A new energy power transmission network risk assessment method based on channel attention multi-scale convolution

Abstract

The application discloses a new energy power transmission network risk assessment method based on channel attention multi-scale convolution, and belongs to the field of new energy power system risk assessment and artificial intelligence power grid safety analysis. According to the new energy penetration rate and the intermittent intensity, the method constructs an operation condition, adopts the median Latin hypercube sampling and the synchronous back substitution method to generate a typical operation scene of combined disturbance of wind power, photovoltaic and load, constructs a multi-dimensional grid side risk index, combines the information entropy prior weight and the trainable gate vector to form a weighted feature, constructs a local subgraph and a sorting adjacency matrix according to the neighbor node degree sorting, utilizes a multi-scale channel attention graph convolution network to output a node continuous risk score, and determines a risk grade through a piecewise linear membership function. The application is favorable to improving the adaptability and the explainability of risk feature extraction under a complex new energy operation scene.

Description

Technical Field

[0001] This invention belongs to the field of risk assessment of new energy power systems and artificial intelligence power grid security analysis technology, and in particular relates to a risk assessment method for new energy transmission networks based on channel attention multi-scale convolution. Background Technology

[0002] As the proportion of new energy sources such as wind power and solar power integrated into transmission networks continues to increase, the randomness and volatility of new energy output, combined with load changes, cause variations in line power flow distribution, node voltage levels, and local topology risks depending on the operating scenario. Existing risk assessments for new energy transmission networks typically employ deterministic limit-crossing criteria, probabilistic risk analysis, or data-driven methods based on neural networks. These methods construct wind power, solar power, and load disturbance scenarios through random sampling, and then evaluate the operational risks of nodes or lines based on line load rates, voltage deviations, or network structure indicators.

[0003] However, existing methods still have the following shortcomings: scenarios generated by random sampling may have high repetition or insufficient coverage of random space, increasing the burden of power flow calculation and model training; using only line over-limit, voltage over-limit, or single structural indicators makes it difficult to simultaneously characterize the power flow transfer, node voltage changes, topological vulnerability, and local risk propagation relationships caused by new energy disturbances; when using fixed weights to fuse multi-dimensional risk indicators, it is difficult to balance objective statistical differences with the actual contribution of different indicators to risk identification results during model training; conventional graph convolution usually directly aggregates node features according to a preset neighborhood, without fully considering the differences in the network connectivity of neighboring nodes, and it is also difficult to stably introduce the node ranking relationship into subsequent feature processing; in addition, single-scale graph feature extraction is difficult to simultaneously reflect local anomalies near the central node and risk coupling within a larger neighborhood. Therefore, it is necessary to develop a risk assessment scheme for new energy transmission networks that can combine typical scenarios, multi-dimensional risk indicators, local topology ranking, and multi-scale feature extraction. Summary of the Invention

[0004] The purpose of this invention is to provide a risk assessment method for new energy transmission networks based on channel attention multi-scale convolution, so as to improve the shortcomings in the expression of operation scenarios, adaptive weighting of risk indicators, organization of local topological features, and extraction of multi-scale risk features under the condition of high proportion of new energy access.

[0005] To achieve the above objectives, this invention provides a risk assessment method for new energy transmission networks based on channel attention multi-scale convolution, comprising: acquiring topology data, branch operation data, node operation data, wind power output data, photovoltaic power output data, and load data of the new energy transmission network; constructing operating conditions based on the new energy penetration rate and the intermittent intensity of wind power and photovoltaic power output; generating typical operating scenarios corresponding to each operating condition; performing power flow calculations under each typical operating scenario to obtain basic risk indicators corresponding to power flow carrying capacity, voltage safety, and structural vulnerability; and positively normalizing and standardizing the basic risk indicators. The basic objective prior weights are obtained from the information entropy, and the normalized basic risk indicators are weighted and summarized according to the basic objective prior weights to obtain the initial risk quantity of the node. The neighborhood risk clustering degree, node risk volatility coefficient, and neighborhood risk difference degree are calculated based on the initial risk quantity of the node, and the obtained local coupling and uncertainty propagation risk indicators are combined with the basic risk indicators to form a node-risk indicator feature matrix. Objective prior weights are obtained from the information entropy of each risk indicator in the node-risk indicator feature matrix. A trainable risk indicator gating vector with an initial value of 1 and a dimension equal to the number of risk indicators is set, and the objective prior weights are then combined with the trainable risk indicator gating vector. The risk indicator gating vector is trained by multiplying element-wise to obtain the entropy weight gating weight. For each central node, its direct neighbors are sorted in descending order of degree value. At most L of the top-ranked direct neighbors are selected to form a local subgraph with the central node. A local node sequence is formed by prioritizing the central node and arranging the selected direct neighbors in descending order of degree value. A normalized sorted adjacency matrix is ​​constructed based on the topological connectivity of the local subgraph. The local node-risk indicator feature matrix is ​​extracted from the node-risk indicator feature matrix according to the local node sequence and weighted using the entropy weight gating weight. The weighted local node-risk indicator feature matrix is ​​then processed. The node-risk indicator feature matrix and the normalized sorted adjacency matrix are input into a multi-scale channel attention graph convolutional network. Graph convolution propagation is performed on the node features of the local subgraph, and parallel feature extraction at different scales is performed along the node dimension corresponding to the local node sequence. The features at different scales are concatenated along the feature channel dimension, and attention recalibration is performed on the concatenated feature channels. The features corresponding to the central node are extracted from the recalibrated features, and the continuous risk score of the central node is output. Among them, the trainable risk indicator gating vector is updated synchronously with the model loss and network parameters during the training of the multi-scale channel attention graph convolutional network.

[0006] Furthermore, the penetration rate of new energy sources is determined based on the ratio of the sum of the rated installed capacity of wind power and the rated installed capacity of photovoltaics to the peak load of the system; the output coefficient of wind power and the output coefficient of photovoltaics have a first fluctuation range and a second fluctuation range, respectively. The range width of the wind power output coefficient and the range width of the photovoltaic output coefficient in the second fluctuation range are respectively greater than the corresponding range width in the first fluctuation range; the first fluctuation range corresponds to the weak intermittent operation condition, and the second fluctuation range corresponds to the strong intermittent operation condition.

[0007] Furthermore, generating typical operating scenarios includes: dividing the cumulative distribution intervals of wind power output, photovoltaic power output, and load levels into equal-probability strata, using the median cumulative probability of each stratified interval as the sampling probability, and obtaining sample values ​​through the inverse distribution function of the corresponding random variables; randomly permuting and combining the wind power output sample sequence, photovoltaic power output sample sequence, and load level sample sequence to form an initial joint disturbance scenario; and using the synchronous back substitution method to reduce the initial joint disturbance scenario as a whole to obtain typical operating scenarios and their corresponding scenario probabilities.

[0008] Furthermore, the basic risk indicators include line load factor, power over-limit risk index, power flow transfer impact, voltage over-limit risk index, improved voltage stability index, node degree centrality, and electrical betweenness; the local coupling and uncertainty propagation risk indicators include neighborhood risk clustering degree, node risk fluctuation coefficient, and neighborhood risk difference degree; wherein, the mean of the branch-type risk indicators corresponding to each branch adjacent to the target node is mapped to the target node, and the node-type risk indicators are used as the risk indicator characteristics of the corresponding nodes.

[0009] Furthermore, the power flow transfer impact is determined based on the change in active power flow of the node-associated lines relative to the baseline operating state under typical operating scenarios and the maximum allowable transmission power of the corresponding lines; the improved voltage stability index is determined based on the actual node voltage, the node rated voltage, and the node allowable voltage limit; the initial risk of a node is obtained by weighting and summing the normalized basic risk indices according to the basic objective prior weights, and does not include the output results of the multi-scale channel attention map convolutional network; the neighborhood risk clustering degree is determined based on the initial risk of the selected neighboring nodes sorted by node degree value; the node risk fluctuation coefficient is determined based on the initial risk of the node under each typical operating scenario, the scenario probability of each typical operating scenario, and the probability weighted average of the initial risk of the node; the neighborhood risk difference degree is determined based on the difference in the initial risk between the central node and the selected neighboring nodes.

[0010] Furthermore, for each typical operating scenario, the information entropy of each risk indicator is calculated based on the difference in normalized values ​​of each risk indicator between different nodes under that typical operating scenario, and the objective prior weight is determined based on the information entropy of each risk indicator; the objective prior weight is multiplied one by one with the element corresponding to the position in the trainable risk indicator gating vector to form the weighting coefficient corresponding to each risk indicator; the weighting coefficient corresponding to each risk indicator is used to weight the corresponding risk indicator column in the local node-risk indicator feature matrix.

[0011] Furthermore, the degree value of a direct neighbor node is the number of branches directly connected to that direct neighbor node. When the number of direct neighbor nodes is greater than L, the top L direct neighbor nodes are selected; when the number of direct neighbor nodes is not greater than L, all direct neighbor nodes are selected. The center node is placed at the beginning of the local node sequence, and the selected direct neighbor nodes are arranged in descending order of their degree values ​​after the center node. A sorted adjacency matrix is ​​constructed based on the actual topological connections between nodes in the local node sequence. The degree sorting is also used to determine the node set of the local subgraph and the input order for multi-scale feature extraction.

[0012] Furthermore, the multi-scale channel attention graph convolutional network includes a graph convolutional propagation layer, a multi-scale convolutional layer, a channel attention layer, and an output layer. The graph convolutional propagation layer propagates and aggregates node features in the local subgraph according to the normalized sorted adjacency matrix. The multi-scale convolutional layer includes at least two parallel convolutional branches with different kernel scales. Each parallel convolutional branch performs feature extraction along the node dimension arranged according to the local node sequence, and the outputs of each parallel convolutional branch are concatenated along the feature channel dimension. The channel attention layer performs global average pooling and fully connected gating on the concatenated features to generate multi-scale feature channel attention weights, and uses the multi-scale feature channel attention weights to recalibrate the concatenated features channel by channel.

[0013] Furthermore, the trainable risk index gating vector is a global parameter jointly trained using training samples corresponding to different typical operating scenarios and different center nodes; the reference risk value of the training sample is obtained by weighting and summing the normalized basic risk index of the corresponding center node according to the corresponding objective prior weight; the model loss is the mean square error between the continuous risk score output by the multi-scale channel attention map convolutional network and the reference risk value; at the beginning of model training, all elements of the trainable risk index gating vector are 1; during model training, the trainable risk index gating vector is updated synchronously according to the model loss and the network parameters of the multi-scale channel attention map convolutional network, and is fixed for use after the model training is completed; the output layer extracts the features corresponding to the center node from the features after channel attention recalibration, and maps the features corresponding to the center node to obtain the continuous risk score of the center node under the corresponding typical operating scenario.

[0014] Furthermore, a piecewise linear membership function is used to determine the membership degree of the continuous risk score of the central node to low-risk, medium-risk, and high-risk levels, respectively. A linear transition interval is set between adjacent risk levels, and based on the principle of maximum membership, the risk level with the highest membership degree is determined as the risk level of the central node in the corresponding typical operating scenario.

[0015] Compared with the prior art, the present invention has the following beneficial effects: (1) The present invention uses median Latin hypercube sampling to generate a joint disturbance scenario of wind power, photovoltaic and load, and reduces the initial scenario as a whole by synchronous back substitution method. While maintaining the hierarchical coverage relationship of different random variables, it reduces the number of repeated scenarios, thereby reducing the number of scenarios to be processed for subsequent power flow calculation and model training.

[0016] (2) This invention constructs a node-risk index feature matrix including power flow carrying risk, voltage safety risk, structural vulnerability risk and local coupling and uncertainty propagation risk, and characterizes power flow changes, voltage status and local risk relationships under new energy disturbances through power flow transfer impact degree, improved voltage stability index, neighborhood risk clustering degree, node risk fluctuation coefficient and neighborhood risk difference degree, so that the risk input features can cover multiple operating characteristics of the power transmission network.

[0017] (3) This invention combines the objective prior weights obtained by the information entropy method with the trainable risk index gating vector element by element, so that the model can modify the degree of effect of different risk indicators according to the training objective while retaining the objective difference information of each risk indicator, thereby reducing the problem of relying entirely on fixed weights.

[0018] (4) The present invention sorts and filters according to the node degree value of the direct neighbor nodes, arranges the central node at the first position of the local node sequence, and constructs a sorted adjacency matrix according to the actual topological connection relationship, so that the neighboring nodes with high connection degree are stably expressed in the local subgraph construction and subsequent feature extraction process, which is conducive to preserving the local topological information of key nodes and key transmission channels.

[0019] (5) After graph convolution propagation, the present invention performs parallel feature extraction at different scales along the sorted local node dimensions, and splices the features at each scale along the feature channel dimensions and then performs channel attention recalibration. This can simultaneously express the local abnormal features near the central node and the risk coupling features in a larger neighborhood, and is beneficial to suppress feature channels that contribute less to risk judgment.

[0020] (6) This invention extracts the features corresponding to the central node from the features after channel attention recalibration, outputs continuous risk scores, and then determines low-risk, medium-risk and high-risk levels through piecewise linear membership functions, so that the risk assessment results retain the risk differences reflected by continuous values ​​and can form risk level results that are easy to run and analyze. Attached Figure Description

[0021] Figure 1 This is a schematic diagram of the overall process of the new energy transmission network risk assessment method based on channel attention multi-scale convolution of the present invention.

[0022] Figure 2 This is a schematic diagram of the network-side risk indicator system for new energy transmission networks according to the present invention.

[0023] Figure 3 This is a schematic diagram of the construction of the local subgraph and the sorted adjacency matrix based on the degree sorting of neighbor nodes in this invention.

[0024] Figure 4 This is a schematic diagram of the multi-scale channel attention map convolutional network structure of the present invention. Detailed Implementation

[0025] The present invention will be further described below with reference to the accompanying drawings and specific embodiments. It should be understood that the following embodiments are for illustrative purposes only and are not intended to limit the scope of protection of the present invention.

[0026] In this invention, the trainable risk index gating vector acts on the risk index dimension before network input, and is used to train and correct the objective prior weights obtained by the information entropy method; the multi-scale feature channel attention weights act on the hidden feature channels after the multi-scale convolution output is concatenated. The two have different objects and stages of action.

[0027] Example 1 This embodiment provides a risk assessment method for new energy power transmission networks based on channel attention multi-scale convolution. For example... Figure 1 As shown, the method includes the following steps.

[0028] S101. Obtain data from the new energy power transmission network and construct its operating conditions. Obtain the grid topology, line parameters, node voltage, branch power flow, wind power access nodes, photovoltaic access nodes, load nodes, wind power rated installed capacity, photovoltaic rated installed capacity, system peak load, wind power output factor, photovoltaic output factor, and load level of the new energy transmission network.

[0029] The penetration rate of new energy sources is defined as:

[0030] In the formula, Indicates the penetration rate of new energy sources. Indicates the rated installed capacity of wind power. Indicates the rated installed capacity of photovoltaic power. This indicates the system's peak load.

[0031] The intermittency intensity is determined based on the fluctuation range of the wind power output coefficient and the photovoltaic output coefficient. The first fluctuation range of the wind power output coefficient and the photovoltaic output coefficient corresponds to the weak intermittent operation condition; the second fluctuation range of the wind power output coefficient and the photovoltaic output coefficient corresponds to the strong intermittent operation condition, and the interval width of the corresponding output coefficient in the second fluctuation range is greater than the interval width of the corresponding output coefficient in the first fluctuation range.

[0032] By combining the penetration rate of new energy sources with the intensity of intermittent power generation, different operating conditions of new energy transmission networks can be formed.

[0033] S102. Generate typical operating scenarios For each operating condition, wind power output, solar power output, and load level are combined to form a scenario random vector:

[0034] In the formula, Indicates wind power output. Indicates photovoltaic power output. Indicates the load level.

[0035] The cumulative distribution intervals of wind power output, photovoltaic power output, and load levels are stratified with equal probability. The random variable in the th... The cumulative probability of the median within each stratified interval is:

[0036] In the formula, Indicates the first The random variable in the th... The cumulative probability of the median within each stratified interval Indicates the initial number of scenes.

[0037] Sample values ​​are obtained through the inverse distribution function of the corresponding random variable:

[0038] In the formula, Indicates the first The inverse distribution function of n random variables, Indicates the first The random variable in the th... The corresponding sample values ​​within each stratified interval.

[0039] Wind power output sample sequences, photovoltaic power output sample sequences, and load level sample sequences are generated separately. Each sample sequence is randomly permuted and combined according to the scenario number to obtain an initial joint disturbance scenario set. The synchronous back-substitution method is used to reduce the initial joint disturbance scenario set as a whole to obtain a typical operating scenario set and the scenario probability corresponding to each typical operating scenario.

[0040] S103. Construct a node-risk indicator feature matrix Power flow calculations are performed under various typical operating scenarios to obtain operational data such as node voltage and active power flow of each transmission line.

[0041] like Figure 2 As shown, the grid-side risk indicator system includes power flow carrying capacity risk indicators, voltage safety risk indicators, structural vulnerability risk indicators, and local coupling and uncertainty propagation risk indicators.

[0042] Among them, line load factor, power over-limit risk index, power flow transfer impact, voltage over-limit risk index, improved voltage stability index, nodal centrality and electrical betweenness are used as basic risk indicators; neighborhood risk clustering degree, nodal risk fluctuation coefficient and neighborhood risk difference degree are used as risk indicators of local coupling and uncertainty propagation.

[0043] For branch risk indicators, the mean value of the corresponding indicator for each branch adjacent to the target node is calculated, and the obtained mean value is mapped to the target node; for node risk indicators, the indicator value of the corresponding node is used as the risk indicator feature of that node.

[0044] Scene Next node The impact of tidal current transfer is defined as:

[0045] In the formula, Represents nodes The collection of connected transmission lines Indicates the number of associated lines. Representing a scene Downline The meritorious trend, Indicates the line under the reference operating condition The meritorious trend, Indicates the line The maximum permissible transmission power.

[0046] Scene Next node The improved voltage stability index is defined as:

[0047] In the formula, Represents a node In the scene The actual voltage amplitude under the condition, Indicates the node's rated voltage. This indicates the permissible voltage limit of the node.

[0048] The basic risk indicators are positiveized and normalized, and the basic objective prior weights are obtained using the same method as the information entropy calculation described later. Based on these basic objective prior weights, the normalized basic risk indicators are weighted and summed to obtain the initial risk amount of each node. The initial risk value for a node does not include the output of the multi-scale channel attention map convolutional network.

[0049] Scene Next node The neighborhood risk clustering degree is:

[0050] In the formula, This indicates the nodes selected based on their degree values. The set of neighboring nodes, This indicates the number of nodes in the neighborhood node set. Representing a scene Next Neighbor Nodes The initial risk level.

[0051] node The probability-weighted average of the initial risk under all typical operating scenarios is:

[0052] The node risk volatility coefficient is:

[0053] In the formula, Indicates the number of typical operating scenarios. Indicates the first The scenario probability corresponding to each typical operating scenario Represents a node The initial risk quantity probability-weighted average, This indicates a positive number used to prevent the denominator from being zero.

[0054] Scene Next node The neighborhood risk difference is:

[0055] In the formula, Representing a scene Next node The neighborhood risk difference is represented by the symbol, and the meanings of the other symbols are the same as before.

[0056] The aforementioned risk indicators of local coupling and uncertainty propagation, together with the basic risk indicators, form a node-risk indicator feature matrix corresponding to each typical operation scenario. The rows of the node-risk indicator feature matrix correspond to the nodes of the new energy power transmission network, and the columns correspond to the risk indicators.

[0057] S104. Calculate the entropy weight gate weight. The risk indicators in the node-risk indicator feature matrix are positiveized and normalized. (Scenario) Next node In the The normalized percentage for each risk indicator is:

[0058] Scene Next The information entropy of each risk indicator is:

[0059] Scene Next The objective prior weights of each risk indicator are:

[0060] In the formula, Representing a scene Next node In the Normalized values ​​for each risk indicator Indicates the number of nodes. This indicates the number of risk indicators.

[0061] Set a trainable risk indicator gating vector Its dimension is the same as the number of risk indicators, and the initial value of each element is 1. The scenario... The objective prior weight vector With trainable risk index gating vector Multiplying element by element, we get the scene. Entropy weight gating weights:

[0062] In the formula, Representing a scene The entropy weight gate weight vector under the following conditions This represents the Hadamard product.

[0063] At the start of model training, all elements of the trainable risk index gating vector are 1, without changing the objective prior weights. During model training, the trainable risk index gating vector is updated synchronously with the network parameters as a global trainable parameter. After model training is completed, the trainable risk index gating vector remains fixed and is used for risk assessment in subsequent typical operating scenarios.

[0064] S105. Constructing local subgraphs and local weighted feature matrices like Figure 3 As shown, for any node to be evaluated, it is designated as the central node. The set of direct neighbor nodes of the central node is obtained, and the node degree value of each direct neighbor node is calculated. The node degree value is the number of branches directly connected to the corresponding direct neighbor node.

[0065] Sort direct neighbor nodes in descending order of degree value. When the number of direct neighbor nodes is greater than... When selecting, choose the one that appears first in the sorted list. The number of direct neighbor nodes; when the number of direct neighbor nodes is not greater than When selecting all direct neighbor nodes, place the central node at the beginning of the local node sequence, and then arrange the selected direct neighbor nodes in descending order of their degree values ​​after the central node.

[0066] The central node and its selected direct neighbors together constitute a local subgraph. An ordered adjacency matrix is ​​constructed based on the actual topological connections of each node in the local subgraph within the renewable energy transmission network. This ordered adjacency matrix is ​​then normalized to obtain a normalized ordered adjacency matrix. .

[0067] According to the local node sequence, from the scene Extracting local node-risk indicator feature matrix from the corresponding node-risk indicator feature matrix Utilizing the scenario The entropy weight is used to weight the input:

[0068] In the formula, This represents the locally weighted feature matrix after entropy-gated correction. This represents a diagonal matrix composed of vectors.

[0069] S106, Training a Multi-Scale Channel Attention Map Convolutional Network Scene Lower central node The corresponding local weighted feature matrix and normalized sorted adjacency matrix The model input is used as a training sample.

[0070] The reference risk value for the training samples is adopted in the scenario. Lower central node The initial risk of the node is given, and expressed as:

[0071] In the formula, Representing a scene Next node Reference risk value, Indicates the number of basic risk indicators. This represents the corresponding normalized basic risk indicator. This represents the corresponding basic objective prior weight.

[0072] The model uses the mean squared error loss function:

[0073] In the formula, Indicates the model loss. Represents the training sample set, Indicates the number of training samples. This represents the continuous risk score output by the model. By minimizing the model loss, the trainable risk index gating vector is updated synchronously with the network parameters of the multi-scale channel attention map convolutional network.

[0074] S107, Perform multi-scale channel attention risk identification like Figure 4 As shown, the multi-scale channel attention graph convolutional network includes a graph convolutional propagation layer, a multi-scale convolutional layer, a channel attention layer, and an output layer.

[0075] The graph convolutional propagation layer propagates and aggregates the features of local subgraph nodes based on the normalized and sorted adjacency matrix:

[0076] In the formula, Indicates the first Layer node feature matrix Indicates the first Layer trainable parameter matrix, This represents the activation function.

[0077] The node features after graph convolution propagation are arranged according to the local node sequence. The rows of the local weighted feature matrix correspond to local nodes, the columns correspond to risk indicators, and the center node is located at the beginning of the local node sequence. Multi-scale convolutional layers perform parallel feature extraction along the node dimensions corresponding to the local node sequence, and each scale branch maintains the correspondence of node positions so as to extract the features corresponding to the center node from the recalibrated features.

[0078] In this embodiment, three parallel convolutional branches of different scales are used to extract risk features within the first-order, second-order, and third-order neighborhood ranges, respectively:

[0079] In the formula, , and These represent the features extracted at the first-, second-, and third-order neighborhood scales, respectively. This indicates splicing along the feature channel dimension. This represents multi-scale splicing features.

[0080] The channel attention layer performs global average pooling and fully connected gating on the multi-scale concatenated features to obtain the multi-scale feature channel attention weights:

[0081] Multi-scale concatenated features are recalibrated channel by channel using multi-scale feature channel attention weights:

[0082] In the formula, Indicates global average pooling. and This represents the weight parameters of the fully connected layer. Represents a non-linear activation function. This represents the attention weights for multi-scale feature channels. This represents the multi-scale features after recalibration.

[0083] Extract the features corresponding to the central nodes from the recalibrated multi-scale features. Mapping is performed through the output layer:

[0084] In the formula, and These represent the output layer weights and biases, respectively. Representing a scene Next node The continuous risk score.

[0085] S108. Determine the risk level of the node. Scene Next node The continuous risk score is used as the comprehensive risk score for the corresponding scenario. A piecewise linear membership function is used to calculate the membership degree of this comprehensive risk score to low-risk, medium-risk, and high-risk levels, respectively. Pre-set linear transition intervals are set between low-risk and medium-risk levels, and between medium-risk and high-risk levels. Within each transition interval, the membership degree of adjacent risk levels changes linearly with the continuous risk score.

[0086] Based on the principle of maximum membership, the risk level with the highest membership degree is determined as the risk level of the node in the corresponding typical operating scenario:

[0087] In the formula, Represents a node The risk level, Represents a node The overall risk score, This indicates that the overall risk score belongs to the first category. Membership degree of each risk level.

[0088] Through the above steps, each typical operating scenario and each central node are processed separately to obtain the continuous risk score and risk level of each node of the new energy transmission network under each typical operating scenario.

[0089] Example 2 This embodiment provides a risk assessment method for renewable energy transmission networks under different renewable energy penetration rates and different intermittency intensities. This embodiment uses the same typical operating scenario generation, risk indicator construction, entropy weighting, local subgraph construction, multi-scale channel attention map convolution processing, and risk level determination process as Embodiment 1. The difference lies in the way the operating conditions are set.

[0090] The penetration rates of new energy sources are set at 20%, 40%, and 60% respectively, forming three levels of new energy access: low penetration, medium penetration, and high penetration.

[0091] For each new energy penetration rate, a first fluctuation range and a second fluctuation range are set. The first fluctuation range corresponds to a weak intermittent operation condition; in the second fluctuation range, the interval widths of the wind power output coefficient and photovoltaic power output coefficient are greater than the interval widths of the corresponding output coefficients in the first fluctuation range, and the second fluctuation range corresponds to a strong intermittent operation condition.

[0092] The corresponding operating conditions are formed by the combination of three new energy penetration rates and two intermittent intensity levels. For each operating condition, median Latin hypercube sampling and synchronous back substitution methods are used to obtain typical operating scenarios. Power flow calculation, risk indicator construction, local subgraph construction, and multi-scale channel attention map convolution risk identification are then performed under each typical operating scenario.

[0093] The continuous risk score and risk level of each node under each operating condition and each typical operating scenario are output respectively, so as to compare the node risk results corresponding to weak intermittent operating conditions and strong intermittent operating conditions under the same new energy penetration rate, as well as the node risk results under different new energy penetration rates.

Claims

1. A risk assessment method for new energy power transmission networks based on channel attention multi-scale convolution, characterized in that, include: The system acquires topology data, branch operation data, node operation data, wind power output data, photovoltaic power output data, and load data of the new energy transmission network. Based on the new energy penetration rate and the intermittent intensity of wind and photovoltaic power output, it constructs operating conditions and generates typical operating scenarios corresponding to each operating condition. Power flow calculations are performed under each typical operating scenario to obtain basic risk indicators corresponding to power flow carrying capacity, voltage safety, and structural vulnerability. These basic risk indicators are then normalized and normalized. Basic objective prior weights are obtained based on information entropy. Finally, the normalized basic risk indicators are weighted and aggregated based on these basic objective prior weights to obtain the initial risk quantity of each node. Based on the initial risk level of each node, the neighborhood risk clustering degree, node risk fluctuation coefficient, and neighborhood risk difference degree are calculated. The resulting local coupling and uncertainty propagation risk indicators, along with the basic risk indicators, constitute a node-risk indicator feature matrix. Objective prior weights are obtained based on the information entropy of each risk indicator in the node-risk indicator feature matrix. A trainable risk indicator gating vector with an initial value of 1 and a dimension equal to the number of risk indicators is set. The objective prior weights are multiplied element-wise by the trainable risk indicator gating vector to obtain the entropy weight gating weights. For each central node, its direct neighbors are arranged in descending order of node degree. At most L of the top-ranked direct neighbors are selected to form a local subgraph with the central node. A local node sequence is formed by arranging the central node first and the selected direct neighbors in descending order of node degree. The local subgraph is then used to determine the sequence of nodes. A normalized sorted adjacency matrix is ​​constructed based on topological connectivity. Local node-risk indicator feature matrices are extracted from the node-risk indicator feature matrix according to the local node sequence, weighted using entropy-gated weights, and then input into a multi-scale channel attention graph convolutional network. Graph convolution propagation is performed on the local subgraph node features, and parallel feature extraction at different scales is performed along the node dimensions corresponding to the local node sequence. Features at different scales are concatenated along the feature channel dimensions, and attention recalibration is performed on the concatenated feature channels. Features corresponding to the central node are extracted from the recalibrated features, and the continuous risk score of the central node is output. The trainable risk indicator gating vector is updated synchronously with the model loss and network parameters during the training of the multi-scale channel attention graph convolutional network.

2. The risk assessment method for new energy transmission networks based on channel attention multi-scale convolution according to claim 1, characterized in that, The penetration rate of new energy sources is determined based on the ratio of the sum of the rated installed capacity of wind power and the rated installed capacity of photovoltaics to the peak load of the system; the output coefficient of wind power and the output coefficient of photovoltaics have a first fluctuation range and a second fluctuation range, respectively. The range width of the wind power output coefficient and the range width of the photovoltaic output coefficient in the second fluctuation range are respectively greater than the corresponding range width in the first fluctuation range. The first fluctuation range corresponds to a weak intermittent operating condition, and the second fluctuation range corresponds to a strong intermittent operating condition.

3. The risk assessment method for new energy transmission networks based on channel attention multi-scale convolution according to claim 1, characterized in that, Generating typical operating scenarios includes: stratifying the cumulative distribution intervals of wind power output, photovoltaic power output, and load levels into equal-probability strata, using the median cumulative probability of each stratified interval as the sampling probability, and obtaining sample values ​​through the inverse distribution function of the corresponding random variables; randomly permuting and combining the wind power output sample sequence, photovoltaic power output sample sequence, and load level sample sequence to form an initial joint disturbance scenario; and using the synchronous back-substitution method to reduce the initial joint disturbance scenario as a whole to obtain typical operating scenarios and their corresponding scenario probabilities.

4. The risk assessment method for new energy transmission networks based on channel attention multi-scale convolution according to claim 1, characterized in that, The basic risk indicators include line load factor, power over-limit risk index, power flow transfer impact, voltage over-limit risk index, improved voltage stability index, nodal centrality, and electrical betweenness; the local coupling and uncertainty propagation risk indicators include neighborhood risk clustering degree, nodal risk fluctuation coefficient, and neighborhood risk difference degree; wherein, the mean of the branch-type risk indicators corresponding to each branch adjacent to the target node is mapped to the target node, and the nodal-type risk indicators are used as the risk indicator characteristics of the corresponding nodes.

5. The risk assessment method for new energy transmission networks based on channel attention multi-scale convolution according to claim 4, characterized in that, The power flow transfer impact is determined based on the change in active power flow of the node-associated lines relative to the baseline operating state under typical operating scenarios and the maximum allowable transmission power of the corresponding lines; the improved voltage stability index is determined based on the actual voltage of the node, the rated voltage of the node, and the allowable voltage limit of the node; the neighborhood risk clustering degree is determined based on the initial risk of the neighboring nodes selected according to the node degree value; the node risk fluctuation coefficient is determined based on the initial risk of the node under each typical operating scenario, the scenario probability of each typical operating scenario, and the probability-weighted average of the initial risk of the node; the neighborhood risk difference degree is determined based on the difference in the initial risk between the central node and the selected neighboring nodes.

6. The risk assessment method for new energy transmission networks based on channel attention multi-scale convolution according to claim 1, characterized in that, For each typical operating scenario, the information entropy of each risk indicator is calculated based on the difference in normalized values ​​of each risk indicator between different nodes under that typical operating scenario, and the objective prior weight is determined based on the information entropy of each risk indicator. The objective prior weight is multiplied one by one with the element corresponding to the position in the gating vector of the trainable risk indicator to form the weighting coefficient corresponding to each risk indicator. The weighting coefficient corresponding to each risk indicator is used to weight the corresponding risk indicator column in the local node-risk indicator feature matrix.

7. The risk assessment method for new energy transmission networks based on channel attention multi-scale convolution according to claim 1, characterized in that, The degree value of a direct neighbor node is the number of branches directly connected to that direct neighbor node. When the number of direct neighbor nodes is greater than L, the top L direct neighbor nodes are selected; when the number of direct neighbor nodes is not greater than L, all direct neighbor nodes are selected. The center node is placed at the beginning of the local node sequence, and the selected direct neighbor nodes are arranged in descending order of their degree values ​​after the center node. A sorted adjacency matrix is ​​constructed based on the actual topological connections between nodes in the local node sequence. The degree sorting is also used to determine the node set of the local subgraph and the input order for multi-scale feature extraction.

8. The risk assessment method for new energy transmission networks based on channel attention multi-scale convolution according to claim 7, characterized in that, The multi-scale channel attention graph convolutional network consists of a graph convolutional propagation layer, a multi-scale convolutional layer, a channel attention layer, and an output layer. The graph convolutional propagation layer propagates and aggregates node features in the local subgraph based on a normalized sorted adjacency matrix. The multi-scale convolutional layer includes at least two parallel convolutional branches with different kernel scales. Each parallel convolutional branch performs feature extraction along the node dimension arranged according to the local node sequence, and the outputs of each parallel convolutional branch are concatenated along the feature channel dimension. The channel attention layer performs global average pooling and fully connected gating on the concatenated features to generate multi-scale feature channel attention weights, and uses the multi-scale feature channel attention weights to recalibrate the concatenated features channel by channel.

9. A risk assessment method for new energy transmission networks based on channel attention multi-scale convolution as described in claim 8, characterized in that, The trainable risk indicator gating vector is a global parameter obtained by jointly training it using training samples corresponding to different typical operating scenarios and different central nodes. The reference risk value of the training samples is obtained by weighting the normalized basic risk indicators of the corresponding center nodes according to the corresponding objective prior weights; the model loss is the mean square error between the continuous risk score output by the multi-scale channel attention map convolutional network and the reference risk value; at the beginning of model training, all elements of the trainable risk indicator gating vector are 1. During model training, the trainable risk indicator gating vector is updated synchronously according to the model loss and the network parameters of the multi-scale channel attention map convolutional network, and is fixed after the model training is completed; the output layer extracts the features corresponding to the center nodes from the features after channel attention recalibration, and maps the features corresponding to the center nodes to obtain the continuous risk score of the center nodes under the corresponding typical operating scenarios.

10. A risk assessment method for new energy transmission networks based on channel attention multi-scale convolution according to claim 9, characterized in that, Piecewise linear membership functions are used to determine the membership degree of the continuous risk score of the central node to low-risk, medium-risk, and high-risk levels, respectively. A linear transition interval is set between adjacent risk levels, and the risk level with the highest membership degree is determined as the risk level of the central node in the corresponding typical operating scenario based on the principle of maximum membership.

Images