Power grid toughness evaluation method and system based on deep learning under extreme disaster

Abstract

The invention discloses a power grid toughness evaluation method and system based on deep learning under an extreme disaster, and belongs to the field of power system toughness evaluation, and the method comprises the following steps: S1, obtaining the operation state of a power system under the extreme disaster; s2, simulating an optimal load reduction amount calculation process of the power system; s3, determining an input feature vector and an output feature vector of the deep neural network model; s4, constructing a regression and classification model based on the deep neural network in combination with the random forest; s5, training is carried out; s6, performing evaluation; and S7, constructing a power grid toughness evaluation index, and outputting the toughness level of the power system by using the toughness evaluation index based on the regression and classification model. According to the power grid toughness evaluation method and system based on deep learning under the extreme disaster, the artificial intelligence algorithm is adopted to replace the solving process of optimal load reduction calculation, the calculation process of toughness evaluation can be accelerated, disaster early warning can be responded in time, and losses caused by the disaster are reduced as much as possible.

Description

Technical Field

[0001] The present invention relates to the technical field of power system resilience assessment, and in particular to a deep learning-based power grid resilience assessment method and system under extreme disasters. Background Art

[0002] As the frequency of typhoon disasters around the world increases significantly, the impact on power systems is becoming more and more serious, highlighting the need to quickly and effectively assess the resilience of power systems in the face of extreme natural disasters.

[0003] The core of power system resilience assessment is to solve a series of optimal power flow problems by minimizing optimal load reduction. Specifically, by using optimal load reduction calculation, a direct correspondence between the power system state and the load is established. Through this mathematical model, we can better understand and predict the performance of the system when it is subjected to external shocks.

[0004] At the same time, in recent years, deep neural networks have shown good performance in solving the problem of safety-constrained DC optimal power flow. In addition, by designing the input feature vector, the key features of the topology change are extracted, further enhancing the accuracy and robustness of the model. Deep neural networks can also effectively handle the non-convergence of the optimal power flow solution, ensuring that a feasible solution is found in a complex network environment. At the same time, the classification algorithm also provides ideas for the identification of unit output in the optimal power flow calculation, which helps to achieve more efficient energy management.

[0005] Although existing research mainly focuses on using artificial intelligence methods to solve the problem of optimal power flow in order to improve the efficiency of complex calculations, the application of artificial intelligence to the optimal load reduction problem in resilience assessment is still insufficient. Summary of the invention

[0006] The purpose of the present invention is to provide a deep learning-based power grid resilience assessment method and system under extreme disasters to solve the above-mentioned technical problems.

[0007] To achieve the above object, the present invention provides a deep learning-based power grid resilience assessment method under extreme disasters, comprising the following steps:

[0008] S1. Obtain the operating status of the power system under extreme disasters;

[0009] S2, based on the operation status of the power system under extreme disasters collected in step S1, simulating the calculation process of the optimal load reduction amount of the power system;

[0010] S3. Determine the input feature vector and the output feature vector of the deep neural network model according to the optimal load reduction amount described in step S2;

[0011] S4. Build regression and classification models based on deep neural networks combined with random forests;

[0012] S5, train and test the regression and classification models, update the network parameters, and output the regression and classification models when the end conditions are met;

[0013] S6, evaluate the regression and classification model, if the threshold is met, output the regression and classification model, otherwise return to step S5;

[0014] S7. Construct grid resilience assessment indicators and use the resilience assessment indicators to output the resilience level of the power system based on the updated regression and classification models.

[0015] Preferably, the operating state of the power system under extreme disasters described in step S1 includes the topological state of the power system, the system load and the node generator power.

[0016] Preferably, step S2 specifically includes the following steps:

[0017] S21, taking the minimum load reduction of node generators as the objective function, and taking the line flow constraint and power balance constraint as the constraint conditions, the optimal load reduction amount LC is calculated by Matpower toolbox;

[0018] The objective function expression is as follows:

[0019]

[0020] In the formula, LC i represents the load reduction of node i; N node represents the number of nodes in the power system;

[0021] The constraint expressions are as follows:

[0022]

[0023] V min,i ≤V i ≤V max,i (5)

[0024] 0≤LC i ≤P D,i (6)

[0025] In the formula, θ ij is the phase angle difference between node i and node j; P G,i represents the active power of the generator at node i when the power system is in a fault state; Q G,i is the reactive power of the generator at node i when the power system is in a fault state; P D,i is the active load of node i in the power system; P F,krepresents the transmission power of line k in the power system under fault conditions; P W,i Represents the active power of the wind turbine at node i; V i and V j Represent the voltage amplitude of node i and node j respectively; G ij represents the conductance between node i and node j; B ij represents the susceptance matrix between node i and node j; Q D,i represents the reactive power of node i; P F,ij represents the transmission power between node i and node j; P Fmin,k and P Fmax,k They represent the minimum transmission power and maximum transmission power of line k in the power system under fault conditions; P Gminx,i and P Gmax,i Respectively represent; Q Gmin,i and Q Gmax,i They represent the minimum reactive power and maximum reactive power of the generator at node i when the power system is in a fault state; P Wmin,i and P Wmax,i Respectively represent the minimum active power and maximum active power of the wind turbine generator at node i; V min,i and V max,i Respectively represent the maximum voltage amplitude and the minimum voltage amplitude of node i;

[0026] S22. Save the injected active power of the node, the self-susceptance matrix, the total load reduction of the power system, and the label value of whether the load reduction occurs. The expression of the injected active power of the node is as follows:

[0027] P inj =P Gi +P wi -P Di (7)

[0028] Where P inj Indicates the node injection power.

[0029] Preferably, the input feature vector of the deep neural network model described in step S3 includes an input feature vector representing topological changes, an input feature vector representing new energy output and load fluctuation uncertainty, wherein the input feature vector representing topological changes is the node self-susceptance matrix B ii , the input feature vector representing the uncertainty of new energy output and load fluctuation is the node injection power P inj ;

[0030] The output feature vector of the deep neural network model includes the regression result Y R And the classification result Y C , where the regression result Y Ris the load loss of the power system to characterize the impact on the power system, and the classification result Y C Whether load reduction occurs to characterize the power system status;

[0031] And the eigenvector expression is as follows:

[0032]

[0033] X in-B =[B ii ] (9)

[0034] Y R =[LC] (10)

[0035] Y C =[S] (11)

[0036] In the formula, X is the input feature vector that represents the uncertainty of renewable energy output and load fluctuation; in-B is the input feature vector that characterizes the topology change; S is the label value. When S=0, it indicates that there is no load shedding in the power system; when S=1, it indicates that there is load shedding in the power system.

[0037] Preferably, in step S4, the regression model of the regression and classification model is a deep neural network model, and the structure of the deep neural network model is determined according to the characteristics of the optimal load reduction calculation problem;

[0038] The classification model of the regression and classification models is a random forest model, and the number of branches and leaves of the random forest model is determined according to the characteristics of the optimal load reduction calculation problem.

[0039] Preferably, the deep neural network model includes a normalization layer, four fully connected layers and an output layer connected in sequence, the first three fully connected layers use a ReLU activation function, and the deep neural network model expression is as follows:

[0040]

[0041] Where, X k is the input feature vector; Y represents the output feature vector; h l 、h l-2 …h 1 Respectively represent the forward propagation activation functions of the lth, l-2…1th layers; σ represents the activation function; W l represents the weight matrix of the lth layer; B l represents the bias matrix of the lth layer;

[0042] The activation function σ is expressed as follows:

[0043] σ(x)=max(0,x) (13)

[0044] Where x represents the input matrix, which includes the node injection power and the node susceptance matrix after batch normalization;

[0045] The deep neural network model uses the mean square error function as the loss function, which is expressed as follows:

[0046]

[0047] Where, L H (y,y pred ) represents the optimization target; D represents the dimension of the predicted value and the true value; y represents the true value of the training; y pred represents the predicted value; y d represents the dth true value;

[0048] Each decision tree in the random forest model performs a perceptual learning vote on the labeled samples, and the output of the random forest classifier is as follows:

[0049]

[0050] In the formula, f RF (a) represents the prediction result of the random forest model for input sample a; f n represents a single decision tree classification model; N represents the number of decision trees included in the random forest model; m = 0 or 1, indicating two categories in the classification problem.

[0051] Preferably, in the model training phase described in step S5, the regression model and the classification model are trained respectively using the same data, and in the model testing phase, the results of the regression model and the classification model are mutually corrected, and the optimal load reduction amount is output;

[0052] The network parameters described in step S5 include weight values ​​and bias values ​​of the regression and classification models, and the training termination condition of the regression and classification models is that the number of iterations reaches a set value or satisfies the early stopping method.

[0053] Preferably, in step S6, the evaluation indicators of the regression and classification models include performance indicators and overall goals, wherein the performance indicator is the proportion of correct classification results A cc , which is expressed as follows:

[0054] A cc =N correct / N total ×100% (16)

[0055] Where N correct Represents the number of samples with correct classification results; N total Indicates the total number of samples for classification;

[0056] The overall goal is the relative error δ in the toughness assessment result, which is expressed as follows:

[0057] δ=|LC real -LC pred | / LC real ×100% (17)

[0058] In the formula, LC real represents the toughness evaluation index obtained by using the optimal trend calculation, LC pred represents the resilience assessment index obtained by using the prediction.

[0059] Preferably, in step S7, the grid resilience evaluation index is defined as the expected value of load reduction in each system state, the probability of power system failure is calculated based on weather forecast information, and the optimal load reduction amount is obtained by regression and classification models;

[0060] Among them, the expected value of load reduction R sys The expression is as follows:

[0061]

[0062] Where P g represents the probability of a failure state caused by a single disaster, which is obtained from weather forecast information; LC g represents the load reduction value under the fault state g caused by the disaster; Ω represents the set of fault states caused by the disaster, which is obtained by the regression and classification model.

[0063] A power grid resilience assessment system based on deep learning under extreme disasters, used to execute a power grid resilience assessment method based on deep learning under extreme disasters, comprising a data acquisition and processing module, an update module, an assessment module and an output module;

[0064] Among them, the data acquisition and processing module is used to obtain the operating status of the power system under extreme disasters, simulate the calculation process of the optimal load reduction of the power system, and determine the input feature vectors and output feature vectors of the regression and classification models and the deep neural network model;

[0065] The update module is used to adjust the parameters of the deep neural network model according to the loss function training results of the deep neural network model;

[0066] The evaluation module is used to evaluate the resilience of the power system under different topological states, new energy sources and load levels;

[0067] The output module is used to output the toughness assessment results.

[0068] Therefore, the present invention adopts the above-mentioned deep learning-based power grid resilience assessment method and system under extreme disasters, which has the following beneficial effects:

[0069] 1. Using neural networks to replace the optimal load reduction calculation process in resilience assessment solves the problem of slow resilience assessment calculation speed in power grids with a high proportion of new energy;

[0070] 2. Combining deep neural networks with random forests to perform regression and classification tasks and achieve mutual correction between the two methods will help to discover potential problems in the power grid in advance, thereby improving the reliability and stability of the power grid and ensuring that people’s lives and work are not affected by power outages.

[0071] In summary, the present invention studies the relationship between input and output characteristics in the optimal load reduction calculation through a neural network, effectively replaces the optimal load reduction model, and achieves the purpose of quickly calculating the system status. The random forest algorithm is then used to perform a secondary check on the load reduction calculation results to ensure the accuracy of the calculation. The rapid resilience assessment can help system operators make timely decisions and effectively reduce the impact of potential power grid failures. Especially in the face of emergencies such as natural disasters, rapid and accurate resilience assessment can help predict the system's response and recovery capabilities, thereby ensuring the stability and reliability of the power system.

[0072] The technical solution of the present invention is further described in detail below through the accompanying drawings and embodiments. BRIEF DESCRIPTION OF THE DRAWINGS

[0073] Figure 1 This is a flow chart of a method for evaluating power grid resilience based on deep learning under extreme disasters according to the present invention;

[0074] Figure 2 This is the topology of the IEEE RTS 79 system described in the simulation experiment. DETAILED DESCRIPTION

[0075] In order to make the purpose, technical scheme and advantages disclosed in the embodiments of the present invention clearer, the embodiments of the present invention are further described in detail in combination with the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are only used to explain the embodiments of the present invention and are not used to limit the embodiments of the present invention. Based on the embodiments in the present invention, all other embodiments obtained by ordinary technicians in this field without making creative work are within the scope of protection of this application. Examples of the embodiments are shown in the accompanying drawings, where the same or similar reference numerals throughout represent the same or similar elements or elements with the same or similar functions.

[0076] It should be noted that the terms "including" and "having" and any variations thereof are intended to cover non-exclusive inclusions. For example, a process, method, system, product or server that includes a series of steps or units is not necessarily limited to those steps or units explicitly listed, but may include other steps or units that are not explicitly listed or inherent to these processes, methods, products or devices.

[0077] The embodiments of the present invention are described in detail below in conjunction with the accompanying drawings.

[0078] like Figure 1 As shown, a deep learning-based power grid resilience assessment method under extreme disasters includes the following steps:

[0079] S1. Obtain the operating status of the power system under extreme disasters;

[0080] The operating state of the power system under extreme disasters described in step S1 includes the topological state of the power system, the system load and the node generator power.

[0081] S2, based on the operation status of the power system under extreme disasters collected in step S1, simulating the calculation process of the optimal load reduction amount of the power system;

[0082] Step S2 specifically includes the following steps:

[0083] S21, taking the minimum load reduction of node generators as the objective function, and taking the line flow constraint and power balance constraint as the constraint conditions, the optimal load reduction amount LC is calculated by Matpower toolbox;

[0084] The objective function expression is as follows:

[0085]

[0086] In the formula, LC i represents the load reduction of node i; N node represents the number of nodes in the power system;

[0087] The constraint expressions are as follows:

[0088]

[0089] V min,i ≤V i ≤V max,i (5)

[0090] 0≤LC i ≤P D,i (6)

[0091] In the formula, θ ij is the phase angle difference between node i and node j; PG,i represents the active power of the generator at node i when the power system is in a fault state; Q G,i is the reactive power of the generator at node i when the power system is in a fault state; P D,i is the active load of node i in the power system; P F,k represents the transmission power of line k in the power system under fault conditions; P W,i Represents the active power of the wind turbine at node i; V i and V j Represent the voltage amplitude of node i and node j respectively; G ij represents the conductance between node i and node j; B ij represents the susceptance matrix between node i and node j; Q D,i represents the reactive power of node i; P F,ij represents the transmission power between node i and node j; P Fmin,k and P Fmax,k They represent the minimum transmission power and maximum transmission power of line k in the power system under fault conditions; P Gminx,i and P Gmax,i Respectively represent; Q Gmin,i and Q Gmax,i They represent the minimum reactive power and maximum reactive power of the generator at node i when the power system is in a fault state; P Wmin,i and P Wmax,i Respectively represent the minimum active power and maximum active power of the wind turbine generator at node i; V min,i and V max,i Respectively represent the maximum voltage amplitude and the minimum voltage amplitude of node i;

[0092] S22. Save the injected active power of the node, the self-susceptance matrix, the total load reduction of the power system, and the label value of whether the load reduction occurs. The expression of the injected active power of the node is as follows:

[0093] P inj =P Gi +P wi -P Di (7)

[0094] Where P inj Indicates the node injection power.

[0095] S3. Determine the input feature vector and the output feature vector of the deep neural network model according to the optimal load reduction amount described in step S2;

[0096] A high proportion of renewable energy power grids not only faces various uncertainties, but also has high-dimensional complex nonlinear characteristics, including: (1) nonlinear characteristics between continuous inputs such as renewable energy output and load and system state; (2) nonlinear characteristics between discrete inputs of system topology and system state. w Situation and load demand P Di The node injection power dimension is small. At the same time, compared with the changes in new energy and load, the change in topology structure has a greater impact on the system state. The node self-susceptance matrix after the topology change can well reflect the system changes. Therefore, the node injection power is constructed as a feature vector to characterize the change in new energy output and load, and the node self-susceptance matrix is ​​constructed as a feature vector to characterize the topology change. The system load loss is selected to characterize the system impact as the output feature vector of the regression model.

[0097] Therefore, the input feature vector of the deep neural network model described in step S3 includes an input feature vector representing topological changes, an input feature vector representing the output of new energy and the uncertainty of load fluctuations, wherein the input feature vector representing topological changes is the node self-susceptance matrix B ii , the input feature vector representing the uncertainty of new energy output and load fluctuation is the node injection power P inj ;

[0098] The output feature vector of the deep neural network model includes the regression result Y R And the classification result Y C , where the regression result Y R is the load loss of the power system to characterize the impact on the power system, and the classification result Y C Whether load reduction occurs to characterize the power system status;

[0099] And the eigenvector expression is as follows:

[0100]

[0101] X in-B =[B ii ] (9)

[0102] Y R =[LC] (10)

[0103] Y C =[S] (11)

[0104] In the formula, X is the input feature vector that represents the uncertainty of renewable energy output and load fluctuation; in-Bis the input feature vector that characterizes the topology change; S is the label value. When S=0, it indicates that there is no load shedding in the power system; when S=1, it indicates that there is load shedding in the power system.

[0105] S4. Construct a regression and classification model based on a deep neural network combined with a random forest, and the goal of the regression and classification model is to establish a mapping relationship between the output feature optimal load reduction amount and the input feature power system state;

[0106] In step S4, the regression model of the regression and classification model is a deep neural network model, and the structure of the deep neural network model is determined according to the characteristics of the optimal load reduction calculation problem; according to the universal approximation principle, the deep neural network can approximate any continuous function with arbitrary precision, so the neural network is selected to mine the optimal load reduction analysis characteristics.

[0107] The deep neural network model includes a standardized layer connected in sequence, four fully connected layers and an output layer. The first three fully connected layers use the ReLU activation function (ReLU helps alleviate the problem of gradient disappearance during neural network training). The deep neural network model expression is as follows:

[0108]

[0109] Where, X k is the input feature vector; Y represents the output feature vector; h l 、h l-2 …h 1 Respectively represent the forward propagation activation functions of the lth, l-2…1th layers; σ represents the activation function; W l represents the weight matrix of the lth layer; B l represents the bias matrix of the lth layer;

[0110] The activation function σ is expressed as follows:

[0111] σ(x)=max(0,x) (13)

[0112] Where x represents the input matrix, which includes the node injection power and the node susceptance matrix after batch normalization;

[0113] The deep neural network model uses the mean square error function as the loss function to intuitively reflect the fitting effect of the deep neural network. Its expression is as follows:

[0114]

[0115] Where, L H (y,y pred ) represents the optimization target; D represents the dimension of the predicted value and the true value; y represents the true value of the training; y predrepresents the predicted value; y d represents the dth true value;

[0116] Due to the stability of the power system, the load reduction value in most system states is 0, and only a few system states have load reduction values ​​greater than 0. Therefore, the system samples that need to be evaluated in the evaluation analysis have obvious unbalanced characteristics, so the random forest classification model is introduced. According to the simplified model of the random forest, the voting results of multiple decision trees are queried, which has high classification accuracy and speed. The classification model of the regression and classification model is the random forest model. The number of branches and leaves of the random forest model is determined according to the characteristics of the optimal load reduction calculation problem.

[0117] Each decision tree in the random forest model performs a perceptual learning vote on the labeled samples, and the output of the random forest classifier is as follows:

[0118]

[0119] In the formula, f RF (a) represents the prediction result of the random forest model for input sample a; f n represents a single decision tree classification model; N represents the number of decision trees included in the random forest model; m = 0 or 1, indicating two categories in the classification problem.

[0120] S5, train and test the regression and classification models, update the network parameters, and output the regression and classification models when the end conditions are met;

[0121] In the model training phase described in step S5, the regression model and the classification model are trained respectively using the same data, and in the model testing phase, the results of the regression model and the classification model are mutually corrected, and the optimal load reduction amount is output;

[0122] The network parameters described in step S5 include weight values ​​and bias values ​​of the regression and classification models, and the training termination condition of the regression and classification models is that the number of iterations reaches a set value or satisfies the early stopping method.

[0123] S6, evaluate the regression and classification model, if the threshold is met, output the regression and classification model, otherwise return to step S5;

[0124] In step S6, the evaluation indicators of the regression and classification models include performance indicators and overall goals, where the performance indicator is the proportion of correct classification results A cc , which is expressed as follows:

[0125] A cc =N correct / N total ×100% (16)

[0126] Where Ncorrect Represents the number of samples with correct classification results; N total Indicates the total number of samples for classification;

[0127] The overall goal is the relative error δ in the toughness assessment result, which is expressed as follows:

[0128] δ=|LC real -LC pred | / LC real ×100% (17)

[0129] In the formula, LC real represents the toughness evaluation index obtained by using the optimal trend calculation, LC pred represents the resilience assessment index obtained by using the prediction.

[0130] S7. Construct grid resilience assessment indicators and use the resilience assessment indicators to output the resilience level of the power system based on the updated regression and classification models.

[0131] In step S7, the grid resilience evaluation index is defined as the expected value of load reduction in each system state, the probability of power system failure is calculated based on weather forecast information, and the optimal load reduction amount is obtained by regression and classification models;

[0132] Among them, the expected value of load reduction R sys The expression is as follows:

[0133]

[0134] Where P g represents the probability of a failure state caused by a single disaster, which is obtained from weather forecast information; LC g represents the load reduction value under the fault state g caused by the disaster; Ω represents the set of fault states caused by the disaster, which is obtained by the regression and classification model.

[0135] A power grid resilience assessment system based on deep learning under extreme disasters, used to execute a power grid resilience assessment method based on deep learning under extreme disasters, comprising a data acquisition and processing module, an update module, an assessment module and an output module;

[0136] Among them, the data acquisition and processing module is used to obtain the operating status of the power system under extreme disasters, simulate the calculation process of the optimal load reduction of the power system, and determine the input feature vectors and output feature vectors of the regression and classification models and the deep neural network model; the update module is used to adjust the parameters of the deep neural network model according to the loss function training results of the deep neural network model; the evaluation module is used to evaluate the resilience level of the power system under different topological states, new energy and load levels; the output module is used to output the resilience evaluation results.

[0137] Simulation test

[0138] The hardware environment for the simulation is Intel(R)Core(TM)i7-13700H CPU@2.4GHz 16GB RAM. Figure 2 The IEEE RTS79 node system shown in the figure is used as an example for simulation analysis and verification. Specifically, the IEEE RTS 79 system includes 32 generators, 33 transmission lines, 5 transformers, 17 load nodes, with a total installed capacity of 3405MW and a maximum system load of 2850MW. In the IEEE RTS 79 system, for the uncertainty of the load, it is assumed that the actual power of the bus load is evenly distributed within the reference value [0.8, 1.2]. Excel is used to generate random numbers, and the active power of the 24 node loads is evenly distributed. Considering the computer calculation burden, 10 groups of loads are sampled in the IEEE RTS 79 system to obtain the current load reduction. On this basis, the uncertainty of new energy is added, the 16-node conventional units are modified to wind turbines, and 10 output conditions are randomly selected.

[0139] In order to introduce changes in the topology of the line fault simulation system, a non-sequential Monte Carlo method was used to generate 500,000 samples for training and testing. In order to speed up the convergence of the algorithm, the RMSprop algorithm (Root Mean Square Propagation) was selected to update the model parameters. Regarding the parameter tuning of the deep neural network, a grid search was performed to optimize the selected deep neural network model hyperparameters. Specifically, the hyperparameters were set to a total of 27 scenarios with learning rate [0.01, 0.001, 0.005], batch number [64, 128, 256], and maximum number of iterations [500, 1000, 1500]. The best hyperparameters with the minimum loss function were selected.

[0140] Meanwhile, for the random forest parameters, the same grid was searched, including the number of parameterized decision trees [100, 300, 500] and the number of leaf nodes in the random forest [3, 5, 10]. The combination with the highest classification accuracy was selected.

[0141] During the model training and testing phases, the training samples were randomly divided into training and test sets in a ratio of 8:2, and the K threshold was set to 10MW. The completed training model obtained by the MCS method (Monte Carlo method) was used to evaluate the resilience of potential scenarios.

[0142] Table 1 Toughness evaluation results

[0143] method Resilience index Regression Accuracy time Classification Accuracy M0 6.035 - 8643.5s 98.63％ M1 5.828 3.415％ 1.64s 98.65％ M2 4.435 26.512％ 1.53s 98.32％ M3 5.879 2.571％ 2.33s 98.75％

[0144] It should be noted that in Table 1, M0 means using optimal power flow calculation for resilience assessment; M1 means using only regression model for resilience assessment; M2 means using mutually corrected regression classification model for resilience assessment, but deleting topological information in the feature vector; M3 means using the mutually corrected regression classification model described in the present invention for resilience assessment.

[0145] As can be seen from Table 1, by comparing M0 with M1, it can be seen that the accuracy of resilience assessment using a regression classification model is higher than that of using only a regression model. That is, in the face of scenarios with topological uncertainty and renewable energy output uncertainty, the deep learning method is significantly superior to the traditional method in terms of resilience assessment time in the test set. Comparing M2 with M3 proves the key importance of integrating topological information into the input features, and the accuracy of the random forest classification model of M3 is between 98% and 99%, which meets the requirements of classification model accuracy, thereby proving that the method proposed in the present invention meets the preset requirements in terms of calculation accuracy and calculation time.

[0146] Finally, it should be noted that the above embodiments are only used to illustrate the technical solution of the present invention rather than to limit it. Although the present invention has been described in detail with reference to the preferred embodiments, those skilled in the art should understand that they can still modify or replace the technical solution of the present invention with equivalents, and these modifications or equivalent replacements cannot cause the modified technical solution to deviate from the spirit and scope of the technical solution of the present invention.

Claims

1. A deep learning-based power grid resilience assessment method under extreme disasters, characterized by: The following steps are involved: S1. Obtain the operating status of the power system under extreme disasters; S2, based on the operation status of the power system under extreme disasters collected in step S1, simulating the calculation process of the optimal load reduction amount of the power system; S3. Determine the input feature vector and the output feature vector of the deep neural network model according to the optimal load reduction amount described in step S2; S4. Build regression and classification models based on deep neural networks combined with random forests; S5, train and test the regression and classification models, update the network parameters, and output the regression and classification models when the end conditions are met; S6, evaluate the regression and classification model, if the threshold is met, output the regression and classification model, otherwise return to step S5; S7. Construct grid resilience assessment indicators and use the resilience assessment indicators to output the resilience level of the power system based on the updated regression and classification models.

2. The method for evaluating power grid resilience under extreme disasters based on deep learning according to claim 1, characterized in that: The operating state of the power system under extreme disasters described in step S1 includes the topological state of the power system, the system load and the node generator power.

3. The method for evaluating power grid resilience under extreme disasters based on deep learning according to claim 2, characterized in that: Step S2 specifically includes the following steps: S21, taking the minimum load reduction of node generators as the objective function, and taking the line flow constraint and power balance constraint as the constraint conditions, the optimal load reduction amount LC is calculated by Matpower toolbox; The objective function expression is as follows: In the formula, LC i represents the load reduction of node i; N node represents the number of nodes in the power system; The constraint expressions are as follows: In min,i ≤V i ≤V max,i (5) 0≤LC i ≤P D,i (6) In the formula, θ ij is the phase angle difference between node i and node j; P G,i represents the active power of the generator at node i when the power system is in a fault state; Q G,i is the reactive power of the generator at node i when the power system is in a fault state; P D,i is the active load of node i in the power system; P F,k represents the transmission power of line k in the power system under fault conditions; P W,i Represents the active power of the wind turbine at node i; V i and V j Represent the voltage amplitude of node i and node j respectively; G ij represents the conductance between node i and node j; B ij represents the susceptance matrix between node i and node j; Q D,i represents the reactive power of node i; P F,ij represents the transmission power between node i and node j; P Fmin,k and P Fmax,k They represent the minimum transmission power and maximum transmission power of line k in the power system under fault conditions; P Gminx,i and P Gmax,i Respectively represent; Q Gmin,i and Q Gmax,i They represent the minimum reactive power and maximum reactive power of the generator at node i when the power system is in a fault state; P Wmin,i and P Wmax,i Respectively represent the minimum active power and maximum active power of the wind turbine generator at node i; V min,i and V max,i Respectively represent the maximum voltage amplitude and the minimum voltage amplitude of node i; S22. Save the injected active power of the node, the self-susceptance matrix, the total load reduction of the power system, and the label value of whether the load reduction occurs. The expression of the injected active power of the node is as follows: P inj =P Gi +P wi -P Di (7) Where P inj Indicates the node injection power.

4. The method for evaluating power grid resilience under extreme disasters based on deep learning according to claim 3 is characterized in that: The input feature vector of the deep neural network model described in step S3 includes an input feature vector representing topological changes, an input feature vector representing the output of new energy and the uncertainty of load fluctuations, wherein the input feature vector representing topological changes is the node self-susceptance matrix B ii , the input feature vector representing the uncertainty of new energy output and load fluctuation is the node injection power P inj ; The output feature vector of the deep neural network model includes the regression result Y R And the classification result Y C , where the regression result Y R is the load loss of the power system to characterize the impact on the power system, and the classification result Y C Whether load reduction occurs to characterize the power system status; And the eigenvector expression is as follows: X in-B =[B ii ] (9) Y R =[LC] (10) AND C =[S] (11) In the formula, X is the input feature vector that represents the uncertainty of renewable energy output and load fluctuation; in-B is the input feature vector that characterizes the topology change; S is the label value. When S=0, it indicates that there is no load shedding in the power system; when S=1, it indicates that there is load shedding in the power system.

5. The method for evaluating power grid resilience under extreme disasters based on deep learning according to claim 4, characterized in that: In step S4, the regression model of the regression and classification model is a deep neural network model, and the structure of the deep neural network model is determined according to the characteristics of the optimal load reduction calculation problem; The classification model of the regression and classification models is a random forest model, and the number of branches and leaves of the random forest model is determined according to the characteristics of the optimal load reduction calculation problem.

6. The method for evaluating power grid resilience under extreme disasters based on deep learning according to claim 5, characterized in that: The deep neural network model includes a sequentially connected normalization layer, four fully connected layers and an output layer. The first three fully connected layers use the ReLU activation function. The deep neural network model expression is as follows: Where, X k is the input feature vector; Y represents the output feature vector; h l 、h l-2 …h 1 Respectively represent the forward propagation activation functions of the lth, l-2…1th layers; σ represents the activation function; W l represents the weight matrix of the lth layer; B l represents the bias matrix of the lth layer; The activation function σ is expressed as follows: σ(x)=max(0,x) (13) Where x represents the input matrix, which includes the node injection power and the node susceptance matrix after batch normalization; The deep neural network model uses the mean square error function as the loss function, which is expressed as follows: Where, L H (y,y pred ) represents the optimization target; D represents the dimension of the predicted value and the true value; y represents the true value of the training; y pred represents the predicted value; y d represents the dth true value; Each decision tree in the random forest model performs a perceptual learning vote on the labeled samples, and the output of the random forest classifier is as follows: In the formula, f RF (a) represents the prediction result of the random forest model for input sample a; f n Represents a single decision tree classification model; N represents the number of decision trees included in the random forest model; m = 0 or 1, indicating two categories in the classification problem.

7. The method for evaluating power grid resilience under extreme disasters based on deep learning according to claim 6, characterized in that: In the model training phase described in step S5, the regression model and the classification model are trained respectively using the same data, and in the model testing phase, the results of the regression model and the classification model are mutually corrected, and the optimal load reduction amount is output; The network parameters described in step S5 include weight values ​​and bias values ​​of the regression and classification models, and the training termination condition of the regression and classification models is that the number of iterations reaches a set value or satisfies the early stopping method.

8. The method for evaluating power grid resilience under extreme disasters based on deep learning according to claim 7, characterized in that: In step S6, the evaluation indicators of the regression and classification models include performance indicators and overall goals, where the performance indicator is the proportion of correct classification results A cc , which is expressed as follows: A cc =N correct / N total ×100% (16) Where N correct Represents the number of samples with correct classification results; N total Indicates the total number of samples for classification; The overall goal is the relative error δ in the toughness assessment result, which is expressed as follows: δ=|LC real -LC pred | / LC real ×100% (17) In the formula, LC real represents the toughness evaluation index obtained by using the optimal trend calculation, LC pred represents the resilience assessment index obtained by using the prediction.

9. The method for evaluating power grid resilience under extreme disasters based on deep learning according to claim 8, characterized in that: In step S7, the grid resilience evaluation index is defined as the expected value of load reduction in each system state, the probability of power system failure is calculated based on weather forecast information, and the optimal load reduction amount is obtained by regression and classification models; Among them, the expected value of load reduction R sys The expression is as follows: Where P g represents the probability of a failure state caused by a single disaster, which is obtained from weather forecast information; LC g represents the load reduction value under the fault state g caused by the disaster; Ω represents the set of fault states caused by the disaster, which is obtained by the regression and classification model.

10. A deep learning-based power grid resilience assessment system under extreme disasters, characterized by: Used to execute a deep learning-based power grid resilience assessment method under extreme disasters as described in any one of claims 1 to 9 above, comprising a data acquisition and processing module, an update module, an assessment module and an output module; Among them, the data acquisition and processing module is used to obtain the operating status of the power system under extreme disasters, simulate the calculation process of the optimal load reduction of the power system, and determine the input feature vectors and output feature vectors of the regression and classification models and the deep neural network model; The update module is used to adjust the parameters of the deep neural network model according to the loss function training results of the deep neural network model; The evaluation module is used to evaluate the resilience of the power system under different topological states, new energy sources and load levels; The output module is used to output the toughness assessment results.

Images