You are on page 1of 9

IEEE TRANSACTIONS ON POWER DELIVERY, VOL. 37, NO.

4, AUGUST 2022 3155

Parameter Identification in Power Transmission


Systems Based on Graph Convolution Network
Zhiwei Wang, Min Xia , Member, IEEE, Min Lu, Lingling Pan, and Jun Liu

Abstract—Parameter Identification plays an important role in At present, methods for obtaining transmission line param-
electric power transmission systems. Existing approaches for pa- eters can be divided into theoretical calculation method, line
rameter identification tasks typically have two limitations: (1) They parameter estimation based on supervisory control and data
generally ignored development trend of historical data, and did
not mine characteristics of corresponding power grid branches. (2) acquisition (SCADA), and line parameter estimation based on
They did not consider the constraints of power grid topology, and phasor measurement unit (PMU). The theoretical calculation
treated different branches independently. Therefore, they could not method is based on the carson model [5], which uses physical
characterize correlations between the center node and its neighbor- parameters such as self-geometric mean distance, mutual geo-
hoods. To overcome these limitations, this work proposes a multi- metric mean distance and wire material structure [6]. However,
task graph convolutional neural network (MT-GCN) which utilizes
the graph convolutional network (GCN) and the fully convolutional the theoretical analysis simplifies the model and ignores the
network (FCN) as building blocks for parameter identification. influence of uncertain factors, such as real-time temperature,
Specially, GCN can extract the structure information to enhance sag and other practical problems [7], which leads to a big gap
local feature extraction. FCN is a decoding module following GCN between the theoretical calculation results and the actual values.
module, and it is used to identify the parameters of each branch As the second method, measurement methods based on the
according to its characteristics. Compared with previous methods,
the proposed method is significantly improved in accuracy. Besides, SCADA data mainly includes two categories: augmented state
this method is robust to measurement noise and errors, and can estimation methods and measurement residual sensitivity anal-
cope with multiple conditions in real power transmission systems. ysis methods [8]. In augmented state estimation methods, the
Index Terms—Graph convolutional network, parameter
parameters are directly estimated as state quantities. Zhao [9] put
identification, power transmission systems. forward a augmented matrix estimation method and a Kalman
filtering method respectively. Zhang [10] proposed a distribu-
tion system state estimation (DSSE) method with considering
I. INTRODUCTION the characteristics of power electronic loads, and introduced
ITH the rapid development of smart grid, the operation running state variables and controllable parameters into DSSE.
W of power grid become very complicated, and the corre-
sponding regulations thereby become very strict [1], as a result,
Rehman [11] and Khazraj [12] proposed constrained weighted
least squares estimation based on virtual measurement for pa-
higher requirements for model parameters accuracy are needed. rameters estimation. Other approaches were the measurement
Due to the limitation of measurement methods, there are errors residual sensitivity analysis methods, which combined param-
in measured parameters of power transmission systems [2]. Pa- eter error sensitivity with weighted least square estimation.
rameter identification is also affected by environmental factors Yang [13] proposed a method of compensating with suspicious
such as geology, climate and uneven delivery of earth resistivity branch power flow instead of estimating parameters. Mousavi’s
along the line [3], even the change of operation mode affects the work [14] combined PMU and SCADA and employed a non-
operation parameters of power grid [4]. linear weighted least-square error algorithm for maximum-
likelihood estimation of parameters.
The third line parameter identification method based on PMU
Manuscript received May 30, 2021; revised August 11, 2021; accepted Octo-
ber 26, 2021. Date of publication November 2, 2021; date of current version July data, which could directly measure the amplitude and phase
25, 2022. This work was supported by the Science and Technology Project of angle of fundamental voltage and current at power frequency,
SGCC, named Data and Model driven Parameter Identification for the branches which provided a new way to identify actual parameters of the
in Power Grid. Paper no. TPWRD-00791-2021. (Corresponding author: Min
Xia.) line [15]. Line parameter identification was generally obtained
Zhiwei Wang and Min Xia are with the Jiangsu Key Laboratory of Big based on the π-type equivalent model of lumped parameters. Es-
Data Analysis Technology, CICAEET, Nanjing University of Information Sci- pecially, in order to reduce the influence of noise and deviation,
ence and Technology, Nanjing 210044, China (e-mail: 532047847@qq.com;
xiamin@nuist.edu.cn). different robust methods were used to improve accuracy. Com-
Min Lu is with the State Grid Zhejiang Electric Power Company, Ltd., bined with Huber criterion and multi-time data superposition, a
Hangzhou 310007, China (e-mail: lu_min@zj.sgcc.com.cn). recursive estimation of window sliding population robust least
Lingling Pan and Jun Liu are with the China Electric Power Research Institute
Company, Ltd., Nanjing 210003, China (e-mail: panlingling@epri.sgcc.com.cn; squares was proposed by Ye [16]. Rahman [17] put forward
liu-jun2@epri.sgcc.com.cn). the improved binary particle swarm (IBPSO) method, which
Color versions of one or more figures in this article are available at converged faster and managed to maximize the measurement
https://doi.org/10.1109/TPWRD.2021.3124528.
Digital Object Identifier 10.1109/TPWRD.2021.3124528 redundancy, and the method was verified in IEEE-30 bus system.

0885-8977 © 2021 IEEE. Personal use is permitted, but republication/redistribution requires IEEE permission.
See https://www.ieee.org/publications/rights/index.html for more information.

Authorized licensed use limited to: NATIONAL INSTITUTE OF TECHNOLOGY WARANGAL. Downloaded on July 27,2022 at 11:34:49 UTC from IEEE Xplore. Restrictions apply.
3156 IEEE TRANSACTIONS ON POWER DELIVERY, VOL. 37, NO. 4, AUGUST 2022

Hu’s work [18] used the adaptive IGG criterion to eliminate the text recognition is a breakthrough. As Euclidean data, voice,
influence of large deviation data. In Routtenberg’s work [19], a text and image can be well adapted to a CNN model. However,
general model for the PMU output with a downsampled negative grid nodes are spatially distributed. As non Euclidean data,
sequence was presented and the method took into account of it is difficult to input them directly into a CNN model [29].
computation complexity and imbalance tolerance. Li [20] and Although the input data can be input into a CNN model by
Bretas [21] combined SCADA data and PMU data to improve being spliced into two-dimension, in fact, this method will split
estimation accuracy. Mahmood’s work [22] proved that using the connection between nodes and introduce noise due to mixed
Kalman filter for real-time data processing in PMU measure- data. To solve the above problems, a multi-task regression model
ments may contain errors and bad data. In Xue’s work [23], based on a graph convolution neural network (MT-GCN) is
a new adaptive robust identification method was proposed for proposed. As an extension of a convolutional neural network,
obtaining accuracy parameters in the case of large random this method can deal with non Euclidean data. The adjacency
noises in PMU measurements, which effectively improved the matrix is constructed according to the topological structure of
accuracy. branches of the power transmission system. Experiments show
Although there are many ways to obtain the transmission line that the proposed deep learning method is superior to existing
parameters, the branch parameter identification method mostly experimental methods, it is able to consider the topology and
adopted the line parameter identification technology based on characteristics of circuit nodes at the same time, and it is able
SCADA data in actual practice [24]. However, the traditional to accurately fit the parameters of each branch. At the same
branch parameter identification methods were mainly based on time, it has good robustness and is not easily affected by noise.
single model-driven calculation, which had some problems, such Besides, the model proposed in this paper can be easily applied
as poor numerical stability, easy divergence and interference or extended to many tasks related to data measurement and
from residual pollution. Yang’s work [25] proposed a bad data processing in modern power systems.
detection method based on spectral clustering, which effectively The rest of this paper is as follows: In the second part,
improved data quality. In addition, the cross section of power we introduce the main structure and principle of the model,
grid branch parameter identification was determined and se- including backbone network and branch network, and briefly
lected by manual experience, which was random and could not introduce the data set. In the third part, the experimental process
effectively generate actual power grid branch parameters. At the is introduced in detail, including the setting of hyperparame-
same time, due to the limitation of traditional branch parameter ters. The proposed method is compared with several machine
identification methods, massive historical data of power system learning methods and deep learning methods, the generalization
was not effectively utilized. performance of the model is also analyzed. In the fourth part,
Due to the limitation of traditional methods, machine learn- the influence of K-order neighborhood of MT-GCN model on
ing methods could mine the characteristics of data and apply model performance is introduced in detail. Finally, our model is
in branch parameter identification, such as linear regression summarized, and the future research direction is put forward.
(LR), support vector regression (SVR) and random forest (RF),
etc [24]. For example, in Zhang’s work [26], the accurate estima- II. PARAMETER IDENTIFICATION BASED ON MULTI-TASK
tion of line parameters is realized by using LR and model driven GRAPH CONVOLUTION NEURAL NETWORK
method. Although these methods are better than the previous This section firstly introduces task requirements, then dis-
methods, they only focused on the node itself, they didn’t paid cusses the idea of spectral convolutions in graph convolution
sufficient attention to connections between grid nodes. These neural networks, and constructs a MT-GCN model for branch
methods had the following shortcomings: firstly, the machine parameter identification for power transmission systems.
learning methods only considered the historical data of the node
itself. As far as LR method and SVR method, their robustness A. Branch Parameter Identification Task
were quite poor due to linearity. When there was a little noise in
the data, there would be a great deviation in prediction results. Given the topology and features of each branch, our goal is
Secondly, even RF, a bagging method, separated the links from to predict the target parameters of each branch in the power
nodes, even though it could improve the robustness of the model transmission systems. For a given branch of a power transmis-
through voting mechanism. However, it would fall into the sion system, i and j represent the nodes at both ends of the
predicament of over-fitting in some noisy branches. It can be k-th branch. The measured active powers Pik and Pjk at both
concluded that considering the connection between nodes in ends of the branch, the measured reactive powers Qki and Qkj ,
the process of parameter identification helps extract the global the measured voltages Uik and Ujk , and the ground susceptance
characteristics of power grid branches, and regression prediction y k of the k-th branch can be obtained. The k-th branch input
on this basis can effectively resist the noise and disturbance of X k = (Pik , Pjk , Qki , Qkj , Uik , Ujk , y k ) ∈ R7k . The input of the
local branches, it is more accurate and robust. proposed model can be expressed as X ∈ Rn×7 , n represents the
Recently, deep learning methods is very popular in various number of nodes. In this study, the loss function1 is constructed
fields [27], however, among existing parameter identification according to the relationship in the π-type equivalent circuit.
methods, there is little literature considers the topology of
power grid, they are more based on single branch historical 1 Loss function is a function constructed according to the difference between
data to predict parameter [28]. The successful implementation predicted data and label data, which is usually used to measure the quality of a
of convolutional neural networks (CNN) in voice, image and model.

Authorized licensed use limited to: NATIONAL INSTITUTE OF TECHNOLOGY WARANGAL. Downloaded on July 27,2022 at 11:34:49 UTC from IEEE Xplore. Restrictions apply.
WANG et al.: PARAMETER IDENTIFICATION IN POWER TRANSMISSION SYSTEMS 3157

Fig. 1. Multi-Task graph convolution neural network structure figure (In the
structure of MT-GCN model, each branch is connected with two full connection
layers after the main backbone graph convolution layers).

The following (1) and (2) can be obtained for the k-th branch
through the derivation of π-type equivalent circuit: and FCN model as the branch network to complete the specific
2 2 node task, thus realizing all branch parameter identification
(Pik
+ Pjk )[Pik
+ (Qki+ Uik y k )2 ]
gk = 2 2 2
together.
(Uik [(Pik + Pjk )2 + [Qki + Qkj + (Uik + Ujk )y k ]2 ]) In order to introduce the MT-GCN model, the spectral con-
(1) volution theory is firstly introduced. Assuming that there is an
2 2 2 2
undirected graph G ∈ (V, E, A), where V is the set of n vertices,
k
−[Qki + Qkj + (Uik + Ujk )y k ][Pik + (Qki + Uik y k )2 ] and Vi ∈ V , E is the set of edges in the graph, and Eij ∈ E.
b = . 
2 2 2
(Uik [(Pik + Pjk )2 + [Qki + Qkj + (Uik + Ujk )y k ]2 ]) A ∈ Rn×n is an adjacency matrix, and Dii = j Aij is the
(2) degree matrix. Δp = D − A represents Laplacian matrix on
the unnormalized graph. The Laplace matrix on the graph is
In the above equations, g k is branch conductance and bk is line symmetrically normalized as follows:
susceptance. According to the above equations derived from 1 1 1 1
Δ = D− 2 Δp D− 2 = I − D− 2 AD− 2 , (3)
π-type equivalent circuit, the true values of branch conductance
and line susceptance can be calculated. Therefore, the branch where I is an identity matrix, the Laplacian matrix can be
parameter identification can be defined as a multi-objective spectrally decomposed after symmetric normalization, and the
regression problem. More specifically, given X k , g̃ k and b̃k of Laplacian matrix can be spectrally decomposed, that is eigen-
a branch are obtained by f (X k ), f (x) represents a mapping of decomposition, and the process of eigendecomposition is de-
input X though graph convolution network (GCN) model. The scribed as follows:
algorithm uses the loss function to optimize the output though ⎛ ⎞
λ1 · · · 0
fully connection network (FCN) and obtains the final branch ⎜ ⎟
parameters. The MT-GCN can effectively use the structural Δ = U ΛU −1 = U ⎝ ... . . . ... ⎠ = U ΛU T , (4)
information of the network, construct the relationship between 0 · · · λn
branches, and implicitly fit the structural constraints of the → → →
where U = (u1 , u2 , . . . , un ) is a matrix with column vectors
network, thus improving the accuracy of branch identification
as unit eigenvectors, Λ is a diagonal matrix composed of n
in power transmission system.
eigenvalues Since U is an orthogonal matrix, symmetric normal-
ization Laplacian matrix can be written as Δ = U ΛU T . Similar
B. MT-GCN Model for Parameter Identification
to Fourier transform in Euclidean space, Fourier transform on
The objective of proposed method is to learn a better rep- graph can be defined by eigenvector of Laplace matrix. With
resentation by integrating local features and global topology respect to the signal X ∈ Rn×7 of the vertex on graph G, the
adequately. As illustrated in Fig. 1, MT-GCN has two parts. The Fourier transform on the graph is expressed as X̂ = U T X. In
input X of each branch has 7-dimensional temporal features, the same way, the inverse Fourier transform of X on the graph
and it can be accumulated by graph convolution layers of GCN. is expressed as X = U X̂. Thus, the convolution operation is
The branch network consists of fully connected layers, after extended to the graph according to (5):
feature extraction and aggregation, the output is integrated with
X ∗ g = U ((U T g) ◦ (U T X)), (5)
global information and will be diverted to the branch network for
parameter identification. That is to say, the proposed MT-GCN where ∗ means convolution, and ◦ means the inner product of
model takes GCN as the backbone to aggregate node features, two vectors. If U T g is regarded as a learnable convolution kernel,

Authorized licensed use limited to: NATIONAL INSTITUTE OF TECHNOLOGY WARANGAL. Downloaded on July 27,2022 at 11:34:49 UTC from IEEE Xplore. Restrictions apply.
3158 IEEE TRANSACTIONS ON POWER DELIVERY, VOL. 37, NO. 4, AUGUST 2022

Fig. 3. Branch topology diagram of power transmission system (structural


representation diagram).

structure of the graph. According to the topological structure


of node i and node j, the adjacency matrix A ∈ Rn×n can be
Fig. 2. Training flow chart of MT-GCN algorithm.
obtained, but as far as power transmission system parameters
are concerned, Pi , Pj , Qi , Qj , Ui , Uj are the parameters on
it is written as gθ . Then the convolution formula on the graph is the nodes, and y is used as the parameter of node connection.
X ∗ g = U gθ U T X. If gθ = U T g is defined as a filter, that is to That target parameters b and g are the parameters of branch
say, gθ = diag(U T g), which is written in the form of a diagonal connection in the power transmission system, they are difficult
matrix, then (6) is obtained: to be modeled by traditional GCN. Our approach is to unify
⎛ ⎞ the parameters Pi , Pj , Qi , Qj , Ui and Uj at the branch Eij ,
ĝ(λ1 ) · · · 0 and to add the feature y of Eij to form seven features for
⎜ . .. ⎟
X ∗ g = U gθ U T X = U ⎜ ⎝ .
. ..
. ⎟ T
. ⎠ U X, (6)
input. In this case, all features converge on the edge, and the
edge is regarded as a node, the node thereby is regarded as a
0 · · · ĝ(λn ) connecting line of the edge. As a result, the adjacency matrix
one of the main disadvantages of (6) is that when the convolution is constructed to complete the model input of the MT-GCN. In
kernel needs n parameters, that is to say, if there are n nodes in addition, the connection between nodes of transmission system
the graph, and the computation of feature decomposition is very is not only single line connection, but also multi line connection,
large, then the computational complexity of the feature matrix which poses a challenge to our model construction. Assuming
U is O(n2 ). This work uses a first-order approximate ChebShev that a node has n branches, we need to divide the node Vi
polynomial [30] to fit graph convolution kernel. According to into Vi1 , Vi2 ,... Vin . We split the multi-connected nodes into
the definition of ChebShev polynomial, the convolution of graph single-connected nodes, and then complete the construction of
neural network can be approximately defined as (7): adjacency matrix. This method is suitable for double or more
branches.
1 1 1 1
X ∗ gθ = θ(IN + D− 2 AD− 2 )X = θ(D̃− 2 ÃD̃− 2 )X, (7)

where à = A + IN , Dii = j Ãij . The spectral convolu- C. Introduction of the Power Transmission System Branch
tion of (7) applied to our graph is H l+1 = f (H l , A) = Parameter Data
1 1
δ(D̃− 2 AD̃− 2 H (l) W (l) ), H is the feture map of each layer. As The data set of power transmission system used in this work
far as graph convolution layers are concerned, the mapping of comes from China Electric Power Research Institute. A topology
1 1
each layer is H (l+1) = δ(D̃− 2 AD̃− 2 H (l) W (l) ). When l = 0, is chosen as an example to show the data composition. We trans-
(l) E×F
H is the input feature X, X ∈ R , in which E reprsents formed its actual topological structure diagram into a structural
the number of edges in the graph, and F represents the feature representation (as shown in Fig. 3). It can be seen from the
number of each edge in the graph. The output of the last layer Fig. 3 that the ring network line consists of eleven topological
is flattened into a vector and transmitted to the fully connection nodes. Yellow nodes are connected by single branch lines, and
layers. Among them the output Y˜k of the k-th branch parameter pink nodes are connected by double branch lines. Topology only
identification of power transmission system can be obtained by indicates how they are connected, not where they are actually
solving Y˜k = F Ck (Xj ), where F Ck represents the mapping located.
function of the k-th branch. Xj is the j-th fully connected layer. The recorded data is in time sequence. Each node contains
The whole MT-GCN algorithm training flowchart is shown in seven features, and the features of all nodes constitute a group
Fig. 2. of graph structure data. The data set consists of 1635 groups of
In addition to setting the graph convolution multi-task model data, and the recorder recorded the data once every 1 minute.
according to parameter identification tasks, there are two other Among them, 1200 sets of data are selected as training sets, 200
innovations. Firstly, in a graph convolution neural network, sets of data are used as verification sets, and the remaining 235
traditional methods construct adjacency matrix according to the sets of data are used as testing sets.

Authorized licensed use limited to: NATIONAL INSTITUTE OF TECHNOLOGY WARANGAL. Downloaded on July 27,2022 at 11:34:49 UTC from IEEE Xplore. Restrictions apply.
WANG et al.: PARAMETER IDENTIFICATION IN POWER TRANSMISSION SYSTEMS 3159

Fig. 4. Visualization diagram of branch data loss (The horizontal axis repre-
sents the first 200 groups of data, and the vertical axis represents 7 features of 6
branches, with a total of 42 features. The picture shows the result of randomly
losing the whole branch feature).

Fig. 5. Visualization diagram of node feature data loss (The horizontal axis
and vertical axis are the same as the above figure, and the picture shows the Fig. 6. Correlation coefficient diagram of 7 features.
result of randomly losing a certain feature of a branch).

between geoelectric y and several other features is very weak.


For our training set and testing set, considering the problems
Besides, we found that Pi and Pj , Qi and Qj , Ui and Uj , Uj
that may be encountered in daily power supply system, we added
and Qi have strong correlation. The proposed MT-GCN model
three kinds of noises. The first kind is Gaussian noise, which
aggregates these features with potential connections, extracts
makes the signal-to-noise ratio (SNR) from 30 dB to 50 dB (i.e.,
these features from the graph, and fits the predicted results.
we design gaussian noise with reference to brown’s work [31]).
We also compared our model with four baseline models,
The second kind noise considers the loss of node data, it means
including three shallow machine learning algorithms commonly
the whole edge data is unavailable. We visualize 200 groups in
used in practice (LinearRegression, SVR, RF) and one deep
Fig. 4. It can be seen that the white part is normal data, and the
learning algorithm FCN. In this work, we have implemented
red part is missing data. We can see that each column consists of
all these algorithms to verify the superior performance of our
6 edges, and each edge contains 7 features. We set the probability
model.
of edge feature loss to be 1/6, that is to say, we set all the data
In this paper, SVR adopted the radial basis convolution kernel
of that edge to 0.
support vector regression machine with C = 100 and γ = 0.1,
The third noise considers feature anomaly, and it means a
and python scikit-learn library [40] is used to realize the support
feature is different from its normal value, and the data loss can be
vector regression machine. The number of trees is set to 300
shown in Fig. 5. We set the probability of data loss to 0.01, and the
for RF, the minimum number of samples per leaf is set to 35,
missing data is set to 0 in data processing, which is represented
and the minimum number of samples required for splitting is
by red in dots (i.e., the above two kinds of noises are inspired by
set to 3. The fully connected neural network is usually used as
chen’s work [32], and this is a fault situation that is likely to occur
the baseline model. The fully connected neural network model
in real practice). According to these three kinds of abnormal
is divided into two layers, with hidden nodes being 128 and 64
conditions, we processed all the data with noise. Results showed
respectively. The linear rectification function (ReLU) is used
that the proposed MT-GCN model achieved higher accuracy and
as the activation function. After five cross-verifications, we got
stronger robustness compared with traditional methods.
the hyper-parameters of SVR, RF. For FCN model, 10% of the
training data was used to calculate the hyperparameters.
III. EXPERIMENT RESULTS AND DISCUSSION In order to prove the reliability of our method in real operation,
In this section, we perform comparison experiments. Com- we add noise and error to the measurement and compare the
pared with the baseline model, our proposed MT-GCN model performance of different models.
has higher accuracy and better robustness.
B. Comparison of Branch Parameter Identification
A. Implementation Details and Baseline Model Performance of Various Models
As illustrated in Fig. 1, the model has two graph convolution The branch parameter identification results of evaluation
layers, the first layer has 512 convolution kernels, and the second experiment are shown in Tables I and II. Table I shows the
layer has only 384 convolution kernels. There are two fully identification results of various models for line susceptance b.
connected layers with 256 and 128 hidden nodes following We use the root mean square error (RMSE) as the evaluation
the second layer. The initial learning rate is 0.0002, the Adam index. For the MT-GCN model, we repeated the test five times
optimizer is used, and the number of training iterations is 100. and recorded the average value of RMSE. According to Table I,
We use PyTorch to implement our model. our proposed MT-GCN yields the best results in most cases,
As shown in Fig. 6, we analyzed the correlation of seven and it shows that the LinearRegression algorithm, which usually
features of the input data, which shows that the correlation used in practice, performs badly due to overfitting.

Authorized licensed use limited to: NATIONAL INSTITUTE OF TECHNOLOGY WARANGAL. Downloaded on July 27,2022 at 11:34:49 UTC from IEEE Xplore. Restrictions apply.
3160 IEEE TRANSACTIONS ON POWER DELIVERY, VOL. 37, NO. 4, AUGUST 2022

TABLE I
PARAMETER IDENTIFICATION ERROR(RMSE) OF LINE SUSCEPTANCE b OF THE MODEL UNDER DIFFERENT NOISE CORRECTION CONDITIONSS

TABLE II
PARAMETER IDENTIFICATION ERROR(RMSE) OF BRANCH CONDUCTANCE g OF THE MODEL UNDER DIFFERENT NOISE CORRECTION CONDITIONS

Table II shows parameter identification results of branch Fig. 7 shows the root mean square error (RMSE) and mean
conductance g. The MT-GCN model can achieve the highest absolute error (MAE) diagrams of line susceptance b and branch
fitting accuracy in most cases. The SVR model and the RF model conductance g. As illustrated in Fig. 7(a)–(f) are the fitting curves
perform well in some error cases, but the lost node features have of line susceptance b with different noise, and (g)–(I) are the
a great influence on RF model, making its accuracy far lower fitting curves of branch conductance g with different noise, the
than the above machine learning models. Similar to Table I, the number of training epochs is set to 100. Fig. 7 shows the gradual
LinearRegression algorithm performs badly in robustness test convergence process of evaluation indexes MAE and RMSE
because of overfitting. It may be because that the complexity with the increase of the number of training epochs, where (f)
of data sets increases, the accuracy of LightGBM model is far and (I) are the fitting curves with adding three mixed noises
lower than those of similar machine learning models. In addition, of Noise50 dB, DropData and DropNode. It can be seen that
it can be found from the above table that the identification result the MT-GCN model has the best results. The method proposed
of CNN model is not ideal. As a deep learning model, we suspect can effectively utilize the structural characteristics of power
that the noise may be caused by splicing data into two-dimension grid, which makes the method robust to noise, information loss
in data processing. and other malicious conditions. The method can improve the
According to Tables I and II, it can be observed that the accuracy and reliability of parameter estimation, so it has many
MT-GCN model has superior performance with errors and advantages in practice.
noises to the data, which indicates that the MT-GCN model
C. Influence of K-Order Neighborhood on Graph Neural
can learn and predict the potential representations of nodes
Network
adjacent to nodes. The experimental results show the robustness
of MT-GCN model, and show that this robustness in MT-GCN According to the above analysis, it can be found that the
model can deal with other types of errors and faults in practice. graph convolution layer is actually a local filter, which is a linear

Authorized licensed use limited to: NATIONAL INSTITUTE OF TECHNOLOGY WARANGAL. Downloaded on July 27,2022 at 11:34:49 UTC from IEEE Xplore. Restrictions apply.
WANG et al.: PARAMETER IDENTIFICATION IN POWER TRANSMISSION SYSTEMS 3161

Fig. 7. Fitting curves of FCN, CNN and GCN on target parmeters under different noise correction conditions. RMSE and MAE are selected by indicators. (a)-(f)
is about line susceptance b, and (g)-(I) is about branch conductance g. (f) represents the superposition of noise50 dB, DropNode and DropData on line susceptance
b. (I) represents the superposition of Noise50 dB, DropNode and DropData on branch conductance g.

Fig. 8. Neighborhood nodes when k = 1 and k = 2.

combination of neighborhood eigenvectors. We can determine


the K-order neighborhood of vertex aggregation according to
the set number of graph convolution layers, as shown in the Fig. 9. Error comparative experiment of k-order neighbors of line susceptance
Fig. 8, when k = 1 and k = 2. We set up a graph convolution b without increasing noise.
neural network with at most 8 graph convolution layers and
analyzed them. The mean absolute error and mean square error
curve are shown in Fig. 9, it shows that the mean absolute error operation can aggregate the features of a vertex and its surround-
and root mean square error curve become larger with the increase ing points, but the aggregation of deep GCN also brings the
of the number of layers in the drawing stack. As can been seen problem of over-fitting, which makes it difficult to distinguish
from Fig. 9 that when k = 1,2 and 3, the MT-GCN model has the the features between nodes. In other words, when the number of
best performance, and its accuracy is the highest. However, when stacked layers is too large, the features of a node will aggregate
the neighborhood is larger than 4, the accuracy of the model will more and more neighbors’ features, making the node and its
decrease and its generalization performance will deteriorate. neighbors similar, and thus increasing the similarity between
According to our analysis, the graph convolution in GCN classes. It will naturally lead to the decline of the accuracy and
model is a special form of Laplace smoothing. This smoothing generalization performance.

Authorized licensed use limited to: NATIONAL INSTITUTE OF TECHNOLOGY WARANGAL. Downloaded on July 27,2022 at 11:34:49 UTC from IEEE Xplore. Restrictions apply.
3162 IEEE TRANSACTIONS ON POWER DELIVERY, VOL. 37, NO. 4, AUGUST 2022

IV. CONCLUSION [13] K. Yang, Y. Gong, P. Zhang, and Z. Liu, “A reactive power compensation
method based on tracing the power flow and loss function of power
With the rapid development of UHV AC/DC hybrid power system,” in Proc. 5th Int. Conf. Electric Utility Deregulation Restructuring
grid system, higher requirements are needed for real-time anal- Power Technol. (DRPT), 2015, pp. 615–619.
[14] S. S. Mousavi-Seyedi, F. Aminifar, and S. Afsharnia, “Parameter estima-
ysis. As the first to apply graph neural network theory to power tion of multiterminal transmission lines using joint PMU and scada data,”
transmission system, this paper proposes a novel multi-task IEEE Trans. Power Del., vol. 30, no. 3, pp. 1077–1085, Jun. 2015.
graph convolutional neural network (MT-GCN) to identify the [15] X. Xie, Y. Xin, J. Xiao, J. Wu, and Y. Han, “WAMS applications in Chinese
power systems,” IEEE Power Energy Mag., vol. 4, no. 1, pp. 54–63,
branch parameters of the whole grid nodes. The MT-GCN uses Jan./Feb. 2006.
graph convolutional network (GCN) and fully convolutional [16] C. Ye, S. Feng, P. Xu, and J. Liu, “Transmission line parameter identifica-
network (FCN) as building blocks. Specially, GCN is used tion considering non-synchronized time of fault recording information,”
in Proc. 2nd IEEE Adv. Inf. Manage., Communicates, Electron. Automat.
to gather the information of neighboring nodes. It can extract Control Conf. (IMCEC), 2018, pp. 1749–1753.
the structure information of the power transmission systems to [17] N. H. Rahman, A. F. Zobaa, and M. Theodoridis, “Improved BPSO for
enhance the local feature extraction. FCN is designed to decode optimal PMU placement,” in Proc. 50th Int. Universities Power Eng. Conf.
(UPEC), 2015, pp. 1–4.
the fusion information of GCN module according to multiple [18] Z. Hu and Y. Chen, “New method of live line measuring the inductance
fully connected neural networks. Experimental results verify the parameters of transmission lines based on GPS technology,” IEEE Trans.
effectiveness and robustness of our proposed MT-GCN. Power Del., vol. 23, no. 3, pp. 1288–1295, Jul. 2008.
[19] T. Routtenberg, R. Concepcion, and L. Tong, “PMU-based detection of
This paper puts forward a flexible and widely applicable voltage imbalances with tolerance constraints,” IEEE Trans. Power Del.,
framework, which can be used in branch parameter identification vol. 32, no. 1, pp. 484–494, Feb. 2017.
for the power transmission system, and it opens up a new [20] J. Li, X. Wang, X. Ren, Y. Zhang, and F. Zhang, “Augmented state esti-
mation method for fault location based on on-line parameter identification
research direction. However, there are still some shortcomings of PMU measurement data,” in Proc. IEEE 2nd Int. Elect. Energy Conf.
in parameter identification of the power transmission system, (CIEEC), 2018, pp. 105–109.
which are as follows: (1) The faults encountered in real practice [21] N. G. Bretas and A. C. Martins, “Transmission line parameter error
detection, identification and correction with geometrical view: Topological
are more complicated, and it is still a problem that if the MT- errors,” in Proc. IEEE Eindhoven PowerTech, 2015, pp. 1–4.
GCN model can effectively deal with; (2) The adjacency matrix [22] F. Mahmood, H. Hooshyar, J. Lavenius, A. Bidadfar, P. Lund,
of GCN model needs to be reconstructed in case of node change and L. Vanfretti, “Real-time reduced steady-state model synthesis of active
distribution networks using PMU measurements,” IEEE Trans. Power Del.,
or network reconfiguration in the power transmission system vol. 32, no. 1, pp. 546–555, Feb. 2017.
topology, and it is not known whether the MT-GCN model can [23] A. Xue et al., “A new robust identification method for transmission line
adapt to this dynamic topology model. parameters based on adaline and IGG method,” IEEE Access, vol. 8,
pp. 132960–132969, 2020.
[24] Z. Yun, X. Cui, and K. Ma, “Online thevenin equivalent parameter identi-
fication method of large power grids using LU factorization,” IEEE Trans.
REFERENCES Power Syst., vol. 34, no. 6, pp. 4464–4475, Nov. 2019.
[25] Z. Yang, H. Liu, T. Bi, and Q. Yang, “Bad data detection algorithm for
[1] A. G. Phadke and J. S. Thorp, Synchronized Phasor Measurements and PMU based on spectral clustering,” J. Modern Power Syst. Clean Energy,
Their Applications. Berlin, Germany: Springer, 2008. vol. 8, no. 3, pp. 473–483, 2020.
[2] J. Zhu, L. Mu, D. Ma, and X. Zhang, “Faulty line identification method [26] J. Zhang, Y. Wang, Y. Weng, and N. Zhang, “Topology identification and
based on Bayesian optimization for distribution network,” IEEE Access, line parameter estimation for non-PMU distribution network: A numer-
vol. 9, pp. 83175–83184, 2021. ical method,” IEEE Trans. Smart Grid, vol. 11, no. 5, pp. 4440–4453,
[3] R. Yao and K. Sun, “Toward simulation and risk assessment of weather- Sep. 2020.
related outages,” IEEE Trans. Smart Grid, vol. 10, no. 4, pp. 4391–4400, [27] L. Duan, J. Hu, G. Zhao, K. Chen, J. He, and S. X. Wang, “Identification
Jul. 2019. of partial discharge defects based on deep learning method,” IEEE Trans.
[4] J. De La Ree, V. Centeno, J. S. Thorp, and A. G. Phadke, “Synchronized Power Del., vol. 34, no. 4, pp. 1557–1568, Aug. 2019.
phasor measurement applications in power systems,” IEEE Trans. Smart [28] G. Hafeez, K. S. Alimgeer, and I. Khan, “Electric load forecasting based
Grid, vol. 1, no. 1, pp. 20–27, Jun. 2010. on deep learning and optimized by heuristic algorithm in smart grid,” Appl.
[5] J. R. Carson, “Wave propagation in overhead wires with ground return,” Energy, vol. 269, 2020, Art. no. 114915.
Bell Syst. Tech. J., vol. 5, no. 4, pp. 539–554, 1926. [29] Z. Wu, S. Pan, F. Chen, G. Long, C. Zhang, and P. S. Yu, “A comprehensive
[6] K. A. Clements, G. W. Woodzell, and R. C. Burchett, “A new method for survey on graph neural networks,” IEEE Trans. Neural Netw. Learn. Syst.,
solving equality-constrained power system static-state estimation,” IEEE vol. 32, no. 1, pp. 4–24, Jan. 2020.
Trans. Power Syst., vol. 5, no. 4, pp. 1260–1266, Nov. 1990. [30] T. N. Kipf and M. Welling, “Semi-supervised classification with graph
[7] Y. Huadong and Q. Yannuo, “Research on target selection simulation convolutional networks,” 2016, arXiv:1609.02907.
based on multidimensional feature parameter identification,” in Proc. Chin. [31] M. Brown, M. Biswal, S. Brahma, S. J. Ranade, and H. Cao, “Character-
Control Decis. Conf. (CCDC), 2019, pp. 533–536. izing and quantifying noise in PMU data,” in Proc. IEEE Power Energy
[8] A. Abur and A. G. Exposito, Power System State Estimation: Theory and Soc. Gen. Meeting (PESGM), 2016, pp. 1–5.
Implementation. Boca Raton, FL, USA: CRC Press, 2004. [32] K. Chen, J. Hu, Y. Zhang, Z. Yu, and J. He, “Fault location in power
[9] J. Zhao et al., “Power system dynamic state estimation: Motivations, distribution systems via deep graph convolutional networks,” IEEE J. Sel.
definitions, methodologies, and future work,” IEEE Trans. Power Syst., Areas Commun., vol. 38, no. 1, pp. 119–131, Jan. 2020.
vol. 34, no. 4, pp. 3188–3198, Jul. 2019. [33] G. Ke et al., “LightGBM: A highly efficient gradient boosting decision
[10] B. Zhang et al., “A distribution system state estimation analysis consid- tree,” in Proc. Adv. Neural Inf. Process. Syst., 2017, pp. 3146–3154.
ering the dynamic load effect,” in Proc. TENCON IEEE Region 10 Conf., [34] T. Chen and C. Guestrin, “XGBoost: A scalable tree boosting system,” in
2018, pp. 2485–2489. Proc. 22nd ACM SIGKDD Int. Conf. Knowl. Discov. Data Mining, 2016,
[11] B. Rehman, M. Ahmad, and J. Hussain, “Analysis of power system pp. 785–794.
harmonics using singular value decomposition, least square estimation [35] Y. Freund et al., “Experiments with a new boosting algorithm,” in Proc.
and FFT,” in Proc. Int. Conf. Energy Syst. Policies (ICESP), 2014, 13th Int. Conf. Int. Conf. Mach. Learn., 1996, pp. 148–156.
pp. 1–5. [36] R. Tibshirani, “Regression shrinkage and selection via the lasso,” J. Roy.
[12] H. Khazraj, F. F. da Silva, and C. L. Bak, “A performance comparison Stat. Soc.: Ser. B. (Methodol.), vol. 58, no. 1, pp. 267–288, 1996.
between extended Kalman filter and unscented Kalman filter in power [37] S. L. Cessie and J. C. Van Houwelingen, “Ridge estimators in logistic
system dynamic state estimation,” in Proc. 51st Int. Universities Power regression,” J. Roy. Stat. Soc.: Ser. C (Applied Statistics), vol. 41, no. 1,
Eng. Conf. (UPEC), 2016, pp. 1–6. pp. 191–201, 1992.

Authorized licensed use limited to: NATIONAL INSTITUTE OF TECHNOLOGY WARANGAL. Downloaded on July 27,2022 at 11:34:49 UTC from IEEE Xplore. Restrictions apply.
WANG et al.: PARAMETER IDENTIFICATION IN POWER TRANSMISSION SYSTEMS 3163

[38] J. Friedman et al., “Additive logistic regression: A statistical view of Min Lu received the master’s degree in computer
boosting (with discussion and a rejoinder by the authors),” Ann. Statist., and its application from North China Electric Power
vol. 28, no. 2, pp. 337–407, 2000. University, BaoDing, China, in 2005. Since 2010,
[39] L. Breiman, “Bagging predictors,” Mach. Learn., vol. 24, no. 2, she has been a Senior Engineer of electric power.
pp. 123–140, 1996. Her current research interests cover power system
[40] F. Pedregosa et al., “Scikit-learn: Machine learning in python,” J. Mach. automation technology, distributed AI technology,
Learn. Res., vol. 12, pp. 2825–2830, 2011. big data analysis.

Lingling Pan received the master’s degree in power


Zhiwei Wang received the B.S. degree in electrical system and automation engineering from the Shang-
engineering and automation from Yanshan Univer- hai University of Electric Power, Shanghai, China,
sity, China. He is currently working toward the mas- in 2010. Since 2012, she has been a Researcher
ter’s degree with the Control science and Engineer- with Dispatching Automation System Department,
ing, Nanjing University of Information Science and China Electric Power Research Institute Company,
Technology (NUIST), Nanjing. His current interests Ltd., Nanjing. Her current research interests cover
include machine learning and its application in power source-grid-load-storage collaborative control, power
system. grid simulation, and data analysis.

Min Xia (Member, IEEE) received the Ph.D. degree Jun Liu received the master’s degree in power system
in cybernetics control engineering from Donghua and automation engineering from the Nanjing Uni-
University, Shanghai, China. He is currently a Pro- versity of Science and Technology, Nanjing, China,
fessor with the Nanjing University of Information in 2008. He is currently engaged in research on dis-
Science and Technology. He is currently the Deputy patching automation system in China Electric Power
Director of Jiangsu Key Laboratory of Big Data Research Institute (CEPRI). His current research
analysis technology. His Principal research interests interests cover parameter identification, power grid
include machine learning theory and its application, analysis, and dispatcher training simulation (DTS).
graph structure data analysis.

Authorized licensed use limited to: NATIONAL INSTITUTE OF TECHNOLOGY WARANGAL. Downloaded on July 27,2022 at 11:34:49 UTC from IEEE Xplore. Restrictions apply.

You might also like