CN116227320A - Double-fed fan control parameter identification method based on LSTM-IPSO - Google Patents
Double-fed fan control parameter identification method based on LSTM-IPSO Download PDFInfo
- Publication number
- CN116227320A CN116227320A CN202211550049.5A CN202211550049A CN116227320A CN 116227320 A CN116227320 A CN 116227320A CN 202211550049 A CN202211550049 A CN 202211550049A CN 116227320 A CN116227320 A CN 116227320A
- Authority
- CN
- China
- Prior art keywords
- value
- lstm
- output
- identification
- doubly
- Prior art date
- Legal status (The legal status is an assumption and is not a legal conclusion. Google has not performed a legal analysis and makes no representation as to the accuracy of the status listed.)
- Pending
Links
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
- G06F17/00—Digital computing or data processing equipment or methods, specially adapted for specific functions
- G06F17/10—Complex mathematical operations
-
- 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
- G06N3/08—Learning methods
-
- 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
- Y02E—REDUCTION OF GREENHOUSE GAS [GHG] EMISSIONS, RELATED TO ENERGY GENERATION, TRANSMISSION OR DISTRIBUTION
- Y02E10/00—Energy generation through renewable energy sources
- Y02E10/70—Wind energy
- Y02E10/76—Power conversion electric or electronic aspects
Abstract
The invention discloses a doubly-fed fan control parameter identification method based on LSTM-IPSO, which comprises the following steps: acquiring doubly-fed fan hardware in-loop experimental data from a real controller by using an RT-LAB semi-physical simulation platform, and building an isomorphic identification model of the doubly-fed fan real controller on a Matlab/simulink platform; adding the dimension of the input feature set, removing irrelevant features, and selecting a feature value with higher correlation as the input feature set of the neural network model; the input feature set and the corresponding control parameter set form a control parameter-input feature set; training and predicting a control parameter-input characteristic set by using an LSTM neural network to obtain a prediction initial value and an optimizing range; the IPSO algorithm is used as a secondary optimizing method for accurate identification, so that the aim of accurate optimizing is fulfilled; judging the reliability of the identification model; the invention solves the technical problem that the traditional identification method is difficult to identify the control parameters of the electromagnetic model of the doubly-fed wind turbine under the low-voltage ride through working condition.
Description
Technical Field
The invention relates to the technical field of new energy power generation parameter identification, in particular to a double-fed fan control parameter identification method based on LSTM-IPSO.
Background
In recent years, with large-scale grid connection of wind turbines, the model problem of wind turbines is increasingly prominent. At present, the power grid analysis mainly depends on model simulation or digital simulation, and model parameters are mostly set to be experience values by manufacturers, wherein grid-side control parameters are difficult to collect from nameplates and technical manuals, but accuracy of the grid-side control parameters directly influences stability and safety of power grid operation, so that from the aspects of fan self-characteristic research and power system stability analysis, it is necessary to construct a fine simulation model of a wind turbine generator set based on measured data and identify the grid-side control parameters.
Based on this, aiming at the problem that the control parameters of the electromagnetic model of the doubly-fed fan are difficult to be identified with high precision by the traditional identification method under the low-voltage ride-through working condition, it is necessary to provide a doubly-fed fan control parameter identification method based on long short-term memory (LSTM) neural network combined with improved particle swarm (Improved Particle Swarm Optimization, IPSO) algorithm.
Disclosure of Invention
The invention aims to overcome the defects and provide a double-fed fan control parameter identification method based on LSTM-IPSO, so as to solve the technical problem that the traditional identification method is difficult to identify the double-fed fan electromagnetic model control parameters with high precision under the low-voltage ride through working condition.
The invention aims to solve the technical problems, and adopts the technical scheme that: a double-fed fan control parameter identification method based on LSTM-IPSO comprises the following steps:
firstly, acquiring ring experiment data of doubly-fed fan hardware from a real controller by using an RT-LAB semi-physical simulation platform, and building an isomorphic identification model of the doubly-fed fan real controller on a Matlab/simulink platform;
step two, increasing the dimension of an input feature set and removing irrelevant features, and selecting a feature value with higher correlation as the input feature set of the neural network model; the input feature set and the corresponding control parameter set form a control parameter-input feature set;
thirdly, training and predicting a control parameter-input characteristic set by using an LSTM neural network to obtain a prediction initial value and an optimizing range;
fourth, based on the prediction initial value and the optimizing range obtained in the fourth step, the IPSO algorithm is used as a secondary optimizing method for accurate identification, and the purpose of accurate optimizing is achieved;
and fifthly, comparing and verifying the identified identification model with hardware-in-loop experimental data, and judging the reliability of the identification model.
Preferably, in the first step, the RT-LAB semi-physical simulation platform is used to obtain the hardware-in-loop experimental data of the doubly-fed wind turbine from the real controller, including the dc bus voltage u dc Dq component i of output current dg And i qg Wherein the d-axis reference current i dg Obtained by DC voltage control, and q-axis reference current i qg Setting to zero, and building an isomorphic identification model of the real controller of the doubly-fed fan on the Matlab/simulink platform.
Preferably, the doubly-fed wind turbine comprises a wind turbine, a gear box, an induction asynchronous generator, a back-to-back converter and a controller thereof;
the control equation related to the isomorphic recognition model of the real controller of the doubly-fed fan is built on the Matlab/simulink platform is as follows:
according to the law of conservation of energy, the KCL equation of the DC capacitor C can be obtained as follows:
wherein P is r For the active power output to C by the machine side converter, P rc Active power injected into a power grid for the grid-side converter;
the dynamic model of the filter reactance is:
wherein u is cd 、u cq And u is equal to sd 、u sq Representing the components of the voltages on the d-axis and q-axis at the outlets of the grid-side converter and the machine-side converter, respectively, X r Is the filter reactance value.
The net side controller control equation can be written as:
wherein u is dcref Is the reference value of direct current voltage, i dgref And i qgref Actual values of net side d-axis and q-axis currents, K p1 Is the proportional coefficient, K of the voltage outer loop PI controller p2 For the q-axis current inner loop proportionality coefficient, K p3 For q-axis current control inner ring proportion coefficient, setting intermediate variables output by PI controller integration link as x 1 、x 2 And x 3 The dynamic equation of the controller is:
wherein K is i1 Integrating coefficient, K for voltage outer loop PI controller i2 For the inner loop integral coefficient of the q-axis current, K i3 Controlling the inner loop integral coefficient for q-axis current, i qgref Typically 0, but under fault conditions i qgref The low pass control module is directly given by:
wherein k is a reactive current support coefficient, U N Rated voltage for grid-connected point of fan, I N Rated current for GSC; when low voltage ride through occurs, the grid-side converter can emit certain reactive power, the value of reactive current can be obtained according to the formula (5), at the moment, the q-axis current inner loop is switched to a low-pass control mode, and the q-axis reference value is given by the low-pass control module.
Preferably, the second step specifically includes the following steps:
step A1, collecting isomorphic recognition model control parameters-input characteristic sets, wherein the input characteristic values are parameters related to a fan network side controller: u (u) dc 、i dg And i qg And the error value F of the hardware-in-loop test data is the control parameter data set of the network-side converter corresponding to the input characteristic value;
a2, eliminating sample data of abnormal data points, so as to avoid affecting identification accuracy;
step A3, dividing the simulation time into 5 intervals according to the modeling guideline of the wind turbine generator electrical simulation model: steady state intervals (1) and (5), low voltage crossing intervals (2), (3) and (4), and inputting i in characteristic values dg And i qg Also by this way the values are taken as feature sets between partitions, where u dc By taking values between partitionsObtaining a characteristic value V a 、V b 、V c 、V d And V e ,i dg Obtaining a characteristic value I through partition da 、I db 、I dc 、I dd And I qe ,i qg Obtaining a characteristic value I through partition qa 、I qb 、I qc 、I qd And I qe ;
A4, carrying out association degree analysis on input control parameters and output characteristic variables by adopting maximum information coefficients (Maximal Information Coefficient, MIC), firstly carrying out a-column b-row gridding on a scatter diagram formed by the control parameters (K) and the characteristic variables (T), solving the maximum mutual information value, normalizing the maximum mutual information value, and finally selecting the maximum mutual information value under different scales as the MIC value;
where mutual information can be seen as the amount of information contained in one random variable about another, the mutual information I (K, T) of K and T is defined as:
wherein K is the sample sum of the variable K, T is the sample sum of the variable T, P (K, T) is the joint probability between the variable K and T, and P (K) and P (T) are the edge probabilities of K and T respectively;
wherein n is the number of samples;
the MIC value range obtained by the formula (8) is between 0 and 1, the greater the MIC value is, the higher the association degree between two variables is, and after the MIC value of an output signal is obtained, a key output signal index is selected as the input characteristic of the LSTM neural network;
and step A5, carrying out normalization processing on the data set, and accelerating the speed of gradient descent to solve the optimal solution after dimensionless data.
Preferably, the third step adopts LSTM neural network according to the existing positionThe method comprises the steps that a data set is processed to conduct preliminary prediction on an initial value of a control parameter of the doubly fed fan and an optimizing range, so that the control parameter approaches to a minimum objective function of an identification model, and after a large amount of debugging work, structural parameters and network training parameters of an LSTM neural network are obtained, and the LSTM neural network model is trained and tested; the core cell state c is controlled by a forgetting gate, an input gate and an output gate, wherein: sigma is a sigmoid function, x t Input to the t-th cell, c t Is the cell state of the t th cell, h t The hidden state of the t-th unit, the sum of the hidden states of the t-th unitSumming the vector elements and integrating the vector elements, c t-1 Is the cell state of the t-1 th cell, h t-1 A hidden state for the t-1 th cell;
the output change of the LSTM neural network output parameter under the influence of the input value can be expressed by the following formula:
wherein f t To forget the gate output value, W i 、W f And W is o Network layer weights for ingress, forget and egress gates, b i 、b f And b o Bias items for input gate, forget gate and output gate, b c Is the network layer bias and,information input to neurons for time t, o t To output the gate output value, ☉ is the hadamard product operator.
Preferably, in the third step, the specific step of training prediction is as follows:
step B1, reasonably selecting and distributing data sets and parameter adjustment, wherein the accuracy of a prediction result is greatly affected, 90% of the data sets are randomly selected as training sets, 10% of the data sets are used as verification sets, and the accuracy of a model is verified;
step B2, setting the iteration training frequency as 500, setting the loss function as MSE, setting the optimizer as Adam, setting the input dimension as 12, setting the output dimension as 6, setting the hidden layer number as 2, setting the hidden layer neuron number as 11 and setting the learning rate as 0.015;
and B3, training and predicting the acquired sequence by adopting an LSTM neural network, and obtaining a parameter prediction initial value and an optimizing range.
Preferably, in the fourth step, the PSO algorithm speed and location update formula is as follows:
V u+1 =ωV u +c 1 r 1 (pbest u -x u )+c 2 r 2 (gbest u -x u ) (10)
x u+1 =x u +V u+1 (11)
wherein V represents the particle update rate, x represents the particle, u represents the current time, u+1 represents the next update time, c 1 、c 2 R is the learning factor 1 、r 2 A random value between 0 and 1, pbest is an individual optimal value, gbest is a global optimal value, and ω is an adaptive inertial weight;
PSO takes the mean square value of the output value of the actual position of each particle and the output value of the optimal position as an objective function of an algorithm to obtain the fitness value of each particle in the current iteration times, and the objective function is expressed as:
in U i,error 、I di,error And I qi,error Is u dc 、i dg And i qg Error of U i 、I di And I qi Is u dc 、i dg And i qg Is the actual measurement data value of U ti 、I dti And I qti Is u dc 、i dg And i qg An output response value obtained by the identification result;
step C1, improving an objective function: according to the self-adaptive weight method, when objective function construction is carried out on multiple target output values, nonlinear dynamic adjustment weight distribution is carried out according to the MIC value and the actual value of each output value, so that the objective function of the multiple output values carries out self-adaptive adjustment on the output values with high association degree or small actual values, and the improvement formula is as follows:
wherein k is U 、k ID And k IQ Is U (U) i 、I di And I qi Adaptive weighting coefficients.
Wherein k is U As shown in equation (14), the adaptive weighting coefficients for the remaining output values may be as shown in equation (14):
wherein mic U Is u dc The sum of MIC values of each interval is calculated by updating the pbest and gbest of each particle using an algorithm in an iterative process, and then calculating the objective function F of each sample I Storing the optimal value and the corresponding parameter value;
step C2, improving inertia weight omega: the inertia weight nonlinear decrementing mode is adopted, so that omega is adaptively adjusted along with the iteration times of particles, the global searching capability of the early stage of the particles is ensured, the local optimizing capability of the later stage is improved, and the following formula (15) is improved:
wherein omega is min And omega max E and G are the current and total iteration times for the minimum and maximum omega values;
step C3, improving learning factor C 1 ,c 2 : learning factor c in particle velocity update formula 1 And c 2 Reflects the optimizing capability of particle individual optimization and population optimization, and the proper learning factors can improve the identification precision and shorten the convergenceTime and probability of reducing the trapping local optimum, learning factors are reduced in a nonlinear manner along with the increase of iteration times in the optimizing process, and a formula (16) is improved:
preferably, the fifth step specifically includes the following steps:
step D1, in order to further verify the effectiveness and practicability of the identification of the proposed model, comparing an LSTM-IPSO algorithm with a PSO algorithm and an LSTM-PSO algorithm, training the model by adopting the same data set, and comparing 3 groups of identification curves with actual measurement curves under 20%, 40%, 60% and 80% of voltage drop degree;
step D2, respectively calculating average deviations of 3 model identification results in the intervals (1), (2), (3), (4) and (5) under the voltage drop degree of 20%, 40%, 60% and 80% according to a formula (17), and then averaging the average deviations under 4 working conditions to obtain errors of PSO, LSTM-PSO and LSTM-IPSO model output results in each interval;
wherein F is Vi Is interval V i Average deviation of u M For actually measuring the DC capacitance voltage, u i To identify the model DC capacitor voltage, K start And K end For the first and last dummy data sequence numbers of the interval.
The invention has the beneficial effects that:
1. according to the invention, the initial value and the optimizing range of the IPSO are obtained based on the LSTM neural network control parameter prediction result, the IPSO algorithm is used as a secondary optimizing method for accurate identification, the purpose of accurate optimizing is achieved, and the technical problem that the control parameters of the electromagnetic model of the doubly-fed fan are difficult to identify with high precision in the traditional identification method under the low-voltage ride-through working condition is solved;
2. in order to remove irrelevant features, performing association degree analysis on input control parameters and output feature variables by adopting a maximum information coefficient, and selecting a key output signal index as an input feature of the LSTM neural network;
3. compared with the traditional parameter identification method, the parameter identification method of the doubly-fed fan controller based on the LSTM-LSTM simulates the input-output characteristics of a fan network side control system through an LSTM training data set, and under the condition that a fan model is not operated, actually measured data is input into an LSTM neural network to obtain a control parameter prediction result;
4. according to the method, when the objective function is constructed on the multi-objective output values according to the self-adaptive weight method, nonlinear dynamic adjustment weight distribution is carried out according to the MIC value and the actual value of each output value, so that the objective function of the multi-output value carries out self-adaptive adjustment on the output values with high association degree or small actual values, and the problem of poor algorithm optimizing effect caused by improper objective function construction is effectively solved. And the self-adaptive inertial weight and the self-adaptive learning factor are also applied, so that the later optimizing range of the identification algorithm is reduced, and the parameter identification precision is improved.
Drawings
FIG. 1 is a schematic diagram of a method for identifying control parameters of a doubly-fed wind turbine according to the present invention;
FIG. 2 is a diagram showing the MIC values of the input characteristic values of each section of the DC capacitor voltage versus each control parameter according to the present invention;
FIG. 3 is a graph showing the MIC values of the d-axis DC input characteristic values of each interval versus each control parameter;
FIG. 4 is a graph showing the MIC values of the input characteristic values of each interval of the q-axis direct current of the present invention for each control parameter;
FIG. 5 is an iteration diagram of a loss function in LSTM model training of the present invention;
FIG. 6 is a graph of the mean square error iterations of the LSTM-IPSO, PSO and LSTM-PSO recognition algorithms of the present invention;
FIG. 7 is a graph showing DC capacitor voltage comparison results of LSTM-IPSO identification curve, PSO identification curve and LSTM-PSO identification curve according to the present invention and actual measurement curve at 20% voltage drop;
FIG. 8 is a graph of d-axis DC comparison results of LSTM-IPSO identification curves, PSO identification curves and LSTM-PSO identification curves according to the present invention with measured curves at 20% voltage sag;
FIG. 9 is a graph of q-axis DC comparison results of LSTM-IPSO identification curves, PSO identification curves and LSTM-PSO identification curves according to the present invention with measured curves at 20% voltage sag;
FIG. 10 is a graph showing DC capacitor voltage comparison results of LSTM-IPSO identification curve, PSO identification curve and LSTM-PSO identification curve according to the present invention with measured curve at 80% voltage drop;
FIG. 11 is a graph of d-axis DC comparison results of LSTM-IPSO identification curves, PSO identification curves and LSTM-PSO identification curves according to the present invention with measured curves at 80% voltage sag;
FIG. 12 is a graph of q-axis DC comparison results of LSTM-IPSO identification curves, PSO identification curves and LSTM-PSO identification curves according to the present invention with measured curves at 80% voltage sag;
FIG. 13 is an error chart of DC capacitor voltage identification results of LSTM-IPSO algorithm, PSO algorithm and LSTM-PSO algorithm of the present invention;
FIG. 14 is an error plot of d-axis DC identification results of LSTM-IPSO algorithm, PSO algorithm and LSTM-PSO algorithm of the present invention;
FIG. 15 is an error plot of the results of the q-axis direct current identification of the LSTM-IPSO algorithm, PSO algorithm and LSTM-PSO algorithm of the present invention.
Detailed Description
The invention is described in further detail below with reference to the drawings and the specific examples.
As shown in fig. 1-15, a doubly-fed fan control parameter identification method based on LSTM-IPSO includes the following steps:
firstly, acquiring ring experiment data of doubly-fed fan hardware from a real controller by using an RT-LAB semi-physical simulation platform, and building an isomorphic identification model of the doubly-fed fan real controller on a Matlab/simulink platform;
step two, increasing the dimension of an input feature set and removing irrelevant features, and selecting a feature value with higher correlation as the input feature set of the neural network model; the input feature set and the corresponding control parameter set form a control parameter-input feature set;
thirdly, training and predicting a control parameter-input characteristic set by using an LSTM neural network to obtain a prediction initial value and an optimizing range;
fourth, based on the prediction initial value and the optimizing range obtained in the fourth step, the IPSO algorithm is used as a secondary optimizing method for accurate identification, and the purpose of accurate optimizing is achieved;
and fifthly, comparing and verifying the identified identification model with hardware-in-loop experimental data, and judging the reliability of the identification model.
Preferably, in the first step, the RT-LAB semi-physical simulation platform is used to obtain the hardware-in-loop experimental data of the doubly-fed wind turbine from the real controller, including the dc bus voltage u dc Dq component i of output current dg And i qg Wherein the d-axis reference current i dg Obtained by DC voltage control, and q-axis reference current i qg Setting to zero, and building an isomorphic identification model of the real controller of the doubly-fed fan on the Matlab/simulink platform.
Preferably, the doubly-fed wind turbine comprises a wind turbine, a gear box, an induction asynchronous generator, a back-to-back converter and a controller thereof; for the grid-side converter, the purpose of the grid-side converter is to keep the direct-current connection voltage stable, ensure good sinusoidal output current and control the power factor. Thus, d-axis reference current i dg Obtained by DC voltage control, and q-axis reference current i dq Typically set to zero.
The control equation related to the isomorphic recognition model of the real controller of the doubly-fed fan is built on the Matlab/simulink platform is as follows:
according to the law of conservation of energy, the KCL equation of the DC capacitor C can be obtained as follows:
wherein P is r For the active power output to C by the machine side converter, P rc Active power injected into a power grid for the grid-side converter;
the dynamic model of the filter reactance is:
wherein u is cd 、u cq And u is equal to sd 、u sq Representing the components of the voltages on the d-axis and q-axis at the outlets of the grid-side converter and the machine-side converter, respectively, X r Is the filter reactance value.
The net side controller control equation can be written as:
wherein u is dcref Is the reference value of direct current voltage, i dgref And i qgref Actual values of net side d-axis and q-axis currents, K p1 Is the proportional coefficient, K of the voltage outer loop PI controller p2 For the q-axis current inner loop proportionality coefficient, K p3 For q-axis current control inner ring proportion coefficient, setting intermediate variables output by PI controller integration link as x 1 、x 2 And x 3 The dynamic equation of the controller is:
wherein K is i1 Integrating coefficient, K for voltage outer loop PI controller i2 For the inner loop integral coefficient of the q-axis current, K i3 Controlling the inner loop integral coefficient for q-axis current, i qgref Typically 0, but under fault conditions i qgref The low pass control module is directly given by:
wherein k is a reactive current support coefficient, U N Rated voltage for grid-connected point of fan, I N Rated current for GSC; when low voltage ride through occurs, the grid-side converter can emit certain reactive power, the value of reactive current can be obtained according to the formula (5), at the moment, the q-axis current inner loop is switched to a low-pass control mode, and the q-axis reference value is given by the low-pass control module.
Preferably, the second step specifically includes the following steps:
step A1, collecting isomorphic recognition model control parameters-input characteristic sets, wherein the input characteristic values are parameters related to a fan network side controller: u (u) dc 、i dg And i qg And the error value F of the hardware-in-loop test data is the control parameter data set of the network-side converter corresponding to the input characteristic value;
a2, eliminating sample data of abnormal data points, so as to avoid affecting identification accuracy;
step A3, dividing the simulation time into 5 intervals according to the modeling guideline of the wind turbine generator electrical simulation model: steady state intervals (1) and (5), low voltage crossing intervals (2), (3) and (4), and inputting i in characteristic values dg And i qg Also by this way the values are taken as feature sets between partitions, where u dc Obtaining a characteristic value V through partition a 、V b 、V c 、V d And V e ,i dg Obtaining a characteristic value I through partition da 、I db 、I dc 、I dd And I qe ,i qg Obtaining a characteristic value I through partition qa 、I qb 、I qc 、I qd And I qe ;
A4, carrying out association degree analysis on input control parameters and output characteristic variables by adopting maximum information coefficients (Maximal Information Coefficient, MIC), firstly carrying out a-column b-row gridding on a scatter diagram formed by the control parameters (K) and the characteristic variables (T), solving the maximum mutual information value, normalizing the maximum mutual information value, and finally selecting the maximum mutual information value under different scales as the MIC value;
where mutual information can be seen as the amount of information contained in one random variable about another, the mutual information I (K, T) of K and T is defined as:
wherein K is the sample sum of the variable K, T is the sample sum of the variable T, P (K, T) is the joint probability between the variable K and T, and P (K) and P (T) are the edge probabilities of K and T respectively;
wherein n is the number of samples;
the MIC value range obtained by the formula (8) is between 0 and 1, the greater the MIC value is, the higher the association degree between two variables is, and after the MIC value of an output signal is obtained, a key output signal index is selected as the input characteristic of the LSTM neural network;
and step A5, carrying out normalization processing on the data set, and accelerating the speed of gradient descent to solve the optimal solution after dimensionless data.
Preferably, the step three adopts an LSTM neural network to perform preliminary prediction according to the initial value and the optimizing range of the control parameter of the doubly fed fan of the processed data set, so that the control parameter approaches to the minimum objective function of the identification model, and after a large amount of debugging work, the structural parameters and the network training parameters of the LSTM neural network are obtained, and the LSTM neural network model is trained and tested; the core cell state c is controlled by a forgetting gate, an input gate and an output gate, wherein: sigma is a sigmoid function, x t Input to the t-th cell, c t Is the cell state of the t th cell, h t The hidden state of the t-th unit, the sum of the hidden states of the t-th unitSumming the vector elements and integrating the vector elements, c t-1 Is the cell state of the t-1 th cell, h t-1 Is t th-hidden state of 1 cell;
the output change of the LSTM neural network output parameter under the influence of the input value can be expressed by the following formula:
wherein f t To forget the gate output value, W i 、W f And W is o Network layer weights for ingress, forget and egress gates, b i 、b f And b o Bias items for input gate, forget gate and output gate, b c Is the network layer bias and,information input to neurons for time t, o t To output the gate output value, ☉ is the hadamard product operator.
Preferably, in the third step, the specific step of training prediction is as follows:
step B1, reasonably selecting and distributing data sets and parameter adjustment, wherein the accuracy of a prediction result is greatly affected, 90% of the data sets are randomly selected as training sets, 10% of the data sets are used as verification sets, and the accuracy of a model is verified;
step B2, setting the iteration training frequency as 500, setting the loss function as MSE, setting the optimizer as Adam, setting the input dimension as 12, setting the output dimension as 6, setting the hidden layer number as 2, setting the hidden layer neuron number as 11 and setting the learning rate as 0.015;
and B3, training and predicting the acquired sequence by adopting an LSTM neural network, and obtaining a parameter prediction initial value and an optimizing range.
Preferably, in the fourth step, the PSO algorithm speed and location update formula is as follows:
wherein V represents the particle update rate, x represents the particle, u represents the current time, u+1 represents the next update time, c 1 、c 2 R is the learning factor 1 、r 2 A random value between 0 and 1, pbest is an individual optimal value, gbest is a global optimal value, and ω is an adaptive inertial weight;
PSO takes the mean square value of the output value of the actual position of each particle and the output value of the optimal position as an objective function of an algorithm to obtain the fitness value of each particle in the current iteration times, and the objective function is expressed as:
in U i,error 、I di,error And I qi,error Is u dc 、i dg And i qg Error of U i 、I di And I qi Is u dc 、i dg And i qg Is the actual measurement data value of U ti 、I dti And I qti Is u dc 、i dg And i qg An output response value obtained by the identification result;
step C1, improving an objective function: according to the self-adaptive weight method, when objective function construction is carried out on multiple target output values, nonlinear dynamic adjustment weight distribution is carried out according to the MIC value and the actual value of each output value, so that the objective function of the multiple output values carries out self-adaptive adjustment on the output values with high association degree or small actual values, and the improvement formula is as follows:
wherein k is U 、k ID And k IQ Is U (U) i 、I di And I qi Adaptive weighting coefficients.
Wherein k is U The rest are as shown in formula (14)The adaptive weighting coefficients for the output values may be as in equation (14):
wherein mic U Is u dc The sum of MIC values of each interval is calculated by updating the pbest and gbest of each particle using an algorithm in an iterative process, and then calculating the objective function F of each sample I Storing the optimal value and the corresponding parameter value;
step C2, improving inertia weight omega: the inertia weight nonlinear decrementing mode is adopted, so that omega is adaptively adjusted along with the iteration times of particles, the global searching capability of the early stage of the particles is ensured, the local optimizing capability of the later stage is improved, and the following formula (15) is improved:
wherein omega is min And omega max E and G are the current and total iteration times for the minimum and maximum omega values;
step C3, improving learning factor C 1 ,c 2 : learning factor c in particle velocity update formula 1 And c 2 The optimization capability of particle individual optimization and population optimization is reflected, the identification precision can be improved, the convergence time can be shortened, the probability of sinking into local optimization can be reduced by a proper learning factor, the learning factor is reduced in a nonlinear manner along with the increase of iteration times in the optimization process, and the formula (16) is improved:
preferably, the fifth step specifically includes the following steps:
step D1, in order to further verify the effectiveness and practicability of the identification of the proposed model, comparing an LSTM-IPSO algorithm with a PSO algorithm and an LSTM-PSO algorithm, training the model by adopting the same data set, and comparing 3 groups of identification curves with actual measurement curves under 20%, 40%, 60% and 80% of voltage drop degree;
step D2, respectively calculating average deviations of 3 model identification results in the intervals (1), (2), (3), (4) and (5) under the voltage drop degree of 20%, 40%, 60% and 80% according to a formula (17), and then averaging the average deviations under 4 working conditions to obtain errors of PSO, LSTM-PSO and LSTM-IPSO model output results in each interval;
wherein F is Vi Is interval V i Average deviation of u M For actually measuring the DC capacitance voltage, u i To identify the model DC capacitor voltage, K start And K end For the first and last dummy data sequence numbers of the interval.
FIG. 6 shows a convergence curve of the mean square error corresponding to the minimum fitness value obtained by optimizing each algorithm, wherein the initial value of the PSO algorithm is larger, the later change is smaller, but the error update speed is obviously slower; the initial value of the LSTM-PSO algorithm is smaller but changes less during the iteration. The LSTM-IPSO algorithm has the advantages of minimum initial value, strongest local searching capability, minimum searched target value in the subsequent extremum searching and optimizing process due to self-adaptive adjustment, and optimization precision and convergence speed.
As shown in fig. 13, 14 and 15, the results u are output for the model of the PSO algorithm, LSTM-PSO algorithm and LSTM-IPSO algorithm dc 、i dg And i qg Average deviations in intervals (1), (2), (3), (4) and (5). Different algorithms for final control parameter K p1 、K i1 、K p2 、K i 、K i3 And K i3 The identification effect of the PSO algorithm is greatly influenced, and the average deviation of the output signal results of the PSO algorithm identification control parameters in the intervals (1), (2), (3), (4) and (5) is larger; adding LSTM algorithm to perform advanced optimization on the objective function, obtaining predictive control parameters and optimizing range, and performing iterative optimization on the objective function by using PSO algorithmThe obtained identification result is more accurate, and the sum of average deviation of the output signal results obtained by identification in the intervals (1), (2), (3), (4) and (5) is smaller than that of a PSO algorithm; after the PSO algorithm is improved, the identification result becomes more accurate, the identification errors of the LSTM-IPSO algorithm in the intervals (1), (2), (3), (4) and (5) are smaller, the average deviation is smaller than 5%, and the feasibility is realized. Through comparative analysis, compared with a PSO algorithm and an LSTM-PSO algorithm, the algorithm provided by the invention has advantages in the aspects of calculation precision, convergence speed and the like, and is suitable for identifying control parameters of a transient model of the fan.
The foregoing embodiments are merely preferred embodiments of the present invention, and should not be construed as limiting the present invention, and the embodiments and features of the embodiments in the present application may be arbitrarily combined with each other without collision. The protection scope of the present invention is defined by the claims, and the protection scope includes equivalent alternatives to the technical features of the claims. I.e., equivalent replacement modifications within the scope of this invention are also within the scope of the invention.
Claims (8)
1. A double-fed fan control parameter identification method based on LSTM-IPSO is characterized in that: it comprises the following steps:
firstly, acquiring ring experiment data of doubly-fed fan hardware from a real controller by using an RT-LAB semi-physical simulation platform, and building an isomorphic identification model of the doubly-fed fan real controller on a Matlab/simulink platform;
step two, increasing the dimension of an input feature set and removing irrelevant features, and selecting a feature value with higher correlation as the input feature set of the neural network model; the input feature set and the corresponding control parameter set form a control parameter-input feature set;
thirdly, training and predicting a control parameter-input characteristic set by using an LSTM neural network to obtain a prediction initial value and an optimizing range;
fourth, based on the prediction initial value and the optimizing range obtained in the fourth step, the IPSO algorithm is used as a secondary optimizing method for accurate identification, and the purpose of accurate optimizing is achieved;
and fifthly, comparing and verifying the identified identification model with hardware-in-loop experimental data, and judging the reliability of the identification model.
2. The method for identifying the control parameters of the doubly-fed wind turbine based on the LSTM-IPSO of claim 1, wherein the method comprises the following steps: in the first step, the real controller-derived doubly-fed fan hardware-in-loop experimental data is obtained by using an RT-LAB semi-physical simulation platform, wherein the real controller-derived doubly-fed fan hardware-in-loop experimental data comprises direct-current bus voltage u dc Dq component i of output current dg And i qg Wherein the d-axis reference current i dg Obtained by DC voltage control, and q-axis reference current i qg Setting to zero, and building an isomorphic identification model of the real controller of the doubly-fed fan on the Matlab/simulink platform.
3. The method for identifying the control parameters of the doubly-fed wind turbine based on the LSTM-IPSO according to claim 2, wherein the method comprises the following steps: the double-fed fan comprises a wind turbine, a gear box, an induction asynchronous generator, a back-to-back converter and a controller thereof;
the control equation related to the isomorphic recognition model of the real controller of the doubly-fed fan is built on the Matlab/simulink platform is as follows:
according to the law of conservation of energy, the KCL equation of the DC capacitor C can be obtained as follows:
wherein P is r For the active power output to C by the machine side converter, P rc Active power injected into a power grid for the grid-side converter;
the dynamic model of the filter reactance is:
wherein u is cd 、u cq And u is equal to sd 、u sq Representing the components of the voltages on the d-axis and q-axis at the outlets of the grid-side converter and the machine-side converter, respectively, X r Is the filter reactance value.
The net side controller control equation can be written as:
wherein u is dcref Is the reference value of direct current voltage, i dgref And i qgref Actual values of net side d-axis and q-axis currents, K p1 Is the proportional coefficient, K of the voltage outer loop PI controller p2 For the q-axis current inner loop proportionality coefficient, K p3 For q-axis current control inner ring proportion coefficient, setting intermediate variables output by PI controller integration link as x 1 、x 2 And x 3 The dynamic equation of the controller is:
wherein K is i1 Integrating coefficient, K for voltage outer loop PI controller i2 For the inner loop integral coefficient of the q-axis current, K i3 Controlling the inner loop integral coefficient for q-axis current, i qgref Typically 0, but under fault conditions i qgref The low pass control module is directly given by:
wherein k is a reactive current support coefficient, U N Rated voltage for grid-connected point of fan, I N Rated current for GSC; when the low voltage crossing occurs, the grid-side converter can generate a certain reactive powerThe value of the reactive current can be obtained according to equation (5), where the q-axis current inner loop is switched to the low-pass control mode and the q-axis reference value is given by the low-pass control module.
4. The method for identifying the control parameters of the doubly-fed wind turbine based on the LSTM-IPSO of claim 1, wherein the method comprises the following steps: the second step specifically comprises the following steps:
step A1, collecting isomorphic recognition model control parameters-input characteristic sets, wherein the input characteristic values are parameters related to a fan network side controller: u (u) dc 、i dg And i qg And the error value F of the hardware-in-loop test data is the control parameter data set of the network-side converter corresponding to the input characteristic value;
a2, eliminating sample data of abnormal data points, so as to avoid affecting identification accuracy;
step A3, dividing the simulation time into 5 intervals according to the modeling guideline of the wind turbine generator electrical simulation model: steady state intervals (1) and (5), low voltage crossing intervals (2), (3) and (4), and inputting i in characteristic values dg And i qg Also by this way the values are taken as feature sets between partitions, where u dc Obtaining a characteristic value V through partition a 、V b 、V c 、V d And V e ,i dg Obtaining a characteristic value I through partition da 、I db 、I dc 、I dd And I qe ,i qg Obtaining a characteristic value I through partition qa 、I qb 、I qc 、I qd And I qe ;
A4, carrying out association degree analysis on input control parameters and output characteristic variables by adopting maximum information coefficients (Maximal Information Coefficient, MIC), firstly carrying out a-column b-row gridding on a scatter diagram formed by the control parameters (K) and the characteristic variables (T), solving the maximum mutual information value, normalizing the maximum mutual information value, and finally selecting the maximum mutual information value under different scales as the MIC value;
where mutual information can be seen as the amount of information contained in one random variable about another, the mutual information I (K, T) of K and T is defined as:
wherein K is the sample sum of the variable K, T is the sample sum of the variable T, P (K, T) is the joint probability between the variable K and T, and P (K) and P (T) are the edge probabilities of K and T respectively;
wherein n is the number of samples;
the MIC value range obtained by the formula (8) is between 0 and 1, the greater the MIC value is, the higher the association degree between two variables is, and after the MIC value of an output signal is obtained, a key output signal index is selected as the input characteristic of the LSTM neural network;
and step A5, carrying out normalization processing on the data set, and accelerating the speed of gradient descent to solve the optimal solution after dimensionless data.
5. The method for identifying the control parameters of the doubly-fed wind turbine based on the LSTM-IPSO of claim 1, wherein the method comprises the following steps: performing preliminary prediction on the initial value and the optimizing range of the control parameters of the doubly fed fan according to the processed data set by adopting an LSTM neural network, enabling the initial value and the optimizing range to approach to the minimum objective function of the identification model, obtaining the structural parameters and the network training parameters of the LSTM neural network after a large amount of debugging work, and training and testing the LSTM neural network model; the core cell state c is controlled by a forgetting gate, an input gate and an output gate, wherein: sigma is a sigmoid function, x t Input to the t-th cell, c t Is the cell state of the t th cell, h t The hidden state of the t-th unit, the sum of the hidden states of the t-th unitSumming vector elements and vector elementProduct, c t-1 Is the cell state of the t-1 th cell, h t-1 A hidden state for the t-1 th cell;
the output change of the LSTM neural network output parameter under the influence of the input value can be expressed by the following formula:
wherein f t To forget the gate output value, W i 、W f And W is o Network layer weights for ingress, forget and egress gates, b i 、b f And b o Bias items for input gate, forget gate and output gate, b c Is the network layer bias and,information input to neurons for time t, o t To output the gate output value, ☉ is the hadamard product operator.
6. The method for identifying the control parameters of the doubly-fed wind turbine based on LSTM-IPSO of claim 5, wherein the method comprises the following steps: in the third step, the specific steps of training and predicting are as follows:
step B1, reasonably selecting and distributing data sets and parameter adjustment, wherein the accuracy of a prediction result is greatly affected, 90% of the data sets are randomly selected as training sets, 10% of the data sets are used as verification sets, and the accuracy of a model is verified;
step B2, setting the iteration training frequency as 500, setting the loss function as MSE, setting the optimizer as Adam, setting the input dimension as 12, setting the output dimension as 6, setting the hidden layer number as 2, setting the hidden layer neuron number as 11 and setting the learning rate as 0.015;
and B3, training and predicting the acquired sequence by adopting an LSTM neural network, and obtaining a parameter prediction initial value and an optimizing range.
7. The method for identifying the control parameters of the doubly-fed wind turbine based on the LSTM-IPSO of claim 1, wherein the method comprises the following steps: in the fourth step, the PSO algorithm speed and position update formula is as follows:
V u+1 =ωV u +c 1 r 1 (pbest u -x u )+c 2 r 2 (gbest u -x u ) (10)
x u+1 =x u +V u+1 (11)
wherein V represents the particle update rate, x represents the particle, u represents the current time, u+1 represents the next update time, c 1 、c 2 R is the learning factor 1 、r 2 A random value between 0 and 1, pbest is an individual optimal value, gbest is a global optimal value, and ω is an adaptive inertial weight;
PSO takes the mean square value of the output value of the actual position of each particle and the output value of the optimal position as an objective function of an algorithm to obtain the fitness value of each particle in the current iteration times, and the objective function is expressed as:
in U i,error 、I di,error And I qi,error Is u dc 、i dg And i qg Error of U i 、I di And I qi Is u dc 、i dg And i qg Is the actual measurement data value of U ti 、I dti And I qti Is u dc 、i dg And i qg An output response value obtained by the identification result;
step C1, improving an objective function: according to the self-adaptive weight method, when objective function construction is carried out on multiple target output values, nonlinear dynamic adjustment weight distribution is carried out according to the MIC value and the actual value of each output value, so that the objective function of the multiple output values carries out self-adaptive adjustment on the output values with high association degree or small actual values, and the improvement formula is as follows:
wherein k is U 、k ID And k IQ Is U (U) i 、I di And I qi Adaptive weighting coefficients.
Wherein k is U As shown in equation (14), the adaptive weighting coefficients for the remaining output values may be as shown in equation (14):
wherein mic U Is u dc The sum of MIC values of each interval is calculated by updating the pbest and gbest of each particle using an algorithm in an iterative process, and then calculating the objective function F of each sample I Storing the optimal value and the corresponding parameter value;
step C2, improving inertia weight omega: the inertia weight nonlinear decrementing mode is adopted, so that omega is adaptively adjusted along with the iteration times of particles, the global searching capability of the early stage of the particles is ensured, the local optimizing capability of the later stage is improved, and the following formula (15) is improved:
wherein omega is min And omega max E and G are the current and total iteration times for the minimum and maximum omega values;
step C3, improving learning factor C 1 ,c 2 : learning factor c in particle velocity update formula 1 And c 2 The optimization capability of particle individual optimization and population optimization is reflected, the identification precision can be improved, the convergence time can be shortened, the probability of sinking into local optimization can be reduced by a proper learning factor, the learning factor is reduced in a nonlinear manner along with the increase of iteration times in the optimization process, and the formula (16) is improved:
8. the method for identifying the control parameters of the doubly-fed wind turbine based on the LSTM-IPSO of claim 1, wherein the method comprises the following steps: the fifth step comprises the following steps:
step D1, in order to further verify the effectiveness and practicability of the identification of the proposed model, comparing an LSTM-IPSO algorithm with a PSO algorithm and an LSTM-PSO algorithm, training the model by adopting the same data set, and comparing 3 groups of identification curves with actual measurement curves under 20%, 40%, 60% and 80% of voltage drop degree;
step D2, respectively calculating average deviations of 3 model identification results in the intervals (1), (2), (3), (4) and (5) under the voltage drop degree of 20%, 40%, 60% and 80% according to a formula (17), and then averaging the average deviations under 4 working conditions to obtain errors of PSO, LSTM-PSO and LSTM-IPSO model output results in each interval;
wherein F is Vi Is interval V i Average deviation of u M For actually measuring the DC capacitance voltage, u i To identify the model DC capacitor voltage, K start And K end For the first and last dummy data sequence numbers of the interval.
Priority Applications (1)
Application Number | Priority Date | Filing Date | Title |
---|---|---|---|
CN202211550049.5A CN116227320A (en) | 2022-12-05 | 2022-12-05 | Double-fed fan control parameter identification method based on LSTM-IPSO |
Applications Claiming Priority (1)
Application Number | Priority Date | Filing Date | Title |
---|---|---|---|
CN202211550049.5A CN116227320A (en) | 2022-12-05 | 2022-12-05 | Double-fed fan control parameter identification method based on LSTM-IPSO |
Publications (1)
Publication Number | Publication Date |
---|---|
CN116227320A true CN116227320A (en) | 2023-06-06 |
Family
ID=86568605
Family Applications (1)
Application Number | Title | Priority Date | Filing Date |
---|---|---|---|
CN202211550049.5A Pending CN116227320A (en) | 2022-12-05 | 2022-12-05 | Double-fed fan control parameter identification method based on LSTM-IPSO |
Country Status (1)
Country | Link |
---|---|
CN (1) | CN116227320A (en) |
Cited By (1)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
CN116796911A (en) * | 2023-08-25 | 2023-09-22 | 国网江苏省电力有限公司淮安供电分公司 | Medium-voltage distribution network optimization regulation and control method and system based on typical scene generation and on-line scene matching |
-
2022
- 2022-12-05 CN CN202211550049.5A patent/CN116227320A/en active Pending
Cited By (1)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
CN116796911A (en) * | 2023-08-25 | 2023-09-22 | 国网江苏省电力有限公司淮安供电分公司 | Medium-voltage distribution network optimization regulation and control method and system based on typical scene generation and on-line scene matching |
Similar Documents
Publication | Publication Date | Title |
---|---|---|
CN102074955B (en) | Method based on knowledge discovery technology for stability assessment and control of electric system | |
CN111523785A (en) | Power system dynamic security assessment method based on generation countermeasure network | |
CN111478314B (en) | Transient stability evaluation method for power system | |
CN111475909B (en) | Wind turbine generator output correlation mapping modeling method based on long-term and short-term memory network | |
CN107590564B (en) | Transient stability constraint-based active power output adjustment method for power system | |
CN112310980B (en) | Safety and stability evaluation method and system for direct-current blocking frequency of alternating-current and direct-current series-parallel power grid | |
CN111523778A (en) | Power grid operation safety assessment method based on particle swarm algorithm and gradient lifting tree | |
CN114006413B (en) | Power system transient stability control method and system based on graph neural network | |
CN110634082A (en) | Low-frequency load shedding system operation stage prediction method based on deep learning | |
CN116227320A (en) | Double-fed fan control parameter identification method based on LSTM-IPSO | |
CN111415010A (en) | Bayesian neural network-based wind turbine generator parameter identification method | |
CN115017787A (en) | Wind power plant voltage ride through characteristic equivalent modeling method and system based on intelligent algorithm | |
CN112821424A (en) | Power system frequency response analysis method based on data-model fusion drive | |
CN111401792A (en) | Dynamic safety assessment method based on extreme gradient lifting decision tree | |
CN115659665A (en) | Voltage sag source signal noise reduction and identification method based on Trojan horse optimization algorithm | |
CN109921462B (en) | New energy consumption capability assessment method and system based on LSTM | |
CN114336632A (en) | Method for correcting alternating current power flow based on model information assisted deep learning | |
CN112821420B (en) | XGboost-based prediction method and system for dynamic damping factor and multidimensional frequency index in ASFR model | |
CN114169395A (en) | Construction method and application of dominant instability mode recognition model of power system | |
CN114036506A (en) | Method for detecting and defending false data injection attack based on LM-BP neural network | |
CN113054653A (en) | Power system transient stability evaluation method based on VGGNet-SVM | |
CN115411775B (en) | Double-fed fan control parameter identification method based on LSTM neural network | |
CN113221465B (en) | Photovoltaic power generation operation and maintenance oriented component model parameter identification method | |
CN111798049B (en) | Voltage stability evaluation method based on integrated learning and multi-target planning | |
CN116702629B (en) | Power system transient stability evaluation method with migration capability |
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 |