CN103457274A - STATCOM current control method of multi-model fuzzy neural network PI controllers - Google Patents

Abstract

The invention relates to a STATCOM current control method of multi-model fuzzy neural network PI controllers. The STATCOM current control method of the multi-model fuzzy neural network PI controllers comprises the steps that S1 a system is divided into a plurality of models Mi (i=1, 2, ..., n) with a load power factor serving as the basis of model dividing, S2 d-axis second-stage fuzzy neural network PI controllers PIdi (i=1, 2, ..., n) and q-axis fuzzy neural network PI controllers PIdi (i=1, 2, ..., n) are respectively designed for each model, and S3 when load current is connected in, corresponding models are selected, parameters kP and ki of the d-axis second-stage fuzzy neural network PI controllers and the q-axis fuzzy neural network PI controllers in each model are set through the fuzzy neural network to achieve the ideal control effect. The STATCOM current control method of the multi-model fuzzy neural network PI controllers can rapidly adapt to the changes of a load and achieve high accuracy.

Description

The STATCOM current control method of multi-model fuzzy neural network PI controller

Technical field

The present invention relates to the static synchroballistic method in the electric power quality reactive power compensation, particularly relate to a kind of static synchroballistic current control method based on multi-model fuzzy neural network PI controller.

Background technology

Utilize STATCOM (STATCOM, Static Synchronous Compensator) to improve the quality of power supply and mainly contain two purposes: improve power factor and regulating system voltage.Yet in some electricity consumption occasions, the variation of load does not cause system voltage to occur obviously to reduce, but larger variation has occurred the power factor of system, therefore the compensation of power factor is just seemed to particularly important.The major control target of STATCOM is to improve the power factor of system by the compensating load reactive power.

The major control method of STATCOM is that two closed loop PI control, although controlling, Direct Current PI there is response speed and reactive power compensation precision preferably, but, when larger variation occurs in load, this control method is difficult to adapt to the variation of load, can cause the reduction of compensation precision and speed.The intelligent control method of other some based on PI mainly contains neuron adaptive PI control, the control of the PI based on genetic algorithm and particle group optimizing PI control etc.But the tuning method of Fuzzy PI Controller need to obtain by field adjustable and expertise, and neuron adaptive PI controls needs the time of growing to train neuron.Although the PI based on genetic algorithm controls and can obtain the PI parameter, to restrain slowlyer, running time is longer.And particle group optimizing PI control can obtain optimal solution fast, but easily be absorbed in local optimum.Said method only can overcome the nonlinear characteristic of STATCOM, when the impact load changes, while causing power factor change, adopt said method, can not adapt to faster the variation of load and obtain higher precision, thereby having affected the compensation precision installed.

Summary of the invention

The object of the present invention is to provide a kind of STATCOM current control method of multi-model fuzzy neural network PI controller, variation in load does not cause system voltage to occur obviously to reduce, but when larger variation has occurred in the power factor of system, make the PI controller can adapt to the variation of the access point power-factor of load, realize the effective compensation to power factor.

To achieve these goals, the STATCOM current control method that the invention provides a kind of multi-model fuzzy neural network PI controller comprises the following steps:

S1: access the size of the power factor of load current according to a distribution system load-side, will be divided into n model M in fuzzy neural network PI controller _i(i=1,2 ..., n);

S2: for a described n model, design respectively d axle second level PI controller PI _di(i=1,2 ..., n) with q axle PI controller PI _qi(i=1,2 ..., n);

S3: access a load current to described distribution system load-side, and select a described n model M _i(i=1,2 ..., n) in corresponding model, and by d axle second level PI controller PI in the selected model of d axle fuzzy neural network module _di(i=1,2 ..., control parameter K n) _pd2and K _id2, by q axle PI controller PI in the selected model of d axle fuzzy neural network module _qi(i=1,2 ..., control parameter K n) _pqand K _iq; Described d axle fuzzy neural network module comprises d axle fuzzy controller and d axle neural net; Described q axle fuzzy neural network module comprises q axle fuzzy controller and q axle neural net;

Wherein, n is positive integer.

As preferably, the value of described n is 3.

As preferably, described d axle fuzzy neural network module and described q axle fuzzy neural network module all adopt the BP neural net.

As preferably, step S3 further comprises:

S31: by given d axle real component reference current I _crefwith static compensator circuit d axle output current I _cddifference e _icand difference e (k) _ic(k) derivative as the input of described d axle fuzzy controller, by described d axle fuzzy controller, exported accordingly; By given d axle real component reference current I _fqwith static compensator circuit q axle output current I _cqdifference e _iqand difference e (k) _iq(k) derivative as the input of described q axle fuzzy neural network controller, by described d axle fuzzy controller, exported accordingly;

S32: the input data using the output in step S31 as described BP neural net are trained respectively described d axle neural net and described q axle neural net, described d axle neural net and described q axle neural net are in full accord, it includes input layer, hidden layer and output layer, wherein input layer has 3 neurons, being input as of described input layer i=1,2,3; Wherein f is described input layer function, and k is input variable; The output of described input layer with the input equate, that is:

S33: the output variable weighting that step S32 is obtained is as the input of the hidden layer of described BP neural net, and the output of acquisition hidden layer, described hidden layer contains 5 neurons, and described input layer is ω ji to the weights of described hidden layer, being input as of wherein said hidden layer: ${\mathrm{input}}_j^{(2-2)}\left(k\right)=\underset{i=1}{\overset{3}{\mathrm{\Σ}}}{\mathrm{\ω}}_{\mathrm{ji}}\mathrm{outpu}{t}_i^{(2-1)}\left(k\right),i=\mathrm{1,2,3};j=\mathrm{1,2}\·\·\·5$

Described hidden layer is output as: ${\mathrm{output}}_j^{(2-2)}\left(k\right)=f\left[{\mathrm{intput}}_j^{(2-1)}\left(k\right)\right],j=\mathrm{1,2}\·\·\·5$

The excitation function of described hidden layer neuron adopts the Sigmoid function of Symmetrical

S34: the output variable weighting that step S33 is obtained is as the input of the output layer of BP neural net, and the output of acquisition output layer, and described output layer has 2 neurons, and hidden layer is ω to the weights of output layer _lj

Being input as of output layer wherein: ${\mathrm{input}}_l^{(2-3)}\left(k\right)=\underset{j=1}{\overset{5}{\mathrm{\Σ}}}{\mathrm{\ω}}_{\mathrm{lj}}\mathrm{outpu}{t}_j^{(2-2)}\left(k\right),l=\mathrm{1,2};j=\mathrm{1,2}\·\·\·5$

Output layer is output as: ${\mathrm{output}}_l^{(2-3)}\left(k\right)=g\left[{\mathrm{intput}}_l^{(2-3)}\left(k\right)\right],l=\mathrm{1,2}$

Wherein because parameter kp, the ki of output layer output are non-negative, therefore excitation function is got non-negative Sigmoid function: $g\left(x\right)=\frac{e^x}{e^x+e^{-x}}$

S35: the parameter K using step S34 output variable kp and ki as the PI controller of corresponding neural net _pqand K _iq;

Here the performance index function that fuzzy neural network is chosen is:

Negative gradient direction according to J is adjusted, an additional minimum inertia coeffeicent of the overall situation that makes search energy Fast Convergent, that is: ${\mathrm{\Δ\ω}}_{\mathrm{lj}}(k+1)=-\mathrm{\η}\frac{\∂J}{{\∂\mathrm{\ω}}_{\mathrm{lj}}}+{\mathrm{\α\ω}}_{\mathrm{lj}}\left(k\right),l=\mathrm{1,2};j=\mathrm{1,2}\·\·\·5$

Wherein η is learning efficiency, and α is inertia coeffeicent;

Hidden layer in the BP neural net to the weights ω lj of output layer is: ${\mathrm{\Δ\ω}}_{\mathrm{lj}}(k+1)=-\mathrm{\η\δoutpu}{t}_j^{(2-2)}\left(k\right)+{\mathrm{\α\ω}}_{\mathrm{lj}}\left(k\right),l=\mathrm{1,2};j=\mathrm{1,2}\·\·\·5$

Wherein, $\mathrm{\δ}=e(k+1)\mathrm{sgn}\left[\frac{\∂y(k+1)}{\∂\mathrm{\Δu}\left(k\right)}\right]\frac{\∂\mathrm{\Δu}\left(k\right)}{\∂{\mathrm{output}}_l^{(2-3)}\left(k\right)}{g}^{\′}\left[\mathrm{intpu}{t}_l^{(2-3)}\left(k\right)\right],l=\mathrm{1,2}$

Wherein, g'(x)=g (x) [1-g (x)].

The present invention, owing to adopting above technical scheme, makes it compared with prior art, has following advantage and good effect:

1) the STATCOM current control method of a kind of multi-model fuzzy neural network PI controller provided by the invention, variation in load does not cause system voltage to occur obviously to reduce, but when larger variation has occurred in the power factor of system, use Direct Current PI control method to make the PI controller can adapt to the variation of the access point power-factor of load, the effective compensation of realization to power factor, can make the power factor of system maintain 1 left and right, thereby obtain satisfied control effect.

2) the STATCOM current control method of a kind of multi-model fuzzy neural network PI controller provided by the invention, but when larger variation occurs in load, the multi-model theory is applied in Direct Current PI control to the reduction that to avoid Direct Current PI control method to be difficult to adapt to the variation of load, can cause compensation precision and speed.

3) the STATCOM current control method of a kind of multi-model fuzzy neural network PI controller provided by the invention, the d axle second level PI controller in each model and q axle PI controller parameter k _pand k _iadjusted the control effect that reaches desirable by fuzzy neural network, can adapt to faster the variation of load and obtain higher compensation precision.

The accompanying drawing explanation

Fig. 1 is the one phase equivalent circuit topological diagram of STATCOM access distribution system;

One phase equivalent circuit when Fig. 2 is distribution system access STATCOM;

Fig. 3 is the structure of fuzzy neural network schematic diagram;

Fig. 4 is the Direct Current control block diagram of p axle or q axle fuzzy neural network controller;

Fig. 5 a is that the power-factor of load is 0.67 o'clock compensation effect figure;

Fig. 5 b is that the power-factor of load is 0.38 o'clock compensation effect figure;

Fig. 5 c is that the power-factor of load is 0.17 o'clock compensation effect figure.

Embodiment

Further illustrate the present invention with specific embodiment with reference to the accompanying drawings.

Referring to the accompanying drawing that the embodiment of the present invention is shown, hereinafter the present invention will be described in more detail.Yet the present invention can be with many multi-form realizations, and should not be construed as the restriction of the embodiment be subject in this proposition.On the contrary, it is abundant and complete open in order to reach proposing these embodiment, and makes those skilled in the art understand scope of the present invention fully.

S1: enter the size of the power factor of load current according to distribution system 1 load 2 side joints, will in fuzzy neural network PI controller, arrange and be divided into n model M _i(i=1,2 ..., n); In the present embodiment, fuzzy neural network PI controller has been divided into to three model M _i(i=1,2,3).

Have response speed and reactive power compensation precision preferably although Direct Current PI controls, when larger variation occurs in load, this control method is difficult to adapt to the variation of load, can cause the reduction of compensation precision and speed.Therefore, the multi-model theory is applied in Direct Current PI control herein.

S2: for n model, design respectively d axle second level PI controller PI _di(i=1,2,3) and q axle PI controller PI _qi(i=1,2,3);

As table 1:

The division of table 1 model

The equivalent topologies structure chart of STATCOM connecting system as shown in Figure 1, wherein U _sfor the equivalent potential of Infinite bus system, R _s+ jX _sserving as reasons load look closely into Infinite bus system equivalence Dai Weinan impedance, r+jx is the equiva lent impedance of STATCOM, U _pccfor the voltage at STATCOM access point place, the supply power voltage of namely loading.

In some electricity consumption occasions, the variation of load is mainly reflected in the variation of power factor.When the load of access impact, one phase equivalent circuit as shown in Figure 2.In the present embodiment model, the impedance of initialization system side is much smaller than the impedance of load-side, that is:

|Z _S|<<|Z _L| (1)

Wherein $\frac{1}{Z_L}=\frac{1}{{\mathrm{<figure-callout\; id="0"\; label="Z\; L"\; filenames="2013103745558100003DEST\_PATH\_IMAGE001.png,2013103745558100003DEST\_PATH\_IMAGE002.png"\; state="\{\{state\}\}">Z</figure-callout>}}_L}+\frac{1}{{\mathrm{<figure-callout\; id="1"\; label="Z\; L"\; filenames="HDA0000371503760000011.png"\; state="\{\{state\}\}">Z</figure-callout>}}_L}+\frac{1}{Z_{L2}}+\&CenterDot;\&CenterDot;\&CenterDot;\&CenterDot;\&CenterDot;\&CenterDot;+\frac{1}{Z_{\mathrm{Li}}}$

When accessing, STATCOM can not be expressed as:

$U_{\mathrm{pcc}}=\frac{U_s{Z}_L}{Z_L+Z_S}---\left(2\right)$

Therefore, the voltage U at STATCOM access point place _pcccan be similar to and think and equate with system voltage, and not change with the variation of access load, that is:

U _pcc≈U _s (3)

Therefore, the offset current of STATCOM input system can be meaned by following formula:

$I_c=\frac{U_c-U_s}{Z_f}-\frac{U_s}{Z_C}---\left(4\right)$

When larger variations occurs the power factor of the load of system, STATCOM need to absorb or send larger reactive current, the reactive current of ability compensating load side.This just needs STATCOM output current I _ccan change in the larger context.Due to the impedance of the system side impedance much smaller than load-side, so the STATCOM output voltage U _cphase approximation equal the system side voltage U _sphase place.According to formula (4), can draw, work as U _cduring increase, I _cincrease, work as U _cwhile reducing, I _creduce.Therefore, STATCOM output voltage U _cvariation determined the reactive current I that STATCOM is sent _cvariation.

As the above analysis, for adapting to the variation needs STATCOM output voltage U of load _cthe amplitude size have a greater change scope.U _ccan mean by following formula.

$\left|U_c\right|=\sqrt{{U_d}^2+{U_q}^2}---\left(5\right)$

Therefore known, the STATCOM output voltage U _camplitude size and active voltage signal U _dwith reactive voltage signal U _qsize relevant.And U _dand U _qby d axle second level PI controller and q axle PI controller, obtained respectively.As can be seen here, regulate d axle second level PI controller PI _di(i=1,2 ..., control parameter K n) _pd2and K _id2with q axle PI controller PI _qi(i=1,2 ..., control parameter K n) _pqand K _iq, can play and regulate active voltage signal U _dwith reactive voltage signal U _qeffect, regulate the STATCOM output voltage U thereby reach _cpurpose, and output voltage U _cdetermined that STATCOM sends the size of reactive current.When electric current one timing of system load, the variation of load-side reactive current can be presented as the variation of the power-factor of load, i.e., when load reactive current is larger, power factor is less, and vice versa.Therefore, when larger variation occurs power factor, can adopt different d axle second level PI controller parameter K _pd2, K _id2with q axle PI controller parameter K _pq, K _iq, make the reactive current that STATCOM is sent follow the tracks of the required reactive current of load.When each component parameters of system is determined, just can carry out partitioning model according to the size of load-side power factor, as shown in Figure 3, determine corresponding d axle second level PI controller parameter and q axle PI controller parameter for different models, controller parameter is adjusted and is obtained by the fuzzy neural network module, and fuzzy neural network controller as shown in Figure 4.

Known according to above-mentioned analysis, along with the variation of load, U _pcccan think approximate constant, and the power factor change of load is larger, foundation that therefore can be using the load-side power factor as model partition, be a plurality of model M by system divides _i(i=1,2 ..., n), then for each model, design respectively d axle second level fuzzy neural network PI controller PI _di(i=1,2 ..., n) with q axle fuzzy neural network PI controller PI _qi(i=1,2 ..., n).Adopt preferred n=3 in the present embodiment.Direct Current control block diagram based on multi-model neural network PI controller as shown in Figure 3.After impact load access, multi-model controller 31 detects load current I _fabc, calculate power factor Q by formula (6) after the adc/dq0 conversion, and select corresponding model, the d axle second level PI controller PI in each model _di(i=1,2 ..., control parameter K n) _pd2and K _id2with q axle PI controller PI _qi(i=1,2 ..., control parameter K n) _pqand K _iqadjusted the control effect that reaches desirable by fuzzy neural network, hereinafter in S3, fuzzy neural network be have been described in detail.

$Q=\frac{I_{\mathrm{Fd}}}{\sqrt{{I_{\mathrm{Fd}}}^2+{I_{\mathrm{Fq}}}^2}}---\left(6\right)$

S3: access a load current to the distribution system load-side, and preference pattern M _i(i=1,2 ..., n) in corresponding model, and by d axle second level PI controller PI in d axle fuzzy neural network controller 32 cover half types _di(i=1,2 ..., control parameter K n) _pd2and K _id2, by q axle PI controller PI in the selected model of d axle fuzzy neural network controller 33 _qi(i=1,2 ..., control parameter K n) _pqand K _iq;

D axle fuzzy neural network controller 32 comprises that d axle fuzzy controller and d axle neural net and q axle fuzzy neural network controller 33 comprise q axle fuzzy controller and q axle neural net.Wherein d axle fuzzy neural network controller 32 and q axle fuzzy neural network controller 33 all adopt preferred BP neural net.

In step S3, also comprise:

S31: by given d axle real component reference current I _crefwith static compensator circuit d axle output current I _cddifference e _ic, and e (k) _ic(k) derivative as the input of d axle fuzzy controller, by d axle fuzzy controller, exported accordingly; By given d axle real component reference current I _fqwith static compensator circuit q axle output current I _cqdifference e _iq, and e (k) _iq(k) derivative

as the input of q axle fuzzy neural network controller, by d axle fuzzy controller, exported accordingly;

Wherein, d axle real component reference current I _creffor balance DC side real component I _dcrefwith load current side real component I _dcand; Circuit d axle output current I _cdfor device output current d axle component, I _cqfor STATCOM output current I _cq axle component, by I _cthrough abc/dq0, conversion obtains, I _fqfor electric current q axle reference quantity, by load current I _fthrough abc/dq0, conversion obtains.

Ambiguity in definition controller rule is as shown in table 2 according to actual needs, to reach better control effect:

Table 2 fuzzy controller rule list

Wherein, d axle real component reference current I _crefwith static compensator circuit d axle output current I _cddifference e _icor d axle real component reference current I (k) _fqwith static compensator circuit q axle output current I _cqdifference e _iq(k), ec represents e _ic(k) derivative

or e _iq(k) derivative

nB representative " negative large ", NM representative " in negative ", NS representative " negative little ", 0 representative " zero ", PS representative " just little ", PM representative " center ", PB represent " honest ", as shown in table 2, when ec is NB, when e is NB, in corresponding table, rule is PB, illustrates that now controlled quentity controlled variable is honest, must be by e and ec to center, just little direction is controlled, and to reach, controls result for making e and ec all be tending towards 0.

S32: the input data using the output in step S31 as the BP neural net are trained respectively d axle neural net and q axle neural net, and d axle neural net and q axle neural net are in full accord, and wherein input layer has 3 neurons, being input as of input layer wherein f is the input layer function, and k is input variable; The output of input layer with the input equate, that is:

S33: the output variable weighting that step S32 is obtained is as the input of the hidden layer 42 of BP neural net, and the output of acquisition hidden layer 42, and hidden layer 42 contains 5 neurons, and input layer is ω ji to the weights of described hidden layer 42, and wherein hidden layer 42 is input as ${\mathrm{input}}_j^{(2-2)}\left(k\right)=\underset{i=1}{\overset{3}{\mathrm{\&Sigma;}}}{\mathrm{\&omega;}}_{\mathrm{ji}}\mathrm{outpu}{t}_i^{(2-1)}\left(k\right),i=\mathrm{1,2,3};j=\mathrm{1,2}\&CenterDot;\&CenterDot;\&CenterDot;5$

Hidden layer 42 is output as: ${\mathrm{output}}_j^{(2-2)}\left(k\right)=f\left[{\mathrm{intput}}_j^{(2-1)}\left(k\right)\right],j=\mathrm{1,2}\&CenterDot;\&CenterDot;\&CenterDot;5$

The neuronic excitation function of hidden layer 42 adopts the Sigmoid function of Symmetrical

S34: the output variable weighting that step S33 is obtained is as the input of the output layer 43 of BP neural net, and the output of acquisition output layer 43, and output layer 43 has 2 neurons, and hidden layer 42 is ω lj to the weights of output layer 43

Being input as of output layer 43 wherein: ${\mathrm{input}}_l^{(2-3)}\left(k\right)=\underset{j=1}{\overset{5}{\mathrm{\&Sigma;}}}{\mathrm{\&omega;}}_{\mathrm{lj}}\mathrm{outpu}{t}_j^{(2-2)}\left(k\right),l=\mathrm{1,2};j=\mathrm{1,2}\&CenterDot;\&CenterDot;\&CenterDot;5$

Output layer 43 is output as: ${\mathrm{output}}_l^{(2-3)}\left(k\right)=g\left[{\mathrm{intput}}_l^{(2-3)}\left(k\right)\right],l=\mathrm{1,2}$

Wherein because parameter kp, the ki of output layer 43 outputs are non-negative, therefore excitation function is got non-negative Sigmoid function: $g\left(x\right)=\frac{e^x}{e^x+e^{-x}}$

S35: the parameter K using step S34 output variable kp and ki as the PI controller of corresponding neural net _pqand K _iq;

Here the performance index function that fuzzy neural network is chosen is:

Negative gradient direction according to J is adjusted, an additional minimum inertia coeffeicent of the overall situation that makes search energy Fast Convergent, that is: ${\mathrm{\&Delta;\&omega;}}_{\mathrm{lj}}(k+1)=-\mathrm{\&eta;}\frac{\&PartialD;J}{{\&PartialD;\mathrm{\&omega;}}_{\mathrm{lj}}}+{\mathrm{\&alpha;\&omega;}}_{\mathrm{lj}}\left(k\right),l=\mathrm{1,2};j=\mathrm{1,2}\&CenterDot;\&CenterDot;\&CenterDot;5$

Wherein η is learning efficiency, and α is inertia coeffeicent;

Hidden layer 42 in the BP neural net to the weights ω lj of output layer 43 is: ${\mathrm{\&Delta;\&omega;}}_{\mathrm{lj}}(k+1)=-\mathrm{\&eta;\&delta;outpu}{t}_j^{(2-2)}\left(k\right)+{\mathrm{\&alpha;\&omega;}}_{\mathrm{lj}}\left(k\right),l=\mathrm{1,2};j=\mathrm{1,2}\&CenterDot;\&CenterDot;\&CenterDot;5$

Wherein, $\mathrm{\&delta;}=e(k+1)\mathrm{sgn}\left[\frac{\&PartialD;y(k+1)}{\&PartialD;\mathrm{\&Delta;u}\left(k\right)}\right]\frac{\&PartialD;\mathrm{\&Delta;u}\left(k\right)}{\&PartialD;{\mathrm{output}}_l^{(2-3)}\left(k\right)}{g}^{\&prime;}\left[\mathrm{intpu}{t}_l^{(2-3)}\left(k\right)\right],l=\mathrm{1,2}$

Wherein, g'(x)=g (x) [1-g (x)].

During the system operation, after impact load access, the multi-model controller detects power-factor of load Q in-scope, and selects corresponding model, d axle second level PI controller PI _di(i=1,2 ..., control parameter K n) _pd2and K _id2with q axle PI controller PI _qi(i=1,2 ..., control parameter K n) _pqand K _iq, by above-mentioned fuzzy neural network controller, adjusted, and d axle second level PI controller output voltage U _d, q axle PI controller output voltage U _qwith residual voltage U _o=0 obtains the control command U of insulated gate bipolar transistor in synchronous static compensator (IGBT, Insulated Gate Bipolar Transistor) after the dq0/abc conversion ^* _abc, the control command U of IGBT ^* _abccarry out space vector pulse width modulation (SVPWM with a triangular carrier, Space Vector Pulse Width Modulation), after modulation, input in synchronous static compensator, make the power-factor of load keep 1 left and right, be that voltage and electric current keep same-phase, to reach desirable control effect, its control block diagram as shown in Figure 3.

After impact load access, load-side power factor Q is 0.67, and the multi-model fuzzy neural network PI Direct Current Control method compensation effect in traditional PI method and the present invention as shown in Figure 5 a, Q is 0.38, multi-model fuzzy neural network PI Direct Current Control method compensation effect in traditional PI method and the present invention as shown in Figure 5 b, Q is 0.17, multi-model fuzzy neural network PI Direct Current Control method compensation effect in traditional PI method and the present invention as shown in Figure 5 c, Fig. 5 a, the abscissa of Fig. 5 b and Fig. 5 a is simulation time, ordinate is system voltage, electric current, and in figure, A is the front system voltage current phase relation of compensation, B is traditional PI method compensation effect figure, C is the multi-model fuzzy neural network PI Direct Current Control method compensation effect figure in the present invention.Can see, STATCOM Direct Current Control method based on multi-model fuzzy neural network PI controller all has compensation speed and higher precision faster under different impact loads, can make the power factor of system maintain 1 left and right, thereby obtain satisfied control effect.

Above-mentioned disclosed be only specific embodiments of the invention, this embodiment is only that clearer explanation the present invention is used, and limitation of the invention not, the changes that any person skilled in the art can think of, all should drop in protection range.

Claims (4)

1. the STATCOM current control method of a multi-model fuzzy neural network PI controller, is characterized in that, comprises the following steps:

S1: access the size of the power factor of load current according to a distribution system load-side, will be divided into n model M in fuzzy neural network PI controller _i(i=1,2 ..., n);

S2: for a described n model, design respectively d axle second level PI controller PI _di(i=1,2 ..., n) with q axle PI controller PI _qi(i=1,2 ..., n);

S3: access a load current to described distribution system load-side, and select a described n model M _i(i=1,2 ..., n) in corresponding model, and by d axle second level PI controller PI in the selected model of d axle fuzzy neural network module _di(i=1,2 ..., control parameter K n) _pd2and K _id2, by q axle PI controller PI in the selected model of d axle fuzzy neural network module _qi(i=1,2 ..., control parameter K n) _pqand K _iq; Described d axle fuzzy neural network module comprises d axle fuzzy controller and d axle neural net; Described q axle fuzzy neural network module comprises q axle fuzzy controller and q axle neural net;

Wherein, n is positive integer.

2. the STATCOM current control method of multi-model fuzzy neural network PI controller as claimed in claim 1, is characterized in that, the value of described n is 3.

3. the STATCOM current control method of multi-model fuzzy neural network PI controller as claimed in claim 1 or 2, is characterized in that, described d axle fuzzy neural network module and described q axle fuzzy neural network module all adopt the BP neural net.

4. the STATCOM current control method of multi-model fuzzy neural network PI controller as claimed in claim 3, is characterized in that, step S3 further comprises:

S31: by given d axle real component reference current I _crefwith static compensator circuit d axle output current I _cddifference e _icand difference e (k) _ic(k) derivative

as the input of described d axle fuzzy controller, by described d axle fuzzy controller, exported accordingly; By given d axle real component reference current I _fqwith static compensator circuit q axle output current I _cqdifference e _iqand difference e (k) _iq(k) derivative

as the input of described q axle fuzzy neural network controller, by described d axle fuzzy controller, exported accordingly;

S32: the input data using the output in step S31 as described BP neural net are trained respectively described d axle neural net and described q axle neural net, described d axle neural net and described q axle neural net are in full accord, it includes input layer, hidden layer and output layer, wherein input layer has 3 neurons, being input as of described input layer

wherein f is described input layer function, and k is input variable; The output of described input layer with the input equate, that is: ${\mathrm{output}}_i^{(2-1)}\left(k\right)=\mathrm{inpu}{t}_i^{(2-1)}\left(k\right),i=\mathrm{1,2,3};$

S33: the output variable weighting that step S32 is obtained is as the input of the hidden layer of described BP neural net, and the output of acquisition hidden layer, described hidden layer contains 5 neurons, and described input layer is ω ji to the weights of described hidden layer, being input as of wherein said hidden layer:

${\mathrm{input}}_j^{(2-2)}\left(k\right)=\underset{i=1}{\overset{3}{\mathrm{\&Sigma;}}}{\mathrm{\&omega;}}_{\mathrm{ji}}\mathrm{outpu}{t}_i^{(2-1)}\left(k\right),i=\mathrm{1,2,3};j=\mathrm{1,2}\&CenterDot;\&CenterDot;\&CenterDot;5$

Described hidden layer is output as: ${\mathrm{output}}_j^{(2-2)}\left(k\right)=f\left[{\mathrm{intput}}_j^{(2-1)}\left(k\right)\right],j=\mathrm{1,2}\&CenterDot;\&CenterDot;\&CenterDot;5$

The excitation function of described hidden layer neuron adopts the Sigmoid function of Symmetrical

S34: the output variable weighting that step S33 is obtained is as the input of the output layer of BP neural net, and the output of acquisition output layer, and described output layer has 2 neurons, and hidden layer is ω to the weights of output layer _lj

Being input as of output layer wherein: ${\mathrm{input}}_l^{(2-3)}\left(k\right)=\underset{j=1}{\overset{5}{\mathrm{\&Sigma;}}}{\mathrm{\&omega;}}_{\mathrm{lj}}\mathrm{outpu}{t}_j^{(2-2)}\left(k\right),l=\mathrm{1,2};j=\mathrm{1,2}\&CenterDot;\&CenterDot;\&CenterDot;5$

Output layer is output as: ${\mathrm{output}}_l^{(2-3)}\left(k\right)=g\left[{\mathrm{intput}}_l^{(2-3)}\left(k\right)\right],l=\mathrm{1,2}$

Wherein because parameter kp, the ki of output layer output are non-negative, therefore excitation function is got non-negative Sigmoid function:

S35: the parameter K using step S34 output variable kp and ki as the PI controller of corresponding neural net _pqand K _iq;

Here the performance index function that fuzzy neural network is chosen is:

Negative gradient direction according to J is adjusted, an additional minimum inertia coeffeicent of the overall situation that makes search energy Fast Convergent, that is:

${\mathrm{\&Delta;\&omega;}}_{\mathrm{lj}}(k+1)=-\mathrm{\&eta;}\frac{\&PartialD;J}{{\&PartialD;\mathrm{\&omega;}}_{\mathrm{lj}}}+{\mathrm{\&alpha;\&omega;}}_{\mathrm{lj}}\left(k\right),l=\mathrm{1,2};j=\mathrm{1,2}\&CenterDot;\&CenterDot;\&CenterDot;5$

Wherein η is learning efficiency, and α is inertia coeffeicent;

Hidden layer in the BP neural net to the weights ω lj of output layer is:

${\mathrm{\&Delta;\&omega;}}_{\mathrm{lj}}(k+1)=-\mathrm{\&eta;\&delta;outpu}{t}_j^{(2-2)}\left(k\right)+{\mathrm{\&alpha;\&omega;}}_{\mathrm{lj}}\left(k\right),l=\mathrm{1,2};j=\mathrm{1,2}\&CenterDot;\&CenterDot;\&CenterDot;5$

Wherein, $\mathrm{\&delta;}=e(k+1)\mathrm{sgn}\left[\frac{\&PartialD;y(k+1)}{\&PartialD;\mathrm{\&Delta;u}\left(k\right)}\right]\frac{\&PartialD;\mathrm{\&Delta;u}\left(k\right)}{\&PartialD;{\mathrm{output}}_l^{(2-3)}\left(k\right)}{g}^{\&prime;}\left[\mathrm{intpu}{t}_l^{(2-3)}\left(k\right)\right],l=\mathrm{1,2}$

Wherein, g'(x)=g (x) [1-g (x)].

Images