CN113792481B - VSG multi-machine system transient stability evaluation method based on artificial neural network - Google Patents

VSG multi-machine system transient stability evaluation method based on artificial neural network Download PDF

Info

Publication number
CN113792481B
CN113792481B CN202111018894.3A CN202111018894A CN113792481B CN 113792481 B CN113792481 B CN 113792481B CN 202111018894 A CN202111018894 A CN 202111018894A CN 113792481 B CN113792481 B CN 113792481B
Authority
CN
China
Prior art keywords
fault
output
neural network
artificial neural
layer
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.)
Active
Application number
CN202111018894.3A
Other languages
Chinese (zh)
Other versions
CN113792481A (en
Inventor
帅智康
赵慧敏
沈阳
赵峰
程慧婕
葛俊
黄文�
沈霞
董雪梅
王钰泉
Current Assignee (The listed assignees may be inaccurate. Google has not performed a legal analysis and makes no representation or warranty as to the accuracy of the list.)
Hunan University
Original Assignee
Hunan University
Priority date (The priority date 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 date listed.)
Filing date
Publication date
Application filed by Hunan University filed Critical Hunan University
Priority to CN202111018894.3A priority Critical patent/CN113792481B/en
Publication of CN113792481A publication Critical patent/CN113792481A/en
Application granted granted Critical
Publication of CN113792481B publication Critical patent/CN113792481B/en
Active legal-status Critical Current
Anticipated expiration legal-status Critical

Links

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
    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06F—ELECTRIC DIGITAL DATA PROCESSING
    • G06F18/00—Pattern recognition
    • G06F18/20—Analysing
    • G06F18/21—Design or setup of recognition systems or techniques; Extraction of features in feature space; Blind source separation
    • G06F18/214—Generating training patterns; Bootstrap methods, e.g. bagging or boosting
    • 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
    • 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
    • Y02—TECHNOLOGIES OR APPLICATIONS FOR MITIGATION OR ADAPTATION AGAINST CLIMATE CHANGE
    • Y02B—CLIMATE CHANGE MITIGATION TECHNOLOGIES RELATED TO BUILDINGS, e.g. HOUSING, HOUSE APPLIANCES OR RELATED END-USER APPLICATIONS
    • Y02B10/00—Integration of renewable energy sources in buildings
    • Y02B10/10—Photovoltaic [PV]

Abstract

The VSG multi-machine system transient stability evaluation method based on the artificial neural network comprises the following steps: (1) determining input features and output features; (2) changing the load level of a VSG multi-machine system, changing the voltage drop of a power grid, changing the active reference value, virtual inertia and damping coefficient of a virtual synchronous machine, changing the current limiting amplitude value of the virtual synchronous machine, changing the fault duration, generating sample data, and dividing a training sample set and a test sample set; (3) carrying out data preprocessing on the input characteristics and the output characteristics; (4) building an artificial neural network; (5) initializing an artificial neural network; (6) training an artificial neural network by using a training sample set to complete the establishment of an evaluation model; (7) carrying out evaluation model test by using the test sample set; (8) the evaluation model is applied online. The method can effectively realize the rapid and accurate evaluation of the multi-machine micro-grid system based on the artificial neural network.

Description

VSG multi-machine system transient stability evaluation method based on artificial neural network
Technical Field
The invention relates to the technical field of micro-grids, in particular to a transient stability evaluation method of a Virtual Synchronous Generator (VSG) multi-machine grid-connected system based on an artificial neural network.
Background
The virtual synchronous machine (VSG) simulates the frequency modulation external characteristic of a synchronous generator, and when the micro-grid is subjected to large disturbance, the virtual synchronous machine has the problem of power angle instability and influences the safe operation of the micro-grid. Therefore, in order to provide information and strive for time for micro-grid transient stability prevention control, the safe and stable operation of the micro-grid is ensured, and accurate and rapid online transient stability evaluation is indispensable.
The existing transient stability evaluation method mainly comprises a time domain simulation method and an energy function method. The time domain simulation method simulates the virtual synchronous machine through numerical calculation and visually displays the change of each variable of the system after disturbance. However, the time domain simulation method value generally can only perform transient stability test on the system so as to perform corresponding planning on the system, for example, the grid-connected stability hardware-in-loop test system and method of the photovoltaic virtual synchronous machine disclosed in CN201710446496.9 are difficult to perform online application on actual systems with variable states. And when multiple machines are accessed, the time-domain simulation method is long in calculation time consumption, extremely high in requirements on hardware equipment and difficult to realize rapid online evaluation. And judging whether the disturbed system can be converged at a balance point or not from an energy angle by an energy function method, and calculating a corresponding attraction domain. However, stability studies based on the energy function method can only perform stability analysis on a simple and ideal specific nonlinear system, for example, the determination method for transient power angle stability of a virtual machine disclosed in CN201811602399.5, and most of the obtained determination results tend to be conservative, and cannot be applied to a multi-machine microgrid system with a high system order. In addition, there is no systematic energy function construction method currently available for virtual synchronous machines with current limiters. In the method for constructing the energy function of the droop inverter including the current limiter disclosed in CN202011155013.8, although the energy function is constructed for the single-machine infinite system of the droop control inverter including the current limiter, the energy function model constructed by the method is very complex, the amount of calculation is large, and the real-time evaluation effect is poor. Therefore, in the field of micro-grids, the existing means can not meet the requirement of online transient stability evaluation of a virtual synchronous machine multi-machine grid-connected system temporarily.
The development of artificial neural networks will provide the possibility of solving the problem. The artificial neural network has strong fitting capability and can effectively solve the nonlinear problem which cannot be solved or is difficult to solve by the traditional method. By means of the advantage, the neural network is widely applied to the fields of image recognition, voice recognition, wind power prediction and the like. The online transient stability evaluation of the virtual synchronous generator multi-machine grid-connected system is a high-dimensional space classification problem for the neural network, complex features can be automatically learned from massive training data by utilizing the nonlinear fitting capability of the artificial neural network, and the mapping relation between system feature variables and stability is established, so that the transient stability classification of the system is realized. However, in the field of micro-grid transient stability evaluation, the application of a neural network is not available at present. Therefore, the difficult problem of online transient stability evaluation of the virtual synchronous machine multi-machine grid-connected system needs to be solved.
Disclosure of Invention
The invention aims to overcome the defects of the background technology and provide a VSG multi-machine system transient stability evaluation method based on an artificial neural network, so that the rapid and accurate evaluation of a multi-machine micro-grid system can be effectively realized.
The technical scheme adopted for solving the technical problem is that the VSG multi-computer system transient stability evaluation method based on the artificial neural network comprises the following steps:
(1) determining input features and output features;
(2) changing VSG multi-machine system load level, changing power grid voltage drop, changing virtual synchronous machine active reference value PrefVirtual inertia J, damping coefficient DpChanging the current limiting amplitude of the virtual synchronous machine, changing the fault duration, generating sample data, and dividing a training sample set and a test sample set;
(3) carrying out data preprocessing on the input characteristics and the output characteristics;
(4) building an artificial neural network;
(5) initializing an artificial neural network;
(6) training an artificial neural network by using a training sample set to complete the establishment of an evaluation model;
(7) carrying out evaluation model test by using the test sample set;
(8) the evaluation model is applied online.
Further, the input characteristics include a system active load level before a fault, a system reactive load level before the fault, a system total active output before the fault, a system total reactive output before the fault, a bus voltage average value before the fault, a maximum relative power angle difference of virtual synchronous machines before the fault, an active output variance of each virtual synchronous machine in the system before the fault, a virtual synchronous machine current amplitude limiting per unit value, a system maximum relative acceleration at the time of the fault, a system minimum relative acceleration at the time of the fault, a virtual synchronous machine acceleration variance at the time of the fault, a system maximum relative angular velocity at the time of the fault, a system maximum relative power angle at the time of the fault, an initial acceleration power average value of each virtual synchronous machine at the time of the fault, a voltage impact at the time of the fault, a maximum active impact, a minimum active impact, a maximum relative acceleration at the time of the fault, a maximum relative angular velocity at the time of the fault, a maximum relative power angle at the time of the fault, a maximum relative acceleration at the time of the fault, a maximum relative angular velocity at the fault, a maximum relative power angle at the fault clearance, a maximum power angle of the fault, a maximum relative acceleration of the fault, a maximum acceleration of the virtual synchronous machine at the fault, a maximum acceleration of the fault, a maximum relative acceleration of the virtual synchronous machine at the fault, a maximum relative acceleration of the fault, a fault clearance, a fault, The virtual synchronous machine outputs an average voltage value during fault clearing, the maximum acceleration power during fault clearing, the average acceleration power during fault clearing, the maximum relative acceleration from the beginning of the fault to the fault clearing, the maximum relative angular velocity from the beginning of the fault to the fault clearing, and the maximum relative power angle from the beginning of the fault to the fault clearing.
Further, in the step (1), the output characteristic is whether the system is unstable, if so, the output is 1, and if so, the output is 0.
Further, in the step (2), three load levels of 50%, 80% and 110%, the grid voltage is reduced to 20% and 50% of two fault conditions, and the active reference value P of the three virtual synchronous machinesref80% of rated value, 1 and 5 of virtual inertia J, and damping coefficient DpCurrent limiting value of the virtual synchronous machine is 1.5 times of rated current value of 10 and 20And 2 times of current rated value and setting the fault duration time to be 0.05s, 0.1s, 0.15s and 0.2s, and performing all permutation and combination to generate sample data.
Further, in the step (2), 90% of sample data is selected as a training sample set, and 10% of sample data is selected as a testing sample set.
Further, in the step (3), the input features and the output features are normalized by adopting a 0-1 normalization method:
Figure BDA0003241025040000041
wherein X is the normalized input characteristic quantity, X is the input characteristic quantity before normalization, and XminIs the minimum value of the input feature, xmaxIs the maximum value of the input feature.
The system further comprises an input layer, a first hidden layer, a second hidden layer and an output layer, wherein the input layer is connected with the first hidden layer, the first hidden layer is connected with the second hidden layer, and the second hidden layer is connected with the output layer; the number of the neurons of the input layer, the first hidden layer, the second hidden layer and the output layer is 26, 100, 50 and 1 respectively.
Further, in the step (5), the learning rate of the neural network is set to be 0.01, and the maximum iteration number is set to be 1000; the activation functions of the first hidden layer and the second hidden layer are both selected from ReLU functions, the activation function of the output layer is selected from Sigmoid functions, the ReLU function is shown as a formula (2), the Sigmoid function is shown as a formula (3),
f1(x)＝max(0,x) (2)
Figure BDA0003241025040000051
where x is the output value of each layer of neurons.
Further, in the step (6), a Levernberg-Marquardt back propagation algorithm is adopted to train the feedforward neural network, and the method specifically comprises the following steps:
(6-1) calculating the forward propagation of the neural network:
Figure BDA0003241025040000052
wherein X is the input feature vector of the input layer, H1Is the output vector of the first hidden layer, H2Is the output vector of the second hidden layer, ω[1]Is the weight of the neuron of the first hidden layer, ω[2]Is the weight, ω, of the neurons of the second hidden layer[Y]Weight of neurons of the output layer, b[1]Bias of neurons of the first hidden layer, b[2]Bias of neurons of the second hidden layer, b[Y]Is the bias of the neurons of the output layer, Y is the output vector of the network, Y ═ Y1,Y2,…,Yi,…,Yn]；
(6-2) calculating the error index of the neural network and the error vector of the output layer:
Figure BDA0003241025040000053
e(o)＝[e1(o),e2(o),…,ei(o),…,en(o)]
wherein E (o) is a neural network error index function, YiIs a calculated value of the ith output characteristic of the neural network, Y'iIs the actual value of the ith output characteristic, ei(o) is the error of the ith output feature of the output layer, n is the number of output features, and e (o) is the error vector of the output layer;
(6-3) reversely correcting the weight and the bias of each layer of neuron according to the error vector of the output layer:
Δok＝[JT(o)J(o)+μI]-1JT(o)e(o) (6)
ok+1＝ok+Δok (7)
wherein J (o) is a Jacobian matrix of vector o partial derivatives, JT(o) is the transpose of J (o), mu is schoolThe learning rate, I is a unit matrix, and o is a vector formed by the weight and the bias; okIs the weight and offset of the kth iteration, ok+1The corrected weight and the offset vector; delta okIs the weight bias increment.
Further, in the step (7), when the model performance is evaluated, if the output result is greater than 0.5, the sample is regarded as a stable sample, and the correction output is 1; the denormalization is regarded as instability, the correction output is 0, and the output result is shown in formula (8):
Figure BDA0003241025040000061
comparing the output value of the neural network with the actual value, and calculating the accuracy rate, wherein the accuracy rate is shown as the formula (9):
Figure BDA0003241025040000062
compared with the prior art, the invention has the following advantages:
(1) the transient stability original input feature set suitable for the virtual synchronous machine multi-machine grid-connected system is constructed for the first time, the feature set can effectively reflect the instability characteristic of the system, can be independent of network topology, cannot increase input features along with the complication of a network, and effectively avoids dimension disasters.
(2) The transient stability evaluation method of the virtual synchronous machine multi-machine grid-connected system based on the artificial neural network is provided, the nonlinear self-learning capability of the artificial neural network is utilized, the complex calculation of the traditional transient stability evaluation is transferred to the offline training, and the problems that the online transient stability evaluation analysis of the multi-machine system is complex, the evaluation is difficult, and the online evaluation speed is difficult to guarantee are solved.
Drawings
Fig. 1 is a typical configuration diagram of a grid-connected system of three virtual synchronous machines in this embodiment.
Fig. 2 shows the internal circuit and control system structure of each virtual synchronous machine in this embodiment.
Detailed Description
The invention is described in further detail below with reference to the figures and the embodiments.
Due to the fact that the order of a virtual synchronous machine multi-machine system model is high, source cross coupling is achieved, and the working state of the system is frequently changed by a current limiter, the transient stability assessment analysis difficulty is large, the assessment difficulty is high, and the existing means cannot meet the online transient stability assessment requirements. Therefore, the invention provides a transient stability evaluation method of a virtual synchronous machine multi-machine grid-connected system based on an artificial neural network, in the implementation, three virtual synchronous machines are adopted, and the transient stability evaluation method comprises the following steps:
the first step is as follows: input features and output features are determined. The input characteristics directly influence the accuracy of the neural network evaluation, and the method is a very important ring in the construction of a neural network model. At present, a machine learning method of a virtual synchronous inverter multi-machine grid-connected system is not considered temporarily, so that an original characteristic structure in the aspect is not considered temporarily. In the embodiment, a group of input feature sets which can effectively reflect the instability characteristics of the virtual synchronous machine multi-machine system and can avoid dimension disasters is constructed for the first time by combining a traditional power system transient stability evaluation input feature construction method, micro-grid evaluation requirements, virtual synchronous machine transient stability characteristics and current limiter influences. The feature set contains 26 input features in total, which are shown in table 1 and specifically includes active load level of a system before fault, reactive load level of the system before fault, total active output of the system before fault, total reactive output of the system before fault, average voltage of bus before fault, maximum relative power angle difference of virtual synchronous machines before fault, active output variance of each virtual synchronous machine before fault, current amplitude limiting per unit value of the virtual synchronous machine, maximum relative acceleration of the system during fault, minimum relative acceleration of the system during fault, acceleration variance of each virtual synchronous machine during fault, maximum relative angular velocity of the system during fault, maximum relative power angle of the system during fault, average initial acceleration power of each virtual synchronous machine during fault, voltage impact during fault, maximum active impact, minimum active impact, maximum relative acceleration during fault clearance, maximum relative angular velocity during fault clearance, The maximum relative power angle during fault clearing, the average value of output voltage of the virtual synchronous machine during fault clearing, the maximum acceleration power during fault clearing, the average acceleration power during fault clearing, the maximum relative acceleration from the beginning of fault to the fault clearing, the maximum relative angular velocity from the beginning of fault to the fault clearing, and the maximum relative power angle from the beginning of fault to the fault clearing can be independent of network topology, and the input characteristic quantity cannot be increased along with the complication of a network. The transient stability evaluation of the virtual synchronous machine multi-machine grid-connected system can be regarded as the classification problem of the strong nonlinear system in a high-dimensional space, the output characteristic is whether the system is unstable or not, if the system is stable, the output is 1, and if the system is unstable, the output is 0.
TABLE 1 original feature set for transient stability evaluation of virtual synchronous machine multi-machine grid-connected system
Figure BDA0003241025040000081
Figure BDA0003241025040000091
The second step is that: and acquiring sample data, and dividing a training sample set and a test sample set. Through simulation or experiment, system fault parameters and system control parameters are changed, the current limiting value of the virtual synchronous machine is changed, the fault duration is changed, and sample data are generated. In the embodiment, the load fluctuation and the fault randomness are considered, three load levels of 50%, 80% and 110% are recorded, and two fault conditions of 20% and 50% of the grid voltage are recorded; recording active reference values P of three virtual synchronous machines in consideration of output change of the virtual synchronous machinesrefVirtual inertia J, damping coefficient DpParameter variation; considering that the current amplitude limiting values of all systems are different, and recording two conditions that the current amplitude limiting value of the virtual synchronous machine is 1.5 times of the rated current value and 2 times of the rated current value; the failure duration is recorded in four conditions of 0.05s, 0.1s, 0.15s and 0.2 s. Namely three load levels of 50%, 80% and 110%, the grid voltage is reduced to 20% and 50% of two fault conditions, and the active reference value P of three virtual synchronous machinesrefIs a forehead80% of constant value, 1 and 5 of virtual inertia J, and damping coefficient DpThe method comprises the following steps of 10 and 20, setting the current limiting value of the virtual synchronous machine to be two conditions of 1.5 times of current rated value and 2 times of current rated value, and setting the fault duration to be 0.05s, 0.1s, 0.15s and 0.2s, and performing all permutation and combination to generate sample data.
Training data and test data were randomly divided. In the implementation, 90% of sample data is selected as a training sample set, and 10% of sample data is selected as a testing sample set. And aiming at the training sample set, judging whether the training sample is unstable or not on the basis of whether the power angle between any two virtual synchronous generators or the virtual synchronous generators and the power grid at the end of simulation is larger than 360 degrees or not. Since the instability problem can be regarded as a classification problem, whether it is unstable or not can be characterized by 0 or 1. If the power angle is greater than 360 degrees, the instability is determined, and the sample is marked as 0. Otherwise, the value is judged to be stable and marked as 1.
The third step: and (4) preprocessing data. And carrying out normalization processing on the input characteristics and the output characteristics, eliminating the difference between different dimensions and accelerating the convergence speed of the model. Common normalization methods are 0-1 normalization and Z-score normalization. Both methods can be well applied to neural network data preprocessing, and the embodiment adopts a 0-1 standardization method. The specific formula is shown in (1), wherein X is the input characteristic quantity after normalization, X is the input characteristic quantity before normalization, and XminIs the minimum value of the input feature, xmaxIs the maximum value of the input feature. Since the output characteristic itself is in the range of 0 to 1, no normalization process is required.
Figure BDA0003241025040000101
The fourth step: and building an artificial neural network. Since the feedforward neural network using the back propagation algorithm has very good nonlinear approximation performance, the feedforward neural network is selected for transient stability evaluation in the embodiment. In order to ensure the speed and accuracy of online evaluation of the neural network, the embodiment adopts a feedforward neural network with two hidden layers. The system specifically comprises an input layer, a first hidden layer, a second hidden layer and an output layer, wherein the input layer is connected with the first hidden layer, the first hidden layer is connected with the second hidden layer, and the second hidden layer is connected with the output layer. Through multiple tests, the effect is better when the neuron number edge division of the first hidden layer and the second hidden layer is set to be 100 and 50. The number of neurons in each layer (input layer-first hidden layer-second hidden layer-output layer) of the whole artificial neural network is 26, 100, 50, 1, respectively.
The fifth step: an artificial neural network is initialized. The neural network learning rate is set to be 0.01, and the maximum iteration number is set to be 1000. The active function of each hidden layer selects a ReLU function, and in order to prevent overfitting, the active function of the output layer selects a Sigmoid function to ensure that the output of the artificial neural network is in a range of 0-1. The ReLU function is shown in equation (2) and the Sigmoid function is shown in equation (3). Where x is the output value of each layer of neurons.
f1(x)＝max(0,x) (2)
Figure BDA0003241025040000111
And a sixth step: and training the artificial neural network by using the training sample set to complete the establishment of the evaluation model. In order to achieve better nonlinear fitting capability, i.e., accurately establish the nonlinear mapping relationship between the input features and the output features, the present embodiment trains the feedforward neural network by using a levenberg-Marquardt back propagation algorithm. The feedforward neural network has the characteristic of forward propagation, and the core of backward propagation is as follows: calculating forward propagation, calculating a neural network error index and an output layer error vector according to the forward propagation, and reversely correcting the weight and the bias of each layer of neuron according to the output layer error vector, thereby finally ensuring that the neural network output value of each sample in a training set reaches an expectation. The specific calculation formula of the neural network forward propagation is shown in formula (4), and the error calculation method of the output value and the actual value is shown in formula (5) and the correction method of the weight value and the bias is shown in formulas (6) and (7).
Figure BDA0003241025040000112
Figure BDA0003241025040000113
e(o)＝[e1(o),e2(o),…,ei(o),…,en(o)]
Δok＝[JT(o)J(o)+μI]-1JT(o)e(o) (6)
ok+1＝ok+Δo (7)
Wherein X is the input feature vector of the input layer, H1Is the output vector of the first hidden layer, H2Is the output vector of the second hidden layer, ω[1]Is the weight of the neuron of the first hidden layer, ω[2]Is the weight, ω, of the neurons of the second hidden layer[Y]Weight of neurons of the output layer, b[1]Bias of neurons of the first hidden layer, b[2]Bias of neurons of the second hidden layer, b[Y]Is the bias of the neurons of the output layer, Y is the output vector of the network, Y ═ Y1,Y2,…,Yi,…,Yn](ii) a E (o) is a neural network error indicator function, YiIs a calculated value of the ith output characteristic of the neural network, Y'iIs the actual value of the ith output characteristic, ei(o) is the error of the ith output characteristic of the output layer, and n is the number of the output characteristics; j (o) a Jacobian matrix of partial derivatives of the vector o, JT(o) is the transpose matrix of J (o), mu is the learning rate, I is the unit matrix, e (o) is the output layer error vector, o is the vector composed of weight and bias; o. okIs the weight and offset of the kth iteration, ok+1The corrected weight and the offset vector; delta okIs the weight bias increment.
The seventh step: and carrying out evaluation model test by using the test sample set. The Sigmoid function only maps the output result in the range of 0-1 to give the instability and instability probability, so that the output result needs to be subjected to interval division during online application. If the output result is greater than 0.5, the sample is regarded as a stable sample, and the correction output is 1; the denormalization is regarded as instability, the correction output is 0, and the output result is shown in formula (8). And comparing the output value of the neural network with the actual value, and calculating the accuracy rate, wherein the accuracy rate is shown as a formula (9).
Figure BDA0003241025040000121
Figure BDA0003241025040000122
Eighth step: the evaluation model is applied online.
The invention has the following advantages:
1. the transient stability original input feature set suitable for the virtual synchronous machine multi-machine grid-connected system is constructed for the first time, the feature set can effectively reflect the instability characteristic of the system, can be independent of network topology, cannot increase input features along with the complication of a network, and effectively avoids dimension disasters.
2. The method for evaluating the transient stability of the virtual synchronous machine multi-machine grid-connected system based on the artificial neural network is provided, the nonlinear self-learning capability of the artificial neural network is utilized, the complex calculation of the traditional transient stability evaluation is transferred to offline training, and the problems that the online transient stability evaluation analysis of the multi-machine system is complex, the evaluation is difficult, and the online evaluation speed is difficult to guarantee are solved.
In order to evaluate the effect of the online transient stability evaluation method provided by the invention, example verification needs to be carried out on a multi-machine grid-connected system to obtain the actual evaluation effect of the provided model. And table 2 shows the evaluation accuracy, the off-line training time consumption and the on-line evaluation time consumption of the proposed model in a three-virtual synchronous machine grid-connected system. It can be seen that the method can rapidly evaluate the stability of a multi-machine system, the online evaluation speed reaches 0.18ms, and the method has good evaluation performance, and the evaluation accuracy is 97.2%.
TABLE 2 evaluation of the Effect
Table 2 Evaluation effect of different online TSA methods
Figure BDA0003241025040000131
Model verification is performed by taking a typical three-virtual-synchronous-machine grid-connected system as an example, as shown in fig. 1. Each virtual synchronous machine is composed of an inverter, a filter and a control system, as shown in fig. 2. Exemplary system parameters are shown in table 3. Considering the fluctuation of the load and the randomness of the fault degree, three load levels of 50%, 80% and 110% are recorded, and two fault conditions of 20% and 40% of the grid voltage are recorded. Considering three virtual synchronous machine control parameter active reference values PrefVirtual inertia J, damping coefficient DpChanging, recording and displaying active reference value P of three virtual synchronous machinesref80% of rated value, 1 and 5 of virtual inertia J, and damping coefficient Dp10 and 20. The influence of the current amplitude limiting value on the stability of the power angle is considered, and two conditions that the current amplitude limiting value of the virtual synchronous machine is 1.5 times of the rated value of the current and 2 times of the rated value of the current are recorded. The failure duration is set to 0.05s, 0.1s, 0.15s, 0.2 s. The simulation software was MATLAB/Simulink, and 3072 valid samples were generated. Through off-line training and on-line testing, model prediction accuracy, off-line training time and on-line evaluation time are obtained, and specific results are shown in table 2.
TABLE 3 System parameters
Table 3 System Parameter
Figure BDA0003241025040000141
Various modifications and variations of the present invention may be made by those skilled in the art, and they are also within the scope of the present invention provided they are within the scope of the claims of the present invention and their equivalents.
What is not described in detail in the specification is prior art that is well known to those skilled in the art.

Claims (9)

1. A VSG multi-machine system transient stability evaluation method based on an artificial neural network is characterized by comprising the following steps: the method comprises the following steps:
(1) determining input features and output features; the input characteristics comprise the active load level of a system before a fault, the reactive load level of the system before the fault, the total active output of the system before the fault, the total reactive output of the system before the fault, the voltage average value of a bus before the fault, the maximum relative power angle difference of a virtual synchronous machine before the fault, the active output variance of each virtual synchronous machine in the system before the fault, the current amplitude limiting per unit value of the virtual synchronous machine, the maximum relative acceleration of the system during the fault, the minimum relative acceleration of the system during the fault, the acceleration variance of each virtual synchronous machine during the fault, the maximum relative angular velocity of the system during the fault, the maximum relative power angle of the system during the fault, the initial acceleration power average value of each virtual synchronous machine during the fault, the voltage impact during the fault, the maximum active impact, the minimum active impact, the maximum relative acceleration during the fault clearance, the maximum relative angular velocity during the fault clearance, the maximum relative power angle, The virtual synchronous machine outputs an average voltage value during fault clearing, the maximum acceleration power during fault clearing, the average acceleration power during fault clearing, the maximum relative acceleration from the beginning of the fault to the fault clearing, the maximum relative angular velocity from the beginning of the fault to the fault clearing, and the maximum relative power angle from the beginning of the fault to the fault clearing;
(2) changing VSG multi-machine system load level, changing grid voltage drop, changing virtual synchronous machine active reference value PrefVirtual inertia J, damping coefficient DpChanging the current limiting amplitude of the virtual synchronous machine, changing the fault duration, generating sample data, and dividing a training sample set and a test sample set;
(3) carrying out data preprocessing on the input characteristics and the output characteristics;
(4) building an artificial neural network;
(5) initializing an artificial neural network;
(6) training an artificial neural network by using a training sample set to complete the establishment of an evaluation model;
(7) carrying out evaluation model test by using the test sample set;
(8) the evaluation model is applied online.
2. The method for transient stability evaluation of a VSG multi-machine system based on an artificial neural network as claimed in claim 1, wherein: in the step (1), the output characteristic is whether the system is unstable, if so, the output is 1, and if not, the output is 0.
3. The method for evaluating transient stability of VSG multi-machine system based on artificial neural network as claimed in claim 1 or 2, wherein: in the step (2), three load levels of 50%, 80% and 110%, the grid voltage is reduced to 20% and 50% of two fault conditions, and the active reference value P of three virtual synchronous machinesref80% of rated value, 1 and 5 of virtual inertia J, and damping coefficient DpThe method comprises the following steps of 10 and 20, setting the current limiting value of the virtual synchronous machine to be two conditions of 1.5 times of current rated value and 2 times of current rated value, and setting the fault duration to be 0.05s, 0.1s, 0.15s and 0.2s, and performing all permutation and combination to generate sample data.
4. The VSG multi-machine system transient stability evaluation method based on the artificial neural network as claimed in claim 1 or 2, wherein: in the step (2), 90% of sample data is selected as a training sample set, and 10% of sample data is selected as a testing sample set.
5. The method for evaluating transient stability of VSG multi-machine system based on artificial neural network as claimed in claim 1 or 2, wherein: in the step (3), the input characteristic and the output characteristic are normalized by adopting a 0-1 standardization method:
Figure FDA0003607256770000021
wherein X is the normalized input characteristic quantity, X is the input characteristic quantity before normalization, and XminIs the minimum value of the input feature, xmaxIs the maximum value of the input feature.
6. The method for evaluating transient stability of VSG multi-machine system based on artificial neural network as claimed in claim 1 or 2, wherein: in the step (4), the artificial neural network is a feedforward neural network, the feedforward neural network comprises an input layer, a first hidden layer, a second hidden layer and an output layer, the input layer is connected with the first hidden layer, the first hidden layer is connected with the second hidden layer, and the second hidden layer is connected with the output layer; the number of the neurons of the input layer, the first hidden layer, the second hidden layer and the output layer is 26, 100, 50 and 1 respectively.
7. The method for evaluating transient stability of a VSG multi-machine system based on an artificial neural network as claimed in claim 6, wherein: in the step (5), setting the neural network learning rate to be 0.01 and the maximum iteration number to be 1000; the activation functions of the first hidden layer and the second hidden layer are both selected from ReLU functions, the activation function of the output layer is selected from Sigmoid functions, the ReLU function is shown as a formula (2), the Sigmoid function is shown as a formula (3),
f1(x)＝max(0,x) (2)
Figure FDA0003607256770000031
where x is the output value of each layer of neurons.
8. The method for transient stability evaluation of a VSG multi-machine system based on an artificial neural network as claimed in claim 6, wherein: in the step (6), a Levernberg-Marquardt back propagation algorithm is adopted to train a feedforward neural network, and the method specifically comprises the following steps:
(6-1) calculating the forward propagation of the neural network:
Figure FDA0003607256770000032
wherein X is the input character of the input layerEigenvector, H1Is the output vector of the first hidden layer, H2Is the output vector of the second hidden layer, ω[1]Is the weight of the neuron of the first hidden layer, ω[2]Is the weight, ω, of the neurons of the second hidden layer[Y]Weight of neurons of the output layer, b[1]Bias of neurons of the first hidden layer, b[2]Bias of neurons of the second hidden layer, b[Y]Is the bias of the neurons of the output layer, Y is the output vector of the network, Y ═ Y1,Y2,L,Yi,L,Yn]；
(6-2) calculating the error index of the neural network and the error vector of the output layer:
Figure FDA0003607256770000041
e(o)＝[e1(o),e2(o),L,ei(o),L,en(o)]
wherein E (o) is a neural network error index function, YiCalculated for the ith output characteristic of the neural network, Yi' is the actual value of the ith output characteristic, ei(o) is the error of the ith output feature of the output layer, n is the number of output features, and e (o) is the error vector of the output layer;
(6-3) reversely correcting the weight and the bias of each layer of neuron according to the error vector of the output layer:
Δok＝[JT(o)J(o)+μI]-1JT(o)e(o) (6)
ok+1＝ok+Δok (7)
wherein J (o) is a Jacobian matrix of vector o partial derivatives, JT(o) is a transposed matrix of J (o), mu is a learning rate, I is a unit matrix, and o is a vector composed of a weight and an offset; okIs the weight and offset of the kth iteration, ok+1The corrected weight and the offset vector; delta okIs the weight bias increment.
9. The method for evaluating transient stability of VSG multi-machine system based on artificial neural network as claimed in claim 1 or 2, wherein: in the step (7), when the model performance is evaluated, if the output result is greater than 0.5, the sample is regarded as a stable sample, and the correction output is 1; the denormalization is regarded as instability, the correction output is 0, and the output result is shown in formula (8):
Figure FDA0003607256770000051
comparing the output value of the neural network with the actual value, and calculating the accuracy rate, wherein the accuracy rate is shown as the formula (9):
Figure FDA0003607256770000052
CN202111018894.3A 2021-09-01 2021-09-01 VSG multi-machine system transient stability evaluation method based on artificial neural network Active CN113792481B (en)

Priority Applications (1)

Application Number Priority Date Filing Date Title
CN202111018894.3A CN113792481B (en) 2021-09-01 2021-09-01 VSG multi-machine system transient stability evaluation method based on artificial neural network

Applications Claiming Priority (1)

Application Number Priority Date Filing Date Title
CN202111018894.3A CN113792481B (en) 2021-09-01 2021-09-01 VSG multi-machine system transient stability evaluation method based on artificial neural network

Publications (2)

Publication Number Publication Date
CN113792481A CN113792481A (en) 2021-12-14
CN113792481B true CN113792481B (en) 2022-05-31

Family

ID=78879355

Family Applications (1)

Application Number Title Priority Date Filing Date
CN202111018894.3A Active CN113792481B (en) 2021-09-01 2021-09-01 VSG multi-machine system transient stability evaluation method based on artificial neural network

Country Status (1)

Country Link
CN (1) CN113792481B (en)

Citations (2)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN110311375A (en) * 2019-07-29 2019-10-08 东北大学 A kind of micro-capacitance sensor transient stability control method containing more virtual synchronous machines
CN111416393A (en) * 2020-04-15 2020-07-14 上海电力大学 Virtual synchronous generator fault current suppression method based on self-adaptive virtual impedance

Family Cites Families (1)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CA2883059A1 (en) * 2007-11-07 2009-11-12 Power Analytics Corporation Systems and methods for real-time forecasting and predicting of electrical peaks and managing the energy, health, reliability, and performance of electrical power systems based onan artificial adaptive neural network

Patent Citations (2)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN110311375A (en) * 2019-07-29 2019-10-08 东北大学 A kind of micro-capacitance sensor transient stability control method containing more virtual synchronous machines
CN111416393A (en) * 2020-04-15 2020-07-14 上海电力大学 Virtual synchronous generator fault current suppression method based on self-adaptive virtual impedance

Non-Patent Citations (3)

* Cited by examiner, † Cited by third party
Title
Transient Stability Prediction for Load Flow Cascading Models Using Random Forests;Jacques Delport等;《2018 IEEE/PES输配电会议暨博览会（T&D）》;20180823;第1-5页 *
基于人工神经网络的暂态稳定性分析;戴仁昶等;《电力系统自动化》;20000625;第1-3页 *
基于人工神经网络的暂态稳定评估技术的研究;张文朝;《中国优秀硕士论文电子期刊全文数据库（电子期刊）》;20030115;第一-四章 *

Also Published As

Publication number Publication date
CN113792481A (en) 2021-12-14

Similar Documents

Publication Publication Date Title
El-Keib et al. Application of artificial neural networks in voltage stability assessment
Li et al. Data-driven optimal control strategy for virtual synchronous generator via deep reinforcement learning approach
Liu et al. Neuro-fuzzy networks for voltage security monitoring based on synchronized phasor measurements
CN112821424B (en) Power system frequency response analysis method based on data-model fusion driving
CN113591379B (en) Auxiliary decision-making method for transient stability prevention and emergency coordination control of power system
CN105403834A (en) Dynamic state evaluation method of generator
CN112800683B (en) System short-circuit current level evaluation method and system based on convolutional neural network
CN109066651B (en) Method for calculating limit transmission power of wind power-load scene
CN114036506A (en) Method for detecting and defending false data injection attack based on LM-BP neural network
CN105353270A (en) Power quality disturbance source fault-tolerant positioning method considering distributed generator connection
CN113792481B (en) VSG multi-machine system transient stability evaluation method based on artificial neural network
CN111125880B (en) Method for generating simulation data of power system under transient stable view angle
La Scala et al. A qualitative approach to the transient stability analysis [of power systems]
CN108054768A (en) Transient stability evaluation in power system method based on principal component analysis
Lai et al. Application of evolutionary programming to transient and subtransient parameter estimation
CN111224427A (en) Micro-grid self-adaptive anti-islanding disturbance load impedance value prediction method and device
CN116227320A (en) Double-fed fan control parameter identification method based on LSTM-IPSO
CN113435575B (en) Gate graph neural network transient stability evaluation method based on unbalanced data
Shi et al. Load clustering and synthetic modeling based on an improved fuzzy C means clustering algorithm
CN115048857A (en) CNN-based transient stability limit transmission power determination method
Zhang et al. Derivation of load model parameters using improved genetic algorithm
Meléndez-Pérez et al. Design of an ANFIS Automatic Voltage Regulator of a Synchronous Generator
Kataoka A smooth power flow model of electric power system with generator reactive power limits taken into consideration
Chen et al. Control Parameter Identification for DFIG based on RFE-BiLSTM
Li et al. A Parameter Optimization Method Based on Eigenvalue Analysis for Microgrid With Multiple VSG Units

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
GR01 Patent grant
GR01 Patent grant