CN110929989B - N-1 safety checking method with uncertainty based on deep learning - Google Patents

Abstract

The present invention discloses an uncertainty-containing N-1 security verification method based on deep learning, the main steps of which are: 1) acquiring power network data, and establishing an input feature vector set X _in ; 2) preprocessing the feature vectors, and Divided into training set and verification set; 3) Established deep neural network DC power flow model; 4) Using the training set and verification set to train the deep neural network DC power flow model; 5) Establishing a test set with real-time data of the power network, and inputting Into the trained deep neural network DC power flow model, the power flow output feature vector Y _out is obtained; 6) Safety verification is performed on the output feature vector Y _out =[θ _i ,P _ij ]. The present invention can be widely applied to online analysis of N-1 security check in uncertain scene, can obtain check accuracy comparable to traditional DC power flow solution method, and improve analysis speed by nearly a hundred times.

Description

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

Technical Field

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

Background Art

The power system essentially operates in an uncertain environment. N-1 safety verification is an important analytical tool to ensure system safety. In recent years, due to the increasing penetration of renewable energy such as photovoltaics and wind power, and the randomness and intermittency of new energy such as wind and solar, their large-scale access has brought more uncertainty to the power system. In order to cope with the impact of growing uncertainty on the operation of the power system, in addition to taking into account branch circuit breaking, N-1 safety verification also needs to consider uncertain scenarios such as new energy fluctuations, which leads to new computational challenges for N-1 safety verification.

At present, N-1 safety verification generally adopts DC power flow model calculation. The DC power flow model avoids the problems of non-convergence of calculations caused by AC power flow model and increases the complexity of economic dispatch, unit combination and other models. However, for N-1 safety verification that needs to consider the uncertainty of new energy scenarios, assuming that M wind farms need to be considered, and each wind farm has 10 uncertainty scenarios, then N-1 safety verification needs to perform N×10M power flow calculations, and its calculation burden will increase exponentially compared with traditional N-1 safety verification. Therefore, scholars have been seeking improved methods to increase the speed of N-1 safety verification.

At present, the research on accelerating the speed of N-1 safety verification mainly reduces the scenarios by model derivation. Some studies reduce the number of scenarios by establishing optimization models, mainly including establishing mixed integer linear programming and two-level programming models, reducing the number of verification scenarios by three-level filtering methods, using iterative boundaries and continuous pruning methods to reduce scenarios, and using GPU to accelerate grid connectivity judgment. Considering the uncertainty of new energy and multi-period problems of power systems, although the above methods can effectively reduce the number of scenarios and speed up the verification of a single scenario, the number of scenarios that need to be verified in N-1 safety analysis is still huge, and its computational efficiency still needs to be further improved. In summary, it is urgent to study an efficient N-1 safety verification method.

Summary of the invention

The purpose of the present invention is to solve the problems existing in the prior art.

The technical solution adopted to achieve the purpose of the present invention is as follows: a deep learning-based N-1 safety verification method with uncertainty mainly includes the following steps:

1) Obtain power network data and establish an input feature vector set _Xin .

Furthermore, the power network data includes the power network topology structure and source load data of each node.

Furthermore, the input feature vector set _Xin = [P _i , △P _ij ] and the output feature vector Y _out = [θ _i ,P _ij ]. Wherein, _Pi is a continuous feature vector, which represents the total active power injected by the new energy node. θ _i is the node phase angle, and _Pij is the active power flow of each branch. △P _ij is a discrete feature vector, which represents the difference in active power of each branch before and after the branch is disconnected.

The difference in active power of each branch before and after the branch is disconnected △P _ij is as follows:

为支路开断前各支路有功功率。

为支路开断后各支路有功功率。支路开断前后各支路的有功功率之差△P_ij的维度为n_branch。n_branch为系统支路数。In the formula,

It is the active power of each branch before the branch is disconnected.

is the active power of each branch after the branch is disconnected. The dimension of the difference △P _ij between the active power of each branch before and after the branch is disconnected is n _branch . n _branch is the number of system branches.

2) Preprocess the feature vector and divide it into training set and validation set.

The data preprocessing method is normalization.

The normalization formula is as follows:

In the formula, x _μ is the sample mean, x _δ is the sample standard deviation. x is the input sample, i.e., the eigenvector. x* is the normalized data. The mean of data x* is 0 and the variance is 1.

3) Established deep neural network DC power flow model.

The deep neural network DC power flow model is shown below:

为第l层神经元的前馈传递函数。l＝1,2,…,n，n为神经网络层数。θ＝{W,b}为深度神经网络待优化参数。In the formula,

is the feedforward transfer function of the lth layer of neurons. l＝1,2,…,n, n is the number of neural network layers. θ＝{W,b} is the parameter to be optimized for the deep neural network.

如下所示：The feedforward transfer function of the l-th layer of neurons

As shown below:

Where Xl ^-1 is the input of the lth layer of neurons. _Wl and _bl are the weight matrix and offset vector between the lth layer of neurons and the l-1th layer of neurons. s is the activation function.

The activation function s is as follows:

4) Use the training set and validation set to train the deep neural network DC power flow model.

The main steps for training a deep neural network DC power flow model are as follows:

4.1) Input the training set into the deep neural network DC power flow model.

4.2) Randomly initialize the parameters θ to be optimized in the deep neural network DC power flow model.

4.3) Use the RMSProp algorithm to update the deep neural network parameters for the tth time, that is:

Where 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 with respect to θ. ρ is the decay rate. t is the number of iterations. The initial value of t is 1.

4.4) Input the validation set into the deep neural network DC power flow model to determine whether the test accuracy of the validation set decreases. If so, stop the iteration. If not, determine whether the number of iterations t>t _max is true. If so, stop the iteration. If not, return to step 2. _{t max} is the maximum number of iterations.

5) A test set is established with real-time data of the power network and input into the trained deep neural network DC power flow model to obtain the power flow output feature vector Y _out .

6) Perform security verification on the output feature vector Y _out =[θ _i ,P _ij ].

The method for verifying the safety of the output characteristic vector Y _out =[θ _i ,P _ij ] is: judging whether the active power flow P _ij >P _max of the branch holds, if so, judging that the branch ij is overloaded, if not, the branch ij is in a safe state.

The technical effect of the present invention is unquestionable. The characteristic vector designed by the present invention can effectively cover the new energy, continuous load change characteristics and topological structure change characteristics possessed by N-1 safety verification. The DC power flow model obtained after sample training can effectively mine the complex characteristics between the input and output of the DC power flow equation, and can directly perform power flow calculations on all uncertain scenarios to be verified, providing technical support for the determination of overloaded lines and the safe design of operation plans.

The deep learning strategy designed by the present invention is suitable for N-1 safety verification. Starting from the data preprocessing method and activation function, the z-score standardization method is used to normalize the power flow samples, and the ReLU activation function and the linear function are combined as the activation function of the deep neural network to effectively extract the power flow characteristics. The RMSProp learning algorithm is selected to train the deep neural network DC power flow model, which effectively realizes the acquisition of the optimal parameter θ.

The present invention can be widely used in online analysis of N-1 safety verification in uncertain scenarios, can obtain verification accuracy comparable to that of traditional DC power flow solution methods, and improve the analysis speed by nearly 100 times.

BRIEF DESCRIPTION OF THE DRAWINGS

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

Figure 2 is a comparison of the convergence speeds of M1, M2, and M3 in Example 1;

Figure 3 is a comparison of the convergence speeds of M1 and M3 in Example 2;

Figure 4 shows the loss function descent process of M1 and M5 in Example 1;

Figure 5 shows the loss function descent process of M1 and M5 in Example 2.

DETAILED DESCRIPTION

The present invention is further described below in conjunction with the embodiments, but it should not be understood that the above subject matter of the present invention is limited to the following embodiments. Without departing from the above technical ideas of the present invention, various substitutions and changes are made according to the common technical knowledge and customary means in the art, which should all be included in the protection scope of the present invention.

Embodiment 1:

Referring to FIG1 , the N-1 safety verification method with uncertainty based on deep learning mainly includes the following steps:

1) Obtain power network data and establish an input feature vector set _Xin .

Furthermore, the power network data includes the power network topology structure and source load data of each node.

Furthermore, the input feature vector set _Xin = [P _i , △P _ij ] and the output feature vector Y _out = [θ _i ,P _ij ]. Wherein, _Pi is a continuous feature vector, which represents the total active power injected by the new energy node. θ _i is the node phase angle, and _Pij is the active power flow of each branch. △P _ij is a discrete feature vector, which represents the difference in active power of each branch before and after the branch is disconnected.

The difference in active power of each branch before and after the branch is disconnected △P _ij is as follows:

为支路开断前各支路有功功率。

为支路开断后各支路有功功率。支路开断前后各支路的有功功率之差△P_ij的维度为n_branch。n_branch为系统支路数。在计算△P_ij时，各节点负荷取为负荷均值，新能源出力取为额定值。该向量的维度为n_branch(其中，n_branch为系统支路数)，随系统规模的增大仅线性增长，同时不同拓扑之间各支路的有功功率之差均发生变化，因而不存在拓扑信息被淹没的问题，其值的大小还有效刻画了支路开断对各支路的影响程度，使得该向量有效地涵盖了拓扑结构变化对电力系统潮流的重要影响。In the formula,

It is the active power of each branch before the branch is disconnected.

is the active power of each branch after the branch is disconnected. The dimension of the difference in active power of each branch before and after the branch is disconnected △P _ij is n _branch . n _branch is the number of system branches. When calculating △P _ij , the load of each node is taken as the load average, and the output of new energy is taken as the rated value. The dimension of this vector is n _branch (where n _branch is the number of system branches). It only grows linearly with the increase of system scale. At the same time, the difference in active power of each branch between different topologies changes, so there is no problem of topological information being submerged. The size of its value also effectively describes the degree of influence of branch disconnection on each branch, so that this vector effectively covers the important influence of topological structure changes on power system flow.

It is worth noting that the feature vector is the material basis for deep neural network function fitting. In order to establish a deep neural network DC power flow model that can effectively extract the N-1 safety verification features considering the uncertainty of new energy scenarios, the input feature vector of the sample should include the new energy, load continuous variability characteristics and topological structure change characteristics of the N-1 safety verification.

In view of the continuous change characteristics, the present invention selects the sum of the node injected active power of each continuous variable (new energy, load, etc.) as the input feature vector. The node injected power can effectively reflect the fluctuation of new energy and load, and its dimension is only the number of system nodes.

In view of the discrete change characteristics, conventional methods for representing topological structures include the admittance matrix and the 0-1 vector representing the branch state (0 represents a branch disconnection and 1 represents a branch closure). The admittance matrix can reflect the branch disconnection status and the node association relationship. However, when the admittance matrix is used as a feature vector, its dimension will grow quadratically with the increase in system scale. Although the dimension of the 0-1 vector representing the branch state only grows linearly with the increase in system scale, it can only reflect the disconnection status of each branch. Moreover, both of the above two methods have the problem of topological information being submerged, so it is difficult for deep neural networks to effectively extract the important impact of branch disconnection on the power system flow. In this regard, the present invention proposes to use the difference in active power of each branch before and after the branch is disconnected △P _ij as the feature vector representing the topological structure.

2) Preprocess the feature vector and divide it into training set and validation set.

The data preprocessing method is normalization.

Taking into account the topological changes caused by branch disconnection, there are a small number of data points in the flow sample that deviate from the sample mean. For example, when a branch is disconnected, the active power of the branch is 0MW, while in most other cases, the active power of the branch is still near the mean. The min-max method uses the minimum and maximum values in the sample to perform normalization, and because the deviated data points will affect the max or min value, the min-max method is susceptible to the influence of a small number of deviated data points. The z-score method uses the overall information of the sample, that is, the sample mean and the sample standard deviation, to perform normalization, which is less affected by the deviated data points and can better preserve the data distribution characteristics while eliminating numerical problems. Therefore, the present invention uses the following normalization method:

In the formula, x _μ is the sample mean, x _δ is the sample standard deviation. x is the input sample, i.e., the eigenvector. x* is the normalized data. The mean of data x* is 0 and the variance is 1.

3) Established deep neural network DC power flow model.

The deep neural network DC power flow model is shown below:

为第l层神经元的前馈传递函数。l＝1,2,…,n，n为神经网络层数。θ＝{W,b}为深度神经网络待优化参数。W表示权重，b表示偏移。In the formula,

is the feedforward transfer function of the lth layer of neurons. l＝1,2,…,n, n is the number of neural network layers. θ＝{W,b} is the parameter to be optimized for the deep neural network. W represents the weight and b represents the offset.

如下所示：The feedforward transfer function of the l-th layer of neurons

As shown below:

Where Xl ^-1 is the input of the lth layer of neurons. _Wl and _bl are the weight matrix and offset vector between the lth layer of neurons and the l-1th layer of neurons. s is the activation function.

The activation function s is as follows:

In this embodiment, the activation function of the top-level feedforward transfer function is designed to be a linear function, and the activation functions of other layers are ReLU activation functions (Rectified Linear Unit).

4) Use the training set and validation set to train the deep neural network DC power flow model.

The main steps for training a deep neural network DC power flow model are as follows:

4.1) Input the training set into the deep neural network DC power flow model.

4.2) Randomly initialize the parameters θ to be optimized in the deep neural network DC power flow model.

4.3) Use the RMSProp algorithm (Root Mean Square Prop, deep learning optimization algorithm) to update the deep neural network parameters for the tth time, that is:

Where 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 with respect to θ. ρ is the decay rate. t is the number of iterations. The initial value of t is 1.

4.4) Input the validation set into the deep neural network DC power flow model to determine whether the test accuracy of the validation set decreases. If so, stop the iteration. If not, determine whether the number of iterations t>t _max is true. If so, stop the iteration. If not, return to step 2. _{t max} is the maximum number of iterations.

5) A test set is established with real-time data of the power network and input into the trained deep neural network DC power flow model to obtain the power flow output feature vector Y _out .

6) Perform security verification on the output feature vector Y _out =[θ _i ,P _ij ].

The method for verifying the safety of the output characteristic vector Y _out =[θ _i ,P _ij ] is: judging whether the active power flow P _ij >P _max of the branch holds, if so, judging that the branch ij is overloaded, if not, the branch ij is in a safe state.

Embodiment 2:

The N-1 safety verification method with uncertainty based on deep learning mainly includes the following steps:

1) Obtain power network data and establish an input feature vector set _Xin .

2) Preprocess the feature vector and divide it into training set and validation set.

3) Established deep neural network DC power flow model.

4) Use the training set and validation set to train the deep neural network DC power flow model.

5) A test set is established with real-time data of the power network and input into the trained deep neural network DC power flow model to obtain the power flow output feature vector Y _out .

6) Perform security verification on the output feature vector Y _out =[θ _i ,P _ij ].

Embodiment 3:

The main steps of the N-1 safety verification method with uncertainty based on deep learning are shown in Example 2, wherein the deep neural network DC power flow model is as follows:

为第l层神经元的前馈传递函数。l＝1,2,…,n，n为神经网络层数。θ＝{W,b}为深度神经网络待优化参数。In the formula,

is the feedforward transfer function of the lth layer of neurons. l＝1,2,…,n, n is the number of neural network layers. θ＝{W,b} is the parameter to be optimized for the deep neural network.

如下所示：The feedforward transfer function of the l-th layer of neurons

As shown below:

Where Xl ^-1 is the input of the lth layer of neurons. ^Wl and ^bl are the weight matrix and offset vector between the lth layer of neurons and the l-1th layer of neurons. s is the activation function.

The activation function s is as follows:

Embodiment 4:

The main steps of the N-1 safety verification method with uncertainty based on deep learning are shown in Example 2, wherein the main steps of training the deep neural network DC power flow model are as follows:

1) Input the training set into the deep neural network DC power flow model.

2) Randomly initialize the parameters θ to be optimized in the deep neural network DC power flow model.

3) Use the RMSProp algorithm to update the deep neural network parameters for the tth time, that is:

Where η is the learning rate. ε is a constant. r is the cumulative squared gradient. ▽ _θ L is the partial derivative of the mean square error loss function with respect to θ. ρ is the decay rate. t is the number of iterations. The initial value of t is 1. Usually ρ＝0.99, η＝0.001, ε＝1×10 ^-8 are set.

4) Input the validation set into the deep neural network DC power flow model to determine whether the test accuracy of the validation set decreases. If so, stop the iteration. If not, determine whether the number of iterations t>t _max is established. If so, stop the iteration. If not, return to step 2. _{t max} is the maximum number of iterations.

Embodiment 5:

The experiment to verify the uncertainty-containing N-1 safety verification method based on deep learning mainly includes the following steps:

1) System construction: In the IEEE 30-node system, the present invention introduces a 10MW photovoltaic power station on nodes 5 and 12, and a 16MW wind farm on nodes 10, 15 and 27, with a new energy penetration rate of 20%. In the IEEE 118-node system, the present invention introduces a 250MW photovoltaic power station on nodes 13, 14, 16 and 23, and a 330MW wind farm on busbars 59, 80 and 90, with a new energy penetration rate of 20%. Among them, the wind speed obeys the two-parameter Weibull distribution, the scale parameter is 2.016, the shape parameter is 5.089, the maximum power is 16MW, the cut-in wind speed, rated wind speed, and cut-out wind speed are 3.5m/s, 15m/s, and 25m/s respectively. The light intensity obeys the Beta distribution, the maximum power is 20MW, and the α and β of the Beta distribution are 2.06 and 2.5 respectively. The load follows a normal distribution with a mean value given by the standard test system and a standard deviation of 10% of the mean value.

2) Construct different comparative examples, as follows:

Example 1: IEEE 30-bus system, renewable energy penetration rate 20%, load standard deviation 10% of the mean.

Example 2: IEEE 118-node system, renewable energy penetration rate of 20%, and load standard deviation of 10% of the mean.

The comparison methods in the simulation include M0-M5. The FDNN DC power flow model used in the present invention has 3 hidden layers, and the number of neurons in each layer is 100 and 500 for Example 1 and Example 2, respectively.

M0: DC power flow, used as verification standard.

M1: FDNN power flow model using the voltage difference of each node proposed in the present invention as the topological structure feature vector.

M2: SDAE power flow model with the admittance matrix as the topological structure eigenvector.

M3: FDNN power flow model with 0-1 vector representing branch status as topological structure feature vector.

M4: M1 with min-max normalization method.

M5: M1 with only ReLU as activation function.

3) The present invention generates samples and performs tests according to source-load distribution and line faults. For different examples, the parameters of the deep neural network DC power flow model used in the present invention are specifically shown in Table 1. All examples of the present invention are tested in the hardware environment of Intel(R) Core(TM) i7-9700K CPU@3.70GHz 32GB RAM.

In order to compare the performance of different methods, the present invention designs the following indicators: the probability that the absolute error of the _Pva voltage phase angle exceeds 0.01rad, and the probability that the absolute error of the _Ppf branch active power exceeds 1MW.

Table 1 Parameter settings of M1 under different examples

Example Hidden layer structure Number of training samples Number of validation samples Number of test samples Example 1 [100 100 100] 20000 10000 10000 Example 2 [500 500 500] 50000 10000 10000

4) The influence of topological structure eigenvector on the results

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

In terms of accuracy, the probability that the absolute error obtained by the three methods M1-M3 is greater than the set value is compared in Table 2. As can be seen from Table 2, the probability that the absolute error of the calculation results of the proposed method M1 is greater than the set value in Example 1 and Example 2 is less than 5%. Although the M2 method (using the admittance matrix as the eigenvector representing the topological structure) can obtain good calculation accuracy in Example 1, as the system scale increases, the input eigenvector grows exponentially. The vector dimension representing the topological structure in Example 2 reaches 27848, which exceeds the calculation cost of the hardware environment used and cannot be trained. The M3 method uses 0-1 vectors as eigenvectors representing the topological structure. Its small change amount is easily submerged by other large amounts of change information, resulting in the deep neural network being unable to effectively mine the power flow characteristics of Example 2. In the calculation results of Example 2, the probability that the branch active power exceeds 1MW exceeds 16%. It can be seen that M1 using the proposed eigenvector has higher model accuracy than M2 and M3.

Table 2 Comparison of power flow calculation accuracy of M1, M2, and M3

In terms of convergence speed, the loss function decline curves of M1, M2, and M3 during training in Example 1 and Example 2 are shown in Figure 2 and Figure 3 respectively. As shown in Figure 2, for Example 1, due to the small scale of the system, the three methods M1, M2, and M3 can all well mine the flow characteristics of Example 1, and the convergence speeds of the three are comparable. As shown in Figure 3, for Example 2, the convergence speed of M1 is much higher than that of M3. At the end of training, the loss functions of M1 and M3 are 0.122 and 0.359 respectively, and the loss function of M1 is reduced by 66.0% compared with that of M3. Therefore, in general, M1 has a faster convergence speed than M2 and M3.

5) The impact of normalization methods on the results

Under the convergence condition (satisfying the early stopping method or the number of iterations does not exceed the iteration threshold of 1000), the accuracy results of the DC power flow calculated by the z-score standardization method M1 designed by the present invention and the min-max standardization method M4 are shown in Table 3. It can be seen from Table 1 that for Example 1 and Example 2, the probability that the absolute error in the calculation results of M1 is greater than the set value is less than that of M4. Among them, for Example 1, in the calculation results of M1, the probability that the absolute error is greater than the set value is less than 0.3%. In the calculation results 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 active power is greater than the set value exceeds 14%. For Example 2, in the calculation results of M1, the probability that the absolute error is greater than the set value is less than 1.9%. In the calculation results of M2, the probability that the absolute error of the voltage phase angle and the branch active power is greater than the set value exceeds 69%. It can be seen that the z-score standardization method adopted by M1 is more suitable for processing power flow samples that take into account topological changes.

Table 3 Comparison of power flow calculation accuracy between M1 and M4

6) The impact of activation function on the results

Figures 4 and 5 show the loss function descent curves when using M1 and M5 for training in Example 1 and Example 2, respectively. It can be seen from the figure that whether in Example 1 or Example 2, the designed activation function M1 can converge to a lower loss function than M5 which only uses the ReLU activation function. In Example 1, the training using the M1 and M5 methods converged after reaching the iteration threshold, and the loss function values dropped to 0.747 and 11.694, respectively. It can be seen that the method M1 proposed in the present invention can effectively reduce the loss function value by 93.6% compared with M5. In Example 2, M1 converged after reaching the iteration threshold of 1000, and M5 stopped training when it met the early stopping method convergence condition when it iterated to 425 times. After the training converged, M1 and M5 reduced the loss function of Example 2 to 0.033 and 27.575 respectively. The method proposed in the present invention also showed obvious advantages and could effectively reduce the loss function value by 99.9% compared with M5. It can be seen that the present invention designs the last layer loss function as a linear function, which enables the deep neural network to capture a wider range of outputs, thereby better matching the data preprocessing method to effectively mine the trend characteristics.

7) N-1 safety verification performance analysis based on deep learning technology

This section analyzes the N-1 safety verification performance based on deep learning technology from the aspects of calculation accuracy and calculation speed. Among them, the training convergence condition of the deep neural network is to meet the early stopping method or the number of iterations does not exceed the iteration threshold of 1000. Assume that each new energy station in Example 1 has 4 scenes to be verified, and each new energy station in Example 2 has 3 scenes to be verified. Then Example 1 and Example 2 need to verify 41986 and 406782 scenes respectively.

The trained deep neural network DC power flow model is used to directly map the power flow of all scenes to be verified and make branch over-limit judgments based on the results. The present invention uses the average calculation error of the phase angle of all nodes in the scene to be verified, the average calculation error of the branch power, and the accuracy of the branch over-limit judgment (verification accuracy) to judge the calculation accuracy of the method proposed in the present invention. In addition, it should be noted that since the IEEE 118 node system does not have a branch power threshold, the present invention uses the branch power threshold of the Boston 118 node system instead. The statistical results of the N-1 safety check calculated by the proposed method M1 are shown in Table 4. From Table 4, it can be seen that the average calculation error of the node phase angle of M1 in Example 1 and Example 2 is -1.0×10-5rad and 3.9×10-5rad respectively, and the average calculation error of the branch power of M1 in Example 1 and Example 2 is -1.7×10-3MW and 2.1×10-3MW respectively. The accuracy of the N-1 safety check performed by the method proposed in the present invention can reach more than 99.9% in both examples. It can be seen that the method proposed in the present invention has high fitting accuracy for power flow calculation, and the N-1 safety verification accuracy can effectively meet the needs of industrial applications.

Table 4 M1 safety check accuracy analysis

Example Mean_Pij(MW) Mean_θ(rad) Calibration 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 is a comparison of the time for calculating the safety check by the method M1 proposed in the present invention and the method M0 currently used in the industry. It can be seen from the table that the method proposed in the present invention can greatly shorten the calculation time in both examples. Among them, in Example 1, the method proposed in the present invention only takes 0.12 seconds, while the industrial method M0 takes 64.16 seconds. The method proposed in the present invention can increase the calculation speed by 535 times. In Example 2, the method proposed in the present invention also shows obvious advantages in terms of calculation speed. The calculation time for calculating the N-1 safety check using M1 and M0 is 8.50 seconds and 820.09 seconds respectively. The N-1 check speed of the method proposed in the present invention is 96 times higher than that of the industrial method. Therefore, the method proposed in the present invention can increase the N-1 safety check speed by nearly 100 times while ensuring the accuracy of the high-quality nuclear. It is worth noting that with the increase of new energy stations and the consideration of multiple time periods, the calculation time of the industrial solution method will increase exponentially, and the advantages of the method proposed in the present invention will become more and more obvious.

Table 5 Comparison of N-1 safety check speed between M0 and M1

The present invention starts from the two aspects of feature vector construction and learning strategy design, and proposes an N-1 fast verification method based on deep learning technology. The node injection power and the branch power difference before and after the branch is disconnected are constructed as input feature vectors to characterize the source load and topological structure changes, so that the deep neural network can effectively extract the important influence of source load and topological structure changes on the tidal current. In addition, after analyzing the mathematical characteristics of conventional data preprocessing methods and activation functions, the z-score normalization method is used to normalize and preprocess the tidal current samples; and then the ReLU activation function and the linear function are combined as the activation function of the deep neural network to effectively extract the tidal current characteristics, thereby forming a set of deep neural network learning strategies suitable for N-1 verification.

Claims (4)

1. A deep learning-based N-1 safety verification method with uncertainty, characterized by comprising the following steps: 1) Obtain power network data and establish an input feature vector set _Xin ; The power network data includes the power network topology and source load data of each node; 2) Preprocess the feature vector and divide it into training set and validation set; 3) Establish a deep neural network DC power flow model; 4) Use the training set and validation set to train the deep neural network DC power flow model; 5) Establish a test set with real-time data of the power network and input it into the trained deep neural network DC power flow model to obtain the power flow output feature vector Y _out ; 6) Perform security verification on the output feature vector Y _out =[θ _i ,P _ij ]; Input feature vector set _Xin = [P _i , △P _ij ] and output feature vector Y _out = [θ _i ,P _ij ]; where _Pi is a continuous feature vector, representing the total active power injected by the new energy node; θ _i is the node phase angle, _Pij is the active power flow of each branch; △P _ij is a discrete feature vector, representing the difference in active power of each branch before and after the branch is disconnected; The difference in active power of each branch before and after the branch is disconnected △P _ij is as follows:

In the formula,

It is the active power of each branch before the branch is disconnected;

is the active power of each branch after the branch is disconnected; the dimension of the difference in active power of each branch before and after the branch is disconnected △P _ij is n _branch ; n _branch is the number of system branches; The method for verifying the safety of the output characteristic vector Y _out =[θ _i ,P _ij ] is as follows: determining whether the active power flow of the branch P _ij >P _max is established, if established, determining that the branch ij is overloaded, if not established, the branch ij is in a safe state; The deep neural network DC power flow model is shown below:

In the formula,

is the feedforward transfer function of the lth layer of neurons; l = 1, 2, ..., n, n is the number of neural network layers; θ = {W, b} is the parameter to be optimized for the deep neural network; The feedforward transfer function of the l-th layer of neurons

As shown below:

Where Xl ^-1 is the input of the lth layer of neurons; _Wl and _bl are the weight matrix and offset vector between the lth layer of neurons and the l-1th layer of neurons; s is the activation function. 2. The N-1 safety verification method with uncertainty based on deep learning according to claim 1 is characterized in that: the data preprocessing method is normalization; The normalization formula is as follows:

In the formula, x _μ is the sample mean, x _δ is the sample standard deviation; x is the input sample, that is, the eigenvector; x* is the normalized data; the mean of the data x* is 0 and the variance is 1. 3. The N-1 security verification method with uncertainty based on deep learning according to claim 1, characterized in that the activation function s is as follows:

4. The N-1 safety verification method with uncertainty based on deep learning according to claim 1 is characterized in that the steps of training the deep neural network DC power flow model are as follows: 1) Input the training set into the deep neural network DC power flow model; 2) Randomly initialize the parameters θ to be optimized in the deep neural network DC power flow model; 3) Use the RMSProp algorithm to update the deep neural network parameters for the tth time, that is:

In the formula,

is the partial derivative of the mean square error loss function with respect to θ; ⊙ is the Hamiltonian multiplier; η is the learning rate; ε is a constant; r is the cumulative square gradient; ρ is the decay rate; t is the number of iterations; the initial value of t is 1; 4) Input the verification set into the deep neural network DC power flow model to determine whether the test accuracy of the verification set decreases. If so, stop the iteration. If not, determine whether the number of iterations t>t _max is established. If so, stop the iteration. If not, return to step 2); t _max is the maximum number of iterations.

Images