Controlling a turbine with a recurrent neural network
원문보기
IPC분류정보
국가/구분
United States(US) Patent
등록
국제특허분류(IPC7판)
G06F-019/00
G05B-013/02
F04D-027/00
G06N-003/04
출원번호
US-0396337
(2013-04-08)
등록번호
US-9639070
(2017-05-02)
우선권정보
DE-10 2012 206 651 (2012-04-23)
국제출원번호
PCT/EP2013/057307
(2013-04-08)
국제공개번호
WO2013/160090
(2013-10-31)
발명자
/ 주소
Düll, Siegmund
Udluft, Steffen
Weichbrodt, Lina
출원인 / 주소
Siemens Aktiengesellschaft
대리인 / 주소
Lempia Summerfield Katz LLC
인용정보
피인용 횟수 :
0인용 특허 :
4
초록▼
A method for controlling a turbine is proposed, which is characterized at any point in the control by a hidden state. The dynamic behavior of the turbine is modeled with a recurrent neural network comprising a recurrent hidden layer. In this case, the recurrent hidden layer is formed from vectors of
A method for controlling a turbine is proposed, which is characterized at any point in the control by a hidden state. The dynamic behavior of the turbine is modeled with a recurrent neural network comprising a recurrent hidden layer. In this case, the recurrent hidden layer is formed from vectors of neurons, which describe the hidden state of the turbine at the time points of the regulation, wherein two vectors are chronologically linked for each time point with a first connection bridging a time and second connection bridging at least two points in time. Short-term effects can be controlled by means of the first connections and long-term effects can be adjusted by means of the second connections. Secondly, emissions and also occurring dynamics in the turbine can be minimized. Furthermore, a regulating device and a turbine with such a regulating device are proposed.
대표청구항▼
1. A method for controlling a turbine, the turbine comprising a plurality of sensors for providing sensor values ascertained on the turbine and a plurality of actuators for actuating the turbine, the turbine being characterized at each instant from a plurality of chronological instants of control by
1. A method for controlling a turbine, the turbine comprising a plurality of sensors for providing sensor values ascertained on the turbine and a plurality of actuators for actuating the turbine, the turbine being characterized at each instant from a plurality of chronological instants of control by a hidden state that is derivable by sensor values and a rating signal for the hidden state and is influenceable by alterable actuator values for the plurality of actuators, the method comprising: modeling dynamic behavior of the turbine with a recurrent neural network comprising an input layer, a recurrent hidden layer, and an output layer based on training data comprising sensor values of the plurality of sensors, actuator values of the plurality of actuators, and rating signals, wherein the input layer is formed from first vectors of neurons that describe sensor values, actuator values, or sensor values and actuator values at the instants, wherein the recurrent hidden layer is formed from second vectors of neurons that describe the hidden state of the turbine at the instants, wherein chronologically, for all the instants, two respective vectors from the second vectors are connected to a first connection that spans one instant, and, chronologically, two respective vectors from the second vectors are connected to a second connection that spans at least two instants, and wherein the output layer is formed from at least one third vector of neurons that describe the rating signal or at least one portion of the sensor values, at least one portion of the actuator values, or the at least one portion of the sensor values and the at least one portion of the actuator values at the instants;performing a learning method, an optimization method, or a learning and optimization method on the hidden states in order to provide a set of rules having optimized actuator values for each of the hidden states;ascertaining the current hidden state using the recurrent neural network and currently ascertained sensor values from the plurality of sensors;ascertaining current actuator values; andactuating the plurality of actuators using the provided set of rules and the current hidden state. 2. The method of claim 1, wherein the recurrent neural network is a deployed recurrent neural network. 3. The method of claim 1, wherein apart from the current hidden state at a current instant, all the hidden states at past and future instants are spanned by a second connection. 4. The method of claim 1, wherein the second connections are arranged symmetrically about the current hidden state. 5. The method of claim 1, wherein the recurrent neural network is divided into two subnetworks that are connected by the current hidden state, and wherein a first subnetwork of the two subnetworks describes the hidden states for past instants, and a second subnetwork of the two subnetworks describes the hidden states for future instants. 6. The method of claim 3, wherein the first connections and the second connections are formed by neural structures comprising matrices and further vectors of neurons. 7. The method of claim 3, wherein a respective length L is set for the second connections, so that a predetermined number of instants is spanned. 8. The method of claim 7, wherein the length of the second connections is set by a heuristic, and wherein the length of the second connections is chosen such that a total number of necessary steps for reaching the current hidden state from the hidden states of the past instants is minimal. 9. The method of claim 1, wherein the learning method, the optimization method, or the learning and optimization method is a computer learning method. 10. The method of claim 9, wherein the learning method, the optimization method, or the learning and optimization method comprises dynamic programming, prioritized sweeping, Q learning, a recurrent control neural network (RCNN), or any combination thereof. 11. The method of claim 1, wherein the rating signal is represented by a rating function that is dependent at least on a portion of the sensor values, the actuator values, or the sensor values and the actuator values. 12. The method of claim 1, wherein the learning method, the optimization method, or the learning and optimization method uses the rating signals in order to implement the actuator values for an optimum rating signal. 13. The method of claim 1, wherein the turbine is a gas turbine, wherein the rating signal is determined at least by an efficiency of the gas turbine, a pollutant emission of the gas turbine, or the efficiency and the pollutant emission of the gas turbine, alternating pressures, mechanical, thermal, or mechanical and thermal loads on a combustion chamber of the gas turbine, or any combination thereof. 14. The method of claim 1, wherein the turbine is a wind turbine, and wherein the rating signal is determined at least by force loading, alternating loading, or force loading and alternating loading on one or more rotor blades of the wind turbine. 15. A computer program product having a program code stored on a non-transitory machine-readable medium and executable by a computer to control a turbine, the turbine comprising a plurality of sensors for providing sensor values ascertained on the turbine and a plurality of actuators for actuating the turbine, the turbine being characterized at each instant from a plurality of chronological instants of control by a hidden state that is derivable by sensor values and a rating signal for the hidden state and is influenceable by alterable actuator values for the plurality of actuators, the program code including instructions comprising: modeling dynamic behavior of the turbine with a recurrent neural network comprising an input layer, a recurrent hidden layer, and an output layer based on training data comprising sensor values of the plurality of sensors, actuator values of the plurality of actuators, and rating signals, wherein the input layer is formed from first vectors of neurons that describe sensor values, actuator values, or sensor values and actuator values at the instants, wherein the recurrent hidden layer is formed from second vectors of neurons that describe the hidden state of the turbine at the instants, wherein chronologically, for all the instants, two respective vectors from the second vectors are connected to a first connection that spans one instant, and, chronologically, two respective vectors from the second vectors are connected to a second connection that spans at least two instants, and wherein the output layer is formed from at least one third vector of neurons that describe the rating signal or at least one portion of the sensor values, at least one portion of the actuator values, or the at least one portion of the sensor values and the at least one portion of the actuator values at the instants;performing a learning method, an optimization method, or a learning and optimization method on the hidden states in order to provide a set of rules having optimized actuator values for each of the hidden states;ascertaining the current hidden state using the recurrent neural network and currently ascertained sensor values from the plurality of sensors;ascertaining current actuator values; andactuating the plurality of actuators using the provided set of rules and the current hidden state. 16. The computer program product of claim 15, wherein the recurrent neural network is a deployed recurrent neural network. 17. The computer program product of claim 16, wherein apart from the current hidden state at a current instant, all the hidden states at past and future instants are spanned by a second connection. 18. The computer program product of claim 17, wherein the second connections are arranged symmetrically about the current hidden state. 19. The method of claim 6, wherein the neural structures comprise matrices of weights or neural connections. 20. The method of claim 9, wherein the computer learning method is a strengthening learning method.
연구과제 타임라인
LOADING...
LOADING...
LOADING...
LOADING...
LOADING...
이 특허에 인용된 특허 (4)
Rajamani Ravi ; Chbat Nicolas Wadih ; Ashley Todd Alan, Controller with neural network for estimating gas turbine internal cycle parameters.
Schäfer, Anton Maximilian; Udluft, Steffen; Zimmermann, Hans-Georg, Method for the computer-assisted control and/or regulation of a technical system where the dynamic behavior of the technical system is modeled using a recurrent neural network.
※ AI-Helper는 부적절한 답변을 할 수 있습니다.