JP5562366B2 - Hybrid intelligent power control system and method - Google Patents

Description

  The present invention relates to a power control method and system, which uses a fuzzy sliding mode and an on-line training recurrent neural network (RNN), and in the system, The method relates to a kind of hybrid intelligent power control method and system using permanent magnet synchronous generator (PMSG) speed control and turbine pitch angle control.

  As is well known, in general, power generation is performed by changing the constant speed and power used by converting with a digital adapter. The variable speed turbine allows operation at the maximum power operating point under various rotational speeds, and maximum efficiency can be achieved by optimizing the rotational shaft speed. These characteristics are the characteristics and advantages of the transmission conversion system, and most of them are in a single control mode, although several control methods have been submitted to achieve maximum power control. .

  Permanent magnet synchronous generators are constantly increasing in recent research, and the ideal characteristics of permanent magnet synchronous generators are a solid structure, high air flow density, high power density, high rotational torque inertia ratio, high rotation Torque. Furthermore, compared with induction motors, permanent magnet synchronous generators have the advantage of high efficiency. This is because the rotor wear is small and the load current is relatively low at the rated speed. The change in coupling control capability to generator parameters is relatively insensitive, and the controllability of the permanent magnet synchronous generator control is predictable.

  The known AC power output without mechanical sensors (wind speed sensor and position sensor) is about 80%. A turbine using a known DFIG series-connected non-linear control for shifting has a considerably large error rate for predicting its rotational speed. The well-known Hill-climbing Algorithm study considers turbine torque, but in any case requires a kind of intelligent memory on-line training process. Power output adjustment by two well-known methods, that is, pitch angle and power load control are both operations to adjust the turbine, but the deviation of the power coefficient is too large. .

  The present invention uses a fuzzy sliding mode controller and applies the fuzzy inference mechanism and the adaptive calculation method to the speed control of a permanent magnet synchronous generator.

  The fuzzy slide mode controller is operated in alignment with the switching surface and feeds back the dynamic behavior of the system to control the robust state of the system when the system is in slide mode. In general, during sliding mode control, there is often an upper bound of uncertainties, and therefore, in actual applications, there are uncertainties including parameter changes and external machine vibrations. Definite boundary values are difficult to obtain. The fuzzy slide mode position controller of the present invention solves the above difficulties and is used to predict a simple fuzzy inference mechanism, the upper bound of uncertainties.

  The present invention, in one embodiment, provides a hybrid intelligent power control system, which implements a fuzzy inference mechanism to control a generator, fuzzy sliding mode controller, and online training A recurrent neural network (RNN) pitch angle controller that achieves maximum power output by performing an on-line training method to control the pitch angle of a turbine coupled to the generator. wherein the turbine responds to changes input to the turbine by adjusting the pitch angle of the blades connected to the turbine, as well as the fuzzy inference mechanism (fuzzy inference mechanism). mechanism) Control shaft speed to achieve maximum power output.

  Another embodiment of the present invention provides a hybrid intelligent power control method, which uses a fuzzy sliding mode controller to perform a fuzzy inference mechanism to control a generator. And performing an on-line training method using a recurrent neural network (RNN) pitch angle controller and the pitch angle of the turbine coupled to the generator To achieve maximum power output by controlling.

  The present invention provides a hybrid fuzzy slide model used for speed control of a permanent magnet synchronous generator in a loss minimization manner, and a high-performance online training recurrent neural network for turbine pitch angle control (on-line training recurrent). Neural network (RNN) design. A dynamic back-propagation learning algorithm is used to adjust the RNN controller. At different rotational speeds, PSG is below rated speed, maximum power point tracking can be used, and maximum energy can be captured in the rotating device. The slide mode controller is matched to the switching surface of the operation, of which fuzzy inference mechanism is used to predict the upper bound of uncertainties.

表示図である。
リカレントニューラルネットワーク（recurrent neural network,RNN）構造図である。 オンライントレーニング（on-line training）のリカレントニューラルネットワーク（recurrent neural network,RNN）フローチャートである。 It is a well-known turbine power generation system display figure. FIG. 2 is a typical power coefficient Cp and tip speed ratio (TSR) / curve diagram. It is a control system display figure of a permanent magnet synchronous generator (PMSG). It is a control simplification system display figure of a permanent magnet synchronous generator (PMSG). Membership functions, S,

It is a display figure.
It is a recurrent neural network (RNN) structure diagram. It is a recurrent neural network (RNN) flowchart of on-line training.

  As shown in FIG. 1, a turbine 10 is axially coupled to a permanent magnet synchronous generator (PMSG) 12 via a gear box 11, the conversion loss of gear acceleration in this portion is ignored, and a permanent magnet synchronous generator ( The PMSG is connected to a power converter 13 and a DC load inverter 14 circuit to convert external energy power (Pw), eg wind energy power, to mechanical power, eg wind turbine output power (Pm). ), Further converted into AC power (Pe), and then converted into DC power (Pdc). This power is supplied directly to the power system and the conversion loss of the acceleration gear can be ignored. The present invention is also used in gearless direct drive systems.

ｐとＡはそれぞれ羽根がスイープする空気密度と面積であり、
Vwは風速（ｍ／ｓｅｃ）であり、
Ｃpはパワー係数（ｐｏｗｅｒ ｃｏｅｆｆｉｃｉｅｎｔ）であり、尖端速度比ｔｉｐ ｓｐｅｅｄ ｒａｔｉｏ（ＴＳＲ）λの非線形方程式とされる。

ｒは風力タービンの羽根半径であり、
ωrはタービン速度であり、
ＣpはＴＳＲλと羽根ピッチ角竈の方程式である。

Power electronics must be installed to capture maximum energy. The device is located between the turbine generator and the grid, and the frequency of this part is constant. If the input to the turbine generator is wind power, the output is the power that rotates the generator rotor. For different speed wind turbines, the output mechanical power obtained from the wind turbine is expressed as:

p and A are the air density and area that the blades sweep respectively,
Vw is the wind speed (m / sec)
Cp is a power coefficient, and is a nonlinear equation having a tip speed ratio (tip speed ratio) (TSR) λ.

r is the blade radius of the wind turbine,
ωr is the turbine speed,
Cp is an equation of TSRλ and blade pitch angle 竈.

As shown in FIG. 2, formula (3) is used to generate a typical Cp and λ curve, where the wind turbine has the best speed ratio λopt and the maximum power coefficient Cpmax. And when β = 0, the tip speed ratio (TSR) of formula (2) is adjusted to the best value λopt = 6.9, and the maximum power coefficient reaches Cpmax = 0.412 Sometimes, the control goal of maximum power acquisition is achieved and is obtained from Formula (1) and Formula (2).

  Formula (4) is the relationship between turbine power and speed at maximum power output. When adjusting the system under the maximum power standard, it is necessary to consider that the turbine power must not be greater than the rated power of the generator. Once the rated power of the generator has achieved the rated wind speed, the output power is Must be restricted. For a variable speed wind power generator, a mechanical actuator is used to change the blade pitch angle to lower the power coefficient and maintain the rated power of the power. When operating at the maximum power coefficient, a wind generator has a rated speed obtained from the wind speed that is lower than the rated power of the generator.

永久磁石同期発電機（ＰＭＳＧ）の動態公式は、

ｗeは電動機角速度
ｎpは極数
Ｊは風力タービン発電機（ＷＴＧ）慣性量
Ｂは発電機摩擦係数
永久磁石同期発電機（ＰＭＳＧ）の機械モジュール中、回転子回転の参考は、

及び

そのうち、
ｖd,ｖq ＝ｄ，ｑ軸固定子電圧
ｉd,ｉq ＝ｄ，ｑ軸固定子電流
Ｌd,Ｌq ＝ｄ，ｑ軸固定子インダクタンス
λd,λq ＝ｄ，ｑ軸固定子磁束鎖
Ｒ＝定子抵抗
Ws＝変換機周波数
Ｉfd＝等価ｄ−軸磁化電流
Ｌmd＝ｄ−軸交互インダクタンス
電気トルク、及び発電機動態は、

である。 The three-phase permanent magnet synchronous generator (PMSG) has its mechanical torque (Tm) and electric torque (Te) expressed as follows:

The dynamic formula of the permanent magnet synchronous generator (PMSG) is

We are the motor angular speed np is the number of poles J is the wind turbine generator (WTG) inertia quantity B is the machine coefficient of the generator friction coefficient permanent magnet synchronous generator (PMSG), the reference for rotor rotation is

as well as

Of which
vd, vq = d, q-axis stator voltage id, iq = d, q-axis stator current Ld, Lq = d, q-axis stator inductance λd, λq = d, q-axis stator flux chain R = constant resistance
Ws = converter frequency Ifd = equivalent d-axis magnetizing current Lmd = d-axis alternating inductance electric torque, and generator dynamics

It is.

トルク電流であり、スピードコントローラより発生する。 FIG. 3 shows the control structure configuration of the permanent magnet synchronous generator (PMSG) 31, current control PWM, voltage source converter (VSC) 32, inverter 33, current controller 34, field direction The field directing machine includes a coordinate translator 35, a fuzzy sliding mode speed controller 36, a pitch control 37, and an AC grid 38. To do. Pw is the external energy power, and Pm is the output power of the wind turbine. Through a field-directed machine, a permanent magnet synchronous generator (PMSG) system can reasonably be displayed in the control block diagram in FIG. Of which

Torque current, generated from the speed controller.

そのうち、
△ωは回転子角度速度差であり、
ωrは発電機回転子角速度である。
同期発電機自身の安定状態パワー出力は以下のようであり、

以下のように表示可能であり、

そのうち、

公式（１９）を考慮すると、方程式は以下のように表示可能であり、

そのうち、△Ａ、△Ｂ、△Ｄは、Ｊ，Ｂ，Ｋｔ，及びＴｍ等のパラメータで表示され得て、
Ｊは回転慣性量（the inertia moment）、
Ｔmは機械トルク（mechanical torque）、
Ｂは発電機の摩擦係数（friction coefficient of the generator）であり、
方程式（２０）は、以下のように表示され得る。

Ｆ（ｔ）は総合変化因子であって、以下のように表示され得る。

方程式（２１）により、全体運転の切換え面は、システムパラメータＡ、Ｂより直接達成され得る。 As shown in FIG. 3, the fuzzy sliding mode speed controller 36 has the following relational variable definitions:

Of which
Δω is the rotor angular speed difference,
ωr is the generator rotor angular velocity.
The steady state power output of the synchronous generator itself is as follows:

Can be displayed as follows:

Of which

Considering the formula (19), the equation can be expressed as

Among them, ΔA, ΔB, ΔD can be displayed with parameters such as J, B, Kt, and Tm,
J is the inertia moment,
Tm is the mechanical torque,
B is the friction coefficient of the generator,
Equation (20) may be expressed as follows:

F (t) is an overall change factor and can be expressed as follows:

According to equation (21), the switching surface of the overall operation can be achieved directly from the system parameters A, B.

そのうち、
Ｃは、正の定数マトリックス集合であり、
Ｋは、状態フィードバックゲインマトリックスである。 The design of the switching surface of the fuzzy sliding mode speed controller 40 is as shown in FIG. 4, which is the fuzzy sliding mode speed controller 40, which provides online training. It includes an RNN pitch controller 41 and a power generator 42 for execution. The switching surface including the integral forward slide mode controller is designed as follows.

Of which
C is a positive constant matrix set,
K is a state feedback gain matrix.

であり、システム方程式（２２）の等価動態は、

である。 From formula (23), if the state trajectory of system equation (22) is captured from the switching surface of equation (23),

And the equivalent behavior of the system equation (22) is

It is.

  In equation (24), when the system pole is located on the right half, the velocity error x1 (t) contracts to zero exponent, so no divergence occurs and the system dynamics are like a state feedback control system. .

そのうち、ｓｇｎ（・）は信号関数であり、

ｆは、

を指す。 Design of sliding mode speed controller:
Based on the switching surface described above, an arithmetic rule that approaches the fuzzy condition and compensates for the presence of the forward slide mode is designed. The forward slide mode rotation speed controller format is as follows.

Of these, sgn (•) is a signal function,

f is

Point to.

切換え面（switching surface）のファジィ推論のメンバシップ関数（membership functions）を、［−２，２］に設定する。 The slide mode controller uses Kf's fuzzy estimation, which includes parameter changes and external machine turbulence, in the upper bound of uncertainties, S,

Set the membership functions for fuzzy inference of the switching surface to [−2, 2].

ファジィスライドモードがコントロールするファジィ規則である。 Using a fuzzy sliding mode speed controller 40, which executes a fuzzy inference mechanism, controls the generator 42, and online training (on- RNN using line training RNN pitch controller 41 to perform an on-line training method and to perform on-line training as shown in FIG. The pitch angle controller 41 controls the pitch angle of the turbine generator to achieve maximum efficiency, and the turbine is coupled to the generator, as shown in FIG. Among them, N: (negative) Negative
ZE, Z: (zero) Zero
P: (Correct) Positive
H: (Huge) Huge
NH: (negative and huge) Negative Huge
NB: (Negative and large) Negative Big
NM: (Negative and Negative) Negative Medium
NS: (Small negative) Negative Small
PH: (positive and huge) Positive Huge
PS: (Positive and Small) Positive Small
PM: (Positive and Medium) Positive Medium
PB: (Positive and large) Positive Big
PH: (positive and huge) Positive Huge
VS: (very small) Very Small
VB: (very large) Very Big
B: (Large) Big
M: (Medium) Medium

Fuzzy rule controlled by fuzzy slide mode.

がＰＢ（正で大）であり、これにより、比較的大きなＫfが切換え面（switching surface）に対して必要である。規則１３はＳが切換え面（switching surface）上にあり、

が零であり、僅かだが非常に小さいＫfが切換え面（switching surface）に対して必要である。
Ｓ，

は、それぞれ、
Ｋfが、

で、重心（ｔｈｅ ｃｅｎｔｅｒ ｏｆ ｇｒａｖｉｔｙ：ＣＯＧ）よりファジー計算により求められ、そのうち、

は調整可能パラメータベクトル（adjusable parameter vector）であり、ｃ1からｃ25はＫfのメンバーシップ関数（membership functions）中心であり、

は起動強度ベクトル（fired stranght vector）である。 The fuzzy inference mechanism has only 25 rules in Table 1. For example, rule 25 states that S is farther than the switching surface,

Is PB (positive and large), which requires a relatively large Kf for the switching surface. Rule 13 is that S is on the switching surface,

Is zero and a small but very small Kf is required for the switching surface.
S,

Respectively
Kf

And obtained by fuzzy calculation from the center of gravity (COG),

Is an adjustable parameter vector, c1 to c25 are the center of Kf membership functions,

Is a fired stranght vector.

Recurrent neural network (RNN):
Like the three-layer neutral network shown in FIG. 6A, the network structure is divided into three layers in total, which are an input layer of the first layer, an intermediate layer of the second layer, and an output layer of the third layer, respectively. . The RNN controller is operated in the turbine, Pw is wind power, Pm is the output power of the wind turbine, and e is the error of the turbine output power.

であり、そのうち、

である。第Ｎ次のサンプル定数の誤差（ｅｒｒｏｒ）は、
ｅ（Ｎ）＝Ｐw−Ｐm と表され得て、エラー変換（change of error）は、
ｃｅ（Ｎ）＝ｅ（Ｎ）−ｅ（Ｎ−１）
と表され得て、そのうち、ｅ（Ｎ−１）は前の一つのサンプル時間の誤差（ｅｒｒｏｒ）値である。
ｅはタービン出力パワーの誤差（ｅｒｒｏｒ）であり、ＲＮＮピッチ角コントロールシステムにより、ピッチ角命令（pitch angle command）βcが評価される。 The RNN input is

Of which

It is. The error of the Nth order sample constant is
e (N) = Pw−Pm and the error change (change of error) is
ce (N) = e (N) -e (N-1)
Where e (N−1) is the error value of the previous one sample time.
“e” is an error of the turbine output power, and the pitch angle command βc is evaluated by the RNN pitch angle control system.

  As shown in FIG. 6A, the normal use unit input layer, intermediate layer, and output layer of the RNN are 2, 9, and 1, respectively, of which the permanent magnet synchronous generator (PMSG) 61, supervised learning and Three layers of a training process (Supervised learning and training process) 62 and a recurrent neural network (RNN) are coupled to each other.

そのうち、

は第１層入力層の入力信号である。

は第１層入力層のノード関数である。

は第１層入力層の出力信号である。
Ｎは反復の次数である。

は第１層入力層動態関数（activation function）、シグモイド関数（sigmoid function）である。 1st layer: Input layer

Of which

Is an input signal of the first layer input layer.

Is a node function of the first layer input layer.

Is an output signal of the first layer input layer.
N is the order of iteration.

Is a first layer input layer activation function, a sigmoid function.

そのうち、
Ｎは反復次数、

は第２層中間層の出力信号、

は第２層中間層の入力信号、

は第２層中間層のノード関数、

は第２層中間層のリカレント単位の重み値

は第２層中、入力層と中間層を連結する重み値、
nは第２層中間層中のニューロン数、

は第２層中間層中の、動態関数（ａｃｔｉｖａｔｉｏｎ ｆｕｎｃｔｉｏｎ），シグモイド関数（ｓｉｇｍｏｉｄ ｆｕｎｃｔｉｏｎ）である。 Second layer: Hidden Layer
There are many forms of the basis function form of the intermediate layer. For example, there are a linear function, a Gaussian function, a logic function, and an exponential function.

Of which
N is the iteration order,

Is the output signal of the second intermediate layer,

Is the input signal of the second layer middle layer,

Is the node function of the second layer middle layer,

Is the weight value of the recurrent unit of the second layer middle layer

Is a weight value connecting the input layer and the intermediate layer in the second layer,
n is the number of neurons in the second interlayer,

Are an activation function and a sigmoid function in the second intermediate layer.

そのうち、
Ｎは反復次数、

は出力層の出力信号、

は第３層出力層中、中間層と出力層間を連結するのに用いられる重み値、

は第３層出力層中の、動態関数（ａｃｔｉｖａｔｉｏｎ ｆｕｎｃｔｉｏｎ）である。

は第３層出力層中のノード関数である。

は第Ｎ次の、該第３層出力層の入力信号である。
Ｋは第３層出力層中のニューロン数である。 Third layer: Output layer

Of which
N is the iteration order,

Is the output signal of the output layer,

Is a weight value used to connect the intermediate layer and the output layer in the third layer output layer,

Is an activation function in the third output layer.

Is a node function in the third output layer.

Is the Nth-order input signal of the third layer output layer.
K is the number of neurons in the third output layer.

をトレーニング形態調整し、リカレント方法を応用し、各１層の誤差を計算更新する。監督式学習（Ｓｕｐｅｒｖｉｓｅｄ ｌｅａｒｎｉｎｇ）の目的は、誤差関数（ｅｒｒｏｒ ｆｕｎｃｔｉｏｎ）Ｅの最小化であり、そのうち、

であり、そのうち、
Ｐwは外在エネルギーパワーであり、
Ｐmはタービンの出力パワーである。 Supervised learning and training process:
Once the RNN is initialized, a gradient descent supervised learning law is used to train the system.
A dynamic back-propagation algorithm is derived as well, and the RNN parameters

The training form is adjusted, the recurrent method is applied, and the error of each layer is calculated and updated. The purpose of supervised learning is to minimize an error function E, of which

Of which
Pw is the external energy power,
Pm is the output power of the turbine.

更新
誤差項は、

そのうち、
ｅmはパワー差値（Ｐw−Ｐm）、

は第３層出力層の出力信号、

は逆伝播された誤差、
Ｅは誤差関数（Ｅｒｒｏｒ ｆｕｎｃｔｉｏｎ）である。

重みは以下の式により調整され、すなわち、

そのうち、

は第３層更新中の、第２層中間層出力と第３層の連結重み値の毎次更新反復差量、

は第２層中間層の出力信号、

は第３層出力層の出力信号、

は第３層出力層中のノード関数、

は逆伝播された誤差、
Eは誤差関数（ｅｒｒｏｒ ｆｕｎｃｔｉｏｎ）である。

重み更新は、

により、そのうち、

は第２層中間層と第３層出力層の連結重み値の毎次修正量であり、

は第３層更新中の、第２層中間層出力と第３層の連結重み値の毎次更新反復差量であり、

は調整パラメータWjkの学習速度である。 3rd layer: Weight value

The update error term is

Of which
em is a power difference value (Pw−Pm),

Is the output signal of the third layer output layer,

Is the backpropagated error,
E is an error function.

The weight is adjusted by the following equation:

Of which

Is the third-order update iteration difference between the second-layer intermediate layer output and the third-layer connection weight value during the third-layer update,

Is the output signal of the second intermediate layer,

Is the output signal of the third layer output layer,

Is the node function in the third output layer,

Is the backpropagated error,
E is an error function.

Weight update is

Of which

Is a correction amount of the connection weight value of the second layer intermediate layer and the third layer output layer,

Is the recursive difference between the second layer intermediate layer output and the third layer connection weight value during the third layer update,

Is the learning speed of the adjustment parameter Wjk.

と

の更新

の適応性規則は、

の適応性規則は、

と

を更新し、

そのうち、
ηjは調整パラメータ

の学習速度であり、
ηijは調整パラメータ

の学習速度であり、

は第２層更新中の、第２層中間層リカレント単位の重み値であり、

は第２層更新中の、第２層中間層重み値の毎次更新反復差量であり、

は第２層更新中の、入力層と中間層を連結する重み値であり、

は第２層更新中の、第２層中間層と第１層出力層を連結する重み値の毎次更新反復差量である。

及び

を調整することで、誤差関数（ｅｒｒｏｒ ｆｕｎｃｔｉｏｎ）

を零に接近させる学習方法を導き出せる。 Second layer:

When

Update

The applicability rule is

The applicability rule is

When

Update

Of which
ηj is the adjustment parameter

The learning speed of
ηij is the adjustment parameter

The learning speed of

Is the weight value of the second layer intermediate layer recurrent unit during the second layer update,

Is the second-order update iteration amount of the second-layer intermediate layer weight value during the second-layer update,

Is a weight value connecting the input layer and the intermediate layer during the second layer update,

Is a recurring difference amount for each update of the weight value connecting the second intermediate layer and the first output layer during the second layer update.

as well as

By adjusting the error function (error function)

A learning method can be derived that approaches to zero.

をトレーニングする。
（ｉ）ステップ６２９：試験を収縮し、誤差関数（ｅｒｒｏｒ ｆｕｎｃｔｉｏｎ）

がすでに最小化されたか否かを決定し、イエスであれば、停止し、ノーであれば、ステップ６２２に進み、反復する。 As shown in the flowchart of FIG. 6B, an on-line training recurrent neural network (RNN) includes the following steps.
(A) Step 621: The amount of change in the recurrent neural network (RNN) is initialized, and the process proceeds to Step 622.
(B) Step 622: Evaluate whether or not to proceed to the structure learning process. If yes, go to step 623, otherwise go to step 628.
(C) Step 623: Evaluate whether or not to increase the node of the new membership functions. If yes, go to step 624, otherwise go to step 628.
(D) Step 624: The new node is increased, and the process proceeds to Step 625.
(E) Step 625: Similarity test is performed on the new node and other nodes.
(F) Step 626: Accept the new node and go to step 628.
(G) Step 627: Delete the new node, and go to Step 628.
(H) Step 628: Using a supervised learning process,

To train.
(I) Step 629: Shrink the test and an error function

If yes, stop and if no, go to step 622 and repeat.

  Accurate prediction of rotational speed is not used in steady state, but is used speedily in traditional proportional-integral (PI) controllers even when uncertain modules of different rotational speeds cannot be determined. obtain.

  In a preferred embodiment, a sliding mode speed controller and an RNN pitch angle controller can be used to control wind energy conversion systems (WECS).

  The above description is only an example of the present invention, and does not limit the scope of the present invention. Any equivalent changes and modifications that can be made based on the scope of the claims of the present invention are all described in the present invention. Shall belong to the scope covered by the rights.

第１層入力層の入力信号

第１層入力層の出力信号

第２層中間層の出力信号

第２層中間層のリカレント単位の重み値

第２層中、入力層と中間層を連結する重み値

第３層中、中間層と出力層を連結する重み値

第３層中間層の出力信号
６１ 永久磁石同期発電機（ＰＭＳＧ）
６２ 監督式学習とトレーニングプロセス（Supervised learning and training process）
Ｐw 外在エネルギーパワー
Ｐm 風力タービンの出力パワー 10 Turbine 11 Gearbox 12 Permanent magnet synchronous generator (PMSG)
13 Power converter
14 DC Load (DC Load) Inverter
31 Permanent magnet synchronous generator (PMSG)
32 Current Control PWM Voltage Source Converter (VSC)
33 Inverter
34 current controller
35 Coordinate translator
36 fuzzy sliding mode speed controller
37 Pitch angle controller (pitch control)
38 AC Grid
40 fuzzy sliding mode speed controller
41 RNN pitch angle controller for on-line training
42 Generator

Input signal of the first layer input layer

Output signal of the first layer input layer

Output signal of the second layer middle layer

Weight of recurrent unit in the second and middle layers

Weight value that connects the input layer and the intermediate layer in the second layer

The weight value that connects the middle layer and the output layer in the third layer

Output signal 61 of the third layer intermediate layer Permanent magnet synchronous generator (PMSG)
62 Supervised learning and training process
Pw External energy power Pm Wind turbine output power

Claims (18)

In the hybrid intelligent power control system,
A fuzzy sliding mode controller that controls the generator by executing a fuzzy inference mechanism;
Recurrent neural network (RNN) pitch angle that achieves maximum power output by performing an on-line training method to control the pitch angle of the turbine coupled to the generator A controller (pitch angle controller),
It encompasses,
A recurrent neural network is a three-layer neural network process that includes an input layer that is a first layer, an intermediate layer that is a second layer, and an output layer that is a third layer,
The base layer form of the intermediate layer is selected from a linear function, a Gaussian function, a logic function, and an exponential function. Power control system. The hybrid intelligent power control system according to claim 1, The pitch angle controller further includes a mechanical actuator, which is used to adjust the pitch angle of the turbine and is an on-line training recurrent neural network ( Hybrid intelligent power, characterized in that the mechanical power output rated value of the turbine is maintained by a recurrent neural network (RNN), and the power of the turbine is not higher than the rated power of the permanent magnet synchronous generator Control system. The hybrid intelligent power control system according to claim 1, The input layer of the first layer is

x _i ¹ (N) is the input signal of the input layer which is the first layer,
net _i ¹ (N) is the node function of the input layer, which is the first layer,
O _i ¹ (N) is the output signal of the input layer, which is the first layer,
N is the iteration order,
f _i ¹ A hybrid intelligent power control system, characterized in that is an activation function or a sigmoid function of the input layer which is the first layer. The hybrid intelligent power control system according to claim 1, The intermediate layer is

Of which
N is the iteration order,
O _j ² Is the output signal of the second layer, the middle layer,
x _i ² (N) is the input signal of the second layer, the middle layer,
net _j ² (N) is the node function of the second layer, the middle layer,
W _j ² Is the weight value of the recurrent unit of the intermediate layer which is the second layer,
W _ij ² Is a weight value connecting the input layer and the intermediate layer in the second layer,
n is the number of neurons in the second layer, the middle layer,
f _j ² Is a hybrid intelligent power control system characterized by being a sigmoid function, a dynamic function in the intermediate layer which is the second layer. The hybrid intelligent power control system according to claim 1, The output layer is

Of which
N is the iteration order,
O _k ^Three (N) is the output signal of the output layer,
W _jk ^Three Is the weight value used to connect between the middle layer and the output layer in the third output layer, f _k ^Three Is the dynamic function in the third layer, the output layer,
net _k ^Three (N) is the node function in the output layer, which is the third layer,
x _j ^Three (N) is the Nth order output signal of the output layer which is the third layer,
k is the number of neurons in the output layer, which is the third layer,
β _c Is the blade pitch angle control command,
A hybrid intelligent power control system characterized by The hybrid intelligent power control system according to claim 2,
The on-line training recurrent neural network (RNN) is a supervised learning and training process that uses a gradient descent supervised learning law to train the system. As well as the dynamic pack propagation algorithm, the RNN parameter W through the training form _jk ^Three , W _j ² , W _ij ² , Applying the recurrent method, calculating and updating the weight value of each layer, the purpose of supervised learning is to minimize the error function E,
Of which

P _w Is the external energy power,
P _m Is the output power of the turbine,
e _m ² Is the square of the power difference ((Pw-Pm) ² ),
A hybrid intelligent power control system characterized by 7. The hybrid intelligent power control system according to claim 6, wherein the third layer has a weight value W. _jk ^Three Update and the error term is

Of which
e _m Is the power difference value (P _w -P _m ),
O _k ^Three Is the output signal of the output layer which is the third layer,
δ _k Is the error due to backpropagation,
E is the error function,
△ W _jk ^Three The weight is adjusted by the following formula:

Of which
△ W _jk ^Three Is the update amount of each successive update of the weight value connecting the intermediate layer of the second layer and the output layer of the third layer during the update of the third layer,
O _j ² Is the output signal of the second intermediate layer,
O _k ^Three Is the output signal of the third layer output layer,
net _k ^Three (N) is the node function in the third output layer,
δ _k Is the error due to backpropagation,
E is the error function,
W _jk ^Three The weight value is W _jk ^Three (N + 1) ＝ W _jk ^Three (N) + η _jk △ W _jk ^Three Updated by (N),
Of which
W _jk ^Three (N + 1) is the weight correction value for connecting the intermediate layer as the second layer and the output layer as the third layer,
η _jk Is the adjustment parameter W _jk A hybrid intelligent power control system, characterized by its learning speed. The hybrid intelligent power control system according to claim 6, wherein the second layer updates W _j ² and W _ij ² ,
The adaptability rule for W _j ² is

The adaptability rule for W _ij ² is

Update W _j ² and W _ij ²
W _j ² (N + 1) ＝ W _j ² (N) + η _j △ W _j ² (N)
W _ij ² (N + 1) = W _ij ² (k) + η _ij △ W _ij ² (N)
Of which
η _j is the learning speed of the adjustment parameter W _j ²
η _ij is the learning speed of the adjustment parameter W _ij ² ,
W _j ² is the weight value of the second layer intermediate layer recurrent unit during the second layer update,
△ W _j ² is the second-order intermediate layer weight value of each second-order update iteration difference during the second layer update,
W _ij ² is a weight value connecting the input layer and the intermediate layer during the second layer update,
△ W _jk ² is the second order update iteration of the second order update iteration difference weight connection weight value of the second layer middle layer and the third layer output layer during the second layer update,
By adjusting W _j ² , W _ij ² and W _jk

A hybrid intelligent power control system, characterized by deriving a learning method that brings the zero close to zero. The hybrid intelligent power control system according to claim 6, wherein a recurrent neural network (RNN) for on-line training is:
(A) Step 621: Initialize a recurrent neural network (RNN) change amount, and proceed to Step 622;
(B) Step 622: Evaluate whether to proceed to the structure learning process. If yes, go to step 623; if no, go to step 628.
(C) Step 623: Evaluate whether to increase the nodes of the new membership functions. If yes, go to step 624; if no, go to step 628.
(D) Step 624: Increase the new node, go to step 625,
(E) Step 625: Similarity test is performed on the new node and other nodes
(F) Step 626: Accept new node, go to step 628,
(G) Step 627: Delete the new node and go to Step 628.
(H) Step 628: Using a supervised learning process, W _j ² , W _ij ² And W _jk Train and
(I) Step 629: Convergence of the test and an error function

Hybrid intelligent power control system comprising the steps of: determining whether or not has already been minimized, stopping if yes, proceeding to step 622 and repeating if no . In the hybrid intelligent power control method,
Controlling the generator by executing a fuzzy inference mechanism using a fuzzy sliding mode controller;
Controlling the pitch angle of a turbine coupled to the generator by performing an on-line training method using a recurrent neural network (RNN) pitch angle controller. And achieving the maximum power output,
Including
The recurrent neural network (RNN) is a three-layer neural network process including an input layer as a first layer, an intermediate layer as a second layer, and an output layer as a third layer.
The base layer form of the intermediate layer is selected from a linear function, a Gaussian function, a logic function, and an exponential function. Power control method. 11. The hybrid intelligent power control method according to claim 10, wherein the pitch angle controller further includes a mechanical actuator, which is used to adjust the pitch angle of the turbine. On-line training, a recurrent neural network (RNN) maintains the turbine's mechanical power output rated value, and the turbine's power is the permanent magnet synchronous generator rating. A hybrid intelligent power control method characterized by not being higher than power. The hybrid intelligent power control method according to claim 10, wherein the input layer of the first layer is:

x _i ¹ (N) is the input signal of the input layer which is the first layer,
net _i ¹ (N) is the node function of the input layer, which is the first layer,
O _i ¹ (N) is the output signal of the input layer, which is the first layer,
N is the iteration order,
f _i ¹ A hybrid intelligent power control method characterized in that is an activation function or a sigmoid function of an input layer as the first layer. The hybrid intelligent power control method according to claim 10, wherein the intermediate layer comprises:

Of which
N is the iteration order,
O _j ² Is the output signal of the second layer, the middle layer,
x _i ² (N) is the input signal of the second layer, the middle layer,
net _j ² (N) is the node function of the second layer, the middle layer,
W _j ² Is the weight value of the recurrent unit of the intermediate layer which is the second layer,
W _ij ² Is a weight value connecting the input layer and the intermediate layer in the second layer,
n is the number of neurons in the second layer, the middle layer,
f _j ² A hybrid intelligent power control method characterized in that is a sigmoid function or a dynamic function in the intermediate layer which is the second layer. The hybrid intelligent power control method according to claim 10, wherein the output layer comprises:

Of which
N is the iteration order,
O _k ^Three (N) is the output signal of the output layer,
W _jk ^Three Is the weight value used to connect between the middle layer and the output layer in the third output layer, f _k ^Three Is the dynamic function in the third layer, the output layer,
net _k ^Three (N) is the node function in the output layer, which is the third layer,
x _j ^Three (N) is the Nth order output signal of the output layer which is the third layer,
k is the number of neurons in the output layer, which is the third layer,
β _c Is the blade pitch angle control command,
A hybrid intelligent power control method characterized by 12. The hybrid intelligent power control method according to claim 11, wherein the on-line training recurrent neural network (RNN) is a supervised learning and training process, The supervisory learning law of gradient descent is used for training of this system, and with the dynamic pack propagation algorithm, the parameter W _jk ^Three , W _j ² , W _ij ² , Applying the recurrent method, calculating and updating the weight value of each layer, the purpose of supervised learning is to minimize the error function E,
Of which

P _w Is the external energy power,
P _m Is the output power of the turbine,
e _m ² Is the square of the power difference ((Pw-Pm) ² ),
A hybrid intelligent power control method characterized by 16. The hybrid intelligent power control method according to claim 15, wherein the third layer has a weight value W. _jk ^Three Update and the error term is

Of which
e _m Is the power difference value (P _w -P _m ),
O _k ^Three Is the output signal of the output layer which is the third layer,
δ _k Is the error due to backpropagation,
E is the error function,
△ W _jk ^Three The weight is adjusted by the following formula:

Of which
△ W _jk ^Three Is the update amount of each successive update of the weight value connecting the intermediate layer of the second layer and the output layer of the third layer during the update of the third layer,
O _j ² Is the output signal of the second intermediate layer,
O _k ^Three Is the output signal of the third layer output layer,
net _k ^Three (N) is the node function in the third output layer,
δ _k Is the error due to backpropagation,
E is the error function,
W _jk ^Three The weight value is W _jk ^Three (N + 1) ＝ W _jk ^Three (N) + η _jk △ W _jk ^Three Updated by (N),
Of which
W _jk ^Three (N + 1) is the weight correction value for connecting the intermediate layer as the second layer and the output layer as the third layer,
η _jk Is the adjustment parameter W _jk A hybrid intelligent power control method characterized by the learning speed of The hybrid intelligent power control method according to claim 15, wherein the second layer updates W _j ² and W _ij ² ,
Of which
The adaptability rule for W _j ² is

The adaptability rule for W _ij ² is

Update W _j ² and W _ij ²
W _j ² (N + 1) ＝ W _j ² (N) + η _j △ W _j ² (N)
W _ij ² (N + 1) = W _ij ² (k) + η _ij △ W _ij ² (N)
Of which
η _j is the learning speed of the adjustment parameter W _j ²
η _jk is the learning speed of the adjustment parameter W _ij ²
W _j ² is the weight value of the second layer intermediate layer recurrent unit during the second layer update,
△ W _j ² is the second-order intermediate layer weight value of each second-order update iteration difference during the second-layer update,
W _ij ² is a weight value connecting the input layer and the intermediate layer during the second layer update,
△ W _jk ² is the second order update iteration of the second order update iteration difference weight connection weight value of the second layer middle layer and the third layer output layer during the second layer update,
By adjusting W _j ² , W _ij ² and W _jk ² , the error function

A hybrid intelligent power control method, characterized by deriving a learning method that brings a squeeze closer to zero. The hybrid intelligent power control method according to claim 15, wherein a recurrent neural network (RNN) for on-line training is:
(A) Step 621: Initialize a recurrent neural network (RNN) change amount, and proceed to Step 622;
(B) Step 622: Evaluate whether to proceed to the structure learning process. If yes, go to step 623; if no, go to step 628.
(C) Step 623: Evaluate whether to increase the nodes of the new membership functions. If yes, go to step 624; if no, go to step 628.
(D) Step 624: Increase the new node, go to step 625,
(E) Step 625: Similarity test is performed on the new node and other nodes
(F) Step 626: Accept new node, go to step 628,
(G) Step 627: Delete the new node and go to Step 628.
(H) Step 628: Using a supervised learning process, W _j ² , W _ij ² And W _jk Train and
(I) Step 629: Convergence of the test and an error function

If yes, stop if yes, go to step 622 and repeat if no
A hybrid intelligent power control method comprising the above steps.

Images