Risk discrimination method and system for participation of scale wind power in grid reconstruction based on MPNN

Abstract

The invention discloses an MPNN-based risk discrimination method and system for participation of large-scale wind power in grid reconstruction, and relates to the technical field of power system wind power access risk assessment, and the method comprises the steps: defining a risk function, and dividing the risk level of a wind power access reconstruction grid; generating a sample data set and reconstructing the sample data set into a graph data form; and training the MPNN model, and judging the risk level after wind power integration. According to the MPNN-based risk discrimination method for participation of large-scale wind power in grid reconstruction, the risk is divided into different levels through the risk function, the recovery speed and reliability of a power grid are improved, a sample set covering various topological structures and wind power access scenes is established, the generalization ability of a model in a changing environment is improved, and the risk discrimination method is suitable for large-scale wind power grid reconstruction. The intelligent level of recovery of the power system is improved, the wind power access risk level is rapidly output through feature normalization and graph data processing, the reliability and the recovery speed of the power system are improved, and better effects are achieved in the aspects of reliability, generalization ability and intelligence.

Description

Technical Field

[0001] The present invention relates to the technical field of risk assessment for wind power integration into power systems, and specifically to a risk discrimination method for large-scale wind power participating in network reconfiguration based on MPNN. Background Art

[0002] Under the background of the global energy structure transformation towards low-carbon, the scale of new energy sources represented by wind power integrated into the power system is continuously expanding, which poses new challenges to the stability and restoration ability of the power system. Reasonably integrating wind power during the power system network reconfiguration stage to improve the system restoration efficiency has become an important research direction for building a new type of power system. In recent years, the combination of deep learning and power systems has become a new path to solve complex problems. Among them, graph neural networks (Graph Neural Network, GNN) have gradually been applied to the risk discrimination of wind power participating in network reconfiguration due to their advantages in processing non-Euclidean data and complex topological structures, providing a new data-driven solution.

[0003] Although certain progress has been made in current research on wind power participating in power grid restoration, during the network reconfiguration process, due to the continuous change of the system topological structure, existing risk discrimination methods mainly rely on simulation modeling and are difficult to adapt to the complex power grid environment with dynamic changes. These methods usually build models based on specific topological structures and need to rebuild the simulation model when the topology changes, resulting in cumbersome calculations, low efficiency, and weak generalization ability. In addition, traditional risk assessment models based on Euclidean data are difficult to describe the complex topological relationships of power systems and cannot effectively utilize graph data characteristics such as buses, lines, and their connectivity.

[0004] In contrast, graph neural networks (such as message passing graph neural network MPNN) can directly handle the changes in the topological structure of power systems and maintain a high discrimination accuracy when the topology changes dynamically. However, existing MPNN models are mostly applied to transient stability analysis and fault location, and there is still less research on the risk discrimination of wind power integration into network reconfiguration. Summary of the Invention

[0005] In view of the above existing problems, the present invention is proposed.

[0006] Therefore, the technical problem solved by the present invention is: existing wind power integration risk discrimination methods have problems such as being unable to adapt to dynamic topological changes, complex and time-consuming calculation processes, insufficient generalization ability, and the optimization problem of how to quickly and accurately discriminate the risk level of wind power integration into network reconfiguration.

[0007] To solve the above technical problems, the present invention provides the following technical solutions: A risk discrimination method for large-scale wind power participating in network reconstruction based on MPNN, including defining a risk function and dividing the risk levels of wind power accessing the reconstructed network; generating a sample data set and reconstructing it into a graph data form; training the MPNN model to discriminate the risk level after wind power access.

[0008] As a preferred solution of the risk discrimination method for large-scale wind power participating in network reconstruction based on MPNN according to the present invention, wherein: the defining of the risk function includes considering the system frequency and voltage deviations caused by the fluctuations of wind power output, and at the same time taking into account the acceptability of transient frequency and voltage offsets, dividing the risk levels of wind power accessing the reconstructed network. During the power restoration process, the negative impacts caused by wind power are the frequency deviation and voltage deviation caused by output fluctuations. The risk function for wind power participating in network reconstruction is defined as:

[0009]

[0010] wherein, Δf max is the maximum system frequency deviation caused by wind power with a capacity of P a connected at time t, ΔU max is the maximum system voltage deviation caused by wind power with a capacity of P a connected at time t, g and h are frequency quality functions, and w is a weight coefficient, representing the importance of voltage quality compared to frequency quality.

[0011] As a preferred solution of the risk discrimination method for large-scale wind power participating in network reconstruction based on MPNN according to the present invention, wherein: the dividing of the risk levels of wind power accessing the reconstructed network includes that when the risk level after wind power access to the reconstructed network is Class I, it means that the impact of wind power access on the system is very small, and the frequency and voltage fluctuations are within the safe range and can be ignored; Class II means that the fluctuations after wind power access interfere with the system stability, but the positive impact of wind power access is greater than the negative impact. At this time, connecting wind power accelerates the system restoration process and the disturbance to the system is within the acceptable range; Class III means that the negative impact after wind power access is greater than the positive impact. At this time, it is comprehensively considered whether to connect wind power; Class IV means that the disturbance caused by wind power access to the system exceeds the tolerance range of the network. At this time, wind power is not connected.

[0012] As a preferred solution of the risk discrimination method for large-scale wind power participating in network reconstruction based on MPNN according to the present invention, wherein: the generated sample data set includes, during the generation of the sample data set, simulating a complete network reconstruction process through PSD-BPA, changing the access position and capacity of wind power in the topological structures at different time steps, obtaining the maximum values, durations of system frequency and voltage fluctuations during the process, and training set labels, that is, the risk levels of wind power access. The data obtained by simulation are all long vectors, and the data are reconstructed into the form of graph data. If there is no wind power access at a node in the node information, it is 0. The input of the message passing graph neural network is the information of nodes and lines. The node data is a three-dimensional array with the array shape of (N, M, 5), where N is the number of samples, M is the number of nodes in the system after the network reconstruction is completed, that is, the largest number of nodes in all samples, and 5 is the feature information on each node. The line data is a four-dimensional array with the array shape of (N, M, M, 3), where the information in the M×M dimension represents the specific position of the line. When it is 0, it means there is no edge between two nodes, and when it is 1, it means there is an edge between two nodes. 3 is the feature information on each line.

[0013] As a preferred solution of the risk discrimination method for large-scale wind power participating in network reconstruction based on MPNN according to the present invention, wherein: the reconstruction into the form of graph data includes simulating and obtaining data in the form of long vectors during the sample generation process, and the graph data processing reconstructs the data into the input graph data required by the MPNN model. The nodes in the graph data represent the buses in the reconstructed network, and the edges in the graph data represent the lines in the reconstructed network in the model. The risk discrimination model for wind power access reconstruction of the network based on MPNN adapts to the increase and decrease of the feature dimensions of nodes or edges in the input variables, and does not frequently adjust the model framework when the input information changes. The model applies Z-score standardization to uniformly convert the data of different magnitudes in the samples into the same magnitude, which is expressed as:

[0014]

[0015] where μ is the mean of the sample data, σ is the standard deviation of the sample data, and the sample data is used as the input of the neural network for training the model after normalization.

[0016] As a preferred solution of the risk discrimination method for large-scale wind power participating in network reconstruction based on MPNN according to the present invention, wherein: the training of the MPNN model includes, after constructing the risk discrimination model, dividing the data obtained during the graph data generation and processing into a training set and a validation set, inputting the training set into the MPNN model for offline training, and screening the model with the best performance as the final model according to the performance in the validation set. During training, by comparing the sample labels and the output risk function values, calculating the loss function, and using the backpropagation algorithm to update the internal weights, biases, and parameters of the model.

[0017] As a preferred solution of the risk discrimination method for large-scale wind power participating in grid reconstruction based on MPNN according to the present invention, wherein: the discrimination of the risk level after wind power access includes obtaining the online operation scenario of the system and the access location and capacity of the wind power. According to the data, input features are selected, namely node information and line information. After being processed and transformed into the neural network input data format, it is input into the trained risk discrimination model, and the output is the risk level of the wind power accessing the system at this time. The impact of the wind power access on the power system restoration at this time is judged. For the same input, the model obtains multiple different prediction outputs passing through the model through random forward propagation, and uses the standard deviation of multiple predictions to quantify the uncertainty of the risk prediction of the wind power accessing the reconstructed grid, and takes the average prediction value of multiple predictions as the result.

[0018] Another object of the present invention is to provide a risk discrimination system for large-scale wind power participating in grid reconstruction based on MPNN, which can train a message passing graph neural network model through a training discrimination module and online discriminate the risk level of the current wind power access, solving the problem of poor generalization ability when facing the dynamically changing grid topology and being difficult to accurately predict the risk level under complex conditions.

[0019] As a preferred solution of the risk discrimination system for large-scale wind power participating in grid reconstruction based on MPNN according to the present invention, wherein: it includes a risk division module, a generation and reconstruction module, and a training discrimination module; the risk division module is used to define a risk function and divide the risk level of the wind power accessing the reconstructed grid; the generation and reconstruction module is used to generate a sample data set and reconstruct it into a graph data form; the training discrimination module is used to train the MPNN model and discriminate the risk level after the wind power access.

[0020] A computer device includes a memory and a processor, and the memory stores a computer program, characterized in that when the processor executes the computer program, the steps of the risk discrimination method for large-scale wind power participating in grid reconstruction based on MPNN are realized.

[0021] A computer-readable storage medium stores a computer program thereon, characterized in that when the computer program is executed by a processor, the steps of the risk discrimination method for large-scale wind power participating in grid reconstruction based on MPNN are realized.

[0022] Advantages of the present invention: The risk discrimination method for large-scale wind power participating in network reconstruction based on MPNN provided by the present invention quantifies the impact of wind power access on the power system according to the system frequency and voltage deviation caused by the fluctuation of wind power output, and divides it into different levels through a risk function, improving the restoration speed and reliability of the power grid. A sample set covering various topological structures and wind power access scenarios is established, enabling the model to have the ability to identify different power grid states and risks, improving the generalization ability of the model in a changing environment, enhancing the intelligent level of power system restoration. The real-time data of wind power access is input into the trained MPNN model, and through feature normalization and graph data processing, the risk level of wind power access is quickly output to achieve online monitoring and discrimination, enhancing the reliability and restoration speed of the power system, reducing the risk of large-scale power outages in the power system. The present invention achieves better results in terms of reliability, generalization ability, and intelligence. BRIEF DESCRIPTION OF THE DRAWINGS

[0023] In order to more clearly illustrate the technical solutions of the embodiments of the present invention, the drawings required for description in the embodiments will be briefly introduced below. Obviously, the drawings in the following description are only some embodiments of the present invention. For those of ordinary skill in the art, other drawings can be obtained based on these drawings without creative efforts. Among them:

[0024] Figure 1 It is the overall flowchart of the risk discrimination method for large-scale wind power participating in network reconstruction based on MPNN provided by the first embodiment of the present invention.

[0025] Figure 2 It is the system frequency and voltage quality function diagram of the risk discrimination method for large-scale wind power participating in network reconstruction based on MPNN provided by the first embodiment of the present invention.

[0026] Figure 3 It is the risk discrimination model diagram of wind power access reconstruction network based on MPNN of the risk discrimination method for large-scale wind power participating in network reconstruction provided by the first embodiment of the present invention.

[0027] Figure 4 It is the schematic diagram of the node message passing graph neural network framework of the risk discrimination method for large-scale wind power participating in network reconstruction based on MPNN provided by the first embodiment of the present invention.

[0028] Figure 5 It is the improved graph data example diagram of the risk discrimination method for large-scale wind power participating in network reconstruction based on MPNN provided by the first embodiment of the present invention.

[0029] Figure 6The structure diagram of the MPNN model for the risk discrimination method of large-scale wind power participating in network reconstruction based on MPNN provided by the first embodiment of the present invention.

[0030] Figure 7 The IEEE 39-node system diagram with a wind farm for the risk discrimination method of large-scale wind power participating in network reconstruction based on MPNN provided by the second embodiment of the present invention.

[0031] Figure 8 The performance diagram of different models of the risk discrimination method of large-scale wind power participating in network reconstruction based on MPNN provided by the second embodiment of the present invention in a single topological structure.

[0032] Figure 9 The performance diagram of different models of the risk discrimination method of large-scale wind power participating in network reconstruction based on MPNN provided by the second embodiment of the present invention in a changing topological structure.

[0033] Figure 10 The overall module diagram of the risk discrimination system of large-scale wind power participating in network reconstruction based on MPNN provided by the third embodiment of the present invention. Detailed implementation manners

[0034] To make the above objects, features, and advantages of the present invention more obvious and understandable, the following detailed description of the specific implementation manners of the present invention is provided in conjunction with the accompanying drawings of the specification. Obviously, the described embodiments are part of the embodiments of the present invention, rather than all embodiments. All other embodiments obtained by those of ordinary skill in the art based on the embodiments of the present invention without creative efforts shall fall within the scope of protection of the present invention.

[0035] Embodiment 1

[0036] Refer to Figures 1 - 6 , which is an embodiment of the present invention, providing a risk discrimination method for large-scale wind power participating in network reconstruction based on MPNN, including:

[0037] S1: Define a risk function and divide the risk levels of wind power access to the reconstructed network.

[0038] Furthermore, defining the risk function includes considering the system frequency and voltage deviations caused by the fluctuations of wind power output.

[0039] It should be noted that while taking into account the acceptability of transient frequency and voltage offsets, the risk levels of wind power access to the reconstructed network are divided. During the power restoration process, the negative impacts caused by wind power are the frequency deviation and voltage deviation caused by output fluctuations. The risk function for wind power participating in network reconstruction is defined as:

[0040]

[0041] Where Δf max Access P at time t a The maximum system frequency deviation caused by wind power capacity, ΔU max Access P at time t a The maximum system voltage deviation caused by wind power of different capacity, g and h are frequency quality functions, and w is the weight coefficient, which indicates the importance of voltage quality compared to frequency quality.

[0042] It should also be noted that according to the national standards for frequency and voltage, when the system frequency deviation is less than 0.2Hz, greater than 0.5Hz and greater than 0.8Hz, the system frequency is called safe, critically safe and unsafe. To represent different frequency states, the frequency quality function g is assigned values ​​of 1, 0 and -1. To simplify the analysis, the quality function is expressed by Figure 2 The piecewise function shown in the figure is used to represent x. 1 、x 2 、x 3 The values ​​of are 0.2, 0.5 and 0.8 respectively, in Hz. Figure 2 It can also be used to represent the voltage quality function h, when x 1 、x 2 、x 3 The values ​​are taken as 0.05, 0.1 and 0.15 respectively according to the national voltage standards, and the unit is pu.

[0043] It should also be noted that if the voltage and frequency offsets exceed the specified risk function value for a short period of time, but the voltage and frequency offset time is very short and within the system's tolerable range, then the access to wind power is still safe for the grid. Therefore, in addition to considering the risk function value, the risk level classification after wind power access also needs to consider the voltage and frequency offset time. Compared with hydropower and thermal power units, wind turbines have a lower tolerance for voltage and frequency offsets. Therefore, the main factor affecting the acceptability of transient voltage and frequency offsets in power systems containing wind power is still the wind turbine.

[0044] It should also be noted that if the voltage and frequency offset acceptability constraints are met, it means that the impact on the system is large and beyond the normal range at the moment of wind power access, but as time goes by, the frequency and voltage tend to stabilize, and the time consumed by the whole process is less than the national standard. In this case, it is considered that such wind power access to reconstruct the grid is safe. On the contrary, if the voltage and frequency offset acceptability constraints are not met, it is necessary to consider the risk function value for comprehensive judgment. In general, considering the acceptability of transient frequency and voltage offsets can not only maximize the system's absorption of wind power and increase the utilization rate of wind power, but also increase the speed of grid recovery and reduce the losses caused by major power outages.

[0045] Furthermore, the risk level classification of wind power integrated into the reconstructed power grid includes the risk level after wind power is integrated into the reconstructed power grid.

[0046] It should be noted that a risk level of Class I after wind power is integrated into the reconstructed power grid indicates that the impact of wind power integration on the system is very small, and the frequency and voltage fluctuations are within the safe range and can be ignored. Class II indicates that the fluctuations after wind power integration interfere with the system stability, but the positive impact of wind power integration is greater than the negative impact. At this time, integrating wind power accelerates the system recovery process and the disturbance to the system is within an acceptable range. Class III indicates that the negative impact after wind power integration is greater than the positive impact. At this time, it is necessary to comprehensively consider whether to integrate wind power. Class IV indicates that the disturbance generated by wind power integration after integration exceeds the tolerance range of the power grid. At this time, wind power should not be integrated.

[0047] It should also be noted that since large-scale power outages in the region are accidents with extremely serious losses, there are significant differences between the power grid reconstruction and normal operating conditions. At this time, the requirements for power quality are much lower than normal. Therefore, whether the frequency and voltage quality meet the standards also needs to be determined in combination with the actual situation on site.

[0048] It should also be noted that MPNN was initially proposed for predicting chemical molecular properties. It directly predicts the corresponding labels through the graph data G=(V, E) to achieve end-to-end learning of graph-structured data. In addition, it can process node and edge features with different structures, making up for the disadvantage of traditional graph convolutional neural networks that cannot process edge features. Figure 4 Schematic diagram showing the prediction of sample graph data using MPNN, and the predicted label is The prediction model f is defined as for each v i ∈V using the function to embed into the p-dimensional node representation vector Then it is processed through multiple message passing steps, where each node representation vector is recursively updated through the aggregation and transformation of the representation vectors of its adjacent nodes and the corresponding edge vectors. e 1,2 represents that the power flow direction is from node v 1 to node v 2 , at the L-th step of message passing, the vector h (L),i is updated according to the message function M, and the update function U is expressed as:

[0049]

[0050] Similar to the node message passing process, the message passing formula for edges is obtained as:

[0051]

[0052] Among them, the message passing function M is a fully connected neural network, and the state update function U is a gated recurrent unit. This network belongs to an improved recurrent neural network and can capture the temporal features in the sequence. Denote the given N-sample training data set, and the objective function of the training prediction model is expressed as:

[0053]

[0054] Among them, L is the loss function of regression. is the average value of each element in. The vector of each node is input into a readout function r, and the readout function r is implemented using the Set2Set network model to output the predicted label of the node. The given predicted value is expressed as:

[0055]

[0056] S2: Generate a sample data set and reconstruct it into the form of graph data.

[0057] Furthermore, generating the sample data set includes simulating a complete network framework reconstruction process through PSD-BPA during the generation of the sample data set.

[0058] It should be noted that in the topological structures at different time steps, change the access position and capacity of wind power to obtain the maximum value, duration, and training set label of the system frequency and voltage fluctuations during the process, that is, the risk level of wind power access. The data obtained from the simulation are all long vectors. Reconstruct the data into the form of graph data. If there is no wind power access at a node in the node information, it is 0. The input of the message passing graph neural network is the information of nodes and lines. The node data is a three-dimensional array with the array shape of (N, M, 5), where N is the number of samples, M is the number of nodes in the system after the network framework reconstruction is completed, that is, the largest number of nodes in all samples, and 5 is the feature information on each node. The line data is a four-dimensional array with the array shape of (N, M, M, 3), where the information in the M×M dimension represents the specific position of the line. When it is 0, it means there is no edge between two nodes, and when it is 1, it means there is an edge between two nodes. 3 is the feature information on each line.

[0059] Furthermore, reconstructing into the form of graph data includes simulating and obtaining data in the form of long vectors during the sample generation process.

[0060] It should be noted that the graph data processing reconstructs the data into the input graph data required by the MPNN model. The nodes in the graph data represent the buses in the reconstructed power grid framework, and the edges in the graph data represent the lines in the model for the reconstructed power grid framework. The risk discrimination model for the wind power access reconstructed power grid framework based on MPNN adapts to the increase or decrease of the feature dimensions on the nodes or edges in the input variables, and does not frequently adjust the model framework when the input information changes. The model applies Z-score standardization to uniformly convert the data of different magnitudes in the sample into the same magnitude, which is expressed as:

[0061]

[0062] Among them, μ is the mean of the sample data, σ is the standard deviation of the sample data, and the sample data is normalized and used as the input of the neural network to train the model.

[0063] It should also be noted that the generated graph data form can most intuitively represent the relationship between nodes and lines. However, there are also some problems. The matrix of line information in the graph data is a very sparse adjacency matrix. In the real power system, the number of nodes is extremely large, and the changes in lines are also more frequent. For the power system, if each line in the graph data is described in this way, then its sample data will be very large, which will not only make it difficult to input the data, but also seriously affect the training speed of the model. Therefore, after obtaining the above graph data, the description of the lines was improved as follows. The specific details are as Figure 5 shown.

[0064] It should also be noted that the edges in the improved graph data are described by an adjacency list. If the lines in the graph change, the adjacency list also changes accordingly, that is, the two are in a one-to-one correspondence relationship. Figure 5 The description of the nodes in remains unchanged. The description of the lines is divided into line feature information and an adjacency list. The line feature information of the model of the present invention is 3-dimensional. Each element in the adjacency list is a tuple. The k-th tuple tuple(i,j) indicates that there is an edge between node i and node j, corresponding to the k-th edge of the graph, and indicates the power flow direction. Splitting the line information into feature information and an adjacency list, converting the originally very sparse matrix into two sets of data not only reduces the difficulty of inputting the graph data, but also increases the training speed of the model, enables the model to adapt to more complex power systems in reality, and enhances the generalization ability of the model.

[0065] S3: Train the MPNN model to discriminate the risk level after wind power access.

[0066] Furthermore, training the MPNN model includes constructing a risk discrimination model.

[0067] It should be noted that the data obtained in the process of graph data generation and processing is divided into a training set and a validation set, and the training set is input into the MPNN model for offline training. According to the performance in the validation set, the model with the best performance is selected as the final model. During training, the loss function is calculated by comparing the sample labels and the output risk function values, and the internal weights, biases, and parameters of the model are updated through the backpropagation algorithm.

[0068] It should also be noted that to make the model differentiable, all component functions of f, M, U, and g in the model are parameterized as neural networks. As Figure 6 shown, the function is a two-layer fully connected network. Its first layer has 150 ReLU units, and the second layer has 48 ReLU units. The node feature information and the line feature information are respectively transformed into 48-dimensional long vectors through the embedding function . The node feature information is 5-dimensional, and the line feature information is 3-dimensional. The function M is a four-layer fully connected network. Its first three layers have 250 ReLU units, and the output of the last layer is a 96-dimensional graph embedding vector after graph pooling. The function U is a recurrent neural network with 50 gated recurrent units. At each time step L, it outputs the current hidden state h ((L-1),i) through the previous hidden state h ((L),i) and the current input m ((L),i) . The number of message passing steps is set to 4. After obtaining the 96-dimensional graph embedding vector, it is input into a four-layer fully connected network. The first three layers consist of 500 ReLU units, and its learning decay rate is 0.1. The last layer has 4 ReLU units for subsequent risk discrimination. Dropout regularizes the model during the training phase and is also used to quantify the prediction uncertainty during the training phase.

[0069] It should also be noted that the loss function used for training is the cross-entropy loss function. The initial learning rate obtained by optimizing the hyperparameters using the Adam optimizer is set to 0.01, and the learning rate decays to 90% every 10 rounds of iteration. The number of samples selected for each training, the batch size, is set to 20.

[0070] Furthermore, discriminating the risk level after the wind power access includes obtaining the online operation scenarios of the system and the access location and capacity of the wind power.

[0071] It should be noted that according to the data, the input features are selected, namely node information and line information. After being processed and transformed into the input data format of the neural network, they are input into the trained risk discrimination model. The output is the risk level of the wind power access system at this time, and the impact of the wind power access on the power system restoration at this time is judged. For the same input, the model obtains multiple different prediction outputs passing through the model through random forward propagation, and uses the standard deviation of multiple predictions to quantify the uncertainty of the risk prediction of the wind power access to reconstruct the grid. The average predicted value of multiple predictions is taken as the result.

[0072] Embodiment 2

[0073] Refer to Figures 7 - 9 , an embodiment of the present invention provides a risk discrimination method for large-scale wind power participating in grid reconstruction based on MPNN. In order to verify the beneficial effects of the present invention, scientific demonstration is carried out through economic benefit calculation and simulation experiments.

[0074] To verify the effectiveness of the risk discrimination model for the wind power access to reconstruct the grid, the algorithm is implemented using Python programming, and the New England 10-machine 39-node system is selected for sample generation. The computing platform configuration is: Intel(R) Xeon(R) GOLD 5218R 2.10GHz CPU, 128GB RAM, and NVIDIA 3080 10G GPU.

[0075] The wind farm access locations are nodes 6, 14, 18, 22, and 28, and each is equipped with 120 DFIG-type wind turbines with a single capacity of 1.5 MW. The 30th unit is selected as the black start power source during the restoration process, assuming that it can be connected to the grid and generate electricity within 5 minutes after starting the auxiliary load. The energized areas of the system restoration at the starting moments of the 2nd, 3rd, and 4th time steps are respectively as Figure 7 shown by the pink line, green line, and light blue line areas in. These 3 moments are respectively denoted as t2, t3, and t4 moments. At the initial stage of grid reconstruction, the system is very fragile and has poor ability to cope with the wind power output fluctuations. Therefore, only the black start unit is connected to the grid and generates electricity at the 1st time step. After forming the initial reconstructed small system, the wind power is connected to supply power to the area to be restored.

[0076] To obtain the risk coefficients of the wind power access to reconstruct the grid under different topological structures, sample data are obtained by PSD-BPA simulation, and are respectively denoted as dataset A and dataset B.

[0077] At a certain time step, change the connection location and capacity of wind power. While the topological structure remains unchanged at the same time step, continuously change the connection capacity and location of wind power to obtain the maximum frequency deviation, maximum voltage deviation, and the deviation time of voltage and frequency of the system after connecting wind power. Then, obtain the corresponding risk level when connecting wind power at this time. Use the grid structure of the fourth time step to simulate and generate 1000 sample data. Among them, 500 are samples with the wind power connection power fluctuating in the range of 60 - 160 MW at node 3, and 500 are samples with the wind power connection power fluctuating in the range of 60 - 160 MW at node 18.

[0078] Change the connection location and capacity of wind power in different time steps. At this time, different time steps correspond to different topological structures. Change the connection capacity and location of wind power in different grids. Similarly, obtain the corresponding risk levels of the wind power connected to the system. Use the grid structures of the second, third, and fourth time steps to simulate and generate 3000 sample data respectively. Each time step corresponds to 1000 data. The connection location of wind power and the fluctuation range are set the same as in dataset A. The fourth time step in dataset B is the same as dataset A.

[0079] Since the original dataset is in the form of a long vector, first reconstruct it into the graph data form required for model input, and then perform normalization processing. Divide dataset A into a training set (80%) and a test set (20%). Build a risk level classifier on the training set and evaluate the model performance on the test set. Select the following three evaluation indicators: macro F1 (F1), overall accuracy (accuracy, Acc.), and Kappa value. The larger the values of these three indicators, the better the model classification performance. Among them, the value range of kappa is [-1, 1], but usually the value of kappa is [0, 1]. Generally, it can be divided into the following five groups to represent different levels of consistency: 0.0 - 0.20 represents very low consistency, 0.21 - 0.40 represents general consistency, 0.41 - 0.60 represents medium consistency, 0.61 - 0.80 represents high consistency, and 0.81 - 1 represents almost perfect consistency. To clearly explain the meanings of the above indicators, establish the confusion matrix of a four-class problem shown in Table 1.

[0080] The four evaluation indicators and related variables are expressed as:

[0081]

[0082]

[0083] Among them, P 1 is the precision rate of class I, which is the probability of samples that are actually positive among all samples predicted to be positive. R 1Recall rate of class I, which is the probability of being predicted as a positive sample among the samples that are actually positive. Macro F1 is the harmonic mean of precision and recall rate, generally used in multi-classification problems. Acc is the accuracy rate, representing the prediction accuracy of the overall samples, P e is the proportion of cells with chance agreement or expected chance agreement. Kappa is an index to measure classification accuracy, and the higher the value, the higher the classification accuracy achieved by the model.

[0084] To verify the performance of the MPNN model constructed by the present invention for the risk assessment of the reconstructed power grid with wind power access, 4 common data-driven methods are selected for comparison, namely logistic regression (LR), random forest (RF), XGBoost (XGB), and Convolutional Neural Network (CNN). Among them, LR uses Logistic Regression in sklearn to perform logistic regression training on the sample matrix data. The training parameter selects newton-cg as the optimization algorithm and selects 1.0 as the penalty coefficient. The number of trees for RF and XGB is 100. The convolutional kernel sizes of the convolutional layers of CNN are set to (32, 32), (24, 24), (16, 16), and (8, 8) respectively. The stride of the convolutional layer is set to 16, 12, 8, 8 in sequence. The pooling sizes of the two pooling layers P1 and P2 in each convolutional module are (2, 2). The dropout rate of the Dropout layer is set to 0.1. The number of neurons in the fully connected layer is set to 1000. The output layer is set to 4, corresponding to the number of risk levels of the reconstructed power grid with wind power access. The number of Epochs for network training is set to 100. The update of model parameters adopts the stochastic gradient descent (SGD) optimizer, and the initial value of the learning rate is set to 0.001 and does not change in each iteration. Except for MPNN, the other 4 models are commonly used for data analysis and prediction under a fixed topological structure. Therefore, the risk assessment classification prediction performance of various methods will be analyzed from two aspects: a single topological grid structure and a changing topological grid structure.

[0085] Prediction of the risk level of the reconstructed power grid with wind power access under a single topological structure. The above five risk level prediction models of LR, RF, XGB, CNN, and MPNN are obtained through the training of dataset A. Dataset A changes the wind power access position and capacity at the fourth time step of grid reconstruction, and the topological structure of the network does not change. The performance indicators of different models on the test set under dataset A are shown in Table 2 and Figure 8 as follows.

[0086] From Table 2 and Figure 8From the results in [dataset A], it can be seen that in the test results of dataset A, the LR model has the worst risk discrimination effect. Moreover, when the number of decision trees in the random forest is too large, it will cause the calculation speed of the model to slow down. The overall accuracy of RF is 3.36% higher than that of LR. The overall accuracies of XGB and CNN have increased by more than 5% compared with the previous two comparison models. The MPNN model performs best in various indicators, and its overall accuracy has increased by 11.5% compared with LR. The evaluation indicators of the performance of the five models are generally not very different under a single topological structure. The performance indicators of the MPNN model are slightly higher than those of the other four comparison models. It can be seen that at a certain fixed time step during the network reconstruction and restoration process, the five models can all predict the risk level of the connected wind power more accurately. However, the MPNN model has the highest prediction accuracy for the risk level of the connected wind power, which also reflects that MPNN can handle graph data well under a single topological structure, capture the key information of the impact of wind power connection on the reconstructed network, and make a good judgment on whether wind power can be connected at a specific time step during the network reconstruction process.

[0087] For the risk level prediction of the reconstructed network with wind power connection under a changing topological structure, use dataset B to train the above five types of risk level prediction models, namely LR, RF, XGB, CNN, and MPNN. Dataset B contains all time steps when wind power participates in the network reconstruction, that is, all time steps except the first time step. The topological structures between different time steps have changed greatly as the system gradually recovers. Moreover, at the same time step, the connection location and capacity of wind power will also change continuously, which more severely tests the adaptability of the model to a system with a continuously changing topological structure. The performance indicators of different models on the test set under dataset B are shown in Table 3 and Figure 9 as follows.

[0088] This dataset contains all time steps when wind power participates in the network reconstruction, and the topological structure changes greatly. From Table 3 and Figure 9 it can be found that in the test results of dataset B, the performance indicators of the MPNN model are much higher than those of the other four comparison models. The performance indicators of the LR, RF, XGB, and CNN models have decreased significantly compared with the single topological structure. This is because the structures of these four models themselves cannot adapt to the changing topological structure and are suitable for a fixed topological structure. However, during the network reconstruction process of system restoration, the topological structure will change continuously. Therefore, the above four traditional models are not applicable to the risk discrimination of the reconstructed network with wind power connection. In contrast, whether it is a single topological structure or a topological structure with drastic changes, the three performance indicators of MPNN are the best and there are no large fluctuations, which fully shows that the MPNN model can effectively process the graph structure data of the topology during the power system network reconstruction process, can learn the topological structure information of the network at different time steps, and is more practical in operation scenarios with variable topologies such as network reconstruction.

[0089] Table 1 Confusion Matrix Table for Four-Class Problems

[0090]

[0091]

[0092] Table 2 Performance Index Table of Different Models under a Single Topology Structure

[0093] Model MacroF1 Acc Kappa LR 91.45 86.52 0.821 RF 93.57 89.88 0.846 XGB 97.05 95.64 0.893 CNN 97.14 95.91 0.911 MPNN 98.70 98.02 0.951

[0094] Table 3 Performance Index Table of Different Models under a Changing Topology Structure

[0095] Model MacroF1 Acc Kappa LR 62.65 59.53 0.501 RF 63.45 59.79 0.526 XGB 69.65 64.53 0.591 CNN 80.55 76.48 0.727 MPNN 97.21 96.74 0.902

[0096] Example 3

[0097] Refer to Figure 10 , which is an embodiment of the present invention, provides a risk discrimination system for large-scale wind power participating in network reconstruction based on MPNN, including: a risk division module, a generation reconstruction module, and a training discrimination module.

[0098] Among them, the risk division module is used to define a risk function and divide the risk levels of wind power access to the reconstructed network; the generation reconstruction module is used to generate a sample data set and reconstruct it into a graph data form; the training discrimination module is used to train the MPNN model and discriminate the risk levels after wind power access.

[0099] If the function is implemented in the form of a software functional unit and sold or used as an independent product, it can be stored in a computer-readable storage medium. Based on such an understanding, the technical solution of the present invention, in essence, or the part that contributes to the prior art or a part of this technical solution, can be embodied in the form of a software product. This computer software product is stored in a storage medium and includes several instructions for causing a computer device (which can be a personal computer, a server, or a network device, etc.) to execute all or part of the steps of the methods of the various embodiments of the present invention. The aforementioned storage medium includes: various media such as USB flash drives, mobile hard disks, read-only memories (ROM, Read-Only Memory), random access memories (RAM, Random Access Memory), magnetic disks, or optical discs that can store program codes.

[0100] The logic and / or steps represented in the flowchart or otherwise described herein can, for example, be considered as a definable sequence list of executable instructions for implementing logical functions, and can be embodied specifically in any computer-readable medium for use by an instruction execution system, apparatus, or device (such as a computer-based system, a system including a processor, or other systems that can fetch instructions from the instruction execution system, apparatus, or device and execute the instructions), or in connection with these instruction execution systems, apparatuses, or devices. For the purposes of this specification, a "computer-readable medium" can be any device that can contain, store, communicate, propagate, or transport a program for use by or in connection with an instruction execution system, apparatus, or device.

[0101] More specific examples (non-exhaustive list) of computer-readable media include the following: an electrical connection portion (electronic device) having one or more wirings, a portable computer diskette (magnetic device), a random access memory (RAM), a read-only memory (ROM), an erasable programmable read-only memory (EPROM or flash memory), an optical fiber device, and a portable compact disc read-only memory (CDROM). Additionally, the computer-readable medium can even be paper or other suitable media on which a program can be printed, since the program can be obtained electronically, for example, by optically scanning the paper or other media, followed by editing, interpretation, or otherwise processing as appropriate, and then stored in a computer memory.

[0102] It should be understood that various parts of the present invention can be implemented using hardware, software, firmware, or a combination thereof. In the above-described embodiments, multiple steps or methods can be implemented using software or firmware stored in a memory and executed by a suitable instruction execution system. For example, if implemented using hardware, as in another embodiment, any one or a combination of the following techniques well known in the art can be used: discrete logic circuits having logic gate circuits for implementing logical functions on data signals, application-specific integrated circuits having appropriate combinational logic gate circuits, programmable gate arrays (PGAs), field programmable gate arrays (FPGAs), etc. It should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention and not to limit them. Although the present invention has been described in detail with reference to the preferred embodiments, those of ordinary skill in the art should understand that the technical solutions of the present invention can be modified or equivalently replaced without departing from the spirit and scope of the technical solutions of the present invention, and they should all be covered within the scope of the claims of the present invention.

[0103] It should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention rather than to limit them. Although the present invention has been described in detail with reference to the preferred embodiments, those of ordinary skill in the art should understand that the technical solutions of the present invention can be modified or equivalently replaced without departing from the spirit and scope of the technical solutions of the present invention, and they should all be covered within the scope of the claims of the present invention.

Claims

1. A risk identification method for large-scale wind power participating in grid reconstruction based on MPNN is characterized by: include: Define the risk function and classify the risk level of wind power access and grid reconstruction; Generate sample data sets and reconstruct them into graph data form; Train the MPNN model to identify the risk level after wind power access.

2. The MPNN-based risk identification method for large-scale wind power participating in grid reconstruction according to claim 1, characterized in that: The risk function definition includes considering the system frequency and voltage deviation caused by wind power output fluctuations, taking into account the acceptability of transient frequency and voltage deviations, and dividing the risk level of wind power access to the grid reconstruction. During the power restoration process, the negative impact caused by wind power is the frequency deviation and voltage deviation caused by output fluctuations. The risk function of wind power participating in grid reconstruction is defined as: Where Δf max Access P at time t a The maximum system frequency deviation caused by wind power capacity, ΔU max Access P at time t a The maximum system voltage deviation caused by wind power of different capacity, g and h are frequency quality functions, and w is the weight coefficient, which indicates the importance of voltage quality compared to frequency quality.

3. The MPNN-based risk identification method for large-scale wind power participating in grid reconstruction according to claim 2, characterized in that: The risk level classification of wind power access and grid reconstruction includes: Class I risk level after wind power access and grid reconstruction, which means that the impact of wind power access on the system is very small, and the frequency and voltage fluctuations are within a safe range and can be ignored; Class II means that the fluctuations after wind power access interfere with the stability of the system, but the positive impact of wind power access outweighs the negative impact. At this time, connecting wind power speeds up the recovery process of the system and the disturbance to the system is within an acceptable range; Class III means that the negative impact after wind power access outweighs the positive impact. At this time, whether to connect wind power is comprehensively considered; Class IV means that the disturbance to the system caused by wind power access exceeds the tolerance of the grid, and wind power is not connected at this time.

4. The MPNN-based risk identification method for large-scale wind power participating in grid reconstruction according to claim 3 is characterized in that: The generating of the sample data set includes, in the process of generating the sample data set, simulating a complete grid reconstruction process by PSD-BPA, changing the access position and capacity of wind power in the topological structure at different time steps, obtaining the maximum value and duration of the system frequency and voltage fluctuations and the training set label, that is, the risk level of wind power access, the simulated data are all long vectors, the data are reconstructed into a graph data form, the node information is 0 if there is no wind power access to the node, the input of the message passing graph neural network is the information of the node and the line, the node data is a three-dimensional array, the array shape is N, M, 5), where N is the number of samples, M is the number of nodes in the system after the grid reconstruction is completed, that is, the maximum number of nodes in all samples, 5 is the feature information on each node, the line data is a four-dimensional array, the array shape is (N, M, M, 3), where the information of the M×M dimension represents the specific position of the line, 0 indicates that there is no edge between the two nodes, 1 indicates that there is an edge between the two nodes, and 3 is the feature information on each line.

5. The MPNN-based risk identification method for large-scale wind power participating in grid reconstruction according to claim 4, characterized in that: The reconstruction into graph data form includes simulating and obtaining long vector form data during sample generation, and graph data processing reconstructs the data into input graph data required by the MPNN model. The nodes in the graph data represent the busbars in the reconstructed grid, and the edges in the graph data represent the lines of the reconstructed grid in the model. The MPNN-based wind power access reconstruction grid risk discrimination model adapts to the increase or decrease of the feature dimensions of the nodes or edges in the input variables, and does not frequently adjust the model framework when the input information changes. The model uses Z-score standardization to uniformly convert data of different magnitudes in the sample into the same magnitude and is expressed as: Among them, μ is the mean of the sample data, σ is the standard deviation of the sample data, and the sample data is normalized and used as the input training model of the neural network.

6. The MPNN-based risk identification method for large-scale wind power participating in grid reconstruction according to claim 5, characterized in that: The training MPNN model includes, after building the risk discrimination model, dividing the data obtained in the graph data generation and processing process into a training set and a validation set, and inputting the training set into the MPNN model for offline training, and selecting the best performing model as the final model based on the performance in the validation set. During training, the loss function is calculated by comparing the sample label and the output risk function value, and the internal weights, biases and parameters of the model are updated through the back propagation algorithm.

7. The MPNN-based risk identification method for large-scale wind power participating in grid reconstruction according to claim 6, characterized in that: The method for determining the risk level after wind power access includes obtaining the system online operation scenario and the access location and capacity of wind power, selecting input features according to the data, namely, node information and line information, converting the data into a neural network input data format through data processing, and inputting the information into a trained risk determination model. The output is the risk level of the wind power access system at this time, and the impact of the wind power access on the power system recovery at this time is determined. For the same input, the model obtains multiple different prediction outputs through the model through random forward transfer, and uses the standard deviation of multiple predictions to quantify the uncertainty of the risk prediction of wind power access to reconstruct the grid, and takes the average prediction value of multiple predictions as the result.

8. A system using the MPNN-based risk identification method for large-scale wind power participating in grid reconstruction as claimed in any one of claims 1 to 7, characterized in that: It includes risk classification module, generation and reconstruction module, and training and discrimination module; The risk classification module is used to define a risk function and classify the risk level of wind power access to the reconstructed grid; The generation and reconstruction module is used to generate a sample data set and reconstruct it into a graph data form; The training and discrimination module is used to train the MPNN model and discriminate the risk level after wind power is connected.

9. A computer device comprising a memory and a processor, wherein the memory stores a computer program, wherein: When the processor executes the computer program, the steps of the risk identification method of large-scale wind power participating in grid reconstruction based on MPNN described in any one of claims 1 to 7 are implemented.

10. A computer-readable storage medium having a computer program stored thereon, characterized in that: When the computer program is executed by a processor, the steps of the risk identification method of large-scale wind power participating in grid reconstruction based on MPNN described in any one of claims 1 to 7 are implemented.

Images