CN110705831A - Power angle instability mode pre-judgment model construction method after power system fault and application thereof - Google Patents
Power angle instability mode pre-judgment model construction method after power system fault and application thereof Download PDFInfo
- Publication number
- CN110705831A CN110705831A CN201910844471.3A CN201910844471A CN110705831A CN 110705831 A CN110705831 A CN 110705831A CN 201910844471 A CN201910844471 A CN 201910844471A CN 110705831 A CN110705831 A CN 110705831A
- Authority
- CN
- China
- Prior art keywords
- power
- group
- measurement data
- instability mode
- fault
- 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
- 238000010276 construction Methods 0.000 title claims description 14
- 238000000034 method Methods 0.000 claims abstract description 70
- 239000011159 matrix material Substances 0.000 claims abstract description 58
- 238000005259 measurement Methods 0.000 claims abstract description 49
- 238000012549 training Methods 0.000 claims abstract description 32
- 238000013528 artificial neural network Methods 0.000 claims abstract description 31
- 238000013527 convolutional neural network Methods 0.000 claims description 24
- 230000001052 transient effect Effects 0.000 claims description 18
- 230000006870 function Effects 0.000 claims description 15
- 210000002569 neuron Anatomy 0.000 claims description 12
- 239000013598 vector Substances 0.000 claims description 11
- 230000008569 process Effects 0.000 claims description 10
- 238000010606 normalization Methods 0.000 claims description 7
- 230000004913 activation Effects 0.000 claims description 5
- 238000012545 processing Methods 0.000 claims description 5
- 238000002759 z-score normalization Methods 0.000 claims description 5
- 238000011478 gradient descent method Methods 0.000 claims description 3
- 238000003672 processing method Methods 0.000 claims description 3
- 230000009466 transformation Effects 0.000 claims description 3
- 238000013135 deep learning Methods 0.000 abstract description 6
- 238000012360 testing method Methods 0.000 description 15
- 238000013016 damping Methods 0.000 description 8
- 230000001360 synchronised effect Effects 0.000 description 8
- 238000004422 calculation algorithm Methods 0.000 description 7
- 230000009286 beneficial effect Effects 0.000 description 6
- 238000004088 simulation Methods 0.000 description 6
- 238000010801 machine learning Methods 0.000 description 5
- 238000013507 mapping Methods 0.000 description 5
- 230000010355 oscillation Effects 0.000 description 5
- 238000010586 diagram Methods 0.000 description 4
- 230000000694 effects Effects 0.000 description 4
- 238000012795 verification Methods 0.000 description 4
- 238000004458 analytical method Methods 0.000 description 3
- 238000004364 calculation method Methods 0.000 description 3
- 230000008859 change Effects 0.000 description 3
- 230000000368 destabilizing effect Effects 0.000 description 3
- 238000005070 sampling Methods 0.000 description 3
- 238000012800 visualization Methods 0.000 description 3
- 230000009471 action Effects 0.000 description 2
- 238000013459 approach Methods 0.000 description 2
- 230000008901 benefit Effects 0.000 description 2
- 230000005540 biological transmission Effects 0.000 description 2
- 238000011161 development Methods 0.000 description 2
- 238000005516 engineering process Methods 0.000 description 2
- 238000011156 evaluation Methods 0.000 description 2
- 238000000605 extraction Methods 0.000 description 2
- 230000000737 periodic effect Effects 0.000 description 2
- 238000011160 research Methods 0.000 description 2
- 238000012706 support-vector machine Methods 0.000 description 2
- 238000007664 blowing Methods 0.000 description 1
- 238000006243 chemical reaction Methods 0.000 description 1
- 238000007796 conventional method Methods 0.000 description 1
- 238000003066 decision tree Methods 0.000 description 1
- 238000013136 deep learning model Methods 0.000 description 1
- 230000006872 improvement Effects 0.000 description 1
- 238000012986 modification Methods 0.000 description 1
- 230000004048 modification Effects 0.000 description 1
- 230000004044 response Effects 0.000 description 1
- 238000011425 standardization method Methods 0.000 description 1
- 230000009897 systematic effect Effects 0.000 description 1
- 238000009827 uniform distribution Methods 0.000 description 1
- 238000010200 validation analysis Methods 0.000 description 1
- 238000007794 visualization technique Methods 0.000 description 1
Images
Classifications
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06Q—INFORMATION AND COMMUNICATION TECHNOLOGY [ICT] SPECIALLY ADAPTED FOR ADMINISTRATIVE, COMMERCIAL, FINANCIAL, MANAGERIAL OR SUPERVISORY PURPOSES; SYSTEMS OR METHODS SPECIALLY ADAPTED FOR ADMINISTRATIVE, COMMERCIAL, FINANCIAL, MANAGERIAL OR SUPERVISORY PURPOSES, NOT OTHERWISE PROVIDED FOR
- G06Q10/00—Administration; Management
- G06Q10/06—Resources, workflows, human or project management; Enterprise or organisation planning; Enterprise or organisation modelling
- G06Q10/067—Enterprise or organisation modelling
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06N—COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
- G06N3/00—Computing arrangements based on biological models
- G06N3/02—Neural networks
- G06N3/04—Architecture, e.g. interconnection topology
- G06N3/045—Combinations of networks
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06Q—INFORMATION AND COMMUNICATION TECHNOLOGY [ICT] SPECIALLY ADAPTED FOR ADMINISTRATIVE, COMMERCIAL, FINANCIAL, MANAGERIAL OR SUPERVISORY PURPOSES; SYSTEMS OR METHODS SPECIALLY ADAPTED FOR ADMINISTRATIVE, COMMERCIAL, FINANCIAL, MANAGERIAL OR SUPERVISORY PURPOSES, NOT OTHERWISE PROVIDED FOR
- G06Q50/00—Information and communication technology [ICT] specially adapted for implementation of business processes of specific business sectors, e.g. utilities or tourism
- G06Q50/06—Energy or water supply
Landscapes
- Engineering & Computer Science (AREA)
- Business, Economics & Management (AREA)
- Human Resources & Organizations (AREA)
- Strategic Management (AREA)
- Physics & Mathematics (AREA)
- Economics (AREA)
- Theoretical Computer Science (AREA)
- Health & Medical Sciences (AREA)
- General Physics & Mathematics (AREA)
- Entrepreneurship & Innovation (AREA)
- Marketing (AREA)
- General Business, Economics & Management (AREA)
- Tourism & Hospitality (AREA)
- General Health & Medical Sciences (AREA)
- Game Theory and Decision Science (AREA)
- Biophysics (AREA)
- Operations Research (AREA)
- Educational Administration (AREA)
- Public Health (AREA)
- Water Supply & Treatment (AREA)
- Development Economics (AREA)
- Primary Health Care (AREA)
- Life Sciences & Earth Sciences (AREA)
- Artificial Intelligence (AREA)
- Biomedical Technology (AREA)
- Quality & Reliability (AREA)
- Computational Linguistics (AREA)
- Data Mining & Analysis (AREA)
- Evolutionary Computation (AREA)
- Molecular Biology (AREA)
- Computing Systems (AREA)
- General Engineering & Computer Science (AREA)
- Mathematical Physics (AREA)
- Software Systems (AREA)
- Supply And Distribution Of Alternating Current (AREA)
Abstract
The invention discloses a method for constructing a pre-judgment model of a power angle instability mode after a power system fault and application thereof, wherein the method comprises the following steps: extracting a group of bus voltage discrete point sets in the observation window corresponding to each group of original measurement data, and constructing a voltage amplitude matrix and a voltage phase angle matrix corresponding to each group of discrete point sets; determining instability mode label information of corresponding working conditions at a preset time period position behind the observation window according to the power angle curve form corresponding to each group of original measurement data; and training to obtain a neural network prejudgment model through supervised learning based on all samples, wherein each sample comprises a voltage amplitude matrix, a voltage phase angle matrix and instability mode label information corresponding to a group of original measurement data. According to the method, the deep learning is applied to the power angle instability mode prejudgment, the original data is measured according to the bus voltage phasor after the large disturbance, the specific category of the stability or the instability can be quickly and accurately given, the safety of the power system is greatly improved, and the practicability is high.
Description
Technical Field
The invention belongs to the field of power system instability prediction, and particularly relates to a prediction model construction method of a power angle instability mode after a power system fault and application thereof.
Background
The safe and stable operation of the power system is important for the national energy safety and the social and economic development. In interconnected large power grids, the stability problem is more prominent, and new and different characteristics are also shown. In early small grids, aperiodic transient instability was the most significant threat to safe and stable operation. With the increase of the interconnection scale of the power grid, the increase of the proportion of renewable energy sources, the increase of the use of power electronic conversion equipment and the heaviness of the load level, the dynamic characteristics of the power system become more and more complex, the instability forms become more, and the oscillation instability also becomes a common power angle instability situation. The method can be used for quickly and accurately predicting the stability and the instability of the power system after the large disturbance and the category of the instability, so that the time is won for emergency control measures, and meanwhile, a basis is provided for which measure is taken. However, the existing method has some problems in aspects of rapidity, accuracy, prediction fineness and adaptability.
For example, the time domain simulation method is developed more mature, the precision and the speed of the time domain simulation method can meet the requirements of planning, designing and operation mode calculation on transient stability analysis, the adaptability is strong, the provided information is rich, but the calculation burden is heavy, the time consumption is long, and the requirements are difficult to meet during online application. The direct method does not solve the time response of the state quantity of the system, can quickly give a judgment result, but is difficult to construct an energy function when the direct method is oriented to a large power grid, or is only suitable for a classical generator model, and the judgment result is slightly conservative. With the development of WAMS, data-driven machine learning methods emerge. The machine learning method has the problems that the feature extraction depends on expert experience seriously, repeated trial and error is needed, important information can be omitted or redundancy can be caused, the performance of the method depends on the selection of the features to a great extent, and in addition, the features extracted by the existing research are not intuitive, the calculation is complex, and the requirement on a sensor is high. The deep learning method developed in recent years can solve the problem, the straight-face original data is deeply learned, the required features are automatically extracted, the dependence on expert experience is eliminated, and some deep learning methods are good in effect. However, the conventional method using deep learning still cannot realize efficient and fine discrimination of the instability mode, and some instability modes are determined as stable modes, which threatens safety and stability of the power system.
Therefore, how to quickly and accurately pre-determine the instability mode is a technical problem to be solved urgently at present.
Disclosure of Invention
The invention provides a method for constructing a pre-judging model of a power angle instability mode after a power system fault and application thereof, which are used for solving the technical problem that the safe operation of the power system cannot be reliably ensured because the existing pre-judging method cannot realize the fine judgment of the instability mode.
The technical scheme for solving the technical problems is as follows: a pre-judgment model construction method for a power angle instability mode after a power system fault comprises the following steps:
step 1, extracting a group of bus voltage discrete point sets in an observation window corresponding to each group of original measurement data from each group of original measurement data, and constructing a voltage amplitude matrix and a voltage phase angle matrix corresponding to each group of discrete point sets;
step 2, determining instability mode label information of corresponding working conditions at a preset time period position behind the observation window according to the power angle curve form corresponding to each group of original measurement data;
and 3, training to obtain a neural network prejudgment model through supervised learning based on all samples, wherein each sample comprises a group of voltage amplitude matrix, voltage phase angle matrix and instability mode label information corresponding to the original measurement data.
The invention has the beneficial effects that: transient instability of the power system can be caused by insufficient synchronous torque, oscillation instability of the power system can be caused by insufficient damping torque, and safety and stability of the power system can be threatened if the instability is judged to be in a stable state. In addition, only if the categories of the two power angle instability modes are judged quickly and accurately, correct emergency control measures can be taken in a targeted manner, and loss caused by large disturbance is reduced. Therefore, the method firstly constructs the pre-judgment model of the power angle instability mode, specifically introduces deep learning based on the convolutional neural network, and trains the pre-judgment for the power angle stability or instability category of the power system. The method is characterized in that original data are measured according to the voltage phasor of the bus after large disturbance, a prejudgment model which can quickly and accurately give out a stable or unstable type (namely transient instability caused by insufficient synchronous torque or dynamic instability caused by insufficient damping torque) is trained, and the prejudgment model can be used for prejudgment of an unstable mode of an actual power system to provide a basis for making an emergency control measure. In addition, the construction method does not need to manually extract features, gets rid of dependence on expert experience, does not need to greedy and unsupervised pre-training layer by layer, and reduces training difficulty. Therefore, the pre-judging model constructed by the method can quickly and accurately pre-judge, can improve the pre-judging fineness, greatly improves the safety of the electric power system and has strong practicability.
On the basis of the technical scheme, the invention can be further improved as follows.
Further, the step 1 comprises:
acquiring multiple groups of original measurement data under different tidal current operating conditions, fault lines, fault positions and/or fault duration; extracting the bus voltage component of each group of original measurement data at each interval of half cycle in the corresponding observation window to form a voltage amplitude initial matrix and a voltage phase angle initial matrix; and normalizing all the voltage amplitude initial matrixes and all the voltage phase angle initial matrixes by adopting a Z-score normalization method to form a voltage amplitude matrix and a voltage phase angle matrix corresponding to each group of original measurement data.
The invention has the further beneficial effects that: in order to obtain a large amount of data for training and better approach to the actual working condition, multiple time domain simulations can be performed by setting different tide operation working conditions, fault lines, fault positions and/or fault duration time to obtain multiple groups of original measurement data. And extracting all bus voltage phasors of each group of original measurement data in the corresponding observation window, sampling a point at every half cycle in all bus voltage vectors corresponding to each group of original measurement data, forming a voltage amplitude initial matrix and a voltage phase angle initial matrix corresponding to the group of original measurement data, and ensuring that real working condition data are adopted to ensure reliable prejudgment. In order to have the same dimension, all voltage amplitude initial matrixes are normalized by a Z-score normalization method, and all voltage angle initial matrixes are normalized.
Further, the step 1 further comprises:
and adding noise data into the voltage amplitude matrix and the voltage phase angle matrix corresponding to each group of discrete point sets to form a new voltage amplitude matrix and a new voltage phase angle matrix.
The invention has the further beneficial effects that: the noise is added into the original matrix data, and the method can be directly used for training a pre-judging model, and can also adopt a sample with the noise to be retrained on the basis of the pre-judging model obtained by training the sample before the noise is added so as to improve the robustness of the pre-judging model to the noise of the measuring system.
Further, in the step 2, the power angle curve form corresponding to each set of original measurement data specifically includes: and forming factors according to the power angle curve form corresponding to each group of original measurement data and each power angle instability mode.
Further, the instability mode tag information includes a stable mode, a transient instability mode, and a dynamic instability mode.
Further, the convolutional neural network prejudgment model comprises:
the convolutional neural network is used for extracting a characteristic vector from each sample and sending the characteristic vector to the fully-connected neural network;
the fully-connected neural network is used for classifying each sample through nonlinear transformation and a softmax function based on the characteristic vector corresponding to the sample.
Further, the convolutional neural network comprises a convolutional layer, a batch normalization layer, an activation function and a Dropout layer which are connected in sequence;
the fully-connected neural network includes: the neuron comprises a full connection layer, a Dropout layer and three neurons, wherein the full connection layer and the Dropout layer are connected in sequence, and the three neurons are connected with the Dropout layer.
Further, the supervised learning specifically comprises:
and performing supervised learning by adopting a random gradient descent method with hot restart and taking a cross entropy loss function as a target function, wherein output results of different layers in the convolutional neural network prejudgment model are visualized in the process of the supervised learning.
The invention has the further beneficial effects that: the method uses the convolutional neural network to automatically extract required characteristics from original measurement data, gets rid of dependence on expert experience, and then carries out classification through the full-connection network, so that a complex mapping relation from bottom layer measurement data to a power angle instability mode is established through a deep neural network. During off-line training, as new technologies such as ReLU activation, Batch Normalization (Batch Normalization), Xavier initialization and the like are used, the model can directly supervise learning without greedy and complicated process of unsupervised pre-training layer by layer; in the process of supervised learning, a random gradient descent algorithm with 'hot restart' is used, so that the model tends to converge to an optimal point of 'flat and wide', and the generalization capability is good.
The invention also provides a power angle instability mode prejudging method after the power system fault, which comprises the following steps:
step 1, collecting original measurement data to be measured of an observation window at the rear section of a fault after the power system to be measured breaks down;
step 2, processing the original measurement data to be measured based on the data processing method in the step 1 in the pre-judging model construction method of the power angle instability mode after any power system fault to obtain a group of voltage amplitude matrixes and voltage phase angle matrixes;
and 3, based on the group of voltage amplitude matrixes and voltage phase angle matrixes, adopting any one of the above pre-judgment model construction methods of the power angle instability mode after the power system fault to construct an obtained neural network pre-judgment model, and pre-judging to obtain the power angle instability mode after the fault.
The invention has the beneficial effects that: the method is used for pre-judging the power angle instability mode for the first time, and specifically introduces a pre-judging model based on a convolutional neural network to pre-judge the power angle stability or instability type of the power system. According to the method, the stability or instability category (namely transient instability caused by insufficient synchronous torque or dynamic instability caused by insufficient damping torque) can be quickly and accurately given according to the original data measured by the voltage phasor of the bus after large disturbance, so that a basis is provided for formulating an emergency control measure. In addition, the method does not need to manually extract features, and gets rid of dependence on expert experience. Therefore, the method can improve the prejudgment fineness while quickly and accurately prejudging, greatly improve the safety of the electric power system and have stronger practicability.
The present invention also provides a storage medium, in which instructions are stored, and when the instructions are read by a computer, the computer is enabled to execute any one of the above methods for predicting a power angle instability mode of a power system after a fault and/or one of the above methods for predicting a power angle instability mode of a power system after a fault.
Drawings
Fig. 1 is a schematic flow chart diagram of a method for constructing a prediction model of a power angle instability mode after a power system failure according to an embodiment of the present invention;
fig. 2 is a flowchart of construction and application of a prediction model of a power angle instability mode after a power system failure according to an embodiment of the present invention;
fig. 3 is a single line diagram of a new england 10 machine 39 node system provided by an embodiment of the present invention;
FIG. 4 is a learning rate variation trend chart of a random gradient descent algorithm with warm restart in the pre-decision model training process provided by the embodiment of the present invention;
FIG. 5 is a schematic diagram of a neural network according to an embodiment of the present invention;
fig. 6 is a visualization result of a test sample in feature spaces of different levels of a neural network according to an embodiment of the present invention.
Detailed Description
In order to make the objects, technical solutions and advantages of the present invention more apparent, the present invention is described in further detail below with reference to the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are merely illustrative of the invention and are not intended to limit the invention. In addition, the technical features involved in the embodiments of the present invention described below may be combined with each other as long as they do not conflict with each other.
Example one
A method 100 for constructing a pre-determination model of a power angle instability mode after a power system fault is shown in fig. 1, which includes:
and step 130, training to obtain a neural network prejudgment model through supervised learning based on all samples, wherein each sample comprises a voltage amplitude matrix, a voltage phase angle matrix and instability mode label information corresponding to a group of original measurement data.
It should be noted that all samples corresponding to all groups of original measurement data are randomly divided into a training set, a verification set and a test set, a prejudgment model based on a convolutional neural network is built, the training set is input into the prejudgment model for training, a hyper-parameter is adjusted according to the effect of the verification set, and finally the test set is used for checking the performance of the prejudgment model. In addition, the unstable mode tag information includes stable state tag information and each of the unstable mode tag information.
With respect to the preset time period after the observation window, it should be noted that the observation window is generally small, for example, within 0.25s after the fault occurs (the specific time is determined by the action characteristic of the relay protection, and is the longest time that the fault may last, which can be obtained from historical data), and based on the data in the observation window, the stability of the fault after the fault occurs in a long period (for example, at 20s after the fault occurs) is predicted. By training the mapping relation of the neural network, the stable or unstable state of the subsequent time period can be predicted through the data of the observation window.
Transient instability of the power system can be caused by insufficient synchronous torque, oscillation instability of the power system can be caused by insufficient damping torque, and if vibration instability is judged to be in a stable state, safety and stability of the power system are threatened. In addition, only if the categories of the two power angle instability modes are judged quickly and accurately, correct emergency control measures can be taken in a targeted manner, and loss caused by large disturbance is reduced. Therefore, the method firstly constructs the pre-judgment model of the power angle instability mode, specifically introduces deep learning based on the convolutional neural network, and trains the pre-judgment for the power angle stability or instability category of the power system. The method is characterized in that original data are measured according to the voltage phasor of the bus after large disturbance, a prejudgment model which can quickly and accurately give out a stable or unstable type (namely transient instability caused by insufficient synchronous torque or dynamic instability caused by insufficient damping torque) is trained, and the prejudgment model can be used for prejudgment of an unstable mode of an actual power system to provide a basis for making an emergency control measure. In addition, the construction method does not need to manually extract features, gets rid of dependence on expert experience, does not need to greedy and unsupervised pre-training layer by layer, and reduces training difficulty. Therefore, the pre-judging model constructed by the method can quickly and accurately pre-judge, can improve the pre-judging fineness, greatly improves the safety of the electric power system, and has strong practicability.
It should be noted that large disturbance and small disturbance are terms of expertise defined in the power industry, which is to facilitate analysis and deep understanding of the stability problem, and power angle stability is divided into small-interference power angle stability and large-interference power angle stability (also called transient stability) according to the magnitude of disturbance. The small disturbance is small enough to linearize the nonlinear differential equation of the power system at a balance point, and the stability problem is researched on the basis of the linearization of the nonlinear differential equation of the power system, so that the precision is not influenced; whereas large perturbation stabilities have to be investigated by means of nonlinear differential equations. Wherein specific how big, how small, there is not quantitative standard, think big disturbance including short circuit, excision transmission line etc. on the general engineering, little disturbance includes: changes in load, changes in distance between overhead lines due to wind blowing, and the like. The large disturbance and the small disturbance are generally separately researched, different methods are adopted, and the method mainly researches the stability after the short-circuit fault, namely the large disturbance stability problem.
Preferably, step 110 includes: acquiring multiple groups of original measurement data under different tidal current operating conditions, fault lines, fault positions and/or fault duration; extracting the bus voltage component of each group of original measurement data at each interval of half cycle in the corresponding observation window to form a voltage amplitude initial matrix and a voltage phase angle initial matrix; and standardizing all the voltage amplitude initial matrixes and all the voltage phase angle initial matrixes by adopting a Z-score standardization method to form a voltage amplitude matrix and a voltage phase angle matrix corresponding to each group of original measurement data.
It should be noted that the amplitude matrix and the phase angle matrix are in the following forms:
in the formula of Ui,jAnd thetai,j(i-1, 2, …, T, j-1, 2, …, B) is the ith sample, respectivelyAnd (3) amplitude and phase angle of jth bus voltage at the moment, T is the number of sampling points corresponding to the length of the observation window, and B is the number of buses in the power system. Two matrices are stacked as two channels of model input. The length of the observation window is set to the maximum time that the fault may last, which can be obtained from historical fault data.
In addition, Z-score normalization is done by: calculating the average value and the standard deviation of each identical position for all the voltage amplitude initial matrixes, subtracting the average value of each position from the data of each position of each voltage amplitude initial matrix, dividing the data by the standard deviation to obtain the value of the voltage amplitude initial matrix after standardization at the position, and processing the same amplitude matrix by the phase angle matrix for standardization to eliminate the dimension.
In order to obtain a large amount of data for training and better approach to the actual working condition, multiple time domain simulations can be performed by setting different tide operation working conditions, fault lines, fault positions and/or fault duration time to obtain multiple groups of original measurement data. And extracting all bus voltage phasors of each group of original measurement data in the corresponding observation window, sampling a point at every half cycle in all bus voltage vectors corresponding to each group of original measurement data, forming a voltage amplitude initial matrix and a voltage phase angle initial matrix corresponding to the group of original measurement data, and ensuring that real working condition data are adopted to ensure reliable prejudgment. In order to have the same dimension, all voltage amplitude initial matrices are normalized and all voltage angle initial matrices are normalized by a Z-score normalization method.
Preferably, step 110 further comprises:
and adding noise data into the voltage amplitude matrix and the voltage phase angle matrix corresponding to each group of discrete point sets to form a new voltage amplitude matrix and a new voltage phase angle matrix.
The noise is added into the original matrix data, and the method can be directly used for training a pre-judging model, and can also adopt a sample with the noise to be retrained on the basis of the pre-judging model obtained by training the sample before the noise is added so as to improve the robustness of the pre-judging model to the noise of the measuring system.
Preferably, in step 120, the power angle curve form corresponding to each set of raw measurement data specifically includes: according to the power angle curve form corresponding to each group of original measurement data and the forming factors of each power angle instability mode; the form factor is the dominant form factor,
preferably, the instability mode tag information includes a stable mode, a transient instability mode, and a dynamic instability mode.
According to the regulations of the safety and stability of the power system (DL 755-supple 2001), the large-disturbance power angle stability can be divided into transient stability and dynamic stability according to different leading factors (insufficient synchronous torque and insufficient damping torque) which cause the instability of the power angle, wherein the transient stability refers to the power angle stability of the first and second swings after the system is subjected to large disturbance, the physical characteristics of the transient stability are mainly related to synchronous torque, and the criterion is that the transient stability does not lose synchronization after passing through the first or second oscillation period; the dynamic stability mainly refers to the long-time power angle stability of a system under the action of a system dynamic element and a control device after the system is subjected to large disturbance, the physical characteristics of the dynamic stability are mainly related to damping torque, and the criterion is that a generator is in a damped oscillation state relative to a power angle in the dynamic swing process. The samples are labeled as DL 755-2001, e.g., stationary 0, transient destabilizing 1, dynamic destabilizing 2.
Preferably, the convolutional neural network prediction model includes: convolutional neural networks and fully-connected neural networks. The convolutional neural network is used for extracting a characteristic vector from each sample and sending the characteristic vector to the fully-connected neural network; the fully-connected neural network is used for classifying each sample through nonlinear transformation and a softmax function based on the corresponding feature vector of the sample.
Preferably, the convolutional neural network comprises a convolutional layer, a batch normalization layer, an activation function and a Dropout layer which are connected in sequence; the fully-connected neural network comprises: the neuron comprises a full connection layer, a Dropout layer and three neurons, wherein the full connection layer and the Dropout layer are connected in sequence, and the three neurons are connected with the Dropout layer.
It should be noted that three neurons are arranged at the top layer, a normalized probability is obtained in each neuron after processing through a softmax function, and the class corresponding to the neuron with the highest probability is used as a model pre-judgment result.
Preferably, the supervised learning is specifically: and performing supervised learning by adopting a random gradient descent method with hot restart and taking a cross entropy loss function as a target function, wherein output results of different layers in the convolutional neural network prejudgment model are visualized in the process of the supervised learning.
The supervised learning adopts random gradient descent with hot restart to control the periodic change of the learning rate; the learning rate is reduced according to the cosine function rule in each period, and the learning rate returns to the maximum value again at the end of each period, so that the model tends to converge to the optimal point of 'wide and flat', and the generalization performance is better. The learning rate periodic variation strategy may be:
in the formula, lr0For the initial learning rate, T is the total number of training iterations, M is the total number of cycles of learning rate change, and epoch is the current number of iterations.
In addition, the performance of the test model not only comprises indexes such as accuracy, but also can perform tSNE (t-distributed stored systematic neighbor embedding) visualization on test sample points in different hierarchical feature spaces of the neural network, and visually observe the classification effect of the model to monitor the classification performance of the model.
The method uses the convolutional neural network to automatically extract required characteristics from original measurement data, gets rid of dependence on expert experience, and then carries out classification through the full-connection network, so that a complex mapping relation from bottom layer measurement data to a power angle instability mode is established through a deep neural network. During off-line training, as new technologies such as ReLU activation, batch normalization layer (batch normalization), Xavier initialization and the like are used, the model can directly supervise learning without greedy and complicated process of unsupervised pre-training layer by layer; in the process of supervised learning, a random gradient descent algorithm with 'hot restart' is used, so that the model tends to converge to an optimal point of 'flat and wide', and the generalization capability is good.
For better illustration of the invention, the following examples are given:
the method flow of the invention is shown in fig. 2, and is explained by taking a 39-node system of a new england 10 machine as an example, wherein a single line of the system is shown in fig. 3, in the figure, G1-G10 represent generators, thick lines 1-39 represent buses, thin lines represent transmission lines, and arrows represent loads. In the generation stage of the sample set, transient stability simulation is carried out, and voltage data collected by a PMU and a WMAS system are simulated. The pre-fault operating conditions include: 70%, 75%, 80%, … …, 130%, 13 load levels, and the output level of the generator is adjusted according to the load levels, and the voltage of each bus is ensured to be in an allowable range. Three-phase metallic short-circuit faults are arranged on all 34 non-transformer lines in the system, the short-circuit positions are respectively 0%, 20%, 50% and 80%, and the faults respectively last for 0.1s, 0.15s, 0.18s, 0.20s and 0.25 s. The simulation time duration was set to 20s, resulting in 8840 labeled samples (voltage magnitude matrix, voltage phase angle matrix, and destabilizing mode label information). The sample set is randomly divided into three parts of 70%, 15% and 15%, which are respectively a training set, a verification set and a test set. Since the maximum time that a fault may last is 0.25s, the fault window is set to 0.25s, i.e., bus voltage magnitude data (one point per half cycle sample) within 0.25s after the fault is extracted from each sample.
In the off-line training stage, the training set data is input into the deep learning model for training after being subjected to Z-score standardization as described in step 110, and the verification set and the test set are standardized according to the same parameters. During training, the weights of the neural network are initialized by using Xavier uniform distribution, further training is carried out according to a random gradient descent algorithm with hot restart, and the learning rate is controlled to change periodically in the training process as shown in figure 4. And adjusting hyper-parameters such as the number of network layers, the size of a convolution kernel, the convolution step length and the like according to the validation set effect to enable the model performance to be optimal, and after training, checking the model with a test set to obtain a neural network framework as shown in figure 5. The confusion matrix of the test results is shown in table 1, and the accuracy reaches 97.74%.
TABLE 1 confusion matrix of test set results
Other machine learning models, including decision trees, discriminant analysis, support vector machines, nearest neighbor classifiers, etc., were trained using the same normalized input data, and the test results are listed in table 2. The test result shows that the power angle instability type pre-judging model based on the convolutional neural network has more obvious advantages in evaluation accuracy compared with other machine learning models, the accuracy is 1.9% higher than that of the highest support vector machine in other algorithms, and in addition, compared with a conventional random gradient descent algorithm, the accuracy can be improved even if the algorithm is continuously restarted.
TABLE 2 comparison of test results of the method based on convolutional neural network with other machine learning methods
In order to visually represent the capability of automatically extracting features and the capability of identifying and distinguishing power angle instability modes in the method, a tSNE visualization method is used, test sample points in high-dimensional feature spaces of different levels in a deep neural network are projected to a two-dimensional plane for visualization, as shown in FIG. 6, the closer sample points in the diagram represent that the similarity is higher, wherein the convolution output feature in FIG. 6 refers to the position where the convolution neural network is connected with a fully-connected neural network, namely the output quantity of the convolution neural network; the "secondary output feature" in fig. 6 refers to the second last layer of the fully-connected neural network, i.e., the layer before softmax. Inputting data, namely, in an original characteristic space, different types of samples are overlapped in a large area and cannot be distinguished; the feature expression processed by the convolutional neural network enables three types of samples to be distinguished to a certain degree, the samples of different types are respectively aggregated into clusters, and the number of overlapped samples is obviously reduced, so that the powerful feature extraction capability of the convolutional network is shown, and the representation method processed by the convolutional network is beneficial to the next classification task; the characteristics of the secondary output layer (the previous layer of Softmax) after nonlinear mapping of the fully-connected neural network are well classified into three types of samples, so that the judgment result output after the topmost layer of mapping has higher accuracy.
In the invention, noise is added into the sample data for retraining, so that the model has better noise robustness. According to the IEEE standard (C37.118.2-2011), Gaussian white noise with a signal-to-noise ratio of 40dB is added to the total sample, retraining is performed, and the test results are listed in Table 2 as accuracy (plus noise) and "difference" from the case where the noise is not considered. The results show that when the noise interference of PMU is considered, the evaluation accuracy of each model is reduced to a certain extent, but at the moment, the model of the invention still can maintain 97% accuracy, which exceeds other comparison models, and in addition, the accuracy of the model is reduced by only 0.7%, and is also in a better level in other comparison models. Therefore, the power angle instability mode pre-judging model based on the convolutional neural network has better robustness on PMU noise.
Example two
A method 200 for predicting a power angle instability mode after a power system fault comprises the following steps:
step 210, after a fault occurs in the power system to be measured, collecting original measurement data to be measured in a first preset time period after the fault occurs;
step 220, processing original measurement data to be measured based on the data processing method in step 110 in the pre-judging model construction method of the post-fault power angle instability mode of the power system according to the first embodiment to obtain a group of voltage amplitude matrixes and voltage phase angle matrixes;
step 230, based on the set of voltage amplitude matrix and voltage phase angle matrix, a neural network pre-judgment model is constructed by using any one of the pre-judgment model construction methods of the power angle instability mode after the power system fault according to the first embodiment, and the power angle instability mode after the fault is obtained through pre-judgment.
For example, the pre-determination model trained and tested in the first embodiment may be deployed to a control center, and after a power system fails, the pre-determination result of the power angle stability and the instability mode may be quickly obtained by performing Z-score standardization according to voltage phasor measurement data in an observation window returned by the WAMS system and inputting the pre-determination model.
The related technical solution is the same as the first embodiment, and is not described herein again.
The method comprises the steps of carrying out prejudgment on a power angle instability mode for the first time, specifically introducing a prejudgment model based on a convolutional neural network, and carrying out prejudgment on power angle stability or instability types of the power system. According to the method, the stability or instability category (namely transient instability caused by insufficient synchronous torque or dynamic instability caused by insufficient damping torque) can be quickly and accurately given according to the original data measured by the voltage phasor of the bus after large disturbance, so that a basis is provided for formulating an emergency control measure. In addition, the method does not need to manually extract features, and gets rid of dependence on expert experience. Therefore, the method can improve the prejudgment fineness while rapidly and accurately prejudging, greatly improves the safety of the electric power system, and has strong practicability.
EXAMPLE III
A storage medium, wherein instructions are stored in the storage medium, and when the instructions are read by a computer, the computer is caused to execute any one of the power system power angle instability mode pre-judging methods described in the first embodiment above after a fault and/or one of the power system power angle instability mode pre-judging methods described in the second embodiment above after a fault.
The related technical solutions are the same as those of the first embodiment and the second embodiment, and are not described herein again.
It will be understood by those skilled in the art that the foregoing is only a preferred embodiment of the present invention, and is not intended to limit the invention, and that any modification, equivalent replacement, or improvement made within the spirit and principle of the present invention should be included in the scope of the present invention.
Claims (10)
1. A pre-judging model construction method for a power angle instability mode after a power system fault is characterized by comprising the following steps:
step 1, extracting a group of bus voltage discrete point sets in an observation window corresponding to each group of original measurement data from each group of original measurement data, and constructing a voltage amplitude matrix and a voltage phase angle matrix corresponding to each group of discrete point sets;
step 2, determining instability mode label information of corresponding working conditions at a preset time period position behind the observation window according to the power angle curve form corresponding to each group of original measurement data;
and 3, training to obtain a neural network prejudgment model through supervised learning based on all samples, wherein each sample comprises a group of voltage amplitude matrix, voltage phase angle matrix and instability mode label information corresponding to the original measurement data.
2. The method for constructing the prediction model of the power angle instability mode after the power system fault according to claim 1, wherein the step 1 includes:
acquiring multiple groups of original measurement data under different tidal current operating conditions, fault lines, fault positions and/or fault duration; extracting the bus voltage component of each group of original measurement data at each interval of half cycle in the corresponding observation window to form a voltage amplitude initial matrix and a voltage phase angle initial matrix; and normalizing all the voltage amplitude initial matrixes and all the voltage phase angle initial matrixes by adopting a Z-score normalization method to form a voltage amplitude matrix and a voltage phase angle matrix corresponding to each group of original measurement data.
3. The method for constructing the prediction model of the power angle instability mode after the power system fault according to claim 1, wherein the step 1 further comprises:
and adding noise data into the voltage amplitude matrix and the voltage phase angle matrix corresponding to each group of discrete point sets to form a new voltage amplitude matrix and a new voltage phase angle matrix.
4. The method as claimed in claim 1, wherein in the step 2, the power angle curve form corresponding to each set of original measurement data is specifically a power angle curve form corresponding to each set of original measurement data and a forming factor of each power angle instability mode.
5. The method according to claim 1, wherein the instability mode label information includes a stable mode, a transient instability mode, and a dynamic instability mode.
6. The method according to any one of claims 1 to 5, wherein the convolutional neural network prediction model comprises:
the convolutional neural network is used for extracting a characteristic vector from each sample and sending the characteristic vector to the fully-connected neural network;
the fully-connected neural network is used for classifying each sample through nonlinear transformation and a softmax function based on the characteristic vector corresponding to the sample.
7. The method according to claim 6, wherein the convolutional neural network comprises a convolutional layer, a batch normalization layer, an activation function and a Dropout layer which are connected in sequence;
the fully-connected neural network includes: the neuron comprises a full connection layer, a Dropout layer and three neurons, wherein the full connection layer and the Dropout layer are connected in sequence, and the three neurons are connected with the Dropout layer.
8. The method according to claim 7, wherein the supervised learning specifically comprises:
and performing supervised learning by adopting a random gradient descent method with hot restart and taking a cross entropy loss function as a target function, wherein output results of different layers in the convolutional neural network prejudgment model are visualized in the process of the supervised learning.
9. A method for prejudging a power angle instability mode after a power system fault is characterized by comprising the following steps:
step 1, collecting original measurement data to be measured of an observation window after a fault occurs in a power system to be measured;
step 2, processing the original measurement data to be measured based on the data processing method in the step 1 in the pre-judging model construction method of the post-fault power angle instability mode of the power system according to any one of claims 1 to 8 to obtain a group of voltage amplitude matrixes and voltage phase angle matrixes;
and 3, based on the group of voltage amplitude matrixes and voltage phase angle matrixes, adopting the pre-judgment model construction method of the post-fault power angle instability mode of the power system as claimed in any one of claims 1 to 8 to construct an obtained neural network pre-judgment model, and pre-judging to obtain the post-fault power angle instability mode.
10. A storage medium, wherein instructions are stored in the storage medium, and when the instructions are read by a computer, the computer is caused to execute the method for constructing the prediction model of the power system power-angle instability mode after the fault according to any one of claims 1 to 8 and/or the method for predicting the power system power-angle instability mode after the fault according to claim 9.
Priority Applications (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN201910844471.3A CN110705831A (en) | 2019-09-06 | 2019-09-06 | Power angle instability mode pre-judgment model construction method after power system fault and application thereof |
Applications Claiming Priority (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN201910844471.3A CN110705831A (en) | 2019-09-06 | 2019-09-06 | Power angle instability mode pre-judgment model construction method after power system fault and application thereof |
Publications (1)
| Publication Number | Publication Date |
|---|---|
| CN110705831A true CN110705831A (en) | 2020-01-17 |
Family
ID=69194664
Family Applications (1)
| Application Number | Title | Priority Date | Filing Date |
|---|---|---|---|
| CN201910844471.3A Pending CN110705831A (en) | 2019-09-06 | 2019-09-06 | Power angle instability mode pre-judgment model construction method after power system fault and application thereof |
Country Status (1)
| Country | Link |
|---|---|
| CN (1) | CN110705831A (en) |
Cited By (4)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN112200694A (en) * | 2020-10-09 | 2021-01-08 | 华中科技大学 | Dominant instability mode identification model construction and application method based on graph neural network |
| CN113435492A (en) * | 2021-06-21 | 2021-09-24 | 华中科技大学 | Power system dominant instability mode discrimination method based on active learning |
| CN113488992A (en) * | 2021-06-21 | 2021-10-08 | 清华大学 | Method for judging large disturbance stability of electric power system |
| CN113507430A (en) * | 2021-05-10 | 2021-10-15 | 湖南大学 | Method and system for detecting information physical cooperation attack of power system |
Citations (3)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| WO2018006499A1 (en) * | 2016-07-06 | 2018-01-11 | 南方电网科学研究院有限责任公司 | Dominant instability mode identification method and system for power system |
| CN108832619A (en) * | 2018-05-29 | 2018-11-16 | 北京交通大学 | Transient stability evaluation in power system method based on convolutional neural networks |
| CN109886604A (en) * | 2019-03-13 | 2019-06-14 | 华北电力大学 | A kind of transient stability evaluation in power system method based on one-dimensional convolutional neural networks |
-
2019
- 2019-09-06 CN CN201910844471.3A patent/CN110705831A/en active Pending
Patent Citations (3)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| WO2018006499A1 (en) * | 2016-07-06 | 2018-01-11 | 南方电网科学研究院有限责任公司 | Dominant instability mode identification method and system for power system |
| CN108832619A (en) * | 2018-05-29 | 2018-11-16 | 北京交通大学 | Transient stability evaluation in power system method based on convolutional neural networks |
| CN109886604A (en) * | 2019-03-13 | 2019-06-14 | 华北电力大学 | A kind of transient stability evaluation in power system method based on one-dimensional convolutional neural networks |
Non-Patent Citations (2)
| Title |
|---|
| GAO HUANG ET AL: "Snapshot Ensembles：Train1，get M for free", 《ARXIV》 * |
| 肖登明 等: "《电气工程概论》", 31 March 2005, 中国电力出版社 * |
Cited By (8)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN112200694A (en) * | 2020-10-09 | 2021-01-08 | 华中科技大学 | Dominant instability mode identification model construction and application method based on graph neural network |
| CN112200694B (en) * | 2020-10-09 | 2023-11-14 | 华中科技大学 | Dominant instability mode identification model construction and application method based on graph neural network |
| CN113507430A (en) * | 2021-05-10 | 2021-10-15 | 湖南大学 | Method and system for detecting information physical cooperation attack of power system |
| CN113507430B (en) * | 2021-05-10 | 2022-04-22 | 湖南大学 | Method and system for detecting information physical cooperation attack of power system |
| CN113435492A (en) * | 2021-06-21 | 2021-09-24 | 华中科技大学 | Power system dominant instability mode discrimination method based on active learning |
| CN113488992A (en) * | 2021-06-21 | 2021-10-08 | 清华大学 | Method for judging large disturbance stability of electric power system |
| CN113435492B (en) * | 2021-06-21 | 2022-04-22 | 华中科技大学 | Power system dominant instability mode discrimination method based on active learning |
| CN113488992B (en) * | 2021-06-21 | 2024-06-04 | 清华大学 | Large disturbance stability judging method for power system |
Similar Documents
| Publication | Publication Date | Title |
|---|---|---|
| Hu et al. | Real-time transient stability assessment in power system based on improved SVM | |
| Li et al. | A hierarchical data-driven method for event-based load shedding against fault-induced delayed voltage recovery in power systems | |
| Wang et al. | Toward the prediction level of situation awareness for electric power systems using CNN-LSTM network | |
| Baltas et al. | A comparative analysis of decision trees, support vector machines and artificial neural networks for on-line transient stability assessment | |
| CN102074955B (en) | Method based on knowledge discovery technology for stability assessment and control of electric system | |
| CN110705831A (en) | Power angle instability mode pre-judgment model construction method after power system fault and application thereof | |
| CN110994604B (en) | Power system transient stability assessment method based on LSTM-DNN model | |
| CN108551167A (en) | A kind of electric power system transient stability method of discrimination based on XGBoost algorithms | |
| CN111401749A (en) | Dynamic safety assessment method based on random forest and extreme learning regression | |
| CN112508442B (en) | Transient stability assessment method and system based on automatic and interpretable machine learning | |
| CN111680875B (en) | Unmanned aerial vehicle state risk fuzzy comprehensive evaluation method based on probability baseline model | |
| CN112069727B (en) | Intelligent transient stability evaluation system and method with high reliability for power system | |
| Li et al. | A hybrid machine learning framework for enhancing PMU-based event identification with limited labels | |
| CN114169395B (en) | Method for constructing dominant instability mode recognition model of power system and application | |
| CN113612237A (en) | Method for positioning resonance-induced subsynchronous oscillation source in offshore wind farm | |
| Gao et al. | Real-time long-term voltage stability assessment based on eGBDT for large-scale power system with high renewables penetration | |
| Shahriyari et al. | A Deep Learning-Based Approach for Comprehensive Rotor Angle Stability Assessment | |
| CN116561573A (en) | Training set circulation splitting training method for voltage stability evaluation | |
| Ren et al. | A universal defense strategy for data-driven power system stability assessment models under adversarial examples | |
| Wang et al. | SVM based imbalanced correction method for Power Systems Transient stability evaluation | |
| CN112564090B (en) | MBLDA-based AC/DC system transient voltage stability monitoring method | |
| Yang et al. | Fast assessment of short-term voltage stability of AC/DC power grid based on CNN | |
| Bouazzi et al. | Deep recurrent neural networks based Bayesian optimization for fault diagnosis of uncertain GCPV systems depending on outdoor condition variation | |
| Ma et al. | Frequency security assessment for receiving-end system based on deep learning method | |
| CN111965442A (en) | Energy internet fault diagnosis method and device under digital twin environment |
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: 20200117 |
|
| RJ01 | Rejection of invention patent application after publication |