiple operation scenarios to form a joint feature dataset.

[0024] S200: Based on the joint feature dataset, a scheduling operation risk assessment model with continuous state evolution equation as its core is constructed.

[0025] S300: A population optimization strategy based on probability distribution adaptive perturbation is adopted to optimize the hyperparameter combination of the risk assessment model and obtain the optimal hyperparameter combination.

[0026] S400: Substitutes the optimal hyperparameter combination into the risk assessment model to perform continuous dynamic risk assessment of distribution network scheduling operations and outputs the risk assessment results.

[0027] It should be noted that distribution network dispatching operations encompass various scenarios, including switching operations, load transfers, and the connection and disconnection of distributed power sources. In each scenario, there are varying degrees of correlation between the three indicators: node voltage, line power flow, and load shedding. If all indicators are directly input into the model without differentiation, redundant features will interfere with the model's extraction of key risk information, reducing assessment accuracy. Furthermore, the distribution network status continuously evolves during dispatching operations. Traditional static classification models can only output discrete risk judgments at a specific moment in the operation, failing to depict the continuous change in risk from low to high, and thus making it difficult to support the dynamic decision-making of dispatchers during operation execution.

[0028] S100 uses grey relational analysis to calculate the pairwise correlation of monitoring indicators under multiple operational scenarios, selects indicator combinations with high correlation to construct a joint feature dataset, and eliminates redundant interference. S200 constructs a NODE risk assessment model based on the joint feature dataset, with the continuous state evolution equation as the core, and uses differential dynamics to characterize the continuous evolution trajectory of the distribution network state throughout the entire scheduling operation process. S300 adopts a population optimization strategy that introduces t-distribution adaptive perturbation to iteratively optimize the combination of model hyperparameters, so that the model can still maintain fast convergence under dynamic changes in the distribution network. S400 substitutes the optimal hyperparameters into the trained model to complete the continuous dynamic risk assessment of the distribution network scheduling operation and output the results.

[0029] Example 2 Reference Figure 1 - Figure 4 This is the second embodiment of the present invention.

[0030] In this embodiment, step S100 involves acquiring multi-source monitoring data of distribution network scheduling operations, and performing quantitative analysis of the correlation between various monitoring indicators under multiple operation scenarios to form a joint feature dataset, including the following A1 steps: A1: Based on quantitative analysis of the correlation between various monitoring indicators under multiple operation scenarios, a joint feature dataset is formed, including the determination of monitoring indicator sets before and after operation for different scheduling operation scenarios. The monitoring indicator sets include node voltage, line power flow and load loss. Grey relational analysis was used to calculate the pairwise correlation between each indicator in the monitoring indicator set. Based on the degree of correlation, select the index combinations with strong correlation, and construct the selected index combinations into a joint feature dataset.

[0031] Specifically, distribution network dispatching operations involve various scenarios such as switching operations, load transfer, fault handling, and the connection and disconnection of distributed power sources. The operating status of the distribution network changes under each scenario. Node voltage, line power flow, and load shedding are the core monitoring quantities characterizing this operating status, and all three exhibit varying degrees of numerical response before and after different dispatching operations. For the aforementioned multiple operating scenarios, a multi-source dataset of distribution network dispatching operations, including switching operations, load transfer, and the connection and disconnection of distributed power sources, is simulated and generated. This multi-source dataset includes node voltages before and after each operation under each scenario. X 1. X5), Line flow ( X 2, X6) and loss of load ( X 3) Collect and record information to fully cover the status change information throughout the entire scheduling operation process.

[0032] The collected multi-source datasets were cleaned. Missing values ​​and outliers were filled with the mean of data before and after the corresponding record location to ensure the dataset's integrity and continuity. After cleaning, the datasets were normalized to map the numerical range of each feature to the [0,1] interval, eliminating feature weight bias caused by different units of measurement. The normalization formula is as follows: in, These are the original actual values. and These are the maximum and minimum values ​​of the sample data, respectively. These are the normalized values. After normalization, the dataset is divided into a training set and a test set according to a fixed ratio. The training set is used for subsequent model construction and parameter learning, while the test set is used for final model performance validation.

[0033] After partitioning the dataset, Grey-Relational Analysis (GRA) was used to quantify the correlation between the monitoring indicators. Grey-Relational Analysis compares the geometric similarity of the indicator sequences, using the correlation degree to measure the strength of the influence relationship between the indicators. A higher correlation degree indicates a stronger linkage between the two indicators and a higher consistency in their joint response during scheduling operations. Specifically, the calculation was performed using node voltage (… X 1) Line flow ( X 2) and loss of load ( X 3) Using the reference sequence, the correlation between each indicator and the remaining indicators was calculated one by one, yielding the following correlation results: X 1 and X The correlation between 2 and 0 is 0.6453. X 1 and X The correlation between 3 and 0 is 0.5123; X 2 and X The correlation between 1 and 0 is 0.7502. X 2 and X The correlation between X3 and X4 is 0.6198; X The correlation between 1 and 0 is 0.5934. X 3 and X The correlation between the two is 0.5912. Common monitoring indicators for risk assessment of distribution network dispatching operations are selected as shown in Table 1. Three detection indicators, namely node voltage, line power flow, and load shedding, are calculated for risk assessment in different dispatching operations.

[0034] Table 1 Risk Assessment Indicators for Distribution Network Dispatch Operations

[0035] Grey relational analysis was used to calculate the correlation degree between various evaluation indicators in the dataset, and the results are shown in Table 2.

[0036] Table 2 Correlation of Risk Assessment Indicators for Distribution Network Dispatch Operations

[0037] The above results indicate that this scheme sets the correlation threshold to 0.5, and indicators with a correlation greater than 0.5 are considered to have a strong correlation. As shown in Table 2, X 1 and X The correlation coefficient between 2 and 0.6453 is 0. X 1 and X The correlation coefficient for 3 is 0.5123. X 2 and X The correlation coefficient for 1 is 0.7502. X 2 and X The correlation coefficient for 3 is 0.6198. X 3 and X The correlation coefficient of 1 is 0.5934. X 3 and X The correlation coefficient of 2 is 0.5912, and the correlation coefficients of all indicator pairs are higher than 0.5, indicating a high degree of correlation between node voltage and line power flow, and between line power flow and load shedding. The three indicators show a certain degree of synergistic change characteristics before and after the operation. Therefore, the joint combination of node voltage, line power flow, and load shedding is used as the model input feature to construct a joint feature dataset, forming training and testing sets for subsequent training and testing of the NODE risk assessment model.

[0038] In this embodiment, step S200 involves constructing a scheduling operation risk assessment model based on the joint feature dataset and a continuous state evolution equation, including the following steps B1-B2: B1: Construct a scheduling operation risk assessment model based on continuous state evolution equations, including building a shared preprocessing layer to map joint features to a unified hidden vector space; Construct multiple symmetric decision subnets. Each subnet performs multi-level differentiable threshold splitting on the hidden vector and outputs a feature representation vector with the same dimension as the hidden vector. After averaging the outputs of each subnet, the predicted probabilities of each risk category are obtained by mapping the output layer to the fully connected layer. A continuous state evolution equation driven by the distribution network state vector and the scheduling operation input vector is established, and the differential change law of the distribution network state with the scheduling operation is fitted by a neural network.

[0039] B2: The continuous state evolution equation fits the differential change law of the distribution network state with a three-layer fully connected neural network, and jointly maps the current state vector of the distribution network, the scheduling operation input vector and the model learnable parameters into the state change rate; By numerically integrating the continuous state evolution equation, the continuous evolution trajectory of the distribution network state during the entire scheduling operation is obtained, enabling the time-series dynamic deduction of risk states.

[0040] Specifically, the Neural-Ordinary-Differential-Equation (NODE) ​​model is a time-series modeling method that uses continuous differential equations to characterize the state evolution process of a system. Its core idea is to extend the discrete residual network layers to the continuous time domain, replacing layer-by-layer forward propagation with the integral of the differential equations, thereby achieving accurate fitting of the continuous state trajectory of a nonlinear dynamic system. This step, based on the joint feature dataset, constructs a complete NODE scheduling operation risk assessment model according to the following structure.

[0041] First, determine the model input dimension D and the number of output categories C. D is the dimension length of the joint indicator feature vector, and C corresponds to the number of categories in the distribution network dispatching operation risk assessment, including three assessment levels: high risk, medium risk, and low risk.

[0042] A shared preprocessing layer is constructed to map the original input feature vectors to a fixed-size hidden vector H, achieving a unified representation of the features and providing a consistent input basis for subsequent decision subnets. The expression for this process is as follows: Where x represents the input feature vector, This is the weight matrix. For the bias vector of the shared mapping layer, The activation function is H, which is the mapped hidden vector.

[0043] Construct a symmetric decision subnet layer. The symmetric decision subnet (NODE tree structure subnet) consists of multiple independent subnets. Each subnet performs a series of differentiable threshold splitting operations based on the hidden vector H, simulating the tree-like decision process, and outputting a feature representation vector with the same dimension as the hidden vector. For the l-th layer soft splitting operation of the k-th subnet, the formula for calculating the intermediate result vector is as follows: in, Denotes the linear mapping matrix of the l-th layer in the k-th subnet. For the corresponding bias vector, This is the intermediate result vector of the l-th layer, which serves as the input for the next Sigmoid transform. Applying the Sigmoid function for a nonlinear transformation maps each component to the (0,1) interval: After L layers of soft splitting, the final output vector of the k-th subnet is: in Let be the feature representation vector output by the k-th subnet after completing soft splitting at all levels, where K represents the total number of subnets. Each subnet, based on the same hidden vector input, independently makes splitting decisions with different combinations of learnable parameters, collectively covering multiple potential risk classification paths.

[0044] Construct a fully connected aggregation layer, average the output feature vectors of all K subnets, and summarize them into an overall feature representation: in, The average value of the output vectors of all subnets is used as the final overall feature representation and passed to the output layer.

[0045] Build the output layer, After linear mapping and transformation using the Softmax function, the predicted probabilities of each risk category are obtained: in, This is the output layer weight matrix. This is the output layer bias vector. Let c be the predicted probability of the model for risk class c, where class c takes the values ​​of three assessment levels: high risk, medium risk, and low risk.

[0046] Based on the above classification structure, a continuous state evolution equation is established, jointly driven by the distribution network state vector and the scheduling operation input vector, to characterize the continuous dynamic evolution of the distribution network state throughout the entire scheduling operation process: Here, x(t) is the distribution network state vector at time t, containing current state information of node voltage, line power flow, and load shedding; u(t) is the dispatch operation input vector at time t, representing the type and parameters of the dispatch operation currently being executed; f is a three-layer fully connected neural network responsible for jointly mapping the current state vector, operation input vector, and model learnable parameters θ to a state change rate, where θ includes the weights and bias parameters of the fully connected layers. By numerically integrating the above continuous state evolution equation over the time interval, the complete continuous evolution trajectory of the distribution network state x(t) as the dispatch operation progresses can be obtained, thereby realizing the time-series dynamic deduction and evaluation of risk states such as voltage overruns, line overloads, and load shedding throughout the entire operation process. Compared with traditional static classification models that can only provide risk judgment at a single point in time, this continuous evolution equation enables the model to track the complete dynamic process of risk from its inception to its accumulation, providing dispatchers with more timely information on risk change trends.

[0047] In this embodiment, step S300 employs a population optimization strategy based on probability distribution adaptive perturbation to optimize the hyperparameter combination of the risk assessment model, obtaining the optimal hyperparameter combination, including the following steps C1-C2: C1: Population optimization strategies also include initializing the population, with each individual in the population corresponding to a set of hyperparameter combinations of a distribution network scheduling operation risk assessment model; Based on the fitness assessment results of individuals in various populations, individuals in the population are divided into three categories: discoverers, followers, and scouts. A t-distribution is constructed with the current iteration number as the degree of freedom parameter, and random perturbations of the t-distribution are introduced into the position update process of the three types of individuals respectively; The position update is iteratively executed until convergence. The hyperparameter combination corresponding to the optimal individual is substituted into the risk assessment model to complete the training. The risk assessment of the distribution network scheduling operation is performed, and the risk assessment result is output.

[0048] Specifically, the performance of the NODE model is highly sensitive to the settings of hyperparameter combinations. The learning rate determines the step size for updating model parameters, the number of network layers affects the depth of the model's representation of complex risk features, the hidden layer dimension determines the representation capacity of features, the tree depth controls the number of split layers in the symmetric decision subnet, and the weight decay coefficient is used to suppress model overfitting. The reference ranges for each hyperparameter are as follows: learning rate ranges from 0.0001 to 0.01, num_layers ranges from 2 to 8, hidden_dim ranges from 32 to 256, tree_depth ranges from 2 to 5, and weight_decay ranges from 0 to 0.001. If the above hyperparameters are configured using fixed values ​​or simple grid search, it is difficult to adapt to the optimization needs under dynamic changes such as sudden changes in distribution network status or topology adjustments. This step adopts the Improved Sparrow Search Algorithm (ISSA) with adaptive t-distribution perturbation to intelligently optimize the above hyperparameter combinations. The optimization population is initialized with a population size of 20, a maximum number of iterations of 300, a warning threshold of 0.6, a discoverer ratio of 70%, and a scout ratio of 20%. The hyperparameter range of the NODE model is shown in Table 3. Table 3. Range of hyperparameters for the NODE model

[0049] Each individual in the population represents a complete set of hyperparameter combinations, namely a specific value of {learning_rate, num_layers, hidden_dim, tree_depth, weight_decay}, which is randomly initialized within the reference range of the above hyperparameters, thus forming an initial diverse population covering the hyperparameter space.

[0050] In each iteration, the evaluation accuracy of the NODE model trained with the current hyperparameter combination is used as the individual's fitness value. Higher fitness indicates better configuration quality of the current hyperparameter combination. Based on the fitness ranking of each individual, the top 70% of individuals are designated as discoverers, responsible for leading the population to search for high-quality hyperparameter regions. Among the remaining individuals, followers are designated proportionally to explore in the direction of the discoverers. At the same time, scouts are designated from individuals with lower fitness. When the overall fitness of the population falls below a preset warning threshold of 0.6, the scouts are guided to perform random perturbation searches near the current global optimum to avoid the population getting trapped in local optima.

[0051] C2: The position update strategies for the three types of individuals are as follows: the discoverer updates its own position based on the current globally optimal position direction; Followers update their location based on the discoverer's location; When the overall fitness of the population is lower than the preset warning value, the scout jumps to the vicinity of the global optimal position to perform a random perturbation search.

[0052] Specifically, during the position update process for the three types of individuals, a t-distributed random perturbation with the current iteration number as the degree of freedom parameter is introduced to adaptively adjust the position of each individual. The position update formula is as follows: in, This represents the position of the updated i-th individual (i.e., the i-th combination of hyperparameters). This represents the position of the current i-th individual. Let `iter` be a t-distributed random variable with the current iteration number as its degree of freedom. In the early stages of iteration, `iter` is small, the t-distribution has low degrees of freedom and a thick tail, corresponding to a larger random perturbation step size. This helps the population conduct a large-scale global exploration in the hyperparameter space, reducing the risk of premature convergence. As the iteration number increases, the t-distribution gradually approaches a standard normal distribution, and the perturbation step size shrinks accordingly, guiding the population to conduct a fine-grained local search near the current optimal region, thus achieving an adaptive balance between exploration and utilization. After convergence after 300 iterations, the hyperparameter combination corresponding to the individual with the best fitness is output and used as the final configuration to complete the NODE model training. Compared with Random-Search (RS) and Bayesian-Optimization (BO) strategies, ISSA, through adaptive t-distribution perturbation, allows the population to balance global diversity and local accuracy during the search process, achieving a better hyperparameter combination configuration with the same number of optimization iterations.

[0053] In this embodiment, step S400 involves substituting the optimal hyperparameter combination into the risk assessment model to perform continuous dynamic risk assessment of the distribution network scheduling operation and outputting the risk assessment results, including the following step D1: D1: After outputting the risk assessment results, select precision, recall and F1 score as evaluation indicators to evaluate the performance of the optimized risk assessment model, and compare the evaluation results with the control model to verify the improvement effect of risk classification performance in distribution network dispatching operations.

[0054] Specifically, based on the optimal hyperparameter combination obtained from S300 {learning_rate=0.0005, num_layers=4, hidden_dim=64, tree_depth=3, weight_decay=0.001}, the final configuration and training of the NODE scheduling operation risk assessment model were completed, resulting in the ISSA-NODE model. The average optimal value was calculated after 100 runs. To verify the optimization effect of the ISSA hyperparameter optimization strategy compared to RS and BO, each of the three methods was run independently 100 times, and the optimal hyperparameter combination corresponding to each convergence was recorded. The average optimal value of the 100 runs was compared, and the results are shown in Table 4.

[0055] Table 4 Optimal values ​​of hyperparameters for different models

[0056] As shown in Table 4, the ISSA-NODE model yields a learning rate of 0.0005, num_layers of 4, hidden_dim of 64, tree_depth of 3, and weight_decay of 0.001. Compared with the optimal hyperparameter configurations of the RS-NODE and BO-NODE models, the ISSA-NODE combination is more balanced in its coordination of learning rate and weight decay, demonstrating the optimization advantage of adaptive t-distribution perturbation guiding the population to converge toward a better hyperparameter region.

[0057] Using the test set divided in S100 as validation data, the trained ISSA-NODE model is tested for inference. The joint feature vector of each scheduling operation sample in the test set is input into the model. Through the shared preprocessing layer, symmetric decision subnet layer, fully connected convergence layer and Softmax output layer, the predicted probabilities of high risk, medium risk and low risk for each sample are obtained. The category with the highest probability is taken as the final risk assessment result.

[0058] To quantify the model's performance, precision (P), recall (R), and F1 score were selected as comprehensive evaluation metrics. Precision measures the proportion of samples predicted as belonging to a specific risk category by the model, directly reflecting the consistency between the model's evaluation results and the true labels. The calculation formula is as follows: Recall measures the proportion of samples that actually belong to a certain risk category that the model correctly identifies, reflecting the model's ability to cover real risk events. The calculation formula is as follows: The F1 score is the harmonic mean of precision and recall, which comprehensively reflects the overall performance of the model on the risk classification task. The calculation formula is as follows: In the above formula, TP is the number of classes correctly evaluated as positive by the model, TN is the number of classes correctly evaluated as negative by the model, FP is the number of classes incorrectly evaluated as positive by the model, and FN is the number of classes incorrectly evaluated as negative by the model.

[0059] In performance comparison and verification, you can refer to Figure 4 The comparison results are shown in the diagram. The random search optimization NODE model (RS-NODE) ​​and the Bayesian optimization NODE model (BO-NODE) ​​were used as control models and evaluated on the same test set to obtain the accuracy of each model. The test set was used to test the optimized model, and the final scheduling operation risk assessment results are shown in Table 5.

[0060] Table 5 Comparison of Evaluation Results

[0061] The above results show that the ISSA-NODE model outperforms the RS-NODE and BO-NODE control models in all three evaluation metrics. The comparison results are as follows: Figure 4 As shown. By Figure 4 It can be intuitively seen that the ISSA-NODE model outperforms the RS-NODE and BO-NODE control models in terms of accuracy, F1 score, and recall rate R. This verifies that the introduction of adaptive t-distribution perturbation enables the hyperparameter optimization process to find a better parameter configuration within the same number of iterations, thereby improving the overall evaluation accuracy of the NODE model in the risk classification task of distribution network scheduling operations. This also verifies the rationality of the ISSA-based hyperparameter adaptive optimization strategy in the technical solution of this invention.

[0062] Example 3 Reference Figure 2 This is the third embodiment of the present invention.

[0063] The above is an illustrative scheme of a distribution network dispatching operation risk assessment method based on grey relational feature screening and an adaptive NODE model. It should be noted that the technical solution of this distribution network dispatching operation risk assessment system based on grey relational feature screening and an adaptive NODE model belongs to the same concept as the aforementioned distribution network dispatching operation risk assessment method based on grey relational feature screening and an adaptive NODE model. Details not described in detail in this embodiment can be found in the description of the aforementioned distribution network dispatching operation risk assessment method based on grey relational feature screening and an adaptive NODE model.

[0064] This embodiment also provides a distribution network dispatching operation risk assessment system based on grey relational feature screening and an adaptive NODE model, including: a feature screening module, used to acquire multi-source monitoring data of distribution network dispatching operations, and to screen and form a joint feature dataset based on quantitative analysis of the correlation between various monitoring indicators under multiple operation scenarios; a model building module, used to construct a dispatching operation risk assessment model with a continuous state evolution equation as its core based on the joint feature dataset; a hyperparameter optimization module, used to optimize the hyperparameter combination of the risk assessment model using a population optimization strategy based on probability distribution adaptive perturbation to obtain the optimal hyperparameter combination; and a risk assessment module, used to substitute the optimal hyperparameter combination into the risk assessment model to perform continuous dynamic risk assessment of distribution network dispatching operations and output the risk assessment results.

[0065] This embodiment also provides an electronic device applicable to risk assessment of distribution network scheduling operations based on grey relational feature screening and adaptive NODE model, comprising: a memory and a processor; the memory is used to store computer-executable instructions, and the processor is used to execute the computer-executable instructions to implement the distribution network scheduling operation risk assessment method based on grey relational feature screening and adaptive NODE model proposed in the above embodiment.

[0066] This embodiment also provides a storage medium storing a computer program that, when executed by a processor, implements the risk assessment method for distribution network scheduling operations based on grey relational feature screening and adaptive NODE model as proposed in the above embodiments.

[0067] The storage medium proposed in this embodiment and the method for risk assessment of distribution network scheduling operations based on grey relational feature screening and adaptive NODE model proposed in the above embodiments belong to the same inventive concept. Technical details not described in detail in this embodiment can be found in the above embodiments, and this embodiment has the same beneficial effects as the above embodiments.

[0068] Based on the above description of the implementation methods, those skilled in the art can clearly understand that the present invention can be implemented using software and necessary general-purpose hardware, and of course, it can also be implemented using hardware. Based on this understanding, the technical solution of the present invention, or the part that contributes to the prior art, can be embodied in the form of a software product. This computer software product can be stored in a computer-readable storage medium, such as a computer floppy disk, read-only memory (ROM), random access memory (RAM), flash memory, hard disk, or optical disk, etc., including several instructions to cause a computer device (which may be a personal computer, server, or network device, etc.) to execute the methods of the various embodiments of the present invention.

[0069] It should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention and not to limit it. Although the present invention has been described in detail with reference to preferred embodiments, those skilled in the art should understand that modifications or equivalent substitutions can be made to the technical solutions of the present invention without departing from the spirit and scope of the technical solutions of the present invention, and all such modifications or substitutions should be covered within the scope of the claims of the present invention.

Claims

1. A method for risk assessment of distribution network scheduling operations based on grey relational feature screening and adaptive NODE model, characterized in that: This includes acquiring multi-source monitoring data of distribution network scheduling operations, performing quantitative analysis of the correlation between various monitoring indicators under multiple operation scenarios, and screening to form a joint feature dataset; Based on the joint feature dataset, a scheduling operation risk assessment model with continuous state evolution equation as its core is constructed. A population optimization strategy based on probability distribution adaptive perturbation is adopted to optimize the hyperparameter combination of the risk assessment model and obtain the optimal hyperparameter combination. Substitute the optimal hyperparameter combination into the risk assessment model to perform continuous dynamic risk assessment of distribution network scheduling operations and output the risk assessment results.

2. The method for risk assessment of distribution network scheduling operations based on grey relational feature screening and adaptive NODE model as described in claim 1, characterized in that: The quantitative analysis of the correlation between various monitoring indicators under multiple operation scenarios, and the screening to form a joint feature dataset, includes determining the monitoring indicator sets before and after the operation for different scheduling operation scenarios, wherein the monitoring indicator sets include node voltage, line power flow and load loss. Grey relational analysis was used to calculate the pairwise correlation between each indicator in the monitoring indicator set. Based on the degree of correlation, select the index combinations with strong correlation, and construct the selected index combinations into a joint feature dataset.

3. The method for risk assessment of distribution network scheduling operations based on grey relational feature screening and adaptive NODE model as described in claim 2, characterized in that: The construction of the scheduling operation risk assessment model based on the continuous state evolution equation includes: constructing a shared preprocessing layer to map joint features to a unified hidden vector space; Construct multiple symmetric decision subnets. Each subnet performs multi-level differentiable threshold splitting on the hidden vector and outputs a feature representation vector with the same dimension as the hidden vector. After averaging the outputs of each subnet, the predicted probabilities of each risk category are obtained by mapping the output layer to the fully connected layer. A continuous state evolution equation driven by the distribution network state vector and the scheduling operation input vector is established, and the differential change law of the distribution network state with the scheduling operation is fitted by a neural network.

4. The method for risk assessment of distribution network scheduling operations based on grey relational feature screening and adaptive NODE model as described in claim 3, characterized in that: The continuous state evolution equation uses a three-layer fully connected neural network to fit the differential change law of the distribution network state, and jointly maps the current state vector of the distribution network, the scheduling operation input vector and the model learnable parameters into the state change rate. By numerically integrating the continuous state evolution equation, the continuous evolution trajectory of the distribution network state during the entire scheduling operation is obtained, thus realizing the time-series dynamic deduction of the risk state.

5. The method for risk assessment of distribution network scheduling operations based on grey relational feature screening and adaptive NODE model as described in claim 4, characterized in that: The population optimization strategy also includes initializing the population, with each individual in the population corresponding to a set of hyperparameter combinations of a distribution network scheduling operation risk assessment model; Based on the fitness assessment results of individuals in various populations, individuals in the population are divided into three categories: discoverers, followers, and scouts. A t-distribution is constructed using the current iteration number as the degree of freedom parameter, and random perturbations of the t-distribution are introduced into the position update process of the three types of individuals respectively; The position update is iteratively executed until convergence. The hyperparameter combination corresponding to the optimal individual is substituted into the risk assessment model to complete the training. The risk assessment of the distribution network scheduling operation is performed, and the risk assessment result is output.

6. The method for risk assessment of distribution network scheduling operations based on grey relational feature screening and adaptive NODE model as described in claim 5, characterized in that: The position update strategies for the three types of individuals are as follows: the discoverer updates its own position based on the current globally optimal position direction; Followers update their location based on the discoverer's location; When the overall fitness of the population is lower than the preset warning value, the scout jumps to the vicinity of the global optimal position to perform a random perturbation search.

7. The method for risk assessment of distribution network scheduling operations based on grey relational feature screening and adaptive NODE model as described in claim 6, characterized in that: After outputting the risk assessment results, precision, recall, and F1 score are selected as evaluation indicators to evaluate the performance of the optimized risk assessment model. The evaluation results are then compared with the control model to verify the improvement effect of risk classification performance in distribution network dispatching operations.

8. A distribution network dispatching operation risk assessment system based on grey relational feature screening and adaptive NODE model, based on the distribution network dispatching operation risk assessment method based on grey relational feature screening and adaptive NODE model as described in any one of claims 1 to 7, characterized in that: It also includes, The feature filtering module is used to acquire multi-source monitoring data of distribution network scheduling operations, and to filter and form a joint feature dataset based on quantitative analysis of the correlation between various monitoring indicators under multiple operation scenarios. The model building module is used to construct a scheduling operation risk assessment model based on the joint feature dataset and with the continuous state evolution equation as its core. The hyperparameter optimization module is used to optimize the hyperparameter combination of the risk assessment model by adopting a population optimization strategy based on probability distribution adaptive perturbation, so as to obtain the optimal hyperparameter combination. The risk assessment module is used to substitute the optimal hyperparameter combination into the risk assessment model to perform continuous dynamic risk assessment of the distribution network scheduling operation and output the risk assessment results.

9. A computer device comprising a memory and a processor, wherein the memory stores a computer program, characterized in that: When the processor executes the computer program, it implements the steps of the distribution network scheduling operation risk assessment method based on grey relational feature screening and adaptive NODE model as described in any one of claims 1 to 7.

10. A computer-readable storage medium having a computer program stored thereon, characterized in that: When the computer program is executed by the processor, it implements the steps of the distribution network scheduling operation risk assessment method based on grey relational feature screening and adaptive NODE model as described in any one of claims 1 to 7.

Images