CN105262358A - Bidirectional grid connected inverter - Google Patents

Abstract

The invention discloses a bidirectional grid connected inverter, including a power supply Udc, a capacitor C and three inversion circuits. The capacitor C and the three inversion circuits are connected in parallel with the power supply Udc. The inversion circuits are provided with three paths of inductors L, capacitors R, inductors L1, resistors R1 and alternating current power supplies connected in series therebtween, and a capacitor C1 is connected in parallel between each capacitor C1 and each inductor L1. The alternating current power supplies comprise Ua, Ub and Uc. According to the invention, as the traditional PID control strategy exerts obvious limitation on the control of voltage, current and power of an inverter, an advanced intelligent control theory is adopted for timely measuring and controlling a large-power and high-voltage bidirectional grid connected inverter system; and a nonlinear prediction model which is achieved by applying a RBF neural network has advantages of fast convergence rate, prevention of convergence at a local extreme point and concise network structure, thereby further improving the dynamic response characteristic of the system.

Description

Two-way combining inverter

Technical field

The present invention relates to a kind of two-way combining inverter, it provide a kind of adopt the Adaptive PID Control algorithm of genetic algorithm to apply to control non-linear and variable element high-power, high-tension two-way combining inverter and based on RTW technological development control routine; It is a kind of bikini IPM inverter structure, and the feature of this inverter to use withstand voltage lower IPM smart power device, composition high-voltage inverter.

Background technology

Non-linear and the variation parameter characteristics of high-power, two-way combining inverter makes to adopt traditional PID to control to be difficult to obtain good control effects.Artificial neural net under certain conditions can arbitrary accuracy Approximation of Arbitrary Nonlinear Function and have stronger self study, self adaptation, self organization ability.It is loaded down with trivial details and the rate height construction cycle that reports an error is long that the line code programming of traditional PID control method has programming.

Summary of the invention

For adapting to uncertainty, non-linear and time variation that most of modern controlled device has, the invention discloses a kind of two-way combining inverter, adopt the Adaptive PID Control strategy of variable element, artificial neural net is combined with PID control law, the Nonlinear Prediction Models that application RBF neural realizes, there is fast convergence rate, can avoid converging on Local Extremum and the succinct advantage of network configuration, thus can improve the dynamic response characteristic of system.

In order to achieve the above object, technical scheme of the present invention is: a kind of two-way combining inverter, comprises power supply U _dc, with power supply U _dcelectric capacity C in parallel and three road inverter circuits, be provided with inductance L, electric capacity R, inductance L that three tunnels are connected in series between inverter circuit ₁, resistance R ₁and AC power, electric capacity R, inductance L ₁between be parallel with electric capacity C ₁; Described AC power comprises U _a, U _b, U _c.

According to Kirchoff s voltage current law and three-phase voltage source type PWM combining inverter operation principle, two-way combining inverter based on the Mathematical Modeling of switch function is:

$c\frac{{\mathrm{du}}_{dc}}{dt}=\underset{k=a,b,c}{\Σ}{i}_{1k}{S}_k-i_{dc}$

$L\frac{{\mathrm{di}}_{1k}}{dt}+R_1{i}_{1k}=u_{ck}-u_{dc}(s_k-\frac{1}{3}\underset{n=a,b,c}{\Σ}{s}_n)$

u _sk＝u _dc*s _k

i _2k＝i _1k-i _ck

$\underset{k=a,b,c}{\Σ}{u}_k=\underset{k=a,b,c}{\Σ}{i}_{1k}=\underset{k=a,b,c}{\Σ}{i}_{ck}=\underset{k=a,b,c}{\Σ}{i}_{2k}=0$

$u_{ck}=1/c_2*\∫i_{ck}dt=u_{sk}+L\frac{{\mathrm{di}}_{ik}}{dt}+{\mathrm{Ri}}_{1k}=u_k-(L\frac{{\mathrm{di}}_{ik}}{dt}+{\mathrm{Ri}}_{1k})$

Described two-way combining inverter utilizes adaptive PID Control to control, described adaptive PID Control is based on RBF neural Nonlinear Prediction Models, comprise Neural network PID, NNC, learning algorithm, RBF forecast model and controlled device, neural net NNC is BP neural net, neural net NNI is RBF neural, and PID controller directly carries out closed-loop control to controlled device.

The method that described PID Controlling model is set up is:

Assuming that in kth time iterative process, the deviation of J controller is:

e _J(k)＝r _J(k)-y _J(k)J＝1，2，3,N

For J controller, the incremental digital PID of its classics controls formula and can be expressed as follows:

u _J ^·(k)＝u ^· _J(k-1)+W _J1(k)·e _J(k)-W _J2(k)·e _J(k-1)+W _J3(k)·e _J(k-2)

Wherein: $W_{J1}\left(k\right)=K_{PJ}+K_{IJ}\·T_s+\frac{K_{DJ}}{T_s};$

$W_{J2}\left(k\right)=K_{PJ}+2\·\frac{K_{DJ}}{T_s};$

$W_{J3}\left(k\right)=\frac{K_{DJ}}{T_s};$

In above formula, K _pJ, K _iJ, K _dJbeing respectively the ratio of J Neural network PID adjuster, integration, differential coefficient, is also discretization variable; W _j1(k), W _j2(k), W _j3k () is respectively the weight function of J Neural network PID; T _sfor controlling the sampling period.

The method of the control of described neural net NNC is:

By W _j0(k), W _j1(k), W _j2when () is considered as the adjustable coefficient depending on system running state k, formula (2) can be described as nonlinear function, can with neural net NNC by training and learning to find an Optimal Control rule;

Neural net NNC is 13 layers of BP neural net, and its mathematic(al) representation is:

u _J(k)＝f(u _J(k-1),W _J1(k),W _J2(k),W _J3(k),e _J(k),e _J(k-1),e _J(k-2))

Neural net NNC has 4 input nodes, 6 hidden nodes, 3 output nodes; Input node corresponds to selected system running state amount, 3 adjustable parameters of the corresponding PID controller of output node difference, W _j0(k), W _j1(k), W _j2(k); Due to PID controller coefficient W _j0(k), W _j1(k), W _j2k () can not be negative value, so the neuronic excitation function of output layer gets the Sigmoid function of non-negative, and the excitation function of hidden layer neuron gets the Sigmoid function of Symmetrical; Choose the following parameters of J controller:

X _J＝[r _J(k),y _J(k),e _J(k),1]

The output of J neural net NNC input layer is

$\begin{array}{cc}{O}_{Jj}^{\left(1\right)}\left(k\right)=x_{Jj}\left(k\right)& j=1,2,\mathrm{....},M\end{array}$

In formula: for the output node of input layer; M is the number of input variable;

The hidden layer input of neural net NNC, output are

$\left\{\begin{array}{c}{n}_{Ji}^{\left(2\right)}\left(k\right)=\underset{j=1}{\overset{M}{\Σ}}{w}_{ij}^{\left(2\right)}\left(k\right)\·O_{Jj}^{\left(1\right)}\left(k\right)\\ \begin{array}{cc}{O}_{Ji}^{\left(2\right)}\left(k\right)=f\left(n_{Ji}^{\left(2\right)}\right(k\left)\right)& i=1,2,\mathrm{...}Q\end{array}\end{array}\right.$

In formula: for hidden layer weight coefficient; F () is excitation function,

f(·)＝tanhx＝(e ^x-e ^-x)/(e ^x+e ^-x)；

Input layer, hidden layer, the output layer of superscript (1), (2), (3) difference correspondence;

Input, the output of the output layer of J neural net NNC are

$\left\{\begin{array}{c}{n}_{Jl}^{\left(3\right)}\left(k\right)=\underset{i=1}{\overset{Q}{\Σ}}{w}_{jl}^{\left(3\right)}{O}_{Ji}^{\left(2\right)}\left(k\right)\\ O_{Jl}^{\left(3\right)}\left(k\right)=g\left(n_{Jl}^{\left(3\right)}\left(k\right)\right)l=1,2,3\\ O_{J1}^{\left(3\right)}\left(k\right)=W_{J1}\left(k\right)\\ O_{J2}^{\left(3\right)}\left(k\right)=W_{J2}\left(k\right)\\ O_{J3}^{\left(3\right)}\left(k\right)=W_{J3}\left(k\right)\end{array}\right.$

In formula, for output layer weight coefficient, excitation function is:

g(·)＝(1+tanh(x))/2＝e ^x/(e ^x-e ^-x)

Assuming that the performance index function getting J NNC is Quadratic Function Optimization be:

$J_J\left(k\right)=\frac{1}{2}{\left(r_J\right(k)-y_J(k\left)\right)}^2$

In formula: r _jk () is system reference input; y _jk output that () is system;

With the weight coefficient of steepest descent method corrective networks, namely press J _j(k), to the negative gradient direction search adjustment of weight coefficient, and add the Inertia that 1 makes search Fast Convergent global minima, then have:

$w_{lj}^{\left(3\right)}\left(k\right)=w_{lj}^{\left(3\right)}(k-1)+\mathrm{\η}\left(k\right)\·\frac{\∂J_J\left(k\right)}{\∂w_{lj}^{\left(3\right)}\left(k\right)}+\mathrm{\α}\·{\mathrm{\Δw}}_{lj}^{\left(3\right)}(k-1)$

In formula: η (k) is learning rate, α is inertia coeffeicent;

$\frac{\∂J_J\left(k\right)}{\∂w_{lj}^{\left(3\right)}\left(k\right)}=\frac{\∂J_J}{\∂y\left(k\right)}\·\frac{\∂y\left(k\right)}{\∂u\left(k\right)}\·\frac{\∂u\left(k\right)}{\∂O_l^{\left(3\right)}\left(k\right)}\·$

$\frac{\∂O_l^{\left(3\right)}\left(k\right)}{\∂n_l^{\left(3\right)}\left(k\right)}\·\frac{\∂n_l^{\left(3\right)}\left(k\right)}{\∂w_{lj}^{\left(3\right)}\left(k\right)}$

Due to the unknown, for obtaining good control effects, should adopt the prediction output valve of forecast model replace ;

Can be in the hope of by above formula

$\left\{\begin{array}{c}\frac{\∂u\left(k\right)}{\∂{O_1}^{\left(3\right)}\left(k\right)}=e\left(k\right)-e(k-1)\\ \frac{\∂u\left(k\right)}{\∂{O_2}^{\left(3\right)}\left(k\right)}=e\left(k\right)\\ \frac{\∂u\left(k\right)}{\∂{O_3}^{\left(3\right)}\left(k\right)}=e\left(k\right)-2e(k-1)+e(k-2)\end{array}\right.$

The weight coefficient computing formula that therefore can obtain neural net NNC output layer is

G ' [*] in formula=g (x) [1-g (x)]

According to above-mentioned projectional technique, the computing formula that can obtain hidden layer weight coefficient is

$\left\{\begin{array}{c}{{\mathrm{\Δw}}_{ij}}^{\left(2\right)}\left(k\right)=\mathrm{\η}\left(k\right){{\mathrm{\δ}}_i}^{\left(2\right)}{O_j}^{\left(1\right)}\left(k\right)+{{\mathrm{\α\Δw}}_{ij}}^{\left(2\right)}(k-1)\\ {{\mathrm{\δ}}_i}^{\left(2\right)}=f^{\′}\[{n_i}^{\left(2\right)}\left(k\right)\]\underset{l=1}{\overset{3}{\Σ}}{{\mathrm{\δ}}_l}^{\left(3\right)}{w_{li}}^{\left(3\right)}\left(k\right)\\ i=1,2,\mathrm{3...},Q\end{array}\right.$

In formula, f ' []=[1-f ²(x)]/2.

The control method of described neural net NNI is: set up a generator system MISO non linear system and describe

y(k)＝f[u(k-1),y(k-1),y(k-2)]；

Adopt 1 Nonlinear Prediction Models of 3 layers of RBF neural NNI as non-linear controlled device three-phase more than hydraulic generator with 3 input nodes, 5 hidden layer nodes and 1 output node;

Choose X=[u (k-1), y (k-1), y (k-2)] ^tfor the input vector of Nonlinear Prediction Models neural net NNI, form the radial basis vector H=[h of the RBF neural of neural net NNI ₁, h ₂..., h _j..., h _m] ^t, wherein, Gaussian bases

$\begin{array}{cc}{h}_j\left(x\right)=\exp (-\frac{\left|\right|X-Cj|{|}^2}{2{b_j}^2})& j=1,2,\mathrm{...},m\end{array}$

In formula: C _jfor the center vector of j node, C _j=[c _j1, c _j2..., c _ji..., c _jn] ^t; Wherein, i=l, 2.., n; b _jfor the base width parameter of node, and for being greater than the number of zero;

The output of RBF neural NNI is

In formula, W _jfor hidden layer neuron and the neuronic connection weight of output layer;

The performance index function of neural net NNI is taken as Quadratic Function Optimization:

$J_1=\frac{1}{2}{\[y\left(k\right)-y_m\left(k\right)\]}^2$

In formula: y (k) is system output; y _mk () is neural network identification output;

According to gradient descent method, the iterative algorithm exporting power, node center and node sound stage width parameter is

w _j(k)＝w _j(k-1)+η′(k)[y(k)-y _m(k)]h _j+α′[w _j(k-1)-w _j(k-2)]

${\mathrm{\Δb}}_j=\[y\left(k\right)-y_m\left(k\right)\]w_j{h}_j\frac{\left|\right|x-C_j|{|}^2}{{b_j}^3}$

b _j(k)＝b _j(k-1)η′(k)Δb _j+α′[b _j(k-1)-b _j(k-2)]

${\mathrm{\Δc}}_{ji}=\[y\left(k\right)-y_m\left(k\right)\]w_j\frac{x_j-c_{ji}}{{b_j}^2}$

c _ji(k)＝c _ji(k-1)+η′(k)Δc _ji+α′[c _ji(k-1)-c _ji(k-2)]

In formula: η ' (k) is learning rate; α ' is inertia coeffeicent;

The output of controlled device to the derivative algorithm of control inputs is

$\frac{\∂y\left(k\right)}{\∂u\left(k\right)}\≈\frac{\∂{\hat{y}}_m\left(k\right)}{\∂u\left(k\right)}=\underset{j=1}{\overset{m}{\Σ}}{w}_j{h}_j\frac{c_{ji}-x_1}{{b_j}^2}$

X in formula ₁=u (k).

Described RBF forecast model is by the electric current of Sample AC side and line voltage, and line voltage obtains the output angle controlled through phase-locked loop, DC voltage utilize after comparing with reference voltage RBF algorithm to grid-connected tracking power, divided by line voltage U _dqafter obtain the reference signal I of desired output electric current _dq, be input to RBF controller with after the shaft current after sample conversion, obtain controlling PWM output signal after computing.

The simulation model of described two-way combining inverter comprises AC voltage detection module, DC voltage detection module, phase-detection, coordinate transformation module, outer shroud voltage control module and inner ring current control module, by detecting rectifier bridge DC voltage U _dcwith given voltage compare, the error signal obtained obtains given value of current value after outer voltage controller RBF computing, flows into inductive current i by detecting inversion AC _abc, obtain current i through coordinate transform _d, i _q, comparing with given value of current value, by obtaining the modulating wave of controller after RBF computing and synthesis, after modulating wave and triangular carrier compare, obtaining the switching signal of IGBT.

Beneficial effect of the present invention: special two-way combining inverter is due to the uncertainty of energy density and spatial and temporal distributions and non-linear, traditional PID control strategy is adopted to create fairly obvious limitation for the control of the voltage of inverter, electric current, power, therefore adopt advanced Intelligent Control Theory, carrying out real-tune TT & C for high-power, the two-way grid-connected inverter system of high voltage is a preferred technical scheme.But, after adopting modern Intelligent Control Theory, for the operational capability of embedded control system and microprocessor system and hardware resource by proposing higher requirement, novel digital signal processor DSP must be adopted as core processor, and the control variables for system carries out observing and controlling at a high speed.High-power, the two-way combining inverter self adaptation PlD of high voltage based on RBF (radial basis) Neural Network Based Nonlinear forecast model control; Utilize simulinkRTW technological development control routine, build after simulation model adjustment obtains suitable parameters and change into control routine, have the advantages that fast Development obtains control effects; Based on the optimal design of the inverter of genetic algorithm

Accompanying drawing explanation

Fig. 1 is the Neural Network PID Control System that the present invention is based on RBF neural forecast model.

Fig. 2 is the input variable of NNC neural net of the present invention.

Fig. 3 is the output schematic diagram of NNC neural net of the present invention.

Fig. 4 is the main circuit topology figure of the two-way combining inverter of the present invention.

Fig. 5 is that the digital signal processor algorithm of the two-way combining inverter of the present invention controls.

Fig. 6 the present invention is based on the two-way grid-connected control principle drawing of RBF.

Embodiment

The present invention is specifically described below by drawings and Examples.

A kind of two-way combining inverter, as shown in Figure 1, Figure 2, shown in Fig. 3, Fig. 4 and Fig. 5, comprises power supply U _dc, with power supply U _dcelectric capacity C in parallel and three road inverter circuits, be provided with inductance L, electric capacity R, inductance L that three tunnels are connected in series between inverter circuit ₁, resistance R ₁and AC power, electric capacity R, inductance L ₁between be parallel with electric capacity C ₁; Described AC power comprises U _a, U _b, U _c.

Different from adopting the combining inverter of inductor filter, adopt the topological structure of LCL type filter to exchange outlet side and add electric capacity and inductance.Assuming that do not consider that combining inverter DC bus both end voltage fluctuates, three-phase power grid voltage is symmetrical and stable, and main circuit switch components and parts are regarded as Ideal switch.According to Kirchoff s voltage current law and three-phase voltage source type PWM combining inverter operation principle, two-way combining inverter based on the Mathematical Modeling of switch function is:

$c\frac{{\mathrm{du}}_{dc}}{dt}=\underset{k=a,b,c}{\Σ}{i}_{1k}{S}_k-i_{dc}$

$L\frac{{\mathrm{di}}_{1k}}{dt}+R_1{i}_{1k}=u_{ck}-u_{dc}(S_k-\frac{1}{3}\underset{n=a,b,c}{\Σ}{s}_n)$

u _sk＝u _dc*s _k

i _2k＝i _1k-i _ck

$\underset{k=a,b,c}{\Σ}{u}_k=\underset{k=a,b,c}{\Σ}{i}_{1k}=\underset{k=a,b,c}{\Σ}{i}_{ck}=\underset{k=a,b,c}{\Σ}{i}_{2k}=0$

$u_{ck}=1/c_2*\∫i_{ck}dt=u_{sk}+L\frac{{\mathrm{di}}_{ik}}{dt}+{\mathrm{Ri}}_{1k}=u_k-(L\frac{{\mathrm{di}}_{ik}}{dt}+{\mathrm{Ri}}_{1k})$

Based on the adaptive PID Control overall structure of RBF neural Nonlinear Prediction Models

Two-way combining inverter utilizes adaptive PID Control to control, described adaptive PID Control is based on RBF neural Nonlinear Prediction Models, comprise Neural network PID, NNC, learning algorithm, RBF forecast model and controlled device, neural net NNC is BP neural net, neural net NNI is RBF neural, and PID controller directly carries out closed-loop control to controlled device.Based on the overall structure block diagram of high-power, high voltage city underground, the light rail motor-car special I PM inverter adaptive PID Control of RBF neural Nonlinear Prediction Models as Fig. 1.In Fig. 1, neural net NNC is BP neural net, and it is according to the running status of system, regulates the parameter of PID controller, to reaching the profile error of controller parameter.PID controller directly carries out closed-loop control to controlled device, and KI, KD are on-line tuning.Neural net NNI is a RBF neural, as the Nonlinear Prediction Models of non-linear controlled device-high-power, the two-way combining inverter of high voltage.

Adaptive PID Control controls to give full play to the Neural Network PID control that the self adaptation of neural net, non-linear mapping capability and learning ability can form the very strong Parameter adjustable of a kind of adaptive ability, for improving the response speed of controlled system, for nerve network control system, generally need to carry out identification to the characteristic of controlled system, because of the nonlinear characteristic of controlled device, therefore set up the Nonlinear Prediction Models based on neural net.Utilize simulinkRTW technological development control routine, build after simulation model adjustment obtains suitable parameters and change into control routine, have the advantages that fast Development obtains control effects.

Based on the structure of the pid control parameter adaptive controller of BP neural net

First suppose that the deviation of J controller can be expressed as follows in kth time iterative process:

e _J(k)＝r _J(k)-y _J(k)J＝1，2，3,N

For J controller, the incremental digital PID of its classics controls formula and can be expressed as follows:

u _J ^·(k)＝u _J ^·(k-1)+W _J1(k)·e _J(k)-W _J2(k)·e _J(k-1)+W _J3(k)·e _J(k-2)

Wherein: $W_{J1}\left(k\right)=K_{PJ}+K_{IJ}\·T_s+\frac{K_{DJ}}{T_s};$

$W_{J2}\left(k\right)=K_{PJ}+2\·\frac{K_{DJ}}{T_s};$

$W_{J3}\left(k\right)=\frac{K_{DJ}}{T_s};$

In formula, KPJ, KIJ, KDJ are respectively ratio, integration, the differential coefficient of J Neural network PID adjuster, are also discretization variablees; WJ1 (k), WJ2 (k), WJ3 (k) are respectively the weight function of J Neural network PID; Ts is for controlling the sampling period, and when adopting discrete control, this parameter is a very important controlled quentity controlled variable.

The method of the control of described neural net NNC is: by W _j0(k), W _j1(k), W _j2when () is considered as the adjustable coefficient depending on system running state k, formula (2) can be described as nonlinear function, can with neural net NNC by training and learning to find an Optimal Control rule;

Neural net NNC is 13 layers of BP neural net, and its mathematic(al) representation is:

u _J(k)＝f(u _J(k-1),W _J1(k),W _J2(k),W _J3(k),e _J(k),e _J(k-1),e _J(k-2))

Neural net NNC has 4 input nodes, 6 hidden nodes, 3 output nodes; Input node corresponds to selected system running state amount, 3 adjustable parameters of the corresponding PID controller of output node difference, W _j0(k), W _j1(k), W _j2(k); Due to PID controller coefficient W _j0(k), W _j1(k), W _j2k () can not be negative value, so the neuronic excitation function of output layer gets the Sigmoid function of non-negative, and the excitation function of hidden layer neuron gets the Sigmoid function of Symmetrical; Choose the following parameters of J controller:

X _J＝[r _J(k),y _J(k),e _J(k),1]

The output of J neural net NNC input layer is

$\begin{array}{cc}{O}_{Jj}^{\left(1\right)}\left(k\right)=x_{Jj}\left(k\right)& j=1,2,\mathrm{....},M\end{array}$

In formula: for the output node of input layer; M is the number of input variable;

The hidden layer input of neural net NNC, output are

$\left\{\begin{array}{c}{n}_{Ji}^{\left(2\right)}\left(k\right)=\underset{j=1}{\overset{M}{\Σ}}{w}_{ij}^{\left(2\right)}\left(k\right)\·O_{Jj}^{\left(1\right)}\left(k\right)\\ \begin{array}{cc}{O}_{Ji}^{\left(2\right)}\left(k\right)=f\left(n_{Ji}^{\left(2\right)}\right(k\left)\right)& i=1,2,\mathrm{...}Q\end{array}\end{array}\right.$

In formula: for hidden layer weight coefficient; F () is excitation function,

f(·)＝tanhx＝(e ^x-e ^-x)/(e ^x+e ^-x)；

Input layer, hidden layer, the output layer of superscript (1), (2), (3) difference correspondence;

Input, the output of the output layer of J neural net NNC are

$\left\{\begin{array}{c}{n}_{Jl}^{\left(3\right)}\left(k\right)=\underset{i=1}{\overset{Q}{\Σ}}{w}_{jl}^{\left(3\right)}{O}_{Ji}^{\left(2\right)}\left(k\right)\\ O_{Jl}^{\left(3\right)}\left(k\right)=g\left(n_{Jl}^{\left(3\right)}\left(k\right)\right)l=1,2,3\\ O_{J1}^{\left(3\right)}\left(k\right)=W_{J1}\left(k\right)\\ O_{J2}^{\left(3\right)}\left(k\right)=W_{J2}\left(k\right)\\ O_{J3}^{\left(3\right)}\left(k\right)=W_{J3}\left(k\right)\end{array}\right.$

In formula, for output layer weight coefficient, excitation function is:

g(·)＝(1+tanh(x))/2＝e ^x/(e ^x-e ^-x)

Assuming that the performance index function getting J NNC is Quadratic Function Optimization be:

$J_J\left(k\right)=\frac{1}{2}{\left(r_J\right(k)-y_J(k\left)\right)}^2$

In formula: r _jk () is system reference input; y _jk output that () is system;

With the weight coefficient of steepest descent method corrective networks, namely press J _j(k), to the negative gradient direction search adjustment of weight coefficient, and add the Inertia that 1 makes search Fast Convergent global minima, then have:

$w_{lj}^{\left(3\right)}\left(k\right)=w_{lj}^{\left(3\right)}(k-1)+\mathrm{\η}\left(k\right)\·\frac{\∂J_J\left(k\right)}{\∂w_{lj}^{\left(3\right)}\left(k\right)}+\mathrm{\α}\·{\mathrm{\Δw}}_{lj}^{\left(3\right)}(k-1)$

In formula: η (k) is learning rate, α is inertia coeffeicent;

$\frac{\∂J_J\left(k\right)}{\∂w_{lj}^{\left(3\right)}\left(k\right)}=\frac{\∂J_J\left(k\right)}{\∂y\left(k\right)}\·\frac{\∂y\left(k\right)}{\∂u\left(k\right)}\·\frac{\∂u\left(k\right)}{\∂O_l^{\left(3\right)}\left(k\right)}\·$

$\frac{\∂O_l^{\left(3\right)}\left(k\right)}{\∂n_l^{\left(3\right)}\left(k\right)}\·\frac{\∂n_l^{\left(3\right)}\left(k\right)}{\∂w_{lj}^{\left(3\right)}\left(k\right)}$

Due to the unknown, for obtaining good control effects, should adopt the prediction output valve of forecast model replace

Can be in the hope of by above formula

$\left\{\begin{array}{c}\frac{\∂u\left(k\right)}{\∂{O_1}^{\left(3\right)}\left(k\right)}=e\left(k\right)-e(k-1)\\ \frac{\∂u\left(k\right)}{\∂{O_2}^{\left(3\right)}\left(k\right)}=e\left(k\right)\\ \frac{\∂u\left(k\right)}{\∂{O_3}^{\left(3\right)}\left(k\right)}=e\left(k\right)-2e(k-1)+e(k-2)\end{array}\right.$

The weight coefficient computing formula that therefore can obtain neural net NNC output layer is

G ' [*] in formula=g (x) [1-g (x)]

According to above-mentioned projectional technique, the computing formula that can obtain hidden layer weight coefficient is

$\left\{\begin{array}{c}{{\mathrm{\Δw}}_{ij}}^{\left(2\right)}\left(k\right)=\mathrm{\η}\left(k\right){{\mathrm{\δ}}_i}^{\left(2\right)}{O_j}^{\left(1\right)}\left(k\right)+{{\mathrm{\α\Δw}}_{ij}}^{\left(2\right)}(k-1)\\ {{\mathrm{\δ}}_i}^{\left(2\right)}=f^{\′}\[{n_i}^{\left(2\right)}\left(k\right)\]\underset{l=1}{\overset{3}{\Σ}}{{\mathrm{\δ}}_l}^{\left(3\right)}{w_{li}}^{\left(3\right)}\left(k\right)\\ i=1,2,\mathrm{3...},Q\end{array}\right.$

In formula, f ' []=[1-f ²(x)]/2.

For convergence speedup speed, native system adopts the learning algorithm becoming learning rate, according to the size of convergence process medial error, and online adaptive ground regularized learning algorithm speed, the size of η (k).For improving the performance of system, adopt Nonlinear Prediction Models pair carry out on-line identification and prediction.High-power in native system, the two-way combining inverter of high voltage

The control method of described neural net NNI is: set up high-power, a high voltage two-way combining inverter MISO non linear system and describe

y(k)＝f[u(k-1),y(k-1),y(k-2)]；

Adopt 1 Nonlinear Prediction Models of 3 layers of RBF neural NNI as non-linear controlled device three-phase more than hydraulic generator with 3 input nodes, 5 hidden layer nodes and 1 output node;

Choose X=[u (k-1), y (k-1), y (k-2)] ^tfor the input vector of Nonlinear Prediction Models neural net NNI, form the radial basis vector H=[h of the RBF neural of neural net NNI ₁, h ₂..., h _j..., h _m] ^t, wherein, Gaussian bases

$\begin{array}{cc}{h}_j\left(x\right)=\exp (-\frac{\left|\right|X-Cj|{|}^2}{2{b_j}^2})& j=1,2,\mathrm{...},m\end{array}$

In formula: C _jfor the center vector of j node, C _j=[c _j1, c _j2..., c _ji..., c _jn] ^t; Wherein, i=l, 2.., n; b _jfor the base width parameter of node, and for being greater than the number of zero;

The output of RBF neural NNI is

In formula, W _jfor hidden layer neuron and the neuronic connection weight of output layer;

The performance index function of neural net NNI is taken as Quadratic Function Optimization:

$J_1=\frac{1}{2}{\[y\left(k\right)-y_m\left(k\right)\]}^2$

In formula: y (k) is system output; y _mk () is neural network identification output;

According to gradient descent method, the iterative algorithm exporting power, node center and node sound stage width parameter is

w _j(k)＝w _j(k-1)+η′(k)[y(k)-y _m(k)]h _j+α′[w _j(k-1)-w _j(k-2)]

${\mathrm{\Δb}}_j=\[y\left(k\right)-y_m\left(k\right)\]w_j{h}_j\frac{\left|\right|x-C_j|{|}^2}{{b_j}^3}$

b _j(k)＝b _j(k-1)η′(k)Δb _j+α′[b _j(k-1)-b _j(k-2)]

${\mathrm{\Δc}}_{ji}=\[y\left(k\right)-y_m\left(k\right)\]w_j\frac{x_j-c_{ji}}{{b_j}^2}$

c _ji(k)＝c _ji(k-1)+η′(k)Δc _ji+α′[c _ji(k-1)-c _ji(k-2)]

In formula: η ' (k) is learning rate; α ' is inertia coeffeicent;

The output of controlled device to the derivative algorithm of control inputs is

$\frac{\∂y\left(k\right)}{\∂u\left(k\right)}\≈\frac{\∂{\hat{y}}_m\left(k\right)}{\∂u\left(k\right)}=\underset{j=1}{\overset{m}{\Σ}}{w}_j{h}_j\frac{c_{ji}-x_1}{{b_j}^2}$

X in formula ₁=u (k).

Based on the two-way cutting-in control mode of RBF

RBF forecast model adopts combining inverter structure as described in Figure 6, by electric current and the line voltage of Sample AC side, line voltage obtains the output angle controlled through phase-locked loop pll, DC voltage utilize after comparing with reference voltage RBF algorithm to grid-connected tracking power, divided by the Ud of line voltage, obtain the reference signal Idq of desired output electric current after Uq, be input to RBF controller with after the dq shaft current after sample conversion, obtain controlling PWM output signal after computing.

Emulation and control block diagram

As shown in Figure 5, the simulation model of two-way combining inverter mainly contains AC voltage detection module, DC voltage detection module, phase-detection and coordinate transformation module, outer shroud voltage control module, inner ring current control module, by detecting rectifier bridge DC voltage U _dcwith given voltage compare, the error signal obtained obtains given value of current value after outer voltage controller RBF computing, flows into inductive current i by detecting inversion AC _abc, obtain i through coordinate transform _d, i _q, comparing with given value of current value, by obtaining the modulating wave of controller after RBF computing and synthesis, after modulating wave and triangular carrier compare, obtaining the switching signal of IGBT.So just reach the object of control.

For adapting to uncertainty, non-linear and time variation that most of modern controlled device has, the Adaptive PID Control strategy of variable element should be adopted.Artificial neural net is combined with PID control law, gives full play to the Neural Network PID control that the self adaptation of neural net, non-linear mapping capability and learning ability can form the very strong Parameter adjustable of a kind of adaptive ability.For improving the response speed of controlled system, for nerve network control system, generally needing to carry out identification to the characteristic of controlled system, because of the nonlinear characteristic of controlled device, setting up the Nonlinear Prediction Models based on neural net.The Nonlinear Prediction Models that application RBF neural realizes, has fast convergence rate, can avoid converging on Local Extremum and the succinct advantage of network configuration, thus can improve the dynamic response characteristic of system.

It is loaded down with trivial details and the rate height construction cycle that reports an error is long that line code programming has programming, and by utilizing real-time code Core Generator case (RTW) Simulink module map to be automatically converted to C code, through rewriting on a small quantity and adding real time control machine system, can from Simulink module transitions to VC Integrated Development Environment, thus achieve the seamless link of design phase and implementation phase, eliminate the tedious work writing model code and Simulation Control code in VC, and greatly facilitate secondary development and the integration of system.

The above; be only the present invention's preferably embodiment, but protection scope of the present invention is not limited thereto, is anyly familiar with those skilled in the art in the technical scope that the present invention discloses; the change that can expect easily or replacement, all should be encompassed within protection scope of the present invention.

Claims (8)

1. a two-way combining inverter, is characterized in that, comprises power supply U _dc, with power supply U _dcelectric capacity C in parallel and three road inverter circuits, be provided with inductance L, electric capacity R, inductance L that three tunnels are connected in series between inverter circuit ₁, resistance R ₁and AC power, electric capacity R, inductance L ₁between be parallel with electric capacity C ₁; Described AC power comprises U _a, U _b, U _c.

2. two-way combining inverter according to claim 1, is characterized in that, according to Kirchoff s voltage current law and three-phase voltage source type PWM combining inverter operation principle, two-way combining inverter based on the Mathematical Modeling of switch function is:

$c\frac{{\mathrm{du}}_{dc}}{dt}=\underset{k=a,b,c}{\Σ}{i}_{1k}{S}_k-i_{dc}$

$L\frac{{\mathrm{di}}_{1k}}{dt}+R_1{i}_{1k}=u_{ck}-u_{dc}(s_k-\frac{1}{3}\underset{n=a,b,c}{\Σ}{s}_n)$

u _sk＝u _dc*s _k

i _2k＝i _1k-i _ck

$\underset{k=a,b,c}{\Σ}{u}_k=\underset{k=a,b,c}{\Σ}{i}_{1k}=\underset{k=a,b,c}{\Σ}{i}_{ck}=\underset{k=a,b,c}{\Σ}{i}_{2k}=0$

$u_{ck}=1/c_2*\∫i_{ck}dt=u_{sk}+L\frac{{\mathrm{di}}_{ik}}{dt}+{\mathrm{Ri}}_{1k}=u_k-(L\frac{{\mathrm{di}}_{ik}}{dt}+{\mathrm{Ri}}_{1k})$

3. two-way combining inverter according to claim 1, it is characterized in that, described two-way combining inverter utilizes adaptive PID Control to control, described adaptive PID Control is based on RBF neural Nonlinear Prediction Models, comprise Neural network PID, NNC, learning algorithm, RBF forecast model and controlled device, neural net NNC is BP neural net, and neural net NNI is RBF neural, and PID controller directly carries out closed-loop control to controlled device.

4. two-way combining inverter according to claim 3, is characterized in that, the method that described PID Controlling model is set up is:

Assuming that in kth time iterative process, the deviation of J controller is:

e _J(k)＝r _J(k)-y _J(k)J＝1，2，3,N

For J controller, the incremental digital PID of its classics controls formula and can be expressed as follows:

$\stackrel{\·}{u_J}\left(k\right)=\stackrel{\·}{u_J}(k-1)+W_{J1}\left(k\right)\·e_J\left(k\right)-W_{J2}\left(k\right)\·e_J(k-1)+W_{J3}\left(k\right)\·e_J(k-2)$

Wherein: $W_{J1}\left(k\right)=K_{PJ}+K_{IJ}\·T_s+\frac{K_{DJ}}{T_s};$

$W_{J2}\left(k\right)=K_{PJ}+2\·\frac{K_{DJ}}{T_s};$

$W_{J3}\left(k\right)=\frac{K_{DJ}}{T_s};$

In above formula, K _pJ, K _iJ, K _dJbeing respectively the ratio of J Neural network PID adjuster, integration, differential coefficient, is also discretization variable; W _j1(k), W _j2(k), W _j3k () is respectively the weight function of J Neural network PID; T _sfor controlling the sampling period.

5. two-way combining inverter according to claim 4, is characterized in that, the method that described PID Controlling model is set up is:

The method of the control of described neural net NNC is:

By W _j0(k), W _j1(k), W _j2when () is considered as the adjustable coefficient depending on system running state k, formula (2) can be described as nonlinear function, can with neural net NNC by training and learning to find an Optimal Control rule;

Neural net NNC is 13 layers of BP neural net, and its mathematic(al) representation is:

u _J(k)＝f(u _J(k-1),W _J1(k),W _J2(k),W _J3(k),e _J(k),e _J(k-1),e _J(k-2))

Neural net NNC has 4 input nodes, 6 hidden nodes, 3 output nodes; Input node corresponds to selected system running state amount, 3 adjustable parameters of the corresponding PID controller of output node difference, W _j0(k), W _j1(k), W _j2(k); Due to PID controller coefficient W _j0(k), W _j1(k), W _j2k () can not be negative value, so the neuronic excitation function of output layer gets the Sigmoid function of non-negative, and the excitation function of hidden layer neuron gets the Sigmoid function of Symmetrical; Choose the following parameters of J controller:

X _J＝[r _J(k),y _J(k),e _J(k),1]

The output of J neural net NNC input layer is

$O_{Jj}^{\left(1\right)}\left(k\right)=x_{Jj}\left(k\right),j=1,2,\mathrm{....},M$

In formula: for the output node of input layer; M is the number of input variable;

The hidden layer input of neural net NNC, output are

$\left\{\begin{array}{c}{n}_{Ji}^{\left(2\right)}\left(k\right)=\underset{j=1}{\overset{M}{\Σ}}{w}_{ij}^{\left(2\right)}\left(k\right)\·O_{Jj}^{\left(1\right)}\left(k\right)\\ O_{Ji}^{\left(2\right)}\left(k\right)=f\left(n_{Ji}^{\left(2\right)}\right(k\left)\right),i=1,2,...Q\end{array}\right.$

In formula: for hidden layer weight coefficient; F () is excitation function,

f(·)＝tanhx＝(e ^x-e ^-x)/(e ^x+e ^-x)；

Input layer, hidden layer, the output layer of superscript (1), (2), (3) difference correspondence;

Input, the output of the output layer of J neural net NNC are

$\left\{\begin{array}{c}{n}_{Jl}^{\left(3\right)}\left(k\right)=\underset{i=1}{\overset{Q}{\Σ}}{w}_{jl}^{\left(3\right)}{O}_{Ji}^{\left(2\right)}\left(k\right)\\ O_{Jl}^{\left(3\right)}\left(k\right)=g\left(n_{Jl}^{\left(3\right)}\right(k\left)\right),l=1,2,3\\ O_{J1}^{\left(3\right)}\left(k\right)=W_{J1}\left(k\right)\\ O_{J2}^{\left(3\right)}\left(k\right)=W_{J2}\left(k\right)\\ O_{J3}^{\left(3\right)}\left(k\right)=W_{J3}\left(k\right)\end{array}\right.$

In formula, for output layer weight coefficient, excitation function is:

g(·)＝(1+tanh(x))/2＝e ^x/(e ^x-e ^-x)

Assuming that the performance index function getting J NNC is Quadratic Function Optimization be:

$J_J\left(k\right)=\frac{1}{2}{\left(r_J\right(k)-y_J(k\left)\right)}^2$

In formula: r _jk () is system reference input; y _jk output that () is system;

With the weight coefficient of steepest descent method corrective networks, namely press J _j(k), to the negative gradient direction search adjustment of weight coefficient, and add the Inertia that 1 makes search Fast Convergent global minima, then have:

$w_{lj}^{\left(3\right)}\left(k\right)=w_{lj}^{\left(3\right)}(k-1)+\mathrm{\η}\left(k\right)\·\frac{\∂J_J\left(k\right)}{\∂w_{lj}^{\left(3\right)}\left(k\right)}+\mathrm{\α}\·{\mathrm{\Δw}}_{lj}^{\left(3\right)}(k-1)$

In formula: η (k) is learning rate, α is inertia coeffeicent;

$\begin{array}{c}\frac{\∂J_J\left(k\right)}{\∂w_{lj}^{\left(3\right)}\left(k\right)}=\frac{\∂J_J\left(k\right)}{\∂y\left(k\right)}\·\frac{\∂y\left(k\right)}{\∂u\left(k\right)}\·\frac{\∂u\left(k\right)}{\∂O_l^{\left(3\right)}\left(k\right)}\·\\ \frac{\∂O_l^{\left(3\right)}\left(k\right)}{\∂n_l^{\left(3\right)}\left(k\right)}\·\frac{\∂n_l^{\left(3\right)}\left(k\right)}{\∂w_{lj}^{\left(3\right)}\left(k\right)}\end{array}$

Due to the unknown, for obtaining good control effects, should adopt the prediction output valve of forecast model replace

Can be in the hope of by above formula

$\left\{\begin{array}{c}\frac{\∂u\left(k\right)}{\∂{O_1}^{\left(3\right)}\left(k\right)}=e\left(k\right)-e(k-1)\\ \frac{\∂u\left(k\right)}{\∂{O_2}^{\left(3\right)}\left(k\right)}=e\left(k\right)\\ \frac{\∂u\left(k\right)}{\∂{O_3}^{\left(3\right)}\left(k\right)}=e\left(k\right)-2e(k-1)+e(k-2)\end{array}\right.$

The weight coefficient computing formula that therefore can obtain neural net NNC output layer is

G ' [*] in formula=g (x) [1-g (x)]

According to above-mentioned projectional technique, the computing formula that can obtain hidden layer weight coefficient is

$\left\{\begin{array}{c}{{\mathrm{\Δw}}_{ij}}^{\left(2\right)}\left(k\right)=\mathrm{\η}\left(k\right){{\mathrm{\δ}}_i}^{\left(2\right)}{O_j}^{\left(1\right)}\left(k\right)+{{\mathrm{\α\Δw}}_{ij}}^{\left(2\right)}\left(k-1\right)\\ {{\mathrm{\δ}}_i}^{\left(2\right)}=f^{\′}\[{n_i}^{\left(2\right)}\left(k\right)\]\underset{l=1}{\overset{3}{\Σ}}{{\mathrm{\δ}}_l}^{\left(3\right)}{w_{li}}^{\left(3\right)}\left(k\right)\\ i=1,2,3...,Q\end{array}\right.$

In formula, f ' []=[1-f ²(x)]/2.

6. two-way combining inverter according to claim 5, is characterized in that, the control method of described neural net NNI is: set up high-power, a high voltage two-way combining inverter MISO non linear system and describe

y(k)＝f[u(k-1),y(k-1),y(k-2)]；

Adopt 1 Nonlinear Prediction Models of 3 layers of RBF neural NNI as non-linear controlled device three-phase more than hydraulic generator with 3 input nodes, 5 hidden layer nodes and 1 output node;

Choose X=[u (k-1), y (k-1), y (k-2)] ^tfor the input vector of Nonlinear Prediction Models neural net NNI, form the radial basis vector H=[h of the RBF neural of neural net NNI ₁, h ₂..., h _j..., h _m] ^t, wherein, Gaussian bases

$h_j\left(x\right)=\exp (-\frac{\left|\right|X-Cj|{|}^2}{2{b_j}^2}),j=1,2,...,m$

In formula: C _jfor the center vector of j node, C _j=[c _j1, c _j2..., c _ji..., c _jn] ^t; Wherein, i=l, 2.., n; b _jfor the base width parameter of node, and for being greater than the number of zero;

The output of RBF neural NNI is

In formula, W _jfor hidden layer neuron and the neuronic connection weight of output layer;

The performance index function of neural net NNI is taken as Quadratic Function Optimization:

$J_1=\frac{1}{2}{\[y\left(k\right)-y_m\left(k\right)\]}^2$

In formula: y (k) is system output; y _mk () is neural network identification output;

According to gradient descent method, the iterative algorithm exporting power, node center and node sound stage width parameter is

w _j(k)＝w _j(k-1)+η′(k)[y(k)-y _m(k)]h _j+α′[w _j(k-1)-w _j(k-2)]

${\mathrm{\Δb}}_j=\[y\left(k\right)-y_m\left(k\right)\]w_j{h}_j\frac{\left|\right|x-C_j|{|}^2}{{b_j}^3}$

b _j(k)＝b _j(k-1)η′(k)Db _j+α′[b _j(k-1)-b _j(k-2)]

${\mathrm{\Δc}}_{ji}=\[y\left(k\right)-y_m\left(k\right)\]w_j\frac{x_j-c_{ji}}{{b_j}^2}$

c _ji(k)＝c _ji(k-1)+η′(k)Dc _ji+α′[c _ji(k-1)-c _ji(k-2)]

In formula: η ' (k) is learning rate; α ' is inertia coeffeicent;

The output of controlled device to the derivative algorithm of control inputs is

$\frac{\∂y\left(k\right)}{\∂u\left(k\right)}\≈\frac{\∂{\hat{y}}_m\left(k\right)}{\∂u\left(k\right)}=\underset{j=1}{\overset{m}{\Σ}}{w}_j{h}_j\frac{c_{ji}-x_1}{{b_j}^2}$

X in formula ₁=u (k).

7. two-way combining inverter according to claim 3, it is characterized in that, described RBF forecast model passes through electric current and the line voltage of Sample AC side, line voltage obtains the output angle controlled through phase-locked loop, DC voltage utilize after comparing with reference voltage RBF algorithm to grid-connected tracking power, divided by line voltage U _dqafter obtain the reference signal I of desired output electric current _dq, be input to RBF controller with after the shaft current after sample conversion, obtain controlling PWM output signal after computing.

8. two-way combining inverter according to claim 3, it is characterized in that, the simulation model of described two-way combining inverter comprises AC voltage detection module, DC voltage detection module, phase-detection, coordinate transformation module, outer shroud voltage control module and inner ring current control module, by detecting rectifier bridge DC voltage U _dcwith given voltage compare, the error signal obtained obtains given value of current value after outer voltage controller RBF computing, flows into inductive current i by detecting inversion AC _abc, obtain current i through coordinate transform _d, i _q, comparing with given value of current value, by obtaining the modulating wave of controller after RBF computing and synthesis, after modulating wave and triangular carrier compare, obtaining the switching signal of IGBT.