CN111160386B - Data-driven power island detection method based on active reactive power disturbance - Google Patents
Data-driven power island detection method based on active reactive power disturbance Download PDFInfo
- Publication number
- CN111160386B CN111160386B CN201911189108.9A CN201911189108A CN111160386B CN 111160386 B CN111160386 B CN 111160386B CN 201911189108 A CN201911189108 A CN 201911189108A CN 111160386 B CN111160386 B CN 111160386B
- Authority
- CN
- China
- Prior art keywords
- island
- reactive power
- neural network
- back propagation
- self
- 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
Links
- 238000001514 detection method Methods 0.000 title claims abstract description 39
- 238000013528 artificial neural network Methods 0.000 claims abstract description 62
- 238000012549 training Methods 0.000 claims abstract description 30
- 238000012706 support-vector machine Methods 0.000 claims abstract description 28
- 230000004927 fusion Effects 0.000 claims abstract description 26
- 238000004422 calculation algorithm Methods 0.000 claims abstract description 15
- 230000001537 neural effect Effects 0.000 claims abstract description 6
- 239000002245 particle Substances 0.000 claims description 36
- 230000006870 function Effects 0.000 claims description 27
- 238000000034 method Methods 0.000 claims description 22
- 238000010248 power generation Methods 0.000 claims description 17
- 230000003044 adaptive effect Effects 0.000 claims description 8
- 238000005457 optimization Methods 0.000 claims description 5
- 238000012360 testing method Methods 0.000 claims description 5
- 238000004364 calculation method Methods 0.000 claims description 3
- 238000013507 mapping Methods 0.000 claims description 2
- 230000000694 effects Effects 0.000 abstract description 16
- 230000007246 mechanism Effects 0.000 abstract description 5
- 230000008859 change Effects 0.000 description 4
- 238000004891 communication Methods 0.000 description 3
- 238000012423 maintenance Methods 0.000 description 2
- 230000009286 beneficial effect Effects 0.000 description 1
- 238000007621 cluster analysis Methods 0.000 description 1
- 238000010586 diagram Methods 0.000 description 1
- 230000005684 electric field Effects 0.000 description 1
- 238000007499 fusion processing Methods 0.000 description 1
- 230000002093 peripheral effect Effects 0.000 description 1
- 230000008569 process Effects 0.000 description 1
- 238000011160 research Methods 0.000 description 1
- 230000004044 response Effects 0.000 description 1
Classifications
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06F—ELECTRIC DIGITAL DATA PROCESSING
- G06F18/00—Pattern recognition
- G06F18/20—Analysing
- G06F18/23—Clustering techniques
- G06F18/232—Non-hierarchical techniques
- G06F18/2321—Non-hierarchical techniques using statistics or function optimisation, e.g. modelling of probability density functions
- G06F18/23213—Non-hierarchical techniques using statistics or function optimisation, e.g. modelling of probability density functions with fixed number of clusters, e.g. K-means clustering
-
- G—PHYSICS
- G01—MEASURING; TESTING
- G01R—MEASURING ELECTRIC VARIABLES; MEASURING MAGNETIC VARIABLES
- G01R21/00—Arrangements for measuring electric power or power factor
- G01R21/001—Measuring real or reactive component; Measuring apparent energy
- G01R21/003—Measuring reactive component
-
- G—PHYSICS
- G01—MEASURING; TESTING
- G01R—MEASURING ELECTRIC VARIABLES; MEASURING MAGNETIC VARIABLES
- G01R31/00—Arrangements for testing electric properties; Arrangements for locating electric faults; Arrangements for electrical testing characterised by what is being tested not provided for elsewhere
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06F—ELECTRIC DIGITAL DATA PROCESSING
- G06F18/00—Pattern recognition
- G06F18/20—Analysing
- G06F18/24—Classification techniques
- G06F18/241—Classification techniques relating to the classification model, e.g. parametric or non-parametric approaches
- G06F18/2411—Classification techniques relating to the classification model, e.g. parametric or non-parametric approaches based on the proximity to a decision surface, e.g. support vector machines
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06N—COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
- G06N3/00—Computing arrangements based on biological models
- G06N3/004—Artificial life, i.e. computing arrangements simulating life
- G06N3/006—Artificial life, i.e. computing arrangements simulating life based on simulated virtual individual or collective life forms, e.g. social simulations or particle swarm optimisation [PSO]
-
- 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
- G06N—COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
- G06N3/00—Computing arrangements based on biological models
- G06N3/02—Neural networks
- G06N3/08—Learning methods
- G06N3/084—Backpropagation, e.g. using gradient descent
Landscapes
- Engineering & Computer Science (AREA)
- Physics & Mathematics (AREA)
- Theoretical Computer Science (AREA)
- Data Mining & Analysis (AREA)
- General Physics & Mathematics (AREA)
- Evolutionary Computation (AREA)
- Life Sciences & Earth Sciences (AREA)
- Artificial Intelligence (AREA)
- General Engineering & Computer Science (AREA)
- General Health & Medical Sciences (AREA)
- Molecular Biology (AREA)
- Software Systems (AREA)
- Mathematical Physics (AREA)
- Computing Systems (AREA)
- Health & Medical Sciences (AREA)
- Biomedical Technology (AREA)
- Biophysics (AREA)
- Computational Linguistics (AREA)
- Evolutionary Biology (AREA)
- Bioinformatics & Cheminformatics (AREA)
- Bioinformatics & Computational Biology (AREA)
- Computer Vision & Pattern Recognition (AREA)
- Probability & Statistics with Applications (AREA)
- Power Engineering (AREA)
- Supply And Distribution Of Alternating Current (AREA)
Abstract
The invention discloses a data driving power island detection method based on active reactive power disturbance, which comprises the following steps: firstly, reactive disturbance is carried out on a single-phase grid-connected inverter controlled by power by utilizing pre-designed intermittent reactive power, and a corresponding data set under the condition that a pseudo island effect exists is obtained; then, carrying out mechanism analysis on the single-phase inverter to obtain a characteristic data set; combining a k-means algorithm with logic operation to obtain fusion data, training the self-adaptive back propagation neural network by using the characteristic data set and the fusion data, taking the output of the self-adaptive back propagation neural network as a support vector machine to obtain a training set, training the support vector machine, and finally obtaining the self-adaptive back propagation neural network-support vector machine (ABPS) classifier. The invention solves the problem of high detection cost of the traditional island detection method, and does not need to apply extra large disturbance and large undetectable area; provides technical references for accurate island detection, subsequent island effect utilization and the like.
Description
Technical Field
The invention belongs to the field of power island effect detection, and particularly relates to a data driving power island detection method based on active reactive power disturbance.
Background
Renewable energy sources such as photovoltaic, wind energy, geothermal energy and wave energy are increasingly widely applied to distributed power generation, and electric energy is provided for daily life of people. In addition, distributed power generation systems capable of powering local loads and feeding back electrical energy to a distribution grid are increasingly important in the electrical field. However, distributed power generation systems still face many challenges in practical applications, such as anti-islanding protection issues. Island refers to the situation where the grid trips due to a fault or maintenance, and the distributed generation system cannot timely detect a power fault. Thus, islanding will result in an uncontrollable self-powered system consisting of a distributed power generation system and a local load. Islanding can jeopardize maintenance personnel's safety, life of electrical systems and electrical facilities, etc., and therefore, research into solutions for islanding protection, reducing islanding impact is of great importance.
Island protection methods in the prior art can be divided into three main categories, namely passive, communication-based and active methods. Passive island detection methods typically rely on measuring key parameters such as voltage, frequency, phase, etc. to determine whether an island event has occurred. If the parameter exceeds a predefined threshold, over/under voltage protection, over/under frequency protection (OFP/UFP) and phase jump protection are used to detect islanding. The voltage threshold is typically set to 85% -110% of normal, while the frequency threshold is set to 49.5Hz 50.5Hz of a 50Hz system. There are other passive methods such as passive methods based on power factor variation and voltage harmonic distortion. The passive detection method has the advantages of no harm to the quality of electric energy and easy realization. However, a larger non-detection region is formed. NDZ is defined as a series of load parameters for which anti-islanding methods cannot detect islanding. Communication-based detection methods also do not compromise power quality, but are superior to passive detection methods because the former do not have NDZ. However, since peripheral devices are required to detect the island state, the cost of the communication-based method is considerable, and thus the application is not wide.
Disclosure of Invention
The invention aims to: the invention provides a data driving power island detection method based on active reactive power disturbance, which aims to solve the problems of high cost, large detection error and the like in the prior art.
The technical scheme is as follows: the invention provides a data driving power island detection method based on active reactive power disturbance, which is used for judging whether an island exists in a distributed power generation system, and specifically comprises the following steps:
step 1: calculating to obtain reactive power injected into the inverter according to the common point frequency output by the phase-locked loop end in the distributed power generation system; taking the common point frequency and the reactive power injected into the inverter as a group of characteristic data sets; collecting the common point frequency and the corresponding reactive power to obtain n groups of characteristic data sets;
step 2: sequentially carrying out data fusion on the n groups of characteristic data sets to obtain n groups of fusion data;
step 3: based on n groups of fusion data, calculating the weight and the threshold of the self-adaptive back propagation neural network by using a particle swarm algorithm; meanwhile, in each particle swarm iterative computation, taking the n groups of characteristic data sets in the step 1 as training sets of the self-adaptive back propagation neural network to obtain a trained self-adaptive back propagation neural network;
step 4: taking the output of the trained self-adaptive back propagation neural network as the input of a support vector machine, and training the support vector machine to obtain a decision function; finally, building a training self-adaptive back propagation neural network-support vector machine classifier;
step 5: the common point frequency fluctuation mode and the PLC load of the inverter are changed to obtain a group of new characteristic data sets, and the group of new characteristic data sets are classified by utilizing the self-adaptive back propagation neural network-support vector machine classifier, so that whether island exists in the power system or not is judged, specifically, the method comprises the following steps: and taking the new test data as a training set of the trained self-adaptive back propagation neural network to obtain the output of the self-adaptive back propagation neural network, and calculating whether the output is island data by utilizing a decision function so as to judge whether the power generation system is in island.
Further, the specific method for calculating the reactive power injected into the inverter in the step 1 is as follows:
wherein f' represents the common point frequency of the phase-locked loop end output, f r For the resonant frequency of the inverter RLC load, Q f Is the quality factor of the inverter, Q DG And P DG Representing reactive power and active power injected into the inverter, respectively.
Further, the specific method for sequentially performing data fusion on the n sets of feature data sets in the step 2 is as follows:
carrying out k-means calculation on the common point frequency in the n groups of characteristic data sets and the reactive power injected into the inverter to obtain a classification result of each common point frequency and the reactive power injected into the inverter; and performing numerical distribution on each classification result;
data fusion is performed on the group D feature data set by using the following formula, d=1, 2, … n;
L＝Cluster R ∧Cluster f
wherein Cluster R A value after numerical distribution is carried out on the classification result of the reactive power; cluster f A value obtained after numerical value distribution is carried out on the classification result of the common point frequency; l= {0,1}, when 0, it indicates that no islanding of the power system under the set of feature data sets occurs; when 1, the island of the power system under the characteristic data set is shown.
Further, the k-means algorithm is utilized to classify the reactive power data set injected into the inverter, the classification result is two types, and the first type is: the reactive power injected into the inverter is not in the intermittent period; the second category is: the injected reactive power is in an intermittent period; according to the value of the common point frequency when island occurs, classifying the common point frequency data set by using a k-means algorithm, wherein the classification result is three types, and the first type is: common point frequency when island does not occur; the second category is: island occurs and the reactive power injected into the inverter is at the common point frequency of the intermittent period; the third category is: island occurs and the reactive power injected into the inverter is not at the common point frequency of the intermittent period; the values of L after numerical assignment based on the above classification result are shown in the following table 1:
TABLE 1
Wherein the value of the reactive power after being allocated by the first class of the numerical value is 0, and the value of the reactive power after being allocated by the second class of the numerical value is 1; the first class of the common point frequency is assigned a value of 0, the second class is assigned a value of 1, and the third class is assigned a value of 2.
Further, the specific steps of the step 3 are as follows:
step 3.1: initializing parameters in a self-adaptive back propagation neural network and a particle swarm algorithm, wherein the scale of the particle swarm is M;
step 3.2, performing t iterative computation, and taking initial particles as weights and thresholds of the self-adaptive back propagation neural networks in sequence, wherein M self-adaptive back propagation neural networks exist;
step 3.3: sequentially inputting n groups of characteristic data sets serving as training sets into an mth self-adaptive back propagation neural network, and performing epochs training on the n groups of characteristic data sets by the neural network to obtain n outputs;
step 3.4: the nth output and the nth fusion data are subjected to mean square error, and n=1, 2, … and N; obtaining n groups of mean square deviations, and taking the sum of the n groups of mean square deviations as a group of fitness values;
step 3.5, judging whether M is larger than M, if so, turning to step 3.6; if not, m+1, turning to step 3.3;
step 3.6: comparing M groups of fitness values, and judging whether the fitness value smaller than Y is in the M groups of fitness values, wherein Y is a preset value; if yes, stopping the particle swarm algorithm, selecting a group of particles corresponding to the smallest fitness value in the fitness value group smaller than Y as a global optimal solution, and taking the global optimal solution as the weight and the threshold of the self-adaptive back propagation neural network to obtain the self-adaptive back propagation neural network with the determined weight and threshold, thus obtaining the trained self-adaptive back propagation neural network; if not, judging whether t reaches the maximum iteration number; if not, taking a group of particles corresponding to the smallest fitness value in the iterative computation as an individual optimal solution of the iterative computation, comparing the individual optimal solution with a historical global optimal solution, selecting a group of particles corresponding to the smallest fitness value as a global optimal solution of the iterative computation, and converting into a step 3.7; if yes, stopping iterative computation, and selecting a global optimal solution as the weight and the threshold of the self-adaptive back propagation neural network to obtain the self-adaptive back propagation neural network with the determined weight and threshold;
and 3.7, updating the speed and the position of each particle according to the global optimal solution and the individual optimal solution calculated in the iteration, taking the updated particles as initial particles, t+1, and turning to the step 3.2.
Further, the specific method in the step 4 is as follows:
output x of trained adaptive back propagation neural network i (i=1, 2, … n) as training data for the support vector machine and creates a hyperplane describing the decision function as:
y(x i )＝ω T φ(x i )+b
where ω represents the normal vector to the hyperplane, b represents the deviation,for mapping functions for solving x i Is an inner product of (2);
if the decision function satisfies the conditionThen x i The corresponding group of characteristic data is non-island data, which indicates that the power generation system under the data has no island; if the decision function satisfies the condition->Then x i The group of corresponding characteristic data is island data, which indicates that the power generation system under the data has island;
based on y i E { -1,1} and each x i In association, initializing regularization parameters C, and representing the hyperplane as a quadratic programming optimization problem:
wherein xi i 0 is ≡0 is ≡x i A corresponding relaxation variable;
and (3) taking the radial basis function as a kernel function of training data, solving the quadratic programming optimization problem by utilizing a Lagrange function, and rewriting a final decision function as follows:
y(x i )＝ω * K(x i ,x)+b *
wherein x= [ x ] 1 ,x 2 ,…,x n ]ω, b are optimal parameters obtained from KKT conditions.
The beneficial effects are that: the method can well solve the problems of overlarge active disturbance, overlarge undetectable interval and overlarge detection cost existing in the island detection of the grid-connected inverter of the power generation system, does not need to additionally apply large disturbance, does not depend on the traditional island detection threshold value, can obtain initial data for island judgment through data fusion, and finally further obtains island and non-island states through a support vector machine, thereby achieving the purpose of accurately detecting island effects.
Drawings
FIG. 1 is a general flow chart of the present invention;
FIG. 2 is a topology of a grid-tie inverter of the power system;
FIG. 3 is a common point frequency response of grid fluctuations, islanding and reactive power disturbances;
FIG. 4 is a schematic diagram of a back propagation neural network;
FIG. 5 is a clustering result regarding the common point frequency using the k-means method.
Fig. 6 is a clustering result regarding reactive power injected into an inverter using the k-means method.
Fig. 7, wherein (a) is an assigned value for the clustered common point frequency and (b) is an assigned value for the clustered reactive power disturbance;
fig. 8 is a result of a logical operation on n sets of feature data sets.
Fig. 9, where (a) is the common point frequency collected under case 1,2 and (b) is the reactive power injected into the inverter collected under case 1, 2;
fig. 10, where (a) is the result of islanding detection for case 1 using only a support vector machine, and (b) is the result of islanding detection for case 2 using only a support vector machine;
fig. 11, where (a) is the result of islanding detection of case 1 using the present invention, and (b) is the result of islanding detection of case 2 using the present invention.
Detailed Description
The accompanying drawings, which are included to provide a further understanding of the invention and are incorporated in and constitute a part of this specification, illustrate embodiments of the invention and together with the description serve to explain the invention.
The embodiment as shown in fig. 1 provides a data driving power island detection method based on active reactive power disturbance, which specifically includes the following steps:
(1) Defining corresponding characteristic variables according to the occurrence mechanism of the island of the inverter, wherein the common point frequency output by the phase-locked loop and the modulated reactive power injected into the inverter are selected as inverter observation variables, and the modulated reactive power is the reactive power obtained according to the common point frequency; and collecting the common point frequency and the corresponding reactive power disturbance (the reactive power injected into the inverter), and taking one common point frequency and the corresponding reactive power disturbance as a group of characteristic data sets to obtain n groups of characteristic data sets.
(2) By means of mechanism analysis, it is known that the reactive power injected into the inverter has a correlation with the frequency of the common point of the inverter when the island occurs, so that in order to avoid false island false detection phenomenon caused by frequency fluctuation, n sets of characteristic data sets can be sequentially calculated through a logic operation, thereby obtaining n sets of fusion data for training an Adaptive Back Propagation Neural Network (ABPNN).
(3) And (3) combining a particle swarm algorithm, taking the n groups of characteristic data sets in the step (1) and the n groups of fusion data in the step (2) as training sets of the neural network, training the self-adaptive back propagation neural network to obtain a trained self-adaptive back propagation neural network, and obtaining the fusion data for a subsequent test by the neural network to obtain a data fusion model based on the self-adaptive back propagation neural network.
(4) An island fault detection model based on a Support Vector Machine (SVM) is established: for the island effect, two states, namely an island state and a non-island state, are generated in the running process of the inverter, the two states are classified by adopting a Support Vector Machine (SVM), and the output of the trained self-adaptive back propagation neural network is used as SVM training data; finally, the adaptive back propagation neural network-support vector machine (ABPS) classifier is obtained.
(5) And changing a common point frequency fluctuation mode and an RLC load through the constructed single-phase inverter to obtain a group of new characteristic data sets, classifying island and non-island states of the new characteristic data sets by using an ABPS classifier, and finally obtaining an island detection result under the condition that a pseudo island effect exists.
As shown in fig. 2, in this embodiment, a single-phase grid-connected inverter constructed by Matlab/Simulink is adopted, and the inverter is operated to obtain island test system parameters as shown in table 1, and n sets of characteristic data sets of reactive power injected into the inverter and common point frequency (common point frequency of phase-locked loop output) changes of the grid-connected inverter before and after the island effect is simulated.
Table 1 island test system parameters
In order to screen characteristic variables which can better describe island change of the grid-connected inverter, the mechanism analysis of the change of related variables caused by reactive power disturbance of the inverter is as follows:
wherein f' represents the common point frequency of the grid-connected inverter, f r For the resonant frequency of the inverter RLC load, Q f Is the quality factor of the inverter, Q DG And P DG Representing reactive power and active power injected into the inverter, respectively. Because the power control of the inverter keeps the active power, the resonant frequency and the quality factor are constant, a correlation exists between the common point frequency f' and the reactive power injected into the inverter, and finally the common point frequency of the parallel inverter and the input reactive power are selected as characteristics;
as shown in fig. 2, a disturbance is applied to the grid-connected inverter by using a triangular intermittent reactive power disturbance (reactive power injected into the inverter), a corresponding common point frequency value is obtained from the phase-locked loop at the PCC, and the reactive power injected into the inverter is obtained through power detection of the control loop.
As shown in fig. 3, the frequency value at the common point may be affected by the voltage fluctuation of the power grid and exceed the conventional detection threshold, resulting in false detection of the conventional island detection method. When the island effect occurs, the frequency at the common point will reach the inverter resonant frequency first, and then will change with the change of the reactive power injected.
According to the analysis, the common point frequency and the reactive power injected into the inverter are finally selected as characteristic variables.
Through mechanism analysis, the reactive power injected into the inverter has correlation with the frequency of the common point of the inverter when the island occurs, so that in order to avoid false island false detection phenomenon caused by frequency fluctuation, data fusion can be carried out on two groups of data sets through logic operation:
carrying out k-means calculation on the common point frequency in the n groups of characteristic data sets and the reactive power injected into the inverter to obtain a classification result of each common point frequency and the reactive power injected into the inverter; and performing numerical distribution on each classification result;
data fusion is performed on the group D feature data set by using the following formula, d=1, 2, … n;
L＝Cluster R ∧Cluster f
wherein Cluster R A value after numerical distribution is carried out on the classification result of the reactive power; cluster f A value obtained after numerical value distribution is carried out on the classification result of the common point frequency; l= {0,1}, when 0, it indicates that no islanding of the power system under the set of feature data sets occurs; when 1, the island of the power system under the characteristic data set is shown.
The result of the clustering analysis of the common point frequencies by k-means is shown in FIG. 5: the common point frequencies can be divided into three categories, the first: frequency when islanding does not occur, second category: island occurs and the reactive power injected into the inverter is at the common point frequency of the intermittent periods, third category: island occurs and the reactive power injected into the inverter is not at the common point frequency of the intermittent period; wherein, the cluster 1 represents the effect when no island occurs and the effect when the grid interference is in the downlink section, the cluster 2 represents the effect of the resonant frequency on the common point frequency and the effect of the grid uplink section interference when the island is contained in the category, and the cluster 3 represents the effect of the grid uplink section interference and the effect of the injected reactive power on the common point frequency.
The results of the cluster analysis of reactive power by k-means are shown in FIG. 6: the first category is: the reactive power injected into the inverter is not in the intermittent period; the second category is: the injected reactive power is in an intermittent period; when no island occurs, namely, the common point frequency is 50Hz, the injected reactive power is triangular reactive power (the waveform is triangular wave); when islanding occurs, the reactive power injected into the inverter will be less than the reactive power injected when islanding does not occur, as the common point frequency will be clamped to the resonant frequency of the inverter first; when the injected reactive power is in the intermittent period, the injected reactive power value is 0. In the figure, cluster 1 represents the intermittent period, and cluster 2 represents the disturbance period.
Fig. 7 shows that (a) is an assigned value of the clustered common point frequency, and it is known that the value of the first class is assigned 0; the value after the value of the second class is allocated is 1, and the value after the value of the third class is allocated is 2;
(b) The distributed values of the reactive power injected into the inverter after clustering can be known that the value after the distribution of the values of the first class is 0 and the value after the distribution of the values of the second class is 1;
fig. 8 is a data fusion result of n sets of feature data sets, the values of L are shown in table 2,
TABLE 2
The specific method for establishing the data fusion model based on the self-adaptive back propagation neural network in the step 3 is as follows: and (3) importing the training data set (n groups of characteristic data sets) and the logic operation output into Matlab 2018b, and calling the self-adaptive back propagation neural network program to train the data fusion model parameters of the neural network. As shown in fig. 4, the input data is two-dimensional frequency and reactive power, and the output data is one-dimensional fusion data. The corresponding operations are:
step 3.1: initializing parameters in a self-adaptive back propagation neural network and a particle swarm algorithm, wherein the scale of the particle swarm is M;
step 3.2, performing t iterative computation, and taking initial particles as weights and thresholds of the self-adaptive back propagation neural networks in sequence, wherein M self-adaptive back propagation neural networks exist;
step 3.3: sequentially inputting n groups of characteristic data sets serving as training sets into an mth self-adaptive back propagation neural network, and performing epochs training on the n groups of characteristic data sets by the neural network to obtain n outputs;
step 3.4: the nth output and the nth fusion data are subjected to mean square error, and n=1, 2, … and N; obtaining n groups of mean square deviations, and taking the sum of the n groups of mean square deviations as a group of fitness values;
step 3.5, judging whether M is larger than M, if so, turning to step 3.6; if not, m+1, turning to step 3.3;
step 3.6: comparing M groups of fitness values, and judging whether the fitness value smaller than Y is in the M groups of fitness values, wherein Y is a preset value; if yes, stopping the particle swarm algorithm, selecting a group of particles corresponding to the smallest fitness value in the fitness value group smaller than Y as a global optimal solution, and taking the global optimal solution as the weight and the threshold of the self-adaptive back propagation neural network to obtain the self-adaptive back propagation neural network with the determined weight and threshold, thus obtaining the trained self-adaptive back propagation neural network; if not, judging whether t reaches the maximum iteration number; if not, taking a group of particles corresponding to the smallest fitness value in the iterative computation as an individual optimal solution of the iterative computation, comparing the individual optimal solution with a historical global optimal solution, selecting a group of particles corresponding to the smallest fitness value as a global optimal solution of the iterative computation, and converting into a step 3.7; if yes, stopping iterative computation, and selecting a global optimal solution as the weight and the threshold of the self-adaptive back propagation neural network to obtain the self-adaptive back propagation neural network with the determined weight and threshold;
and 3.7, updating the speed and the position of each particle according to the global optimal solution and the individual optimal solution calculated in the iteration, taking the updated particles as initial particles, t+1, and turning to the step 3.2.
The parameters of the adaptive back propagation neural network for the data fusion process are:
net.trainParam.epochs＝2500，
net.trainParam.goal＝5e-4，
net.trainParam.lr＝0.1，
xSize＝26,
maxgen＝300,
netTrainParam. Epochs is the maximum training frequency (epochs frequency) of the neural network, netTrainParam. Gol is the required accuracy of the neural network training, netTrainParam. Lr is the learning rate of the neural network, xSize represents M, which is the population number of particle swarms in the adaptive neural network, and maxgen represents the maximum iteration frequency of the adaptive algorithm.
The specific method of the step 4 is as follows: island fault detection model based on Support Vector Machine (SVM): output x of trained adaptive back propagation neural network i (i=1, 2, … n) as input to the support vector machine of step 4, the specific steps are:
(1) Importing training data x i (i=1, 2, n), a super-plane is established and a plane is formed, and describes the class decision function as:
y(x i )＝ω T φ(x i )+b
where ω represents the normal vector to the hyperplane and b represents the deviation.
(2) If the decision function satisfies the conditionThen x i The corresponding group of characteristic data is non-island data, which indicates that the power generation system under the data has no island; if the decision function satisfies the condition->Then x i The group of corresponding characteristic data is island data, which indicates that the power generation system under the data has island;
(3) Target y i E { -1,1} and each x i And (5) associating. The regularization parameter c is initialized. The problem of finding a hyperplane appears to be a Quadratic Programming (QP) optimization problem.
Wherein, xi i Not less than 0 meterShow AND x i A corresponding relaxation variable that measures the distance between the margin and the inseparable sample.
(4) A Radial Basis Function (RBF) is employed as a kernel function of the training data. The lagrangian function is used to solve the QP problem. The optimal parameters ω, b are obtained from Karush-Kuhn-turner (KKT) conditions, and then the final decision function is rewritten as:
y(x i )＝ω * K(x i ,x)+b *
wherein x= [ x ] 1 ,x 2 ,…,x n ]。
As shown in fig. 9, where (a) is the frequency at the common point acquired at case 1, case 2, and (b) is the reactive power injected into the inverter acquired at case 1, case 2.
Case 1 is that islanding occurs at 0.2s, the resonant frequency is 50.15hz, the rlc load is set to 15.5 Ω, and the reactive power injected is between 0.7 and 0.8 seconds.
Case 2 shows that islanding occurs at 0.3s, the resonant frequency is 50.15hz, the rlc load is 15.5 Ω, and the reactive power injected is 0.6-0.7 seconds.
As shown in fig. 10, wherein (a) is island detection of case 1 using only a support vector machine; (b) The island detection is performed on case 2 by using only a support vector machine.
As shown in fig. 11, where (a) is the result of island detection for case 1 using the ABPS of the present embodiment, and (b) is the result of island detection for case 2 using the ABPS of the present embodiment.
Comparing fig. 10 with fig. 11, it can be seen that the influence caused by the pseudo islanding effect cannot be avoided by using only the support vector machine for detection; the invention can normally detect the real island state when the pseudo island effect occurs.
In addition, the specific features described in the above embodiments may be combined in any suitable manner without contradiction. The various possible combinations of the invention are not described in detail in order to avoid unnecessary repetition.
Claims (3)
1. The data driving power island detection method based on active reactive power disturbance is characterized by being used for judging whether an island exists in a distributed power generation system or not and specifically comprising the following steps of:
step 1: calculating to obtain reactive power injected into the inverter according to the common point frequency output by the phase-locked loop end in the distributed power generation system; taking the common point frequency and the reactive power injected into the inverter as a group of characteristic data sets; collecting the common point frequency and the corresponding reactive power to obtain n groups of characteristic data sets;
step 2: sequentially carrying out data fusion on the n groups of characteristic data sets to obtain n groups of fusion data;
step 3: based on n groups of fusion data, calculating the weight and the threshold of the self-adaptive back propagation neural network by using a particle swarm algorithm; meanwhile, in each particle swarm iterative computation, taking the n groups of characteristic data sets in the step 1 as training sets of the self-adaptive back propagation neural network to obtain a trained self-adaptive back propagation neural network;
step 4: taking the output of the trained self-adaptive back propagation neural network as the input of a support vector machine, and training the support vector machine to obtain a decision function; finally, building a training self-adaptive back propagation neural network-support vector machine classifier;
step 5: the common point frequency fluctuation mode and the PLC load of the inverter are changed to obtain a group of new characteristic data sets, and the group of new characteristic data sets are classified by utilizing the self-adaptive back propagation neural network-support vector machine classifier, so that whether island exists in the power system or not is judged, specifically, the method comprises the following steps: the new test data is used as a training set of the trained self-adaptive back propagation neural network, the output of the self-adaptive back propagation neural network is obtained, and whether the output is island data or not is calculated by utilizing a decision function, so that whether island exists in the power generation system or not is judged;
the specific method for calculating the reactive power injected into the inverter in the step 1 is as follows:
wherein f' represents the common point frequency of the phase-locked loop end output, f r For the resonant frequency of the inverter RLC load, Q f Is the quality factor of the inverter, Q DG And P DG Representing reactive power and active power injected into the inverter respectively;
the specific steps of the step 3 are as follows:
step 3.1: initializing parameters in a self-adaptive back propagation neural network and a particle swarm algorithm, wherein the scale of the particle swarm is M;
step 3.2, performing t iterative computation, and taking initial particles as weights and thresholds of the self-adaptive back propagation neural networks in sequence, wherein M self-adaptive back propagation neural networks exist;
step 3.3: sequentially inputting n groups of characteristic data sets serving as training sets into an mth self-adaptive back propagation neural network, and performing epochs training on the n groups of characteristic data sets by the neural network to obtain n outputs;
step 3.4: the nth output and the nth fusion data are subjected to mean square error, and n=1, 2, … and N; obtaining n groups of mean square deviations, and taking the sum of the n groups of mean square deviations as a group of fitness values;
step 3.5, judging whether M is larger than M, if so, turning to step 3.6; if not, m+1, turning to step 3.3;
step 3.6: comparing M groups of fitness values, and judging whether the fitness value smaller than Y is in the M groups of fitness values, wherein Y is a preset value; if yes, stopping the particle swarm algorithm, selecting a group of particles corresponding to the smallest fitness value in the fitness value group smaller than Y as a global optimal solution, and taking the global optimal solution as the weight and the threshold of the self-adaptive back propagation neural network to obtain the self-adaptive back propagation neural network with the determined weight and threshold, thus obtaining the trained self-adaptive back propagation neural network; if not, judging whether t reaches the maximum iteration number; if not, taking a group of particles corresponding to the smallest fitness value in the iterative computation as an individual optimal solution of the iterative computation, comparing the individual optimal solution with a historical global optimal solution, selecting a group of particles corresponding to the smallest fitness value as a global optimal solution of the iterative computation, and converting into a step 3.7; if yes, stopping iterative computation, and selecting a global optimal solution as the weight and the threshold of the self-adaptive back propagation neural network to obtain the self-adaptive back propagation neural network with the determined weight and threshold;
step 3.7, updating the speed and the position of each particle according to the global optimal solution and the individual optimal solution calculated in the iteration, taking the updated particles as initial particles, t+1, and turning to step 3.2;
the specific method of the step 4 is as follows:
output x of trained adaptive back propagation neural network i (i=1, 2, … n) as training data for the support vector machine and creates a hyperplane describing the decision function as:
y(x i )＝ω T φ(x i )+b
where ω represents the normal vector to the hyperplane, b represents the deviation,for mapping functions for solving x i Is an inner product of (2);
if the decision function satisfies the conditionThen x i The corresponding group of characteristic data is non-island data, which indicates that the power generation system under the data has no island; if the decision function satisfies the condition->Then x i The group of corresponding characteristic data is island data, which indicates that the power generation system under the data has island;
based on y i E { -1,1} and each x i In association, initializing regularization parameters C, and representing the hyperplane as a quadratic programming optimization problem:
wherein xi i 0 is ≡0 is ≡x i A corresponding relaxation variable;
and (3) taking the radial basis function as a kernel function of training data, solving the quadratic programming optimization problem by utilizing a Lagrange function, and rewriting a final decision function as follows:
y(x i )＝ω * K(x i ,x)+b *
wherein x= [ x ] 1 ,x 2 ,…,x n ]ω, b are optimal parameters obtained from KKT conditions.
2. The method for detecting the island of the data-driven power based on the active reactive power disturbance according to claim 1, wherein the specific method for sequentially performing data fusion on n groups of characteristic data sets in the step 2 is as follows:
carrying out k-means calculation on the common point frequency in the n groups of characteristic data sets and the reactive power injected into the inverter to obtain a classification result of each common point frequency and the reactive power injected into the inverter; and performing numerical distribution on each classification result;
data fusion is performed on the group D feature data set by using the following formula, d=1, 2, … n;
L＝Cluster R ∧Cluster f
wherein Cluster R A value after numerical distribution is carried out on the classification result of the reactive power; cluster f A value obtained after numerical value distribution is carried out on the classification result of the common point frequency; l= {0,1}, when 0, it indicates that no islanding of the power system under the set of feature data sets occurs; when 1, the island of the power system under the characteristic data set is shown.
3. The method for detecting the island of the data-driven power based on the active reactive power disturbance according to claim 1, wherein the k-means algorithm is used for classifying the reactive power data set injected into the inverter, and the classification result is two types, namely: the reactive power injected into the inverter is not in the intermittent period; the second category is: the injected reactive power is in an intermittent period; according to the value of the common point frequency when island occurs, classifying the common point frequency data set by using a k-means algorithm, wherein the classification result is three types, and the first type is: common point frequency when island does not occur; the second category is: island occurs and the reactive power injected into the inverter is at the common point frequency of the intermittent period; the third category is: island occurs and the reactive power injected into the inverter is not at the common point frequency of the intermittent period; the values of L after numerical assignment based on the above classification result are shown in the following table 1:
TABLE 1
Wherein the value of the reactive power after being allocated by the first class of the numerical value is 0, and the value of the reactive power after being allocated by the second class of the numerical value is 1;
the first class of the common point frequency is assigned a value of 0, the second class is assigned a value of 1, and the third class is assigned a value of 2.
Priority Applications (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN201911189108.9A CN111160386B (en) | 2019-11-28 | 2019-11-28 | Data-driven power island detection method based on active reactive power disturbance |
Applications Claiming Priority (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN201911189108.9A CN111160386B (en) | 2019-11-28 | 2019-11-28 | Data-driven power island detection method based on active reactive power disturbance |
Publications (2)
| Publication Number | Publication Date |
|---|---|
| CN111160386A CN111160386A (en) | 2020-05-15 |
| CN111160386B true CN111160386B (en) | 2023-11-10 |
Family
ID=70556182
Family Applications (1)
| Application Number | Title | Priority Date | Filing Date |
|---|---|---|---|
| CN201911189108.9A Active CN111160386B (en) | 2019-11-28 | 2019-11-28 | Data-driven power island detection method based on active reactive power disturbance |
Country Status (1)
| Country | Link |
|---|---|
| CN (1) | CN111160386B (en) |
Families Citing this family (2)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN114705947A (en) * | 2022-03-16 | 2022-07-05 | 广东电网有限责任公司 | Island detection model training method, device, equipment and medium |
| CN114565058A (en) * | 2022-03-16 | 2022-05-31 | 广东电网有限责任公司 | Training method, device, equipment and medium for island detection model |
Citations (2)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN104316786A (en) * | 2014-10-10 | 2015-01-28 | 湖南大学 | Mixed isolated island detection method |
| CN105515044A (en) * | 2015-12-22 | 2016-04-20 | 国家电网公司 | Black-start assistant decision-making system based on DTS |
Family Cites Families (1)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| US8527223B2 (en) * | 2009-09-11 | 2013-09-03 | University Of Cincinnati | Methods and systems for energy prognosis |
-
2019
- 2019-11-28 CN CN201911189108.9A patent/CN111160386B/en active Active
Patent Citations (2)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN104316786A (en) * | 2014-10-10 | 2015-01-28 | 湖南大学 | Mixed isolated island detection method |
| CN105515044A (en) * | 2015-12-22 | 2016-04-20 | 国家电网公司 | Black-start assistant decision-making system based on DTS |
Non-Patent Citations (2)
| Title |
|---|
| 李洋，李坤.并网逆变器新型控制策略的研究.《新能源发电控制技术》.2016,全文. * |
| 谢星宇.光伏并网逆变器孤岛检测技术研究.《中国优秀硕士论文全文数据库》.2019,全文. * |
Also Published As
| Publication number | Publication date |
|---|---|
| CN111160386A (en) | 2020-05-15 |
Similar Documents
| Publication | Publication Date | Title |
|---|---|---|
| Zhao et al. | Data-driven correction approach to refine power curve of wind farm under wind curtailment | |
| Zhang et al. | Short-term wind speed interval prediction based on artificial intelligence methods and error probability distribution | |
| Trivedi et al. | Implementation of artificial intelligence techniques in microgrid control environment: Current progress and future scopes | |
| Baltas et al. | A comparative analysis of decision trees, support vector machines and artificial neural networks for on-line transient stability assessment | |
| CN107153845A (en) | A kind of isolated island detection method of the adaptive grid-connected photovoltaic system based on machine learning | |
| Li et al. | Islanding fault detection based on data-driven approach with active developed reactive power variation | |
| CN111160386B (en) | Data-driven power island detection method based on active reactive power disturbance | |
| Huang et al. | A hybrid method for one-day ahead hourly forecasting of PV power output | |
| da Cunha et al. | Small-signal stability analysis in smart grids: An approach based on distributed decision trees | |
| Asghar et al. | RECLAIM: Renewable energy based demand-side management using machine learning models | |
| Shahriyari et al. | A Deep Learning-Based Approach for Comprehensive Rotor Angle Stability Assessment | |
| CN113222067B (en) | Island intelligent detection method based on SVM-Adaboost algorithm | |
| Muzumdar et al. | Analyzing the feasibility of different machine learning techniques for energy imbalance classification in smart grid | |
| KR20190071174A (en) | Method and system for short-term wind speed prediction based on pressure data | |
| Zhang et al. | Prediction modeling of frequency response characteristic of power system based on historical data | |
| Vidya et al. | A review on the hybrid approaches for wind speed forecasting | |
| Sun et al. | Short-term photovoltaic power prediction modeling based on AdaBoost algorithm and Elman | |
| CN116896093A (en) | Online analysis and optimization method for grid-connected oscillation stability of wind farm | |
| Vantuch et al. | Power quality prediction designed as binary classification in AC coupling Off-Grid system | |
| Wang et al. | Model-Data Integration Driven Based Power System Frequency Response Model | |
| Bitaraf et al. | A novel SVM approach of islanding detection in micro grid | |
| Yu et al. | Study on a comprehensive indicator and online classification of early warning of low frequency oscillation in power system | |
| Ramirez-Gonzalez et al. | Power System Inertia Estimation Using A Residual Neural Network Based Approach | |
| Yin et al. | Forecasting the intrinsic viscosity of polyester based on improved extreme learning machine | |
| Li et al. | A short-term photovoltaic output prediction method based on improved PSO-RVM algorithm |
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 |