CN112072634B - Load flow calculation method based on load flow embedding technology - Google Patents

Abstract

The invention discloses a load flow calculation based on a load flow embedding technologyThe method comprises the following steps: determining the maximum node number N, and constructing a training set K, a verification set V and a test set T; constructing corresponding positive samples K for the training data in the step 1)⁺And negative sample K^‑(ii) a Based on the training sample K and the positive sample K in the step 2)⁺Negative sample K^‑And using a triplet-based twin neural network to obtain a tide embedded layer P after full training. The calculation method is a direct method, and when the method is finally used, only known parameters are required to be arranged according to rules and then serve as the input of a deep neural network, and the final load flow value can be obtained through multiplication of a plurality of matrixes and nonlinear operation of neurons.

Description

A Load Flow Calculation Method Based on Power Flow Embedding Technology

technical field

The invention relates to the technical field of power systems, in particular to a power flow calculation method based on a power flow embedding technology.

Background technique

With the continuous development of new energy technologies, a new energy utilization system, the Energy Internet, has emerged. The role of the Energy Internet is to integrate a series of power grid operation data with the support of artificial intelligence technologies such as big data and machine learning to predict various situations, and finally enable all machines, equipment, and systems to be dynamically adjusted in real time to improve the power grid. overall operational efficiency.

With the advent of the era of big data and the improvement of computer technology, neural networks, especially deep learning, have greatly outperformed other machine learning models in the field of artificial intelligence, and have been used in speech recognition, image classification, natural language processing and other fields. It has a wide range of applications and remarkable achievements.

Applying deep learning technology to power flow calculation is a new exploration of power flow calculation, aiming to expand the application of deep learning in power system. Power flow calculation is a very important analysis and calculation of power system, which can accurately and quickly estimate the power flow value of power grid. It is the basis of each calculation link in the power system, and it is also the premise of the stability and reliability analysis of the power system.

The application of deep learning technology to power flow calculation is a supplement to the existing traditional power flow calculation methods. Under the new situation of the energy Internet, the power grid structure involved in power flow calculation becomes more and more complex, and the algorithm is fast and convergent. The requirements are more stringent. The Gauss-Seidel iteration method is simple in principle and occupies less computer memory. However, when it is applied to a large-scale power system, the number of iterations will increase and the convergence will not occur; the Newton-Raphson method has fast convergence and high precision. high, but the Jacobian matrix needs to be recalculated in each iteration process, resulting in problems such as occupying too much computer memory and slowing the calculation speed; the fast decoupling method improves the calculation speed and memory consumption, but when Certain ill-conditioned conditions may result in non-convergence.

Power flow calculation is essentially the solution of a set of nonlinear equations, and deep learning is a tool with powerful nonlinear fitting ability to a certain extent. Therefore, it is feasible to use deep learning to solve power flow. , for this purpose, we propose a power flow calculation method based on power flow embedding technology.

SUMMARY OF THE INVENTION

The main purpose of the present invention is to provide a power flow calculation method based on the power flow embedding technology, which is simple and convenient, has a relatively fast calculation speed, can be used for online power flow calculation, and has no convergence problem, and can calculate any topology network within N nodes. trend value.

To achieve the above object, the technical scheme adopted in the present invention is:

The invention discloses a power flow calculation method based on the power flow embedding technology, comprising the following steps:

1) Determine the maximum number of nodes N and the maximum number of PV nodes K, and construct a training set K, a verification set V, and a test set T;

2) For the training data in step 1), construct corresponding positive samples K ⁺ and negative samples K ⁻ ;

3) Based on the training sample K, positive sample K ⁺ , and negative sample K ⁻ in step 2), using the triplet-based Siamese neural network, after sufficient training, the power flow embedding layer P is obtained;

4) using the power flow embedding layer in step 3) as the first hidden layer of the deep neural network, training the deep neural network on this basis, and retaining the trained deep neural network parameters;

5) For any power grid that needs to perform power flow calculation, use the deep neural network trained in step 4) to obtain the corresponding power flow solution through forward calculation.

Preferably, the step 1) further comprises:

1-1) Determine the data set size, for each data, perform the following steps:

① Determine the number of nodes n and the number of PV nodes k in the form of random numbers, and randomly number each node;

② According to graph theory and depth-first search algorithm, an n-node connected graph is generated, the impedance of each connected edge is randomly generated, and the admittance matrix is obtained and saved;

③ Define node 1 as a balanced node, nodes 2 to (k-1) as PV nodes, and the remaining nodes are not PQ nodes, and randomly generate the voltage amplitude v and phase angle θ of each node;

④ Calculate the active power p and reactive power q of each node according to the power flow equation. The description of the power flow equation is as follows:

Among them, pi and qi _represent the active and reactive power of node _i , respectively, vi represents the voltage amplitude of node _i , and G _ij and B _ij represent the element in the node admittance matrix at row i and column j, respectively. The real part and the imaginary part, θ _ij =θ _i -θ _j , represent the voltage phase difference between node i and node j, the symbol j∈i means that node i and node j are connected, including i=j Happening;

⑤ The active power p and reactive power q of the PQ node, the active power p and voltage amplitude v of the PV node, the voltage amplitude v and voltage phase angle θ of the balance node, the upper triangular matrix G of the real part of the admittance matrix and The upper triangular matrix B of the imaginary part is used as the input part of the data set, and it is used as a row in the input matrix;

⑥ Take the voltage amplitude v and voltage phase angle θ of all nodes as a row of the label matrix;

1-2) Divide the generated data set into training set K, validation set V, and test set T.

Preferably, the step 2) further comprises:

2-1) For the data ( _xi , t _i ) in each training set, perform the following steps:

保留原始导纳矩阵，根据潮流方程计算每个节点的有功功率

和无功功率

(1) Apply a small perturbation to _ti to get

Keep the original admittance matrix and calculate the active power of each node according to the power flow equation

and reactive power

保留原始导纳矩阵，根据潮流方程计算每个节点的有功功率

和无功功率

(2) Apply a large perturbation to _ti to get

Keep the original admittance matrix and calculate the active power of each node according to the power flow equation

and reactive power

以及负样本

(3) Output the corresponding positive sample

and negative samples

作为正样本集中输入矩阵以及标签矩阵的某一行，

作为负样本集中输入矩阵以及标签矩阵的某一行，输出训练集的正样本集K+与负样本集K^-。2-2) will

As a row of the input matrix and label matrix in the positive sample set,

As a row of the input matrix and the label matrix in the negative sample set, the positive sample set K+ and the negative sample set K ^- of the training set are output.

Preferably, the step 3) further comprises:

负样本集K-中

的作为孪生神经网络的输入，如附图1所示的基于三元组损失函数的孪生神经网络模型可描述为：3-1) Based on the twin neural network model of the triplet loss function as shown in Figure 1, the x i in the training set K, the x _i in the positive sample set K+

Negative sample set K-medium

As the input of the Siamese neural network, the Siamese neural network model based on the triplet loss function shown in Figure 1 can be described as:

y ₁ =Wx _i +b,

d ₁ =||y ₁ -y ₂ ||,

d ₂ =||y ₁ -y ₃ ||,

Among them, the embedding layer coefficient P consists of the final (W, b);

3-2) After the Adam algorithm is fully trained, the parameter P of the embedding layer is output, and the rules for updating the parameter x by the Adam algorithm are as follows:

m _n =β ₁ ·m _n-1 +(1-β ₁ )·g _n ,

分别表示梯度的一阶矩估计以及修正后的一阶矩估计，v_n和

分别表示梯度的二阶矩估计以及修正后的二阶矩估计。Among them, the subscript n indicates that it is in the nth iteration process, g _n indicates the gradient of f(x) at x _n , m _n and

represent the first-order moment estimates of the gradient and the modified first-order moment estimates, respectively, v _n and

represent the second-order moment estimate of the gradient and the revised second-order moment estimate, respectively.

Preferably, the step 4) further comprises:

4-1) Select a deep neural network, take the _xi in the training set sample K as its input, and the t _i in K as the label, initialize the parameters of the first hidden layer of the network as the embedding layer parameter P, and the other layers The parameters of the initialization method are selected according to the selected deep neural network. The forward calculation process of the network can be described as:

o _i = f(p, W ₁ , b ₁ , W ₂ , b ₂ , . . . , W _n , _bn , x _i ),

4-2) Use the Adam optimization method to minimize the Loss loss function, output and retain the parameters of each layer (W ₁ , b ₁ , W ₂ , b ₂ , . . . , W _n , _bn ) after sufficient training.

Preferably, the step 5) further comprises:

5-1) For any power grid, add the active power and reactive power of the PQ node, the active power and voltage amplitude of the PV node, the voltage amplitude and voltage phase angle of the balance node, and the real and imaginary parts of the admittance matrix to The input x _i of the neural network is brought into o _i =f(p,W ₁ ,b ₁ ,W ₂ ,b ₂ ,...,W _n ,b _n , _xi ), and the output o _i , that is, the voltage of all nodes Amplitude and voltage angle.

Compared with the prior art, the present invention has the following beneficial effects:

The calculation method of the present invention is a direct method. In the final use, it is only necessary to arrange the known parameters according to the rules and use them as the input of the deep neural network. Through the multiplication of several matrices and the nonlinear operation of neurons, The final power flow value can be obtained, so the method is relatively simple, the calculation speed is fast, and it can be used for online power flow calculation, and there is no convergence problem; the calculation method of the present invention can not only be used for grids with fixed topology It is used to solve the power flow of any variable topology power grid with no more than N nodes.

Description of drawings

FIG. 1 is a twin neural network model based on triplet loss function in a power flow calculation method based on power flow embedding technology of the present invention.

FIG. 2 is a fourteen-node power grid involved in a power flow calculation method based on the power flow embedding technology of the present invention.

FIG. 3 is a flow chart of generating a power grid topology in a power flow calculation method based on the power flow embedding technology of the present invention.

FIG. 4 is a five-node power grid involved in a power flow calculation method based on the power flow embedding technology of the present invention.

FIG. 5 is a residual network model with an embedded layer in a power flow calculation method based on the power flow embedding technology of the present invention.

Detailed ways

The embodiments of the present invention will be described in further detail below with reference to the accompanying drawings.

The following will take N as 14 as an example, in conjunction with the relevant drawings in the embodiments of the present invention, to clearly and completely describe the technical solutions in the embodiments of the present invention. Obviously, the described embodiments are only a part of the embodiments of the present invention. not all examples. Based on the embodiments of the present invention, all other embodiments obtained by those of ordinary skill in the art without creative efforts shall fall within the protection scope of the present invention.

The invention provides a power flow calculation method of a variable topology power grid in an N node based on deep learning, which is simple and convenient, has a fast calculation speed, can be used for online power flow calculation, and has no convergence problem. The method mainly includes the following steps:

Determine the maximum number of nodes N=14, the maximum number of PV nodes K=1, construct training set, validation set, test set, for each data, perform the following steps:

1-1) Take a random integer between 3 and N (in this example, N here is 14), and take the random integer as the number n of the grid nodes, and between 0 and K (in this example, K is taken as 1) A random integer is taken between, and the random integer is taken as the number k of PV nodes in the grid, and each node is randomly numbered. The following steps take n=5, k=1 as an example;

1-2) Generate an n-node connected graph according to graph theory and depth-first search algorithm, randomly generate the impedance of each connected edge, and save it in the form of an admittance matrix. Taking the power grid described in step 1-1) as an example, this The topology of the network can be obtained according to the algorithm shown in Figure 3, and the admittance matrix of the network can be expressed as:

最终的导纳矩阵可表示为：Among them, Y _ij =Y _ji ,

The final admittance matrix can be expressed as:

1-3) Define node 1 as a balanced node, nodes 2 to (k+1) as PV nodes, and the remaining nodes are not PQ nodes, generate the voltage amplitude v and phase angle θ of each node, and follow step 1-1) Taking the power grid described in ~1-2) as an example, the node 1 of the network is defined as a balance node, node 2 is a PV node, and nodes 3 to 5 are PQ nodes, and the voltage amplitude v and phase angle θ of each node are randomly generated. , the values of v and θ are as follows:

v=[v ₁ v ₂ v ₃ v ₄ v ₅ ],

θ=[θ ₁ θ ₂ θ ₃ θ ₄ θ ₅ ],

1-4) Calculate the active power p and reactive power q of each node according to the power flow equation, where the power flow equation is described as follows:

Taking the power grid described in steps 1-1) to 1-3) as an example, substituting v and θ of each node in step 1-3) into the power flow equation, the obtained values of active power p and reactive power q are as follows :

p=[p ₁ p ₂ p ₃ p ₄ p ₅ ],

q=[q ₁ q ₂ q ₃ q ₄ q ₅ ],

1-5) Calculate the active power p and reactive power q of the PQ node, the active power p and voltage amplitude v of the PV node, the voltage amplitude v and voltage phase angle θ of the balance node, and the upper triangle of the real part of the admittance matrix. The matrix G and the imaginary upper triangular matrix B are used as the input part of the data set, and it is used as a row in the input matrix. There are two points to note:

The admittance matrix is symmetric, and the elements on the diagonal can be obtained from other elements in the row or column, so it is only necessary to take the upper triangular element of the admittance matrix;

Since the deep learning model used is a fully connected model, the input dimensions must be consistent, and the training data will include grids with different numbers of nodes, different topologies, different numbers of grid nodes, and PV nodes. The difference of the number will cause the dimension of each data to be inconsistent. For power grids with different node numbers, we deal with the unknown power flow state by setting it to zero. For power grids with different PV nodes, we deal with K triples. Each triplet includes three quantities (p, q, v). For a power grid with k PV nodes, the k triples are taken as (p, 0, v), and the rest are taken as (p, q, 0), taking the power grid described in steps 1-1) to 1-4) as an example, its node 2 is a PV node, and this row of the input matrix can be expressed as:

Input_14=[v ₁ θ ₁ p ₂ 0v ₂ p ₃ p ₄ p ₅ 000000000q ₃ q ₄ q ₅ 000000000

G ₁₂ to G ₁₍₁₄₎ G ₂₃ to G ₂₍₁₄₎ G ₃₄ to G ₃₍₁₄₎ G ₄₅ to G ₄₍₁₄₎ G ₅₆ to G ₅₍₁₄₎ G ₆₇ to G ₆₍₁₄₎ G ₇₈ ～G ₇₍₁₄₎ G ₈₉ ～G ₈₍₁₄₎ G ₉₍₁₀₎ ～G ₉₍₁₄₎ G _(10)(11) ～G _(10)(14) G _(11)(12) ～G _{( 11)(14)} G _(12)(13) ～G _(12)(14) G _(13)(14) ～G _(13)(14) B ₁₂ ～B ₁₍₁₄₎ B ₂₃ ～B _{2(14 )} B ₃₄ to B ₃₍₁₄₎ B ₄₅ to B ₄₍₁₄₎ B ₅₆ to B ₅₍₁₄₎ B ₆₇ to B ₆₍₁₄₎ B ₇₈ to B ₇₍₁₄₎ B ₈₉ to B ₈₍₁₄₎ B ₉₍₁₀₎ ~B ₉₍₁₄₎ B _(10)(11) ~B _(10)(14) B _(11)(12) ~B _(11)(14) B _(12)(13) ~B _{( 12)(14)} B _(13)(14) ～B _(13)(14) ”

When k=0 is taken in step 1-1), that is, when node 2 is a PQ node, the row of the input matrix can be represented

shown as:

Input_14=[v ₁ θ ₁ p ₂ q ₂ 0p ₃ p ₄ p ₅ 000000000q ₃ q ₄ q ₅ 000000000

G ₁₂ to G ₁₍₁₄₎ G ₂₃ to G ₂₍₁₄₎ G ₃₄ to G ₃₍₁₄₎ G ₄₅ to G ₄₍₁₄₎ G ₅₆ to G ₅₍₁₄₎ G ₆₇ to G ₆₍₁₄₎ G ₇₈ ～G ₇₍₁₄₎ G ₈₉ ～G ₈₍₁₄₎ G ₉₍₁₀₎ ～G ₉₍₁₄₎ G _(10)(11) ～G _(10)(14) G _(11)(12) ～G _{( 11)(14)} G _(12)(13) ～G _(12)(14) G _(13)(14) ～G _(13)(14) B ₁₂ ～B ₁₍₁₄₎ B ₂₃ ～B _{2(14 )} B ₃₄ to B ₃₍₁₄₎ B ₄₅ to B ₄₍₁₄₎ B ₅₆ to B ₅₍₁₄₎ B ₆₇ to B ₆₍₁₄₎ B ₇₈ to B ₇₍₁₄₎ B ₈₉ to B ₈₍₁₄₎ B ₉₍₁₀₎ ~B ₉₍₁₄₎ B _(10)(11) ~B _(10)(14) B _(11)(12) ~B _(11)(14) B _(12)(13) ~B _{( 12)(14)} B _(13)(14) ～B _(13)(14) ],

1-6) Take the voltage amplitude v and voltage phase angle θ of all nodes as a certain row of the label matrix. Label

This row of the matrix can be represented as:

Target_14=[v ₁ v ₂ v ₃ v ₄ v ₅ v ₆ v ₇ v ₈ v ₉ v ₍₁₀₎ v ₍₁₁₎ v ₍₁₂₎ v ₍₁₃₎ v ₍₁₄₎ θ ₁ θ ₂ θ ₃ θ ₄ θ ₅ θ ₆ θ ₇ θ ₈ θ ₉ θ ₍₁₀₎ θ ₍₁₁₎ θ ₍₁₂₎ θ ₍₁₃₎ θ ₍₁₄₎ ],

Taking the power grid described in steps 1-1) to 1-4) as an example, this row of the label matrix can be expressed as:

Target_5=[v ₁ v ₂ v ₃ v ₄ v ₅ 0 0 0 0 0 0 0 0 0 θ ₁ θ ₂ θ ₃ θ ₄ θ ₅ 0 0 0 0 0 0 00 0],

The generated dataset is divided into training set, validation set, and test set. In this embodiment, the training set, the verification set and the test set are respectively 400000, 60000 and 40000;

For each training data ( _xi , t _i ) in step 2), perform the following steps:

得到

保留原始导纳矩阵，根据潮流方程计算每个节点的有功功率

和无功功率

经小扰动后得到的

有功功率

和无功功率

可表示为：3-1) Apply a small _perturbation to ti

get

Keep the original admittance matrix and calculate the active power of each node according to the power flow equation

and reactive power

obtained after a small perturbation

Active power

and reactive power

can be expressed as:

得到

保留原始导纳矩阵，根据潮流方程计算每个节点的有功功率

和无功功率

经大扰动后得到的

有功功率

和无功功率

可表示为：3-2) Apply a large _perturbation to ti

get

Retain the original admittance matrix and calculate the active power of each node according to the power flow equation

and reactive power

obtained after a large disturbance

Active power

and reactive power

can be expressed as:

分别作为正样本集中输入矩阵以及标签矩阵的某一行，

分别作为负样本集中输入矩阵以及标签矩阵的某一行；3-3) will

as a row of the input matrix and the label matrix in the positive sample set, respectively,

Respectively as a row of the input matrix and the label matrix in the negative sample set;

The twin neural network model based on triplet loss function uses Tensorflow as the platform, python as the programming language, and stochastic gradient algorithm as the optimization method. After sufficient training, the embedding layer coefficient P is obtained, as shown in Figure 1 based on triplet loss. The Siamese neural network model of the function can be described as:

y ₁ =Wx _i +b,

d ₁ =||y ₁ -y ₂ ||,

d ₂ =||y ₁ -y ₃ ||,

Among them, the embedding layer coefficient P consists of the final (W, b);

The power flow embedding layer in step 2) is used as the first hidden layer of the deep neural network, and on this basis, a deep neural network is trained to calculate the variable topology power grid with the number of nodes not greater than N. Taking the residual network as an example, the final residual network model structure with an embedded layer is shown in Figure 5;

For any power grid that needs to perform power flow calculation, use the deep neural network trained in step 5) to obtain the corresponding power flow solution. Taking the five-node power grid shown in Figure 4 as an example, perform the following steps:

6-1) After sampling the network, obtain relevant parameters, the node parameters of the five-node power grid are shown in Table 1, and the line parameters are shown in Table 2;

Table 1: Five-node grid node parameters

Table 2: Five-node grid line parameters

6-2) According to Table 1-Table 2, obtain the real part G, imaginary part B and the input Input of the deep neural network of its admittance matrix Y:

Input=[1.0501.050523.71.60000000001.01.30.8000000000

0000000000000000000000000-0.8299-0.6240000000000-0.75470000000000000000000000000000000000000000000000000000

0033.3333000000000066.6667000000000003.11203.900200000000002.64150000000000000000000000000000000000000000000000000000],

6-3) The input is used as the input of the deep neural network, and the corresponding output can be obtained after forward calculation.

The feasibility of the present invention is illustrated by the following data. Table 3 shows the training set and test set errors of several stacked autoencoders with N=50 and K=3 without embedding layers.

Table 3: Load flow calculation results based on stacked autoencoders

According to Table 3, it can be found that a network with a large number of neurons in the hidden layer and a large number of hidden layers tends to have better training and testing results. For example, by comparison, we can find the stacked autoencoder with the network structure of 2554-2000-1000-500-100 in Table 3, and we can also find the stacked autoencoder with the network structure of 2554-5000-1000-500-100 The latter has better results on both the training set and the test set; by comparing the stacked autoencoder network with the network structure of 2554-5000-100 in Table 3 and the network structure of 2554-2000-1000- 100 stack auto-encoding network, the latter is slightly lower than the former in terms of training error and test error, but the parameters required for training of the latter are much smaller than the former.

Table 4: Stacked autoencoder training and test times

Table 4 shows the training and testing time of two different structures of the three-hidden-layer stacked autoencoder, where the testing time is the time required to calculate the power flow value of the fourteen-node power grid shown in Fig. 2, as shown in Fig. 2 The 14-node power grid is shown, the grid node parameters are shown in Table 5, and the power grid line parameters are shown in Table 6;

Table 5: Fourteen-node grid node parameters

Table 6: Fourteen-node grid line parameters

For the 14-node power grid shown in Fig. 2, the time required by the Newton-Raphson method is about 0.113314s. According to the results shown in Table 4, the stack autoencoder can be used to obtain faster results. For the 14-node power grid shown in Figure 2, when the resistance value is increased by 100 times, when the Newton-Raphson method is used, the Jacobian matrix will have a singularity, which will cause its error to oscillate up and down and fail to converge. When using the stack autoencoder with the network structure of 2554-2000-1000-500-100 in Table 3 to calculate the power flow value of this network and bring the calculated result back into the power flow equation, the maximum active power can be obtained. The error is 0.03 (KVA), and the maximum error of reactive power is 0.02 (KVA). It can be seen that the present invention can provide the reference value of ill-conditioned power flow to a certain extent.

Through this embodiment, on the one hand, it proves that the deep neural network has the ability to solve the problem of power flow calculation; In order to further illustrate the capability of the embedding layer mentioned in the present invention, we will take N=14, K=1 as an example to train 4 different deep neural networks with embedded layers, namely shallow BP neural network, deep BP neural network network with a deep ReLU neural network and a residual network, and applied it to a five-node power grid as shown in Figure 4.

Tables 7-10 show the simulation results of shallow BP neural network with embedding layer, deep BP neural network, deep ReLU neural network and residual network, respectively. The unit of voltage amplitude and voltage amplitude error in the table is p.u., and the unit of voltage angle and voltage angle error is degree.

Table 7: Experimental results of power flow calculation of shallow BP neural network with embedding layer

Table 8: Experimental results of power flow calculation of deep BP neural network with embedding layer

Table 9: Experimental results of power flow calculation of deep ReLU neural network with embedding layer

Table 10: Experimental results of power flow calculation of residual network with embedded layer

Table 7 shows the experimental results of applying a trained shallow BP neural network with embedding layer to the five-node grid shown in Fig. 4. According to the voltage amplitude error and the voltage angle error, it can be considered that the shallow neural network can be used for the power flow calculation of the power grid to a certain extent, but the accuracy is poor;

Table 8 shows the experimental results of applying a trained deep BP neural network with embedding layers (ten layers) to the five-node grid shown in Fig. 4. By comparing the results between Table 7 and Table 8, the simulation effect of the deep BP neural network is far inferior to that of the shallow BP neural network, because the gradient disappearance problem of the deep BP neural network. At the same time, it also shows that the power flow embedding technology cannot solve the problem of gradient disappearance;

Table 9 shows the experimental results of applying a trained deep ReLU neural network with embedding layers (ten layers) to the five-node grid shown in Fig. 4. By comparing Table 8 and Table 9, it can be found that the error of the deep ReLU neural network is much smaller than that of the deep BP neural network of the same structure;

Table 10 shows the experimental results of applying a trained deep residual network with embedding layers (ten layers) to a five-node grid as shown in Fig. 4. By comparing Tables 7 to 10, the residual network is the best in performance.

Table 11: Simulation results of each network without embedding layer

Table 11 shows the simulation results for each network without the embedding layer. By comparing Tables 7 to 11, with the support of the power flow embedding technology, the performance of each network except the deep BP neural network has been improved to a certain extent. Even the simplest shallow BP neural network can be used. The relatively good experimental results show that the power flow embedding technology can mine the potential features of the grid input data to a certain extent, thereby further improving the expressiveness of the model.

The above are only preferred embodiments of the present invention, it should be pointed out that for those skilled in the art, without departing from the concept of the present invention, several improvements and modifications can also be made, and these improvements and modifications should also be regarded as within the protection scope of the present invention.

Claims (7)

1. A power flow calculation method based on a power flow embedding technology is characterized by comprising the following steps: the method comprises the following steps:

1) determining the maximum node number N, and constructing a training set K, a verification set V and a test set T;

2) constructing a corresponding positive sample set K for the training set K in the step 1)⁺And negative sample set K^-；

2-1) for data (x) in each training set_i，t_i) The following steps are executed:

first pair t_iApplying a small perturbation to obtain

Reserving an original admittance matrix, and calculating the active power of each node according to a load flow equation

And reactive power

Die pair of_iApplying a large perturbation

Reserving an original admittance matrix, and calculating the active power of each node according to a load flow equation

And reactive power

Outputting corresponding positive samples

And negative examples

2-2) mixing

As a certain row of the input matrix and the label matrix in the positive sample set,

as a certain row of the input matrix and the label matrix in the negative sample set, a positive sample set K of the training set K is output⁺And negative sample set K^-；

3) Based on the training set K and the positive sample set K in the step 2)⁺Negative sample set K^-And obtaining the trend after full training by using the triplet-based twin neural networkAn embedding layer P;

3-1) based on triple twin neural network model, training x in set K_iPositive sample set K⁺In

Negative sample set K^-In

As an input of the twin neural network, the forward process of the twin neural network is described as:

y₁＝Wx_i+b，

d₁＝||y₁-y₂||，

d₂＝||y₁-y₃||，

3-2) minimizing a Loss function by adopting an Adam optimization method, and outputting a weight value from an input layer to a hidden layer and an offset (W, b) as an embedded layer parameter P after full training;

4) taking the trend embedded layer in the step 3) as a first hidden layer of the deep neural network, training the deep neural network on the basis, and reserving parameters of the trained deep neural network;

5) and (4) for any power grid needing load flow calculation, using the deep neural network trained in the step 4) to obtain a corresponding load flow solution through forward calculation.

2. The power flow calculation method based on the power flow embedding technology as claimed in claim 1, wherein the step 4) further comprises:

4-1) selecting a deep neural network, and selecting a sample x in a training set K_iAs its input, t in the training set K_iAs a label, initializing parameters of a first hidden layer of the network into embedded layer parameters P, selecting an initialization mode for parameters of other layers according to the selected deep neural network, and describing a forward calculation process of the network as follows:

o_i＝f(p，W₁，b₁，W₂，b₂，…，W_n，b_n，x_i)，

4-2) adopting Adam optimization method to minimize Loss function, outputting after full training and reserving parameter (W) of each layer₁，b₁，W₂，b₂，…，W_n，b_n)。

3. The power flow calculation method based on the power flow embedding technology as claimed in claim 1, wherein the step 5) further comprises:

5-1) for any power grid, the active power and the reactive power of a PQ node, the active power and the voltage amplitude of a PV node, the voltage amplitude and the voltage phase angle of a balance node, and the input x of a neural network on the real part and the imaginary part of an admittance matrix_iBring into o_i＝f(p，W₁，b₁，W₂，b₂，…，W_n，b_n，x_i) In, output o_iI.e. voltage magnitude and voltage angle for all nodes.

4. The power flow calculation method based on the power flow embedding technology as claimed in claim 1, wherein: all training samples as well as test samples were obtained from the simulation.

5. The power flow calculation method based on the power flow embedding technology as claimed in claim 1, wherein: the twin neural network in the step 3) adopts a single hidden layer feedforward neural network.

6. The power flow calculation method based on the power flow embedding technology as claimed in claim 1, wherein: the related neural network comprises a shallow BP neural network, a deep ReLU neural network, a stacked self-encoder and a residual error network, wherein the stacked self-encoder adopts a layer-by-layer initialization method to initialize parameters and then carries out fine tuning through an Adam algorithm, the parameters of the shallow BP neural network, the deep BP neural network and the deep ReLU neural network are optimized through a back propagation algorithm, and the parameters of the residual error network are directly optimized through the Adam algorithm.

7. The power flow calculation method based on the power flow embedding technology as claimed in claim 1, wherein: the trained deep neural network can be used for calculating the tidal current value of the power grid with any topological structure in the N nodes.

Images