CN109240085B - non-Gaussian system dynamic data correction and system control performance optimization method - Google Patents
non-Gaussian system dynamic data correction and system control performance optimization method Download PDFInfo
- Publication number
- CN109240085B CN109240085B CN201811199510.0A CN201811199510A CN109240085B CN 109240085 B CN109240085 B CN 109240085B CN 201811199510 A CN201811199510 A CN 201811199510A CN 109240085 B CN109240085 B CN 109240085B
- Authority
- CN
- China
- Prior art keywords
- gaussian
- data
- function
- model
- probability
- 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
Images
Classifications
-
- G—PHYSICS
- G05—CONTROLLING; REGULATING
- G05B—CONTROL OR REGULATING SYSTEMS IN GENERAL; FUNCTIONAL ELEMENTS OF SUCH SYSTEMS; MONITORING OR TESTING ARRANGEMENTS FOR SUCH SYSTEMS OR ELEMENTS
- G05B13/00—Adaptive control systems, i.e. systems automatically adjusting themselves to have a performance which is optimum according to some preassigned criterion
- G05B13/02—Adaptive control systems, i.e. systems automatically adjusting themselves to have a performance which is optimum according to some preassigned criterion electric
- G05B13/04—Adaptive control systems, i.e. systems automatically adjusting themselves to have a performance which is optimum according to some preassigned criterion electric involving the use of models or simulators
- G05B13/042—Adaptive control systems, i.e. systems automatically adjusting themselves to have a performance which is optimum according to some preassigned criterion electric involving the use of models or simulators in which a parameter or coefficient is automatically adjusted to optimise the performance
Landscapes
- Engineering & Computer Science (AREA)
- Health & Medical Sciences (AREA)
- Artificial Intelligence (AREA)
- Computer Vision & Pattern Recognition (AREA)
- Evolutionary Computation (AREA)
- Medical Informatics (AREA)
- Software Systems (AREA)
- Physics & Mathematics (AREA)
- General Physics & Mathematics (AREA)
- Automation & Control Theory (AREA)
- Complex Calculations (AREA)
Abstract
The invention relates to a method for optimizing system control performance, in particular to a method for correcting dynamic data of a non-Gaussian system and optimizing system control performance, which solves the problems that the data measurement is not close to a real value after being corrected and the optimization effect of a control strategy is not obvious under the condition that the wind power generation process is influenced by non-Gaussian process noise and non-Gaussian measurement noise, and comprises the following steps: firstly, describing a system model under non-Gaussian disturbance; secondly, deducing an iterative formula in the problem description by utilizing an EM algorithm; thirdly, solving the corrected output by using an iterative formulay r (ii) a And fourthly, selecting the performance indexes based on the statistical information to obtain an optimal control law. The advantages are that: 1. considering the influence of non-Gaussian random noise on the system; 2. the dynamic characteristics of the process are considered, and the actual process is better expressed; 3. the calculation efficiency is high, and the requirements of the actual industrial process are met; 4. and (3) fully depicting the random characteristic of the non-Gaussian random quantity by adopting entropy statistical information, and establishing a tracking control index.
Description
Technical Field
The invention relates to a system control performance optimization method, in particular to a non-Gaussian system dynamic data correction and system control performance optimization method.
Background
In practical systems, random disturbances are widespread, especially in industrial processes, such as random factors of wind speed variations in wind power systems, random factors of measurement errors of sensor measurement data.
At present, the correction of the measurement noise of a control system is optimized by controlling the mean value and the variance of corrected data based on the assumption that the measurement noise is gaussian. The measurement error of the actual industrial process is usually non-gaussian, and if the measurement error is corrected by using a method under the traditional gaussian assumption, the optimization control effect is not obvious.
A control strategy can achieve good control results, and it is an important premise that accurate measurement data is obtained in the control process, and for a system containing measurement noise, even if an existing optimal performance index is constructed, the control effect is not reliable. In industrial processes, the measurement data inevitably contains errors (including random errors and gross errors) and the measurement noise is often non-gaussian. Gross errors can be quite large, and both errors can adversely affect process control for systems that require accurate state and measurement estimates. Therefore, in order to ensure that the system maintains a good control effect, more accurate measurement data should be provided, i.e. the raw measurement data should be corrected to be as close to the real data as possible. At present, the correction of the measurement noise of a control system is optimized by controlling the mean value and the variance of corrected data based on the assumption that the measurement noise is gaussian. Such an optimization strategy has the following drawbacks: 1. the control theory for stochastic control systems does not explicitly consider the measurement noise present in the actual sensor device, but even if it is considered, it is discussed on the assumption that it is gaussian distributed; 2. in the conventional data correction method, the dynamic characteristics of the process are not taken into consideration, and therefore, the accessibility of the set value cannot be ensured; 3. in terms of non-gaussian random control, in the existing scientific achievements, the set value is usually taken as a constant value or a preset reference track, but in fact, the set value/reference track needs to be set according to the actual situation, for example, the randomness of the wind speed is a motion track which changes at any time, so that the processing scheme needs to be further considered.
Because the above-mentioned drawbacks and problems in the prior art are difficult to solve, it is necessary to design a method for optimizing the dynamic data correction and system control performance of a non-gaussian system.
Disclosure of Invention
The technical problem solved by the invention is as follows: aiming at the problems that the data measurement is not close to the true value after being corrected and the optimization effect of the control strategy is not obvious under the condition that the wind power generation process is influenced by non-Gaussian process noise (random wind speed) and non-Gaussian measurement noise, the method for correcting the dynamic data of the non-Gaussian system and optimizing the control performance of the system is provided.
The invention is realized by the following operation steps: the method for correcting the dynamic data of the non-Gaussian system and optimizing the control performance of the system comprises the following operation steps:
step one, describing a system model under non-Gaussian disturbance:
in a univariate system, the output at time t can be decomposed into two parts: predictable portionAnd an unpredictable portion (δ (t)):
where ω (t) is the process noise, δ (t) and the measurement noise ε (t) are set to be the Gaussian mixture model, and the probability density function is as follows:
in the formula, g [. C]Is the density of a Gaussian distribution ckAnd fnIs taken as the mean value of the average value,andis the variance;
measurement output ymThe relation to the real output y is:
ym(t)=y(t)+ε(t),
corrected output yrThe posterior distribution of (t) is output by measurement ym(t) and predictable portionGiving out; according to BayesCriterion, yr(t) posterior distribution and ymLikelihood function of (t) andis proportional to the product of the likelihood functions of (a):
wherein, the mean value of each Gaussian component satisfies the following relation:
p(yr(t)) is yr(t), which is a constant, then:
based on the maximum a posteriori distribution principle, yrThe estimated value of (t) can be obtained by the following formula:
i.e. to maximizeAnd L [ y ]m(t)|yr(t)]To obtain the optimum correction signal yr(t) of (d). For measurement data y containing non-Gaussian noisem(t) directly using the maximum likelihood estimation method to find the optimum correction signal yr(t), it is obvious that the problem of too large calculation amount will be faced, so an implicit variable which cannot be observed is defined firstly, a probability model depending on the implicit variable is established, and parameter maximum likelihood estimation is searched in the probability model, namely, the EM algorithm is adopted to finish correcting the signal yr(t) solving;
step two, deducing an iterative formula in the problem description by utilizing an EM algorithm:
(1) and (3) defining hidden variables, writing a log-likelihood function of the complete data:
imagine observation data ym(ym1,ym2,…,ymJ) Is generated as follows: first according to the probability akSelecting the kth Gaussian distribution partial model g (y)m|θk) (ii) a Then according to the probability distribution g (y) of the k-th partial modelm|θk) Generating data ym(ii) a At this time, observation data ymjJ ═ 1,2, …, J, known; observation data of reaction ymjData from the K-th partial model is unknown, K1, 2jkWhich is defined as follows:
obtaining observation data in the same wayCorresponding hidden variableN ═ 1,2,., N is defined as follows:
then, the log-likelihood function of the full data is:
(2) e, determining a Q function in the EM algorithm;
and (3) expecting the log-likelihood function of the complete data to obtain a Q function:
The same principle is that:
is to observe data y under the current model parametersmjThe probability from the kth partial model becomes the partial model k for the observation data ymjThe responsivity of (a);
is to observe data under the current model parametersThe probability from the nth partial model is the n pairs of observation data of the partial modelThe responsivity of (a);
according to the above calculation, the method can obtain
(3) M steps of the EM algorithm: maximizing the Q function to obtain an iterative formula;
the M steps of iteration are maximum value of Q function to parameter, namely model parameter of new iteration is solved:
by usingDenotes theta(i+1)Each parameter of (a); to findOnly the Q function needs to be respectively paired with lambdak,ηk,μn,ρnObtaining a deviation derivative and making the deviation derivative be 0; to findIs at Σ ak=1,∑bnThe partial derivative is calculated and made 0 under the condition of 1, and the result is as follows:
obtaining an iterative formula for solving each parameter; repeating the above calculations until the log-likelihood function values no longer change significantly;
step three, solving and correcting the output y by using the iterative formular:
The gaussian components have the following relationship:
solving the equation set can obtain the optimal corrected output yr;
Selecting performance indexes based on statistical information, which specifically comprises the following steps:
because the process is affected by non-Gaussian random interference, more general statistical information except the mean and the variance is adopted for measuring the randomness; in order to characterize the tracking performance and the interference suppression performance, the average value of the squared tracking error and the entropy of the tracking error are minimized; furthermore, the control energy is also minimized; thus, the controller is obtained by minimizing the following performance indicators:
here, a general measure of uncertainty, i.e., entropy, is used instead of variance in a gaussian system; considering the calculation efficiency, the selected entropy measure is quadratic arbitrary entropyAs is known from the definition of Information Potential (IP),the quadratic arbitrary entropy is a monotonic function of the quadratic information potential, so the maximization of the quadratic arbitrary entropy is equivalent to the minimization of the information potential, and a random gradient method is adopted, namely, the optimal control method is realized.
Compared with the prior art, the invention has the following advantages:
(1) because the actual industrial process is inevitably influenced by random noise and is generally non-Gaussian random noise, the influence caused by the random noise is considered when the distribution of the measurement noise and the process noise is determined, and the method has more generality and practical significance than the method in which the influence of the noise is ignored or the noise is assumed to obey Gaussian distribution in the prior art;
(2) the invention fully considers the dynamic characteristics of the process and better expresses the actual process;
(3) on the premise of reasonable initial parameter selection, the method has higher calculation efficiency and meets the requirements of the actual industrial process;
(4) the invention adopts entropy statistical information to fully depict the random characteristic of non-Gaussian random quantity and establishes a tracking control index.
Drawings
FIG. 1 is a flow chart of the single step algorithm of the present invention.
FIG. 2 is a graph comparing the calibration output of the method of the present invention with that of the prior art. The horizontal axis is time, the vertical axis is output response of the non-Gaussian system, and the three tracks respectively represent a set value, output corrected by the existing Gaussian method and output corrected by the non-Gaussian method in the invention. It is demonstrated by the figure that the system output after the method is adopted is obviously superior to the existing method in accuracy and rapidity.
Detailed Description
The process of the invention is described below with reference to specific examples: for a typical wind energy conversion system, the equivalent load of a permanent magnet synchronous generator is made up of a constant inductance LsAnd a variable resistor RsAre connected in parallel. Under the application example, the method for correcting the dynamic data of the non-Gaussian system and optimizing the control performance of the system comprises the following operation steps:
step one, the permanent magnet synchronous generator model can be described as follows:
wherein R is stator resistance, udAnd uqVoltages of d-and q-components of the stator, LdAnd LqInductances of the stator, i, being d-and q-components, respectivelydAnd iqCurrents, phi, of d-and q-components of the stator, respectivelymIs the magnetic flux (which is a constant), p is the number of pole pair pairs, ΩGIs the rotational speed of the generator, JhIs the total inertia of the turbine, called the high speed shaft, RtIs the turbine radius, i is the gear ratio, u-RsIs a control input, y ═ ΩGIs the system output;
the state space model of the permanent magnet synchronous generator based on the wind power generation system is established as follows:
wherein:
in order to realize real-time control, a state space model of the system is discretized:
wherein x isk∈R3×1Is a state vector, and all variables in F (-) and G (-) are continuous, bounded, and differentiable to the first order, ωkIs a non-gaussian perturbation of random wind speed;
step two, obtaining the measurement output y after the influence of the measured noise in the control processmThe steps of correcting it are as follows:
for measurement data y containing non-Gaussian noisemDirectly using maximum likelihood estimation to obtain its correction signal yrIt is obvious that the problem of too large calculation amount is faced, so that firstly, defining an invisible variable which cannot be observed, establishing a probability model depending on the invisible variable, and searching parameter maximum likelihood estimation in the probability model, namely, adopting EM algorithm to complete ymAnd solving the correction signal.
(2) E, step E: calculating the responsivity according to the current model parameters;
(3) and M: the model parameters for a new iteration are calculated using the following iteration formula:
(4) repeating the step (2) and the step (3) until convergence;
A system of linear equations can be obtained:
solving the equation set to obtain corrected output yrIs afterContinuous system dynamic optimization and control provide measurement data closer to the true value;
selecting performance indexes based on statistical information, which specifically comprises the following steps:
because the process is affected by non-Gaussian random interference, more general statistical information except the mean and the variance is adopted for measuring the randomness; in order to characterize the tracking performance and the interference suppression performance, the average value of the squared tracking error and the entropy of the tracking error are minimized; furthermore, the control energy is also minimized; thus, the controller is obtained by minimizing the following performance indicators:
here, a general measure of uncertainty, i.e., entropy, is used instead of variance in a gaussian system; considering the calculation efficiency, the selected entropy measure is quadratic arbitrary entropyAs is known from the definition of Information Potential (IP),the quadratic arbitrary entropy is a monotonic function of the quadratic information potential, so the maximization of the quadratic arbitrary entropy is equivalent to the minimization of the information potential, and a random gradient method is adopted, namely, the optimal control method is realized.
Claims (1)
1. A non-Gaussian system dynamic data correction and system control performance optimization method is characterized in that: the method comprises the following operation steps:
step one, describing a system model under non-Gaussian disturbance:
in a univariate system, the output at time t can be decomposed into two parts: predictable portionAnd an unpredictable portion (δ (t)):
where ω (t) is the process noise, δ (t) and the measurement noise ε (t) are set to be the Gaussian mixture model, and the probability density function is as follows:
in the formula, g [. C]Is the density of a Gaussian distribution ckAnd fnIs taken as the mean value of the average value,andis the variance;
measurement output ymThe relation to the real output y is:
ym(t)=y(t)+ε(t),
corrected output yrThe posterior distribution of (t) is output by measurement ym(t) and predictable portionGiving out; according to Bayesian criterion, yr(t) posterior distribution and ymLikelihood function of (t) andis proportional to the product of the likelihood functions of (a):
wherein, the mean value of each Gaussian component satisfies the following relation:
p(yr(t)) is yr(t), which is a constant, then:
based on the maximum a posteriori distribution principle, yrThe estimated value of (t) can be obtained by the following formula:
i.e. to maximizeAnd L [ y ]m(t)|yr(t)]To obtain the optimum correction signal yr(t) of (d). For measurement data y containing non-Gaussian noisem(t) finding the optimum by directly using the maximum likelihood estimation methodCorrection signal y ofr(t), it is obvious that the problem of too large calculation amount will be faced, so an implicit variable which cannot be observed is defined firstly, a probability model depending on the implicit variable is established, and parameter maximum likelihood estimation is searched in the probability model, namely, the EM algorithm is adopted to finish correcting the signal yr(t) solving;
step two, deducing an iterative formula in the problem description by utilizing an EM algorithm:
(1) and (3) defining hidden variables, writing a log-likelihood function of the complete data:
imagine observation data ym(ym1,ym2,…,ymJ) Is generated as follows: first according to the probability akSelecting the kth Gaussian distribution partial model g (y)m|θk) (ii) a Then according to the probability distribution g (y) of the k-th partial modelm|θk) Generating data ym(ii) a At this time, observation data ymjJ ═ 1,2, …, J, known; observation data of reaction ymjData from the K-th partial model is unknown, K1, 2jkWhich is defined as follows:
then, the log-likelihood function of the full data is:
(2) e, determining a Q function in the EM algorithm;
and (3) expecting the log-likelihood function of the complete data to obtain a Q function:
The same principle is that:
is to observe data y under the current model parametersmjThe probability from the kth partial model becomes the partial model k for the observation data ymjThe responsivity of (a);
is to observe data under the current model parametersThe probability from the nth partial model is the n pairs of observation data of the partial modelThe responsivity of (a);
according to the above calculation, the method can obtain
(3) M steps of the EM algorithm: maximizing the Q function to obtain an iterative formula;
the M steps of iteration are maximum value of Q function to parameter, namely model parameter of new iteration is solved:
by usingDenotes theta(i+1)Each parameter of (a); to findOnly the Q function needs to be respectively paired with lambdak,ηk,μn,ρnObtaining a deviation derivative and making the deviation derivative be 0; to findIs at Σ ak=1,∑bnThe partial derivative is calculated and made 0 under the condition of 1, and the result is as follows:
obtaining an iterative formula for solving each parameter; repeating the above calculations until the log-likelihood function values no longer change significantly;
step three, solving and correcting the output y by using the iterative formular:
The gaussian components have the following relationship:
solving the equation set to obtain the optimal corrected equationOutput yr;
Selecting performance indexes based on statistical information, which specifically comprises the following steps:
because the process is affected by non-Gaussian random interference, more general statistical information except the mean and the variance is adopted for measuring the randomness; in order to characterize the tracking performance and the interference suppression performance, the average value of the squared tracking error and the entropy of the tracking error are minimized; furthermore, the control energy is also minimized; thus, the controller is obtained by minimizing the following performance indicators:
here, a general measure of uncertainty, i.e., entropy, is used instead of variance in a gaussian system; considering the calculation efficiency, the selected entropy measure is quadratic arbitrary entropyAs is known from the definition of Information Potential (IP),the quadratic arbitrary entropy is a monotonic function of the quadratic information potential, so the maximization of the quadratic arbitrary entropy is equivalent to the minimization of the information potential, and a random gradient method is adopted, namely, the optimal control method is realized.
Priority Applications (1)
Application Number | Priority Date | Filing Date | Title |
---|---|---|---|
CN201811199510.0A CN109240085B (en) | 2018-10-15 | 2018-10-15 | non-Gaussian system dynamic data correction and system control performance optimization method |
Applications Claiming Priority (1)
Application Number | Priority Date | Filing Date | Title |
---|---|---|---|
CN201811199510.0A CN109240085B (en) | 2018-10-15 | 2018-10-15 | non-Gaussian system dynamic data correction and system control performance optimization method |
Publications (2)
Publication Number | Publication Date |
---|---|
CN109240085A CN109240085A (en) | 2019-01-18 |
CN109240085B true CN109240085B (en) | 2021-07-27 |
Family
ID=65053529
Family Applications (1)
Application Number | Title | Priority Date | Filing Date |
---|---|---|---|
CN201811199510.0A Active CN109240085B (en) | 2018-10-15 | 2018-10-15 | non-Gaussian system dynamic data correction and system control performance optimization method |
Country Status (1)
Country | Link |
---|---|
CN (1) | CN109240085B (en) |
Families Citing this family (2)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
CN110703599B (en) * | 2019-09-17 | 2022-06-07 | 太原理工大学 | Organic Rankine cycle system control performance optimization method based on dynamic data correction |
CN111812984B (en) * | 2020-07-20 | 2022-06-03 | 温州大学 | Model-based robust filtering method for inverter control system |
Citations (3)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
CN102360418A (en) * | 2011-09-29 | 2012-02-22 | 山东大学 | Method for detecting eyelashes based on Gaussian mixture model and maximum expected value algorithm |
CN104021289A (en) * | 2014-06-04 | 2014-09-03 | 山西大学 | Non-Gaussian unsteady-state noise modeling method |
CN106683122A (en) * | 2016-12-16 | 2017-05-17 | 华南理工大学 | Particle filtering method based on Gaussian mixture model and variational Bayes |
Family Cites Families (1)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
US9390065B2 (en) * | 2012-07-23 | 2016-07-12 | University Of Southern California | Iterative estimation of system parameters using noise-like perturbations |
-
2018
- 2018-10-15 CN CN201811199510.0A patent/CN109240085B/en active Active
Patent Citations (3)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
CN102360418A (en) * | 2011-09-29 | 2012-02-22 | 山东大学 | Method for detecting eyelashes based on Gaussian mixture model and maximum expected value algorithm |
CN104021289A (en) * | 2014-06-04 | 2014-09-03 | 山西大学 | Non-Gaussian unsteady-state noise modeling method |
CN106683122A (en) * | 2016-12-16 | 2017-05-17 | 华南理工大学 | Particle filtering method based on Gaussian mixture model and variational Bayes |
Non-Patent Citations (2)
Title |
---|
非高斯噪声理论模型分析;祁雷等;《辽宁工学院学报》;19961225;第16卷(第4期);15-16 * |
非高斯系统的控制及滤波方法研究;任密蜂;《中国博士学位论文全文数据库 信息科技辑》;20141215(第12期);14-19 * |
Also Published As
Publication number | Publication date |
---|---|
CN109240085A (en) | 2019-01-18 |
Similar Documents
Publication | Publication Date | Title |
---|---|---|
CN108134549A (en) | A kind of method for improving permanent magnet synchronous motor speed estimate stability | |
CN102779238B (en) | Brushless DC (Direct Current) motor system identification method on basis of adaptive Kalman filter | |
CN110739893B (en) | Improved self-adaptive trackless Kalman filtering rotational inertia identification method | |
CN107742885B (en) | Power distribution network voltage power sensitivity estimation method based on regular matching pursuit | |
CN109240085B (en) | non-Gaussian system dynamic data correction and system control performance optimization method | |
CN109921707B (en) | Switched reluctance hub motor position-free prediction control method | |
CN107658881A (en) | Voltage stability critical point determination methods based on Thevenin's equivalence method | |
CN113285481B (en) | Grid-connected converter inductance parameter online estimation method, prediction control method and system | |
CN108390597A (en) | Permanent magnet synchronous motor nonlinear predictive controller design with disturbance observer | |
CN110888365A (en) | Asynchronous sampling fundamental wave data synchronization method for power grid system | |
CN111208425B (en) | Method for constructing high-precision asynchronous motor system state model and asynchronous motor state detection method | |
CN112054731B (en) | Permanent magnet synchronous motor parameter identification method based on model predictive control | |
CN106208859B (en) | Permanent magnet synchronous motor control method based on interference observer and repetitive controller | |
Alsofyani et al. | Using NSGA II multiobjective genetic algorithm for EKF-based estimation of speed and electrical torque in AC induction machines | |
CN111835251B (en) | Permanent magnet synchronous motor high-performance control method based on speed-free sensing | |
Zhou et al. | A predictive functional control algorithm for multivariable systems with time delay | |
CN115459335A (en) | Inverter model prediction control method for improving stability of direct-current micro-grid | |
CN113641193B (en) | Accurate tracking control method for non-minimum phase system | |
CN113949265B (en) | Self-adaptive backstepping control method for Buck type converter with uncertain parameters | |
Zhu et al. | State estimation based on improved cubature Kalman filter algorithm | |
CN104881077A (en) | Tracking control method of maximum power point in photovoltaic system | |
CN113553538A (en) | Recursive correction hybrid linear state estimation method | |
CN113850425A (en) | Power distribution network state estimation method based on improved generalized maximum likelihood estimation | |
Yang et al. | State and parameter estimation algorithm for state space model based on linear neural network and Kalman filter | |
Chang et al. | Integrated thrust ripple identification and compensation for linear servo system using an MP 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 |