CN110929989A

CN110929989A - N-1 safety checking method with uncertainty based on deep learning

Info

Publication number: CN110929989A
Application number: CN201911040124.1A
Authority: CN
Inventors: 杨知方; 杨燕; 余娟; 雷江龙; 余红欣; 李�杰
Original assignee: Chongqing University; Electric Power Research Institute of State Grid Chongqing Electric Power Co Ltd; State Grid Corp of China SGCC
Current assignee: Chongqing University; Electric Power Research Institute of State Grid Chongqing Electric Power Co Ltd; State Grid Corp of China SGCC
Priority date: 2019-10-29
Filing date: 2019-10-29
Publication date: 2020-03-27
Anticipated expiration: 2039-10-29
Also published as: CN110929989B

Abstract

The invention discloses a deep learning-based safety checking method containing uncertainty N-1, which mainly comprises the following steps: 1) acquiring power network data and establishing an input feature vector set X_in(ii) a 2) Preprocessing the characteristic vector, and dividing the characteristic vector into a training set and a verification set; 3) establishing a deep neural network direct current power flow model; 4) training a direct current power flow model of the deep neural network by utilizing a training set and a verification set; 5) establishing a test set by using real-time data of the power network, and inputting the test set into the trained deep neural network direct current power flow model to obtain a power flow output characteristic vector Y_out(ii) a 6) For output feature vector Y_out＝[θ_i,P_ij]And carrying out security verification. The invention canThe method is widely applied to the N-1 security check online analysis of the uncertain scene, can obtain check precision comparable to that obtained by a traditional direct current power flow solving method, and improves the analysis speed by about one hundred times.

Description

N-1 safety checking method with uncertainty based on deep learning

Technical Field

The invention relates to the field of electric power systems and automation thereof, in particular to a safety checking method containing uncertainty N-1 based on deep learning.

Background

Power systems essentially operate in an uncertain environment. The N-1 security check is an important analysis tool for ensuring the system security. In recent years, as the permeability of renewable energy sources such as photovoltaic energy, wind energy and the like is higher and higher, and new energy sources such as wind energy and the like are random and intermittent, large-scale access of the new energy sources brings more uncertainty to a power system. In order to cope with the influence of increasing uncertainty on the operation of the power system, the N-1 security check needs to consider uncertainty scenes such as new energy fluctuation besides branch disconnection, so that the N-1 security check faces new calculation challenges.

At present, the N-1 security check generally adopts direct current power flow model calculation. The direct current power flow model well avoids the problems of calculation non-convergence, increased complexity of models such as economic dispatching and unit combination and the like caused by the alternating current power flow model, but for N-1 security check needing to consider new energy uncertainty scenes, N-10M times of power flow calculation are needed for N-1 security check on the assumption that M wind power plants need to be considered and each wind power plant has 10 uncertainty scenes, and the calculation burden of the N-1 security check is exponentially increased compared with that of the traditional N-1 security check. Therefore, researchers are constantly seeking improved methods to increase the speed of N-1 security checks.

At present, research aiming at accelerating the N-1 security check speed is mainly carried out scene reduction in a model derivation mode. Some researches reduce the number of scenes by establishing an optimization model, and mainly comprise establishing a mixed integer linear programming and two-stage programming model, reducing the number of checking scenes by a three-stage filtering method, reducing the scenes by using an iteration boundary and continuous pruning method, accelerating the judgment of the power grid connectivity by using a GPU and the like. Under the conditions of considering the uncertainty of new energy, the multi-period problem of the power system and the like, although the method can effectively reduce the number of scenes and accelerate the checking speed of a single scene, the number of scenes needing to be checked in the N-1 security analysis is still huge, and the calculation efficiency of the method still needs to be further improved. In summary, it is necessary to develop an efficient N-1 security check method.

Disclosure of Invention

The present invention is directed to solving the problems of the prior art.

The technical scheme adopted for realizing the purpose of the invention is that the N-1 safety checking method with uncertainty based on deep learning mainly comprises the following steps:

1) acquiring power network data and establishing an input feature vector set X_in。

Further, the power network data includes a power network topology and node source load data.

Further, a feature vector set X is input_in＝[P_i,△P_ij]And outputting the feature vector Y_out＝[θ_i,P_ij]. Wherein, P_iAnd representing the sum of the active power injected by the new energy node for continuous eigenvectors. Theta_iIs a nodal phase angle, P_ij△ P being the active power flow of each branch_ijThe discrete characteristic vector represents the difference of active power of each branch before and after the branch is disconnected.

Active power difference △ P of branches before and after branch disconnection_ijAs follows:

in the formula (I), the compound is shown in the specification,

the active power of each branch circuit before the branch circuit is cut off.

The difference △ P between the active power of each branch before and after the branch is cut off_ijHas a dimension of n_branch。n_branchThe number of system branches.

2) The feature vectors are preprocessed and divided into a training set and a validation set.

The data preprocessing method is normalization.

The normalization formula is as follows:

in the formula, x_μIs the mean value of the samples, x_δIs the sample standard deviation. x is the input sample, i.e. the feature vector. And x is normalized data. Data x mean 0 and variance 1.

3) And establishing a deep neural network direct current power flow model.

The deep neural network direct current power flow model is as follows:

in the formula (I), the compound is shown in the specification,

is the feed forward transfer function of layer i neurons. And l is 1,2, …, n and n are the number of the neural network layers. And theta is a parameter to be optimized of the deep neural network.

Feedforward transfer function for layer I neurons

As follows:

in the formula, X^l-1Is the input to layer I neurons. W_lAnd b_lAre the weight matrix and offset vector between the l-layer neurons and the l-1 layer neurons. s is the activation function.

The activation function s is as follows:

4) and training the direct current power flow model of the deep neural network by utilizing the training set and the verification set.

The method for training the deep neural network direct current power flow model mainly comprises the following steps:

4.1) inputting the training set into the deep neural network direct current power flow model.

And 4.2) randomly initializing a parameter theta to be optimized of the deep neural network direct current power flow model.

4.3) utilizing the RMSProp algorithm to update the parameter of the deep neural network for the t time, namely:

where η is the learning rate ε is a constant and r is the cumulative squared gradient._▽θLIs the partial derivative of the mean square error loss function to theta. ρ is the decay rate. And t is the iteration number. the initial value of t is 1.

4.4) inputting the verification set into the deep neural network direct current power flow model, judging whether the testing precision of the verification set is reduced, if so, stopping iteration, and if not, judging the iteration times t>t_maxAnd if the result is positive, stopping iteration, otherwise, returning to the step 2. t is t_maxIs the maximum number of iterations.

5) Establishing a test set by using real-time data of the power network, and inputting the test set into the trained deep neural network direct current power flow model to obtain a power flow output characteristic vector Y_out。

6) For output feature vector Y_out＝[θ_i,P_ij]And carrying out security verification.

For output feature vector Y_out＝[θ_i,P_ij]The method for carrying out security verification comprises the following steps: judging active power flow P of branch_ij>P_maxAnd if the judgment result is positive, judging that the branch ij is overloaded, and if the judgment result is negative, judging that the branch ij is in a safe state.

The technical effect of the present invention is undoubted. The feature vector designed by the invention can effectively cover new energy, load continuous type change features and topological structure change features of N-1 security check. The direct current power flow model obtained after sample training can effectively mine complex characteristics between input and output of the direct current power flow equation, power flow calculation can be directly carried out on all uncertain scenes to be checked, and technical support is provided for determination of overload circuits and safety design of operation schemes.

The deep learning strategy suitable for N-1 security check is designed by the invention, based on a data preprocessing method and an activation function, a z-score standardization method is adopted to carry out normalization preprocessing on a power flow sample, a ReLU activation function and a linear function are combined to be used as an activation function of a deep neural network to effectively extract power flow characteristics, an RMSProp learning algorithm is selected to train a direct current power flow model of the deep neural network, and the acquisition of an optimal parameter theta is effectively realized.

The method can be widely applied to the N-1 security check online analysis of the uncertain scene, can obtain check precision comparable to that obtained by the traditional direct current power flow solving method, and improves the analysis speed by about one hundred times.

Drawings

FIG. 1 is a DC power flow model based on a deep neural network;

FIG. 2 is a comparison of the convergence rates of M1, M2 and M3 in example 1;

FIG. 3 is a comparison of the convergence rates of M1 and M3 in example 2;

FIG. 4 shows the loss function degradation process of M1 and M5 in example 1;

FIG. 5 shows the loss function degradation process of M1 and M5 in example 2.

Detailed Description

The present invention is further illustrated by the following examples, but it should not be construed that the scope of the above-described subject matter is limited to the following examples. Various substitutions and alterations can be made without departing from the technical idea of the invention and the scope of the invention is covered by the present invention according to the common technical knowledge and the conventional means in the field.

Example 1:

referring to fig. 1, the method for safety check with uncertainty N-1 based on deep learning mainly comprises the following steps:

in the formula (I), the compound is shown in the specification,

the active power of each branch circuit before the branch circuit is cut off.

The active power of each branch is obtained after the branch is disconnected. Before branch is cut offDifference △ P between active power of rear branches_ijHas a dimension of n_branch。n_branchFor the number of system branches, △ P is calculated_ijAnd in time, the load of each node is taken as the mean value of the load, and the output of the new energy is taken as a rated value. The vector has a dimension n_branch(wherein, n_branchThe number of the branches of the system) is only increased linearly along with the increase of the scale of the system, and meanwhile, the difference of the active power of each branch between different topologies is changed, so that the problem that topology information is submerged does not exist, the influence degree of the branch on each branch is effectively described by the value of the number of the branches, and the vector effectively covers the important influence of the change of the topology structure on the power flow of the power system.

Notably, the feature vector is the material basis for the function fitting by the deep neural network. In order to establish a deep neural network direct current power flow model and effectively extract N-1 security check characteristics considering a new energy uncertainty scene, the input characteristic vector of a sample is required to cover new energy, load continuous variability characteristics and topological structure variation characteristics of N-1 security check.

Aiming at continuous type change characteristics, the invention selects the sum of active power injected by nodes of each continuous type variable (new energy, load and the like) as an input characteristic vector. The node injection power can effectively reflect the fluctuation of new energy and load, and the dimensionality of the node injection power is only the number of nodes of the system.

For discrete variation characterization, the conventional method of representing topology includes admittance matrix and 0-1 vector representing branch state (branch open is represented by 0 and branch closed is represented by 1). The admittance matrix can reflect the branch disconnection condition and the node incidence relation, however, when the admittance matrix is taken as the characteristic vector, the dimension of the admittance matrix increases in a square mode along with the increase of the system scale. Although the dimension of the 0-1 vector representing the branch state only grows linearly with increasing system size, it can only reflect the open condition of each branch. In addition, the two methods have the problem that topological information is submerged, so that the deep neural network is difficult to effectively extract the important influence of the branch disconnection on the power flow of the power system. In view of the above, the invention proposes the active power of each branch before and after the branch is cut offPower difference △ P_ijAs feature vectors representing topological structure

The data preprocessing method is normalization.

And considering the topological structure change caused by the broken line of the branch, and a small number of data points deviating from the sample mean value exist in the power flow sample. For example, when a branch is disconnected, the active power of the branch is 0MW, and in most cases, the active power of the branch is still around the mean value. The min-max method is normalized using the minimum and maximum values in the sample, and is susceptible to small amounts of deviating data points because the deviating data points affect the max or min values. The z-score method performs normalization by using sample global information, i.e. sample mean and sample standard deviation, is less affected by deviated data points, and can better save data distribution characteristics while eliminating numerical problems, so the following normalization methods are selected in the invention:

3) And establishing a deep neural network direct current power flow model.

The deep neural network direct current power flow model is as follows:

in the formula (I), the compound is shown in the specification,

is the feed forward transfer function of layer i neurons. And l is 1,2, …, n and n are the number of the neural network layers. And theta is a parameter to be optimized of the deep neural network. W represents a weight, and b represents an offset.

Feedforward transfer function for layer I neurons

As follows:

The activation function s is as follows:

in this embodiment, the activation function of the top feedforward transfer function is designed as a Linear function, and the activation functions of the other layers are ReLU activation functions (Linear rectification functions).

4.3) utilizing RMSProp algorithm (Root Mean Square Prop, deep learning optimization algorithm) to update the parameters of the deep neural network for the t time, namely:

4.4) mixingInputting the verification set into a deep neural network direct current power flow model, judging whether the testing precision of the verification set is reduced, if so, stopping iteration, and if not, judging the iteration times t>t_maxAnd if the result is positive, stopping iteration, otherwise, returning to the step 2. t is t_maxIs the maximum number of iterations.

Example 2:

the method for checking the safety of the N-1 with uncertainty based on deep learning mainly comprises the following steps:

3) And establishing a deep neural network direct current power flow model.

6) For output feature vector Y_out＝[θ_i,P_ij]And carrying out security verification. A

Example 3:

the method for checking the safety of the N-1 with uncertainty based on deep learning mainly comprises the following steps of embodiment 2, wherein a deep neural network direct current power flow model is as follows:

in the formula (I), the compound is shown in the specification,

Feedforward transfer function for layer I neurons

As follows:

in the formula, X^l-1Is the input to layer I neurons. W^lAnd b^lAre the weight matrix and offset vector between the l-layer neurons and the l-1 layer neurons. s is the activation function.

The activation function s is as follows:

example 4:

the method for checking the safety of the N-1 with uncertainty based on deep learning mainly comprises the following steps of embodiment 2, wherein the method for training the deep neural network direct current power flow model mainly comprises the following steps:

1) and inputting the training set into a direct current power flow model of the deep neural network.

2) And randomly initializing a parameter theta to be optimized of the deep neural network direct current power flow model.

3) And (3) utilizing the RMSProp algorithm to update the parameters of the deep neural network for the t time, namely:

wherein η is the learning rate, ε is a constant, r is the cumulative squared gradient, ▽_θL is the partial derivative of the mean square error loss function on θ, ρ is the decay rate, t is the number of iterations, t is initially 1, ρ is typically set to 0.99, η to 0.001, and e is set to 1 × 10^-8。

4) Inputting the verification set into a deep neural network direct current power flow model, judging whether the testing precision of the verification set is reduced, if so, stopping iteration, and if not, judging the iteration times t>t_maxAnd if the result is positive, stopping iteration, otherwise, returning to the step 2. t is t_maxIs the maximum number of iterations.

Example 5:

an experiment for verifying the safety checking method containing uncertainty N-1 based on deep learning mainly comprises the following steps:

1) the method comprises the steps of building a system, introducing 10MW photovoltaic power stations on nodes 5 and 12 in an IEEE 30 node system, introducing 16MW wind power plants on nodes 10, 15 and 27, and leading the penetration rate of new energy to be 20%, introducing 250MW photovoltaic power stations on nodes 13, 14, 16 and 23 in an IEEE 118 node system, introducing 330MW wind power plants on buses 59, 80 and 90, wherein the penetration rate of the new energy is 20%, the wind speed two parameters Weibull distribution, the scale parameter of 2.016, the shape parameter of 5.089, the maximum power of 16MW, the cut-in wind speed, the rated wind speed and the cut-out wind speed are respectively 3.5m/s, 15m/s and 25m/s, the illumination intensity obeys Beta distribution, the maximum power of 20MW, the α and β of the Beta distribution are respectively 2.06 and 2.5, the mean value is a given value test system, and the standard deviation is 10% of the mean value.

2) Different comparative examples were constructed, as follows:

example 1: and in the IEEE 30 node system, the new energy permeability is 20%, and the load standard deviation is 10% of the mean value.

Example 2: and in the IEEE 118 node system, the new energy permeability is 20%, and the load standard deviation is 10% of the mean value.

The comparison methods in the simulation included M0-M5. The FDNN direct current power flow model adopted by the invention has 3 layers of hidden layers in total, and the neuron number of each layer is respectively 100 and 500 for the calculation examples 1 and 2.

M0: and D, direct current flow is used as a verification standard.

M1: the difference of the voltages of the nodes is used as an FDNN power flow model of the characteristic vector of the topological structure.

M2: and taking the admittance matrix as the SDAE power flow model of the topological structure characteristic vector.

M3: and taking the 0-1 vector representing the branch state as the FDNN power flow model of the topological structure characteristic vector.

M4: m1 using min-max normalization method.

M5: the activation function only employs M1 for ReLU.

3) According to the method, samples are sampled according to source load distribution and line sequential faults to generate samples and test, and for different calculation examples, parameters of a direct current power flow model of a deep neural network adopted by the method are specifically shown in a table 1. All the calculations of the present invention were tested in the hardware environment of Intel (R) core (TM) i7-9700K CPU @3.70GHz 32GB RAM.

In order to compare the performances of different methods, the invention designs the following indexes: p_vaProbability that the absolute error of voltage phase angle exceeds 0.01rad, P_pfThe active absolute error of the branch exceeds the probability of 1 MW.

TABLE 1 parameter settings of M1 under different examples

Examples of the design	Hidden layer structure	Number of training samples	Verifying the number of samples	Number of samples tested
					EXAMPLE 1	[100 100 100]	20000	10000	10000
EXAMPLE 2	[500 500 500]	50000	10000	10000

4) Effect of topological structure feature vectors on results

This section is intended to verify that M1 using the eigenvectors proposed in the present invention has higher accuracy of power flow calculation and faster convergence speed of training than M2 using admittance matrix and M3 using 0-1 vector under the same number of iterations (200).

In terms of accuracy, the probabilities of the absolute error being greater than the set value by the three methods M1-M3 are compared with Table 2. As can be seen from table 2, the probability that the absolute error of the calculation result of the proposed method M1 is greater than the set value is less than 5% in both of the calculation examples 1 and 2. Although the M2 method (using admittance matrix as the feature vector for representing the topological structure) can obtain better calculation accuracy in the example 1, the input feature vector grows exponentially with the increase of the system scale, the vector dimension for representing the topological structure in the example 2 reaches 27848, which exceeds the calculation cost of the used hardware environment and cannot be trained. The M3 method adopts 0-1 vector as the characteristic vector for representing the topological structure, the variation is less easily submerged by other large variation information, so that the deep neural network can not effectively mine the power flow characteristic of the example 2, and in the calculation result of the example 2, the probability that the branch active power exceeds 1MW exceeds 16%. It can be seen that M1 using the proposed feature vectors has a higher model accuracy than M2, M3.

Tables 2 comparison of load flow calculation accuracy of M1, M2 and M3

In terms of convergence rate, the loss function decline curves of M1, M2, and M3 in the training in the cases of example 1 and example 2 are shown in fig. 2 and fig. 3, respectively. As shown in fig. 2, for the example 1, due to the small system scale, the flow characteristics of the example 1 can be well mined by the three methods M1, M2 and M3, and the convergence rates of the three methods are equivalent. As shown in fig. 3, for example 2, the convergence rate of M1 was greatly improved compared to M3, and at the end of training, the loss functions of M1 and M3 were 0.122 and 0.359, respectively, and the loss function of M1 was reduced by 66.0% compared to that of M3. Therefore, in summary, M1 has a faster convergence rate than M2 and M3.

5) Effect of normalization methods on results

Under the convergence condition (the early-stop method is met or the iteration number does not exceed the iteration threshold 1000), the precision results of calculating the direct current power flow by adopting the z-score standardization method M1 designed by the invention and the min-max standardization method M4 are shown in Table 3. As can be seen from table 1, in both of the calculation examples 1 and 2, the probabilities that the absolute error of the calculation result of M1 is larger than the set value are smaller than M4. In the calculation results of M1 in example 1, the probability that the absolute error is greater than the set value is 0.3% or less. In the calculation result of M4, the probability that the absolute error of the voltage phase angle is greater than the set value is 3.5%, and the probability that the absolute error of the branch is greater than the set value exceeds 14%. In the calculation results of M1, the probability that the absolute error is larger than the set value is 1.9% or less in example 2. In the calculation result of M2, the probability that the absolute error of the voltage phase angle and the branch path is greater than the set value exceeds 69%. It can be seen that the z-score normalization method adopted by M1 is more suitable for processing power flow samples considering topological structure changes.

Comparison of load flow calculation accuracies of tables 3M 1 and M4

6) Effect of activation function on results

Fig. 4 and 5 show the loss function reduction curves of training examples 1 and 2 using M1 and M5, respectively. It can be seen from the figure that the activation function M1 designed in both example 1 and example 2 can converge to a lower loss function than M5 which uses only the ReLU activation function. In example 1, training using both the M1 and M5 methods converged upon reaching the iteration threshold, with the loss function values falling to 0.747 and 11.694, respectively. Therefore, compared with the M5, the method M1 provided by the invention can effectively reduce the loss function value by 93.6%. In example 2, M1 converged upon reaching the iteration threshold 1000, and M5 stopped training when it satisfied the early-stop convergence condition by iteration to 425 times. The M1 and M5 after the training convergence reduce the loss function of the operator 2 to 0.033 and 27.575, and the method provided by the invention also has obvious advantages, and can effectively reduce the loss function value by 99.9% compared with the M5. Therefore, the last layer of loss function is designed into a linear function, so that the deep neural network can capture wider output, and the trend characteristics can be effectively mined by better matching with a data preprocessing method.

7) N-1 security check performance analysis based on deep learning technology

The section analyzes the N-1 safety checking performance based on the deep learning technology from two aspects of calculation precision and calculation speed. The training convergence condition of the deep neural network is that an early-stopping method is met or the iteration number does not exceed an iteration threshold 1000. Assume that there are 4 scenarios to be verified for each new energy station in example 1, and there are 3 scenarios to be verified for each new energy station in example 2. Then 41986 and 406782 scenes need to be verified in case 1 and case 2, respectively.

And directly mapping the power flow of all scenes to be checked by adopting the trained deep neural network direct current power flow model and judging the out-of-limit of the branch according to the result. The method judges the calculation precision of the method by using the average calculation error of all node phase angles, the average calculation error of branch power and the out-of-limit judgment accuracy (checking accuracy) of the branch in the scene to be checked. Additionally, it is noted that since the IEEE 118 node system does not have a branch power threshold, the present invention employs the branch power threshold of the Boston 118 node system instead. The statistical results of the N-1 security checks calculated using the proposed method M1 are shown in Table 4. From Table 4, it can be seen that M1 has node phase angle average calculation errors of-1.0X 10-5rad and 3.9X 10-5rad in abacus 1 and abacus 2, respectively, and M1 has branch power average calculation errors of-1.7X 10-3MW and 2.1X 10-3MW in abacus 1 and abacus 2, respectively. The accuracy rate of the N-1 security check by the method provided by the invention can reach more than 99.9% in two calculation examples. Therefore, the method provided by the invention has high fitting precision on load flow calculation, and the N-1 safety check accuracy can effectively meet the industrial application requirements.

TABLE 4M 1 Security check accuracy analysis

Examples of the design	Mean_Pij(MW)	Mean_θ(rad)	Checking accuracy
				EXAMPLE 1	-1.7×10^-3	-1.0×10^-5	99.96％
EXAMPLE 2	2.1×10^-3	3.9×10^-5	99.99％

Table 5 shows the comparison between the method M1 of the present invention and the method M0 currently used in the industry for calculating the time for security check. As can be seen from the table, the method provided by the invention can greatly shorten the calculation time in both the examples. In example 1, the method of the present invention only requires 0.12 seconds, whereas 64.16 seconds is required by the industrial method M0, which can increase the calculation speed by 535 times. In example 2, the method of the present invention also shows a significant advantage in terms of computation speed. The calculation time for calculating the N-1 safety check by adopting M1 and M0 is 8.50 seconds and 820.09 seconds respectively, and the N-1 check speed by adopting the method provided by the invention is improved by 96 times by adopting the method in the industry. Therefore, the method provided by the invention can improve the safety checking speed of the N-1 by nearly one hundred times under the condition of ensuring high checking accuracy. It is worth noting that with the increase of new energy stations and the consideration of multi-period problems, the calculation time of the solution method in the industry will increase by times, and the advantages of the method provided by the invention will become more obvious.

TABLE 5 comparison of safety check speed of M0 and M1 for N-1

The invention starts from two aspects of feature vector construction and learning strategy design, and provides an N-1 rapid checking method based on a deep learning technology. The node injection power and the branch power difference before and after the branch is disconnected are constructed into input characteristic vectors representing source load and topological structure changes, so that the deep neural network can effectively extract the important influence of the source load and the topological structure changes on the power flow. In addition, after the conventional data preprocessing method and the mathematical characteristics of the activation function are analyzed, a z-score standardization method is adopted to carry out normalization preprocessing on the power flow sample; and then the ReLU activation function and the linear function are combined to be used as the activation function of the deep neural network to effectively extract the trend characteristics, so that a set of deep neural network learning strategy suitable for N-1 checking is formed.

Claims

1. The method for checking the safety of the N-1 with uncertainty based on deep learning is characterized by mainly comprising the following steps:

2) Preprocessing the characteristic vector, and dividing the characteristic vector into a training set and a verification set;

3) establishing a deep neural network direct current power flow model;

4) training a direct current power flow model of the deep neural network by utilizing a training set and a verification set;

5) establishing a test set by using real-time data of the power network, and inputting the test set into the trained deep neural network direct current power flow model to obtain a power flow output characteristic vector Y_out；

2. The deep learning based uncertainty-containing N-1 security check method according to claim 1, wherein the power network data comprises power network topology and node source load data.

3. The deep learning-based N-1 security check method with uncertainty according to claim 1 or 2, characterized in that a feature vector set X is input_in＝[P_i,△P_ij]And outputting the feature vector Y_out＝[θ_i,P_ij](ii) a Wherein, P_iRepresenting the sum of active power injected by the new energy node for continuous eigenvectors; theta_iIs a nodal phase angle, P_ij△ P as the active power flow of each branch_ijThe discrete characteristic vector represents the difference of active power of each branch circuit before and after the branch circuit is disconnected;

in the formula (I), the compound is shown in the specification,

active power of each branch circuit before the branch circuit is cut off;

the active power of each branch after the branch is cut off, and the difference △ P between the active power of each branch before and after the branch is cut off_ijHas a dimension of n_branch；n_branchThe number of system branches.

4. The deep learning based uncertainty-containing N-1 security check method according to claim 1, characterized in that: the data preprocessing method is normalization;

the normalization formula is as follows:

in the formula, x_μIs the mean value of the samples, x_δIs the sample standard deviation; x is an input sample, i.e. a feature vector; x is normalized data; data x mean 0 and variance 1.

5. The deep learning-based N-1 safety checking method with uncertainty according to claim 1, characterized in that the deep neural network direct current power flow model is as follows:

in the formula (I), the compound is shown in the specification,

a feed-forward transfer function for layer I neurons; 1,2, …, n, n is the number of neural network layers; theta is a parameter to be optimized of the deep neural network;

feedforward transfer function for layer I neurons

As follows:

in the formula, X^l-1An input for layer I neurons; w_lAnd b_lWeight matrix and offset vector between layer I neuron and layer I-1 neuron; s is the activation function.

6. The deep learning-based uncertainty-containing N-1 security check method according to claim 1, wherein an activation function s is as follows:

7. the deep learning-based N-1 safety checking method with uncertainty according to claim 1, characterized in that the training of the deep neural network DC power flow model comprises the following main steps:

1) inputting the training set into a direct current power flow model of the deep neural network;

2) randomly initializing a parameter theta to be optimized of a deep neural network direct current power flow model;

in the formula, ▽_θL is the partial derivative of the mean square error loss function to theta, ⊙ is a Hamiltonian, η is the learning rate, epsilon is a constant, r is the cumulative squared gradient, rho is the decay rate, t is the iteration number, and t is the initial value of 1;

4) inputting the verification set into a deep neural network direct current power flow model, judging whether the testing precision of the verification set is reduced, if so, stopping iteration, and if not, judging the iteration times t>t_maxIf yes, stopping iteration, otherwise, returning to the step 2; t is t_maxIs the maximum number of iterations.

8. The deep learning-based N-1 security check method with uncertainty according to claim 1, characterized in that the output feature vector Y is checked_out＝[θ_i,P_ij]The method for carrying out security verification comprises the following steps: judging active power flow P of branch_ij>P_maxAnd if the judgment result is positive, judging that the branch ij is overloaded, and if the judgment result is negative, judging that the branch ij is in a safe state.