CN115146538A - Power system state estimation method based on message passing graph neural network - Google Patents
Power system state estimation method based on message passing graph neural network Download PDFInfo
- Publication number
- CN115146538A CN115146538A CN202210818253.4A CN202210818253A CN115146538A CN 115146538 A CN115146538 A CN 115146538A CN 202210818253 A CN202210818253 A CN 202210818253A CN 115146538 A CN115146538 A CN 115146538A
- Authority
- CN
- China
- Prior art keywords
- node
- graph
- neural network
- data
- section
- Prior art date
- Legal status (The legal status is an assumption and is not a legal conclusion. Google has not performed a legal analysis and makes no representation as to the accuracy of the status listed.)
- Pending
Links
- 238000000034 method Methods 0.000 title claims abstract description 70
- 238000013528 artificial neural network Methods 0.000 title claims abstract description 23
- 238000005259 measurement Methods 0.000 claims abstract description 89
- 238000003062 neural network model Methods 0.000 claims abstract description 47
- 238000010586 diagram Methods 0.000 claims abstract description 39
- 238000012549 training Methods 0.000 claims abstract description 37
- 238000012360 testing method Methods 0.000 claims abstract description 16
- 238000010606 normalization Methods 0.000 claims abstract description 15
- 238000012545 processing Methods 0.000 claims abstract description 13
- 239000013598 vector Substances 0.000 claims description 44
- 238000012546 transfer Methods 0.000 claims description 25
- 239000011159 matrix material Substances 0.000 claims description 21
- 239000000126 substance Substances 0.000 claims description 12
- 238000011176 pooling Methods 0.000 claims description 11
- 230000004913 activation Effects 0.000 claims description 10
- 238000004422 calculation algorithm Methods 0.000 claims description 9
- 230000008569 process Effects 0.000 claims description 9
- 238000004088 simulation Methods 0.000 claims description 7
- 238000006243 chemical reaction Methods 0.000 claims description 4
- 230000006835 compression Effects 0.000 claims description 4
- 238000007906 compression Methods 0.000 claims description 4
- 238000005065 mining Methods 0.000 claims description 4
- 230000002093 peripheral effect Effects 0.000 claims description 4
- 238000007781 pre-processing Methods 0.000 claims description 3
- 238000011056 performance test Methods 0.000 claims description 2
- 239000013589 supplement Substances 0.000 abstract description 2
- 230000006870 function Effects 0.000 description 30
- 238000004364 calculation method Methods 0.000 description 12
- 239000000243 solution Substances 0.000 description 6
- 238000002347 injection Methods 0.000 description 5
- 239000007924 injection Substances 0.000 description 5
- 230000008859 change Effects 0.000 description 4
- 238000004458 analytical method Methods 0.000 description 3
- 238000013527 convolutional neural network Methods 0.000 description 3
- 238000013135 deep learning Methods 0.000 description 3
- 238000010801 machine learning Methods 0.000 description 2
- 230000007246 mechanism Effects 0.000 description 2
- 238000011160 research Methods 0.000 description 2
- 238000013473 artificial intelligence Methods 0.000 description 1
- 238000010219 correlation analysis Methods 0.000 description 1
- 230000000694 effects Effects 0.000 description 1
- 238000005516 engineering process Methods 0.000 description 1
- 238000012986 modification Methods 0.000 description 1
- 230000004048 modification Effects 0.000 description 1
- 230000008447 perception Effects 0.000 description 1
- 238000010248 power generation Methods 0.000 description 1
- 230000009467 reduction Effects 0.000 description 1
- 230000009466 transformation Effects 0.000 description 1
- 230000001052 transient effect Effects 0.000 description 1
Images
Classifications
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06F—ELECTRIC DIGITAL DATA PROCESSING
- G06F30/00—Computer-aided design [CAD]
- G06F30/20—Design optimisation, verification or simulation
- G06F30/27—Design optimisation, verification or simulation using machine learning, e.g. artificial intelligence, neural networks, support vector machines [SVM] or training a model
-
- H—ELECTRICITY
- H02—GENERATION; CONVERSION OR DISTRIBUTION OF ELECTRIC POWER
- H02J—CIRCUIT ARRANGEMENTS OR SYSTEMS FOR SUPPLYING OR DISTRIBUTING ELECTRIC POWER; SYSTEMS FOR STORING ELECTRIC ENERGY
- H02J3/00—Circuit arrangements for ac mains or ac distribution networks
-
- H—ELECTRICITY
- H02—GENERATION; CONVERSION OR DISTRIBUTION OF ELECTRIC POWER
- H02J—CIRCUIT ARRANGEMENTS OR SYSTEMS FOR SUPPLYING OR DISTRIBUTING ELECTRIC POWER; SYSTEMS FOR STORING ELECTRIC ENERGY
- H02J2203/00—Indexing scheme relating to details of circuit arrangements for AC mains or AC distribution networks
- H02J2203/20—Simulating, e g planning, reliability check, modelling or computer assisted design [CAD]
-
- Y—GENERAL TAGGING OF NEW TECHNOLOGICAL DEVELOPMENTS; GENERAL TAGGING OF CROSS-SECTIONAL TECHNOLOGIES SPANNING OVER SEVERAL SECTIONS OF THE IPC; TECHNICAL SUBJECTS COVERED BY FORMER USPC CROSS-REFERENCE ART COLLECTIONS [XRACs] AND DIGESTS
- Y04—INFORMATION OR COMMUNICATION TECHNOLOGIES HAVING AN IMPACT ON OTHER TECHNOLOGY AREAS
- Y04S—SYSTEMS INTEGRATING TECHNOLOGIES RELATED TO POWER NETWORK OPERATION, COMMUNICATION OR INFORMATION TECHNOLOGIES FOR IMPROVING THE ELECTRICAL POWER GENERATION, TRANSMISSION, DISTRIBUTION, MANAGEMENT OR USAGE, i.e. SMART GRIDS
- Y04S10/00—Systems supporting electrical power generation, transmission or distribution
- Y04S10/50—Systems or methods supporting the power network operation or management, involving a certain degree of interaction with the load-side end user applications
Landscapes
- Engineering & Computer Science (AREA)
- Physics & Mathematics (AREA)
- Evolutionary Computation (AREA)
- Theoretical Computer Science (AREA)
- Computer Vision & Pattern Recognition (AREA)
- Medical Informatics (AREA)
- Software Systems (AREA)
- Artificial Intelligence (AREA)
- Computer Hardware Design (AREA)
- Geometry (AREA)
- General Engineering & Computer Science (AREA)
- General Physics & Mathematics (AREA)
- Power Engineering (AREA)
- Supply And Distribution Of Alternating Current (AREA)
Abstract
The invention discloses a power system state estimation method based on a message passing diagram neural network, which comprises the following steps: acquiring multi-section historical measurement information from a historical database, carrying out normalization processing, and acquiring topological information of the power system; constructing multi-section historical map data and dividing the multi-section historical map data into a training set and a test set; setting a network structure and model parameters of a message passing graph-based neural network Model (MPNN); performing off-line training on the model, and outputting a state estimation value of a system node voltage amplitude and a branch phase angle difference by a full connection layer; and constructing graph data of the current section according to the acquired current section measurement information after normalization processing and the topological information of the power system, and inputting the graph data of the current section into the model to obtain a system state estimation value of the current section. The invention utilizes the thought of graph theory to analyze and mine the topological space information, supplements the space characteristic for the traditional measuring database and improves the generalization performance of the model.
Description
Technical Field
The invention relates to a power system state estimation method based on a message passing diagram neural network, and belongs to the technical field of power systems.
Background
The scale of new energy power generation is gradually enlarged, and the structure is also gradually complicated, which brings a series of challenges to the safe and stable operation of a power system. While safe operation and coordinated control of the power system require state estimation to provide effective and accurate base data for it. However, the randomness of load fluctuation, access of a large number of power electronic devices, errors of measurement equipment and the like all increase the uncertainty of the operation state of the power grid, and cause troubles to the scheduling of the power system.
The novel power system has the characteristics of being large and complex, at present, a state estimation method based on Weighted Least Square (WLS) is widely applied, WLS estimation results have the characteristics of minimum variance and unbiased, and the WLS estimation method is an optimal state estimation method when only Gaussian white noise exists. However, in actual power grid operation, the 'raw data' collected by the measurement system contains non-gaussian noise and bad data, so that the WLS estimation convergence is poor, and the estimation result is not available. Therefore, in the conventional physical model, an estimation method based on a non-quadratic criterion is proposed. There are also documents that a Weighted Least Absolute Value (WLAV) estimation is converted into a nonlinear transformation and quadratic programming problem, so that the speed of WLAV solution is increased, but a bilinear algorithm causes a reduction in state estimation redundancy, and affects estimation accuracy.
In order to guarantee estimation performance and simultaneously take account of calculation efficiency, many scholars introduce artificial intelligence technology into the field of state estimation. In the literature, a deep neural network is utilized and correlation analysis is applied to extract strong correlation measurement, so that distributed estimation is realized, and the robustness and efficiency of the model are improved. The deep learning method based on the physical guide machine learning modeling is characterized in that a learner provides a deep learning method based on the physical guide through a deep neural network learning time correlation, and compared with a simple data driving method, the deep learning method based on the physical guide machine learning modeling has better interpretability.
The traditional data driving method only excavates potential relation between system measurement information and state quantity, and structural information between power grid nodes and lines is not considered, so that the method cannot adapt to the situation of topology change during actual operation of a power system.
Disclosure of Invention
Aiming at the problem that the topological change of a large-scale power system causes the result of a data-driven state estimation model to be unavailable, the invention provides a Message Passing Neural Network (MPNN) -based power system state estimation method which comprehensively measures the data characteristics and the power grid structural characteristics, and the invention firstly establishes a topological graph model containing a power system; then, combining the topological graph and the measurement information and converting the topological graph and the measurement information into graph data describing node characteristics and edge characteristics so as to bring the power grid structure information into state estimation calculation; finally, power system state estimation is achieved based on the model.
The invention specifically adopts the following technical scheme to solve the technical problems:
a power system state estimation method based on a message passing graph neural network comprises the following steps:
step (1): data acquisition and preprocessing, including: acquiring multi-section historical state quantity from a historical database, acquiring multi-section tidal flow data serving as a true value through simulation, adding Gaussian white noise on the basis of the multi-section tidal flow data to generate multi-section historical measurement information, adding 3-10 sigma mixed Gaussian noise to part of the multi-section historical measurement information to obtain a historical measurement data set containing a bad data sample, performing normalization processing on the historical measurement data set to obtain a normalized historical measurement data set, and acquiring topological information of a power system; constructing multi-section historical graph data describing node characteristics and edge characteristics according to the historical measurement data set and the power system topology information after normalization processing; dividing the multi-section historical map data into a training set and a test set according to corresponding proportions;
step (2): setting a network structure and model parameters of a message passing diagram-based neural network Model (MPNN), wherein the network structure of the message passing diagram-based neural network Model (MPNN) consists of two GNN hidden layers, a diagram pooling layer and a full connection layer;
and (3): performing offline training on the message passing graph-based neural network model MPNN, comprising: inputting a training set in multi-section historical graph data, inputting a feature matrix containing node features and edge features and an adjacent matrix representing a topological connection relation in the training set into GNN hidden layers, using a ReLU activation function between the two GNN hidden layers, mining topological structure information through the two GNN hidden layers to obtain a graph embedding vector, performing pooling compression on the graph embedding vector through a graph pooling layer to obtain a final graph embedding vector, inputting the final graph embedding vector into a full-connection layer, using a Sigmoid activation function by the full-connection layer, and outputting a state estimation value of a system node voltage amplitude and a branch phase angle difference by the full-connection layer; in the training process, comparing the state estimation values of the system node voltage amplitude and the branch phase angle difference output by the message passing diagram based neural network model MPNN with multi-section historical measurement information samples in a training set, calculating a loss function, updating the weight and parameters of the message passing diagram based neural network model MPNN through a back propagation algorithm, and obtaining the message passing diagram based neural network model MPNN which is suitable for different topologies through multiple iterations; performing performance test by using the message passing diagram-based neural network model MPNN obtained by inputting the data of the multi-section historical diagram in the test set;
and (4): the method for online estimation by using the message passing graph-based neural network model MPNN comprises the following steps: and (3) constructing graph data of the current section according to the acquired current section measurement information after normalization processing and the topological information of the power system, inputting the graph data of the current section into the message transfer graph-based neural network model MPNN finished by offline training in the step (3), and obtaining the state estimation values of the system node voltage amplitude and the branch phase angle difference of the current section.
Further, as a preferred technical solution of the present invention, the topology information of the power system acquired in step (1) includes branch admittance information and a connection relationship between nodes.
Further, as a preferred technical solution of the present invention, the node characteristics in the multi-section historical graph data in step (1) include node voltage amplitude and node injection active and reactive power, and the edge characteristics include branch head and end active and reactive power and branch admittance information.
Further, as a preferred technical solution of the present invention, the step (1) of normalizing the historical measurement data set specifically includes:
for the ith historical measurement value of the kth section, carrying out data normalization on the historical measurement value, wherein a specific formula is as follows:
wherein u is k ' is the k-th normalized historical measurement value, u k For the kth original historical measurement, min (-) is the minimum value in a set of data, and max (-) is the maximum value in a set of data.
Further, as a preferred technical solution of the present invention, in the step (3), the branch phase angle difference instead of the node phase angle based on the message passing graph neural network model MPNN is used as the state estimation value, and a specific conversion formula is as follows:
θ＝(U T U) -1 U T Δθ
and the theta is a node voltage phase angle estimation vector, the delta theta is a branch phase angle difference estimation vector, and the U is a correlation matrix.
Further, as a preferred technical solution of the present invention, the node message transfer function adopted by the message transfer graph-based neural network model MPNN in the step (3) is as follows:
wherein the content of the first and second substances,
information received for node v at time t + 1; n (v) is all the adjacency points of the node v;
is the feature vector of the node v at time t;
is a peripheral node omega state;
an edge feature vector of the node v and the node omega at the time t is obtained; e.g. of the type vω Edge features for node v and node ω;
is a learnable parameter matrix; f. of t An activation function that is a fully connected layer;
the edge message transfer function adopted by the message transfer graph-based neural network model MPNN is as follows:
wherein the content of the first and second substances,
information received by the node v and the node ω at time t + 1;
is a learnable parameter matrix;
and the state updating function based on the message passing graph neural network model MPNN is realized by adopting a gating cycle unit GRU, and the specific formula is as follows:
wherein the content of the first and second substances,
a feature vector for node v at time t + 1;
the edge feature vectors of the node v and the node ω at the time t + 1.
By adopting the technical scheme, the invention can produce the following technical effects:
the method for estimating the state of the power system based on the message transfer diagram neural network, provided by the invention, aims at the problem that a data-driven state estimation model is unavailable when the topology of a large-scale power system changes frequently, analyzes and excavates data characteristics of measurement information and power system topology information based on the message transfer diagram neural network, learns topological space information by using the diagram neural network, learns time sequence characteristics of the measurement information by using a neural network structure, and finally, the model can realize accurate perception of the system state under different topologies.
Therefore, the method can simultaneously analyze and mine the topological parameter information and the measurement information, simultaneously analyze the time-space correlation characteristics of the data, estimate the real-time state of different topological structures of the same system, analyze and mine the topological space information by utilizing the thought of the graph theory, supplement the space characteristics for the traditional measurement database, and improve the generalization performance of the data driving state estimation model on the basis of keeping the original robustness and real-time performance. Moreover, the feasibility and the accuracy of the method are verified through a large amount of simulation and actual measurement data, and the method has important significance for improving the generalization performance of the data-driven state estimation model.
Drawings
Fig. 1 is an internal schematic diagram of a conventional graph convolutional neural network.
FIG. 2 is a flow chart of a method for estimating the state of a power system of the message passing diagram neural network according to the present invention.
FIG. 3 is a network architecture diagram of a message passing graph neural network model in accordance with the present invention.
FIG. 4 shows the state estimation results of the IEEE118 node system under different topologies.
FIG. 5 is a graph showing a comparison of bad data branch flow errors in an IEEE118 node system under different topologies.
Detailed Description
The present invention is further illustrated by the following examples, which are intended to be purely exemplary and are not intended to limit the scope of the invention, as various equivalent modifications of the invention will occur to those skilled in the art upon reading the present disclosure and fall within the scope of the appended claims.
The invention relates to a power system state estimation method based on a message passing diagram neural network, the specific flow of which is shown in figure 2, and the method comprises the following steps:
step (1): data acquisition and preprocessing, including: firstly, acquiring multi-section historical state quantity from a historical database, acquiring multi-section tidal flow data as a true value through simulation, adding Gaussian white noise on the basis of the multi-section tidal flow data to generate multi-section historical measurement information, adding 3 sigma-10 sigma mixed Gaussian noise to part of the multi-section historical measurement information to obtain a historical measurement data set containing bad data samples in order to improve the robustness of a model, normalizing the multi-section historical measurement information in the historical measurement data set to obtain a normalized historical measurement data set in consideration of the influence of data dimension on model training and avoiding the phenomenon of overfitting of the model training, and acquiring topological information of a power system. When the system topology changes, in order to reflect the disconnection condition of the system branch, the disconnected branch measurement is filled with zero. In order to improve the model training efficiency, the node voltage amplitude and the branch phase angle difference are respectively modeled. The multi-section historical measurement information comprises node voltage amplitude measurement, node injection active power and reactive power measurement, and branch head and tail end active power and reactive power measurement; the power topology parameter information comprises branch admittance and topology connection relation.
The normalization processing is performed on the historical measurement data set, and specifically includes:
for the ith historical measurement value of the kth section, carrying out data normalization on the historical measurement value, wherein a specific formula is as follows:
wherein u is k ' is the k-th normalized historical measurement value, u k For the kth original historical measurement, min (-) is the minimum value in a set of data, and max (-) is the maximum value in a set of data. The above formula can be used to distribute the measured values in [0,1]And the influence of the data magnitude on the model training is avoided.
And secondly, constructing graph data, namely constructing multi-section historical graph data describing node characteristics and edge characteristics according to the historical measurement data set and the power system topology information after normalization processing, wherein each node in the power system topology is taken as a point in the graph data, a branch connecting each node in the power system topology is taken as an edge of the graph data, and each characteristic forms the graph data by using a topology connection relation. The node characteristics of the graph data comprise node voltage amplitude values and node injection active and reactive power, and the edge characteristics comprise active and reactive power of the head end and the tail end of the branch circuit and branch circuit admittance information.
And finally, processing the processed multi-section historical map data according to the following steps of 7:3, dividing the corresponding proportion into a training set and a test set; the training set is used for determining various hyper-parameters and weights of the message passing graph-based neural network model MPNN, and the test set is used for evaluating the performance of the message passing graph-based neural network model MPNN after training is completed.
In the step (1) of the invention, the traditional state estimation result is the node state quantity, but the node phase angle is a relative value due to the existence of the balance node, and during off-line training, the state estimation result is inaccurate due to the lack of physical connection among data, but the accurate phase angle difference estimation result is more meaningful for the calculation of the system load flow. In order to better reflect the electrical connection between data and enable the final load flow calculation to be more accurate, the message transfer diagram-based neural network model MPNN utilizes branch phase angle difference to replace node phase angle as a state estimation value, so that the data driving model can better mine the electrical connection and the space-time connection between measurement information, and the calculation performance of the data driving state estimation model is improved. The specific conversion formula is as follows:
θ＝(U T U) -1 U T Δθ
wherein theta is a node voltage phase angle estimation vector, delta theta is a branch phase angle difference estimation vector, and U is a correlation matrix.
Step (2): before model training, setting a network structure and model parameters of a message passing graph-based neural network Model (MPNN), wherein the network structure and the model parameters comprise: the overall number of layers of the model, the number of GNN hidden layers, an optimizer, the number of iterations and the like.
The invention is based on Message Passing Neural Network Model (MPNN), and the principle is as follows: the model is a model framework based on graph theory, and compared with a graph convolution neural network structure shown in FIG. 1, the model describes the propagation process of node features on a graph, and is easier to expand to modeling of opposite sides. The forward propagation process of MPNN is mainly divided into two phases of message passing and reading. In the messaging phase, a graph G = (V, E) is defined, where V represents a set of nodes in the graph; e represents the set of edges in the graph. Multiple information delivery processes may be performed during the messaging phase. For a particular node v, the following expression is possible:
wherein the content of the first and second substances,
information received for node v at time t + 1; n (v) is all the adjacency points of the node v;
is the feature vector of the node v at time t; e.g. of the type vω Is the edge characteristics of node v and node ω;
as a function of the message at the node at time t. The formula shows that the information received by the node v is derived from the state of the node v
And peripheral node status
And edge features e connected thereto vω . After the information is generated, the nodes need to be updated:
wherein the content of the first and second substances,
updating a function for the node, the function relating the original node state
And the received information
As input, a new node state is obtained
Is the feature vector for node v at time t + 1.
Based on the above node modeling, the edges in the graph can be similarly modeled, and the specific formula is as follows:
wherein the content of the first and second substances,
and
a message transfer function and a state update function for the edge at time t;
an edge feature vector of the node v and the node omega at the time t is obtained;
the information received by node v and node ω at time t + 1.
In the read phase, a feature vector based on the whole graph is calculated using a read function R (-) and can be expressed as:
wherein, the first and the second end of the pipe are connected with each other,
is the final output vector; r (-) is a reading function which can read graph embedding vectors from the last layer of hidden states of nodes of the whole graph; t is the last time step in the messaging phase. Representing the last GNN hidden layer in the actual model, outputting the vector
Can be used for the task of estimating the system state later.
As shown in fig. 3, the network structure based on the message passing graph neural network model MPNN adopted by the method of the present invention is composed of two GNN hidden layers, a graph pooling layer and a full connection layer, wherein the GNN hidden layers are used for mining topological space information of the information to obtain a graph embedding vector, and the graph pooling layer is used for pooling compression operation to obtain a final graph embedding vector so as to facilitate the subsequent full connection layer access; the full connection layer is used for analyzing the time sequence characteristics of the measured data and finally outputting a state estimation result.
And, the message transfer function of the node used by the message transfer graph-based neural network model MPNN of the present invention is as follows:
wherein the content of the first and second substances,
information received for node v at time t + 1; n (v) is all the adjacency points of the node v;
is the feature vector of the node v at time t;
in the state of the peripheral node ω,
an edge feature vector of the node v and the node omega at the time t is obtained; e.g. of a cylinder vω Edge features for node v and node ω;
is a learnable parameter matrix; f. of t An activation function that is a fully connected layer;
the message transfer graph-based neural network model MPNN of the invention adopts the following edge message transfer functions:
wherein the content of the first and second substances,
information received by the node v and the node ω at time t + 1;
is a learnable parameter matrix;
the state updating function based on the message passing graph neural network model MPNN of the present invention is implemented by using a Gated Round Unit (GRU), and the specific formula is as follows:
wherein, the first and the second end of the pipe are connected with each other,
is the feature vector of the node v at time t + 1;
the feature vectors of node v and node ω at time t + 1.
And, in the reading phase, the invention can also use the reading function R (-) to calculate the whole graph-based feature vector, which can be expressed as:
in the invention, the read function R (-) is a full connection layer, and the parameter setting is obtained by network training.
And (3): performing offline training on the message passing graph-based neural network model MPNN, comprising:
firstly, a training set in the multi-section historical graph data is input, and the input training set contains a feature matrix of node features and edge features and an adjacency matrix representing topological connection relation.
Then, inputting a feature matrix containing node features and edge features in a training set and an adjacent matrix representing a topological connection relation into a GNN hidden layer, using a ReLU activation function between the two GNN hidden layers, mining topological structure information through the two GNN hidden layers to obtain a graph embedding vector, performing pooling compression on the graph embedding vector through a graph pooling layer to obtain a final graph embedding vector, inputting the final graph embedding vector into a full connection layer, using a Sigmoid activation function by the full connection layer, analyzing time sequence features of multi-section historical measurement information by the full connection layer, and outputting state estimation values of system node voltage amplitude and branch phase angle difference;
and comparing the state estimation value output by the message transfer diagram-based neural network model MPNN with the multi-section historical measurement information samples in the training set and calculating a loss function in the training process, optimizing the model parameters through the loss function, updating the weight and parameters of the message transfer diagram-based neural network model MPNN through a back propagation algorithm, and finally obtaining the message transfer diagram-based neural network model MPNN suitable for different topologies through multiple iterations. And in the model testing stage, a message passing diagram-based neural network model MPNN obtained by inputting the multi-section historical diagram data in the test set is used for performing performance testing, namely the multi-section historical diagram data in the test set is input into the trained message passing diagram-based neural network model MPNN, and whether the model performance is good or not is judged according to an error value obtained by comparing a state estimation value output by the model with a true value of a multi-section historical measurement information sample in the test set so as to obtain the trained message passing diagram-based neural network model MPNN.
And (4): the method for online estimation by using the message passing graph-based neural network model MPNN comprises the following steps: and (3) constructing graph data of the current section according to the acquired current section measurement information after normalization processing and the topological information of the power system, inputting the graph data of the current section into the message transfer graph-based neural network model MPNN finished by offline training in the step (3), and obtaining the state estimation values of the node voltage amplitude and the branch phase angle difference of the system of the current section.
Compared with the traditional estimation method, the power system state estimation method based on the message transfer diagram neural network provided by the invention has the following advantages:
1) In the aspect of research field, the method provided by the invention researches the field of power system state estimation.
2) In terms of data type, the present invention is used in the field of power system state estimation. And secondly, the SCADA data used by the method does not have phase angle measurement, and the side modeling is carried out by utilizing the branch power, so that the method is closer to the real data type of the actual power grid.
3) In terms of a data driving model, the method is different from a model used by a traditional method, and the model used by the traditional method generally obtains a transient power angle stability discrimination confidence coefficient S of the current power system operation scene by utilizing a softmax layer 0 And instability discrimination confidence S 1 The model of the method directly utilizes the Dense full connection layer to output the final system state estimation information, and finally obtains a precise real-time data driving state estimation result under the condition that the topology of the large power grid changes frequently.
Therefore, the method analyzes and mines topological space information by using the graph theory idea, and the overall model framework can simultaneously analyze the space-time correlation characteristics of data to obtain the state estimation value of the system because the characteristics in the input graph data comprise the space characteristics in the topological information of the power system and the time sequence characteristics in the multi-section measurement information. The space characteristics are supplemented for the traditional measurement database, so that the generalization performance of the data-driven state estimation model is improved on the basis of keeping the original robustness and real-time performance.
In order to verify the superiority of the method provided by the present invention, simulation test and detailed description are now performed on the IEEE118 node system, which are specifically as follows:
referring to measurement configuration of actual power system state estimation, wherein historical measurement information of multiple sections comprises node voltage amplitude measurement, node injection active and reactive power measurement, and branch head and tail end active and reactive power measurement; the power topology parameter information comprises branch admittance and topology connection relation. In the simulation, the load curve of the actual power system is utilized to carry out simulation to obtain multi-section tidal flow data as a true value, gaussian white noise is added on the basis of tidal flow to simulate normal measurement data, and bad data conditions occurring in the operation process of the actual power system are simulated by increasing or reducing the normal power measurement data by 50% -150% and increasing or reducing the normal voltage amplitude measurement by 15% -25%.
In order to embody the estimation performance of the method under the condition of frequent change of system topology, the method is shown by an example based on the active power of the head end of the branch 1 and the reactive power of the head end under different topology conditions. Wherein, the topology 1 is a complete topology structure of the IEEE118 node system; the topology 2 structure is that the branches 1-3 are disconnected; topology 3 is a branch 12-16 disconnection; topology 4 is structured with branches 5-6 and 23-24 disconnected; topology 5 is a 2-12,4-11 and 62-66 split. The branch disconnection conditions of topology 2 and topology 5 have stronger correlation with the branches displayed by the example, and the branch disconnection conditions of topology 3 and topology 4 have smaller influence on the branches displayed by the example.
Processing is performed according to the method of the present invention, first, in step (1), the measurement information is preprocessed as follows:
for the ith historical measurement value of the kth section, carrying out data normalization on the historical measurement value, wherein a specific formula is as follows:
wherein u is k ' is the k-th normalized historical measurement value, u k For the kth original historical measurement, min (-) is the minimum value in a set of data, and max (-) is the maximum value in a set of data. The above formula can be used to distribute the measured values in [0,1]And the influence of the data magnitude on the model training is avoided.
In the step (1), the traditional state estimation result is a node state quantity, but because of the existence of a balance node, a node phase angle is a relative value, and during off-line training, the state estimation result is inaccurate due to the lack of physical connection among data, but the accurate phase angle difference estimation result is more meaningful for the calculation of the system power flow. In order to better reflect the electrical connection among data and enable the final load flow calculation to be more accurate, the method utilizes branch phase angle difference to replace node phase angle as state quantity, so that the data driving model can better mine the electrical connection and the space-time connection among measurement information, and the calculation performance of the data driving state estimation model is improved. The specific conversion formula is as follows:
θ＝(U T U) -1 U T Δθ
and the theta is a node voltage phase angle estimation vector, the delta theta is a branch phase angle difference estimation vector, and the U is a correlation matrix.
In order to improve the model training efficiency, the node voltage amplitude and the branch phase angle difference are respectively modeled. The measurement information comprises node voltage amplitude measurement, node injection active and reactive power measurement, and branch head and tail end active and reactive power measurement; the topology parameter information comprises branch admittance and topology connection relation.
In steps (2) and (3), the network structure of the proposed message passing graph neural network model MPNN is shown in fig. 3, and the forward propagation process of MPNN is mainly divided into two stages of message passing and reading. In the messaging phase, a graph G = (V, E) is defined, where V represents a set of nodes in the graph; e represents the set of edges in the graph. Multiple information delivery processes may be performed during the messaging phase.
The node message transfer function used by the present invention is as follows:
wherein the content of the first and second substances,
is a learnable parameter matrix; f. of t Is the activation function of the fully connected layer.The edge message transfer function is as follows:
wherein the content of the first and second substances,
is a learnable parameter matrix. The state updating function of the invention is realized by using a Gated Recurrentunit (GRU), and the specific formula is as follows:
in the reading phase, the present invention uses a reading function R (-) to compute a feature vector based on the whole graph, which can be expressed as:
in the invention, the read function R (-) is a full connection layer, and the parameter setting is obtained by network training.
In order to verify the superiority of the present invention, an example test was performed based on the IEEE118 node system. In order to embody the estimation performance of the method under the condition of frequent change of system topology, the method is shown by an example based on the active power of the head end of the branch 1 and the reactive power of the head end under different topology conditions. Wherein topology 1 is the complete topology of the IEEE118 node system; the topology 2 structure is that the branches 1-3 are disconnected; topology 3 is a branch 12-16 disconnection; topology 4 is structured with branches 5-6 and 23-24 disconnected; topology 5 is a 2-12,4-11 and 62-66 split. The branch circuit breaking conditions of the topology 2 and the topology 5 have stronger correlation with the branch circuit shown by the example, and the branch circuit breaking conditions of the topology 3 and the topology 4 have smaller influence on the branch circuit shown by the example.
In order to facilitate the intuitive embodiment of different estimator results, the invention carries out state estimation calculation on the test set data and compares the result with the true value of the power flow to obtain the maximum absolute error and the average absolute error. Because the estimation results of the estimators have magnitude difference, the invention adopts a logarithm form to display, and the system estimation value passing line is at a logarithm value of 3. The specific formula is as follows:
wherein, the first and the second end of the pipe are connected with each other,
for the f-th state estimation result in the l-th cross section, h f (l) Is the f-th true value in the l-th section, U is the total section number, phi fmax Is the maximum absolute error value, phi fave Is the average absolute error. Table 1 shows state estimation errors in different topologies of the IEEE118 node system.
TABLE 1 IEEE118 node System State estimation errors under different topologies
It can be seen from table 1 that, when the measurement information only contains gaussian white noise, compared with the state estimator driven by the traditional physical model, the method of the present invention ignores the strict analysis on the internal mechanism of the object under study and reflects the essential characteristics of data due to the analysis of the potential relation between the measurement information and the system state quantity, so that the estimation performance is better expressed, taking the active power at the head end of the branch in topology 1 as an example, the estimation performance is respectively improved by 63.51% and 72.17% compared with WLS and WLAV. Compared with other data-driven algorithms, the comparison algorithm only analyzes the relation between the measurement information and the state quantity and does not analyze the relation between the power grid topological structure and the state quantity, so that the estimation result of the comparison algorithm does not reach the qualified standard when the topology changes frequently, particularly when the branch connection and disconnection condition and the branch correlation shown by the example are strong.
The traditional state estimation method is based on the iteration of a Jacobian matrix to seek a state vector meeting the precision requirement, and when the scale of a system is increased, the required calculation time of the method is increased along with the increase of the system. The method adopts a data driving idea, and utilizes test and historical data to establish the potential association between the graph data and the state information, thereby avoiding strict analysis on the internal mechanism of the object to be researched and improving the online estimation efficiency of the system. Table 2 shows the comparison of the calculation time of the traditional state estimation algorithm and the method of the invention under different scale systems.
TABLE 2 calculation times under different algorithms
The main innovation point and key point of the method are that topological space information is learned by utilizing the characteristics of the graph neural network, and the generalization performance of the data-driven state estimation model is improved. In order to better reflect the performance of the model, fig. 4 and fig. 5 are respectively a comparison between the state estimation result of the IEEE118 node system under different topologies and the bad data branch flow error of the IEEE118 node system under different topologies.
Fig. 4 is a diagram illustrating the system state estimation result of IEEE118 nodes under different topologies. It can be seen from fig. 4 that the method of the present invention can accurately estimate the system state when the measured data only contains gaussian white noise, and has a better performance in the state estimation precision when the topology changes. The method provided by the invention not only can mine the potential relation between the measurement information and the system state quantity, but also can analyze the spatial relation between the power grid topological structure and the system state quantity, so that the data-driven state estimation model can be more suitable for the real operation condition of the system.
As shown in fig. 5, a graph is shown for comparing flow errors of bad data branches of IEEE118 node systems under different topologies. As can be seen from fig. 5, when the system measurement information includes a certain proportion of bad data, the method of the present invention has better performance in the estimation accuracy of different topologies. Compared with other data-driven state estimation models, when the topology of a Deep Neural Network (DNN) and a Convolutional Neural Network (CNN) changes frequently, the estimation error exceeds the precision established requirement range under the condition that the branch disconnection condition and the branch correlation shown by the example are strong. The method of the invention analyzes the measurement information and the topology information of the power system at the same time, so that the estimation result with higher quality can still be obtained when the system measurement contains bad data and the topology changes.
The embodiments of the present invention have been described in detail with reference to the drawings, but the present invention is not limited to the above embodiments, and various changes can be made within the knowledge of those skilled in the art without departing from the gist of the present invention.
Claims (6)
1. The method for estimating the state of the power system based on the message transfer diagram neural network is characterized by comprising the following steps of:
step (1): acquisition and preprocessing of data, including: acquiring multi-section historical state quantity from a historical database, acquiring multi-section tidal flow data serving as a true value through simulation, adding Gaussian white noise on the basis of the multi-section tidal flow data to generate multi-section historical measurement information, adding 3 sigma-10 sigma mixed Gaussian noise to part of the multi-section historical measurement information to obtain a historical measurement data set containing bad data samples, then performing normalization processing on the historical measurement data set to obtain a normalized historical measurement data set, and acquiring topological information of a power system; constructing multi-section historical graph data describing node characteristics and edge characteristics according to the historical measurement data set and the power system topology information after normalization processing; dividing the multi-section historical map data into a training set and a test set according to corresponding proportion;
step (2): setting a network structure and model parameters of a message passing diagram-based neural network Model (MPNN), wherein the network structure of the message passing diagram-based neural network Model (MPNN) consists of two GNN hidden layers, a diagram pooling layer and a full connection layer;
and (3): performing offline training on the message passing graph-based neural network model MPNN, comprising: inputting a training set in multi-section historical graph data, inputting a feature matrix containing node features and edge features and an adjacent matrix representing a topological connection relation in the training set into GNN hidden layers, using a ReLU activation function between the two GNN hidden layers, mining topological structure information through the two GNN hidden layers to obtain a graph embedding vector, performing pooling compression on the graph embedding vector through a graph pooling layer to obtain a final graph embedding vector, inputting the final graph embedding vector into a full-connection layer, using a Sigmoid activation function by the full-connection layer, and outputting a state estimation value of a system node voltage amplitude and a branch phase angle difference by the full-connection layer; in the training process, comparing the state estimation values of the system node voltage amplitude and the branch phase angle difference output by the message passing diagram based neural network model MPNN with multi-section historical measurement information samples in a training set, calculating a loss function, updating the weight and parameters of the message passing diagram based neural network model MPNN through a back propagation algorithm, and obtaining the message passing diagram based neural network model MPNN which is suitable for different topologies through multiple iterations; performing performance test by using the message passing diagram-based neural network model MPNN obtained by inputting the data of the multi-section historical diagram in the test set;
and (4): the method for online estimation by using the message passing graph-based neural network model MPNN comprises the following steps: and (4) constructing graph data of the current section according to the acquired current section measurement information after normalization processing and the topological information of the power system, inputting the graph data of the current section into the message passing graph-based neural network model MPNN finished by offline training in the step (3), and outputting to obtain a state estimation value of the system node voltage amplitude and the branch phase angle difference of the current section.
2. The message passing graph neural network-based power system state estimation method according to claim 1, wherein the power system topology information acquired in the step (1) includes branch admittance information and connection relationships between nodes.
3. The message passing graph neural network based power system state estimation method of claim 1, wherein the node characteristics in the multi-section historical graph data of step (1) comprise node voltage amplitude and node injected active and reactive power, and the edge characteristics comprise branch head and end active and reactive power and branch admittance information.
4. The method according to claim 1, wherein the step (1) of normalizing the historical measurement data set comprises:
for the ith historical measurement value of the kth section, carrying out data normalization on the historical measurement value, wherein a specific formula is as follows:
wherein u is k ' is the k-th normalized historical measurement value, u k For the kth original historical measurement, min (-) is the minimum value in a set of data, and max (-) is the maximum value in a set of data.
5. The method according to claim 1, wherein the message transfer diagram neural network-based power system state estimation method in step (3) uses branch phase angle difference instead of node phase angle as the state estimation value based on the message transfer diagram neural network model MPNN, and the specific conversion formula is as follows:
θ＝(U T U) -1 U T Δθ
wherein theta is a node voltage phase angle estimation vector, delta theta is a branch phase angle difference estimation vector, and U is a correlation matrix.
6. The message-passing graph neural network-based power system state estimation method according to claim 1, wherein the node message-passing function adopted by the message-passing graph neural network model MPNN in the step (3) is as follows:
wherein the content of the first and second substances,
information received for node v at time t + 1; n (v) is all the adjacent points of the node v;
is the feature vector of the node v at time t;
the state of the peripheral node ω at time t,
an edge feature vector of the node v and the node omega at the time t is obtained; e.g. of a cylinder vω Edge features for node v and node ω;
is a learnable parameter matrix; f. of t An activation function that is a fully connected layer;
the edge message transfer function adopted by the message transfer graph-based neural network model MPNN is as follows:
wherein the content of the first and second substances,
received by node v and node ω at time t +1The information of arrival;
is a learnable parameter matrix;
and the state updating function based on the message passing graph neural network model MPNN is realized by adopting a gating cycle unit GRU, and the specific formula is as follows:
wherein, the first and the second end of the pipe are connected with each other,
is the feature vector of the node v at time t + 1;
the edge feature vectors of the node v and the node ω at the time t + 1.
Priority Applications (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN202210818253.4A CN115146538A (en) | 2022-07-11 | 2022-07-11 | Power system state estimation method based on message passing graph neural network |
Applications Claiming Priority (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN202210818253.4A CN115146538A (en) | 2022-07-11 | 2022-07-11 | Power system state estimation method based on message passing graph neural network |
Publications (1)
| Publication Number | Publication Date |
|---|---|
| CN115146538A true CN115146538A (en) | 2022-10-04 |
Family
ID=83411626
Family Applications (1)
| Application Number | Title | Priority Date | Filing Date |
|---|---|---|---|
| CN202210818253.4A Pending CN115146538A (en) | 2022-07-11 | 2022-07-11 | Power system state estimation method based on message passing graph neural network |
Country Status (1)
| Country | Link |
|---|---|
| CN (1) | CN115146538A (en) |
Cited By (2)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN113762625A (en) * | 2021-09-09 | 2021-12-07 | 国网山东省电力公司经济技术研究院 | Power distribution network state evaluation method and system based on graph convolution network |
| CN115499849A (en) * | 2022-11-16 | 2022-12-20 | 国网湖北省电力有限公司信息通信公司 | Wireless access point and reconfigurable intelligent surface cooperation method |
Citations (4)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN110443724A (en) * | 2019-07-19 | 2019-11-12 | 河海大学 | A kind of electric system fast state estimation method based on deep learning |
| US20210287067A1 (en) * | 2020-03-11 | 2021-09-16 | Insilico Medicine Ip Limited | Edge message passing neural network |
| CN113553538A (en) * | 2021-05-14 | 2021-10-26 | 河海大学 | Recursive correction hybrid linear state estimation method |
| CN114330486A (en) * | 2021-11-18 | 2022-04-12 | 河海大学 | Power system bad data identification method based on improved Wasserstein GAN |
-
2022
- 2022-07-11 CN CN202210818253.4A patent/CN115146538A/en active Pending
Patent Citations (4)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN110443724A (en) * | 2019-07-19 | 2019-11-12 | 河海大学 | A kind of electric system fast state estimation method based on deep learning |
| US20210287067A1 (en) * | 2020-03-11 | 2021-09-16 | Insilico Medicine Ip Limited | Edge message passing neural network |
| CN113553538A (en) * | 2021-05-14 | 2021-10-26 | 河海大学 | Recursive correction hybrid linear state estimation method |
| CN114330486A (en) * | 2021-11-18 | 2022-04-12 | 河海大学 | Power system bad data identification method based on improved Wasserstein GAN |
Non-Patent Citations (2)
| Title |
|---|
| 王铮澄 等: "考虑电力系统拓扑变化的消息传递图神经网络暂态稳定评估" * |
| 黄蔓云;孙国强;卫志农;臧海祥;陈通;陈胜;: "基于脉冲神经网络伪量测建模的配电网三相状态估计" * |
Cited By (2)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN113762625A (en) * | 2021-09-09 | 2021-12-07 | 国网山东省电力公司经济技术研究院 | Power distribution network state evaluation method and system based on graph convolution network |
| CN115499849A (en) * | 2022-11-16 | 2022-12-20 | 国网湖北省电力有限公司信息通信公司 | Wireless access point and reconfigurable intelligent surface cooperation method |
Similar Documents
| Publication | Publication Date | Title |
|---|---|---|
| CN109117951B (en) | BP neural network-based probability load flow online calculation method | |
| CN115146538A (en) | Power system state estimation method based on message passing graph neural network | |
| CN112348292B (en) | Short-term wind power prediction method and system based on deep learning network | |
| WO2022021726A1 (en) | Pmu-based power system state estimation performance evaluation method | |
| CN108448585B (en) | A kind of electric network swim equation solution method of linearization based on data-driven | |
| CN114091816B (en) | Power distribution network state estimation method of gate control graph neural network based on data fusion | |
| CN109088407B (en) | Power distribution network state estimation method based on deep belief network pseudo-measurement modeling | |
| CN111628494B (en) | Low-voltage distribution network topology identification method and system based on logistic regression method | |
| LU500551B1 (en) | Virtual load dominant parameter identification method based on incremental learning | |
| CN114006370B (en) | Power system transient stability analysis and evaluation method and system | |
| CN113612237A (en) | Method for positioning resonance-induced subsynchronous oscillation source in offshore wind farm | |
| CN106372440B (en) | A kind of adaptive robust state estimation method of the power distribution network of parallel computation and device | |
| CN114611676A (en) | New energy power generation system impedance model identification method and system based on neural network | |
| CN115659844B (en) | Simulation method and device based on wind power plant dynamic model and computer equipment | |
| CN116722653A (en) | Dynamic state detection method and system for electric power system | |
| CN116908619A (en) | Harmonic source positioning method based on space-time diagram attention convolution network | |
| CN111861256A (en) | Active power distribution network reconstruction decision method and system | |
| CN116565856A (en) | Power distribution network state estimation method considering unknown topology change | |
| CN116316629A (en) | Data-driven optimal power flow calculation method considering topology feature learning | |
| Shi et al. | A fault location method for distribution system based on one-dimensional convolutional neural network | |
| CN116418000A (en) | Bayesian state estimation method for unobservable distribution network based on embedded tide neural network | |
| CN111651939B (en) | Permanent magnet wind power plant dynamic equivalent modeling method considering control parameter difference of converter | |
| CN114372418A (en) | Wind power space-time situation description model establishing method | |
| CN114221334A (en) | Rapid state estimation method based on graph neural network | |
| CN113673786A (en) | Effective wind speed estimation system and method of wind driven generator based on RBF and LSTM |
Legal Events
| Date | Code | Title | Description |
|---|---|---|---|
| PB01 | Publication | ||
| PB01 | Publication | ||
| SE01 | Entry into force of request for substantive examination | ||
| SE01 | Entry into force of request for substantive examination | ||
| RJ01 | Rejection of invention patent application after publication |
Application publication date: 20221004 |
|
| RJ01 | Rejection of invention patent application after publication |