Unsteady state system control method and related device
By extracting and training the data features of the non-steady-state system and using the forward propagation and backpropagation functions, the optimal state feedback matrix is quickly calculated, which solves the problems of large computational complexity and low efficiency in the existing technology and achieves efficient control of the stability of the non-steady-state system.
Patent Information
- Application Number
- CN202510843585.1
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2025-06-23
- Publication Date
- 2025-09-12
AI Technical Summary
In existing technologies, reinforcement learning methods based on learning have large computational complexity and low efficiency, making it difficult to efficiently control the stability of non-steady-state systems, especially in the presence of additive noise interference.
By extracting features from the data of the non-steady-state system, the state vector is obtained, the system model is trained using the forward propagation and backpropagation functions, and the optimal state feedback matrix is calculated to reduce the amount of calculation and improve the control efficiency.
The optimal state feedback matrix can be quickly obtained under a small number of state vectors, which reduces the amount of calculation, improves the control efficiency of the non-steady-state system, and improves the applicability of the model.
Smart Images
Figure CN120630701A_ABST
Abstract
Description
Technical Field
[0001] The present application relates to the field of control technology, and in particular to a control method and related devices for an unsteady-state system. Background Art
[0002] Solving the linear quadratic regulator (LQR) problem involves calculating a state feedback matrix and, based on this state feedback matrix, adjusting the initial unstable system to a stable state at minimal cost, thereby ensuring stable operation of the system.
[0003] Currently, learning-based reinforcement learning can generate a state feedback matrix and regulate an unsteady system to a stable state. However, this method requires a large-scale search of the problem space using a large amount of sample data to find the optimal state feedback matrix. This process is computationally intensive and reduces the efficiency of controlling unsteady systems.
[0004] Based on this, a solution is urgently needed to solve the above technical problems. Summary of the Invention
[0005] The present application provides a control method and related devices for an unsteady-state system, the purpose of which is to improve the control efficiency of the unsteady-state system.
[0006] To achieve the above objectives, this application adopts the following technical solutions:
[0007] First aspect: An embodiment of the present application provides a control method for a non-steady-state system, which obtains a state vector of the non-steady-state system by extracting features from data of the non-steady-state system; uses the state vector of the non-steady-state system as a system model input to determine an optimal state feedback matrix corresponding to the non-steady-state system; the system model uses a random state vector as an initial system model input, and an initial state feedback matrix as an initial parameter of the initial system model, and trains the initial system model based on a forward propagation function, a backward propagation function, and a weight update function; the initial state feedback matrix is obtained by calculating a discrete-time algebraic Riccati equation solution function using a system matrix, an input matrix, a state weight matrix, and a control weight matrix as inputs; and the non-steady-state system is controlled based on the optimal state feedback matrix.
[0008] In the method provided in the embodiment of the present application, based on the system model with the initial state feedback matrix as the initial parameter, the process of controlling the non-steady-state system only requires a small number of state vectors to quickly obtain the optimal state feedback matrix corresponding to the non-steady-state system, which effectively reduces the amount of calculation and improves the control efficiency of the non-steady-state system.
[0009] In a possible implementation, the step of training the system model includes:
[0010] Inputting a random state vector into an initial system model, using an initial state feedback matrix as an initial parameter of the initial system model, and generating a first matrix based on a forward propagation function;
[0011] Obtaining gradient data based on the first matrix and the back propagation function;
[0012] Based on the gradient data, the weight update function value and the initial state feedback matrix are updated to generate the trained system model.
[0013] In one possible implementation, updating the network weights and the initial state feedback matrix based on the gradient data to generate the trained system model includes:
[0014] Based on the gradient data, the weight update function value and the initial state feedback matrix are updated to generate an updated state feedback matrix, an updated weight update function value, and an updated system model;
[0015] The random state vector, the updated state feedback matrix, the system matrix, the input matrix, the state weight matrix and the control weight matrix are used as inputs of the forward propagation function to generate the second matrix;
[0016] Obtaining updated gradient data based on the second matrix and the back propagation function;
[0017] Based on the updated gradient data, the updated weight update function value and the updated state feedback matrix are updated to generate the trained system model.
[0018] In a possible implementation, obtaining gradient data based on the first matrix and the back propagation function includes:
[0019] Determining a loss function based on the first matrix;
[0020] The loss function is used as the input of the back propagation function to obtain gradient data.
[0021] In a possible implementation, determining a loss function based on the first matrix includes:
[0022] Obtaining a minimized quadratic cost function based on the sum of the squares of the norms of each element in the first matrix, wherein each element is represented based on a state weight matrix and a control weight matrix;
[0023] The minimized quadratic cost function multiplied by 1 / (N+1) is determined as a loss function; wherein N is an integer greater than or equal to 0.
[0024] In a possible implementation, before inputting the random state vector into the initial system model and using the initial state feedback matrix as the initial parameter of the initial system model and generating the first matrix based on the forward propagation function, the method further includes:
[0025] The unsteady system is modeled to generate an initial system model; the initial system model is represented by a z-dimensional state vector of the unsteady system at time k, an m-dimensional input vector of the unsteady system at time k, a system matrix, an input matrix, and an uncertainty vector; wherein k is an integer greater than or equal to; z and m are positive integers.
[0026] In a possible implementation, controlling the non-steady-state system based on the optimal state feedback matrix includes:
[0027] Based on the optimal state feedback matrix, feedback adjustment is performed on the non-steady-state system to adjust the non-steady-state system to a steady-state system.
[0028] In a second aspect, an embodiment of the present application provides a control device for an unsteady-state system, comprising:
[0029] an extracting unit, a determining unit, and a controlling unit;
[0030] The extraction unit is used to extract features from the data of the non-steady-state system to obtain a state vector of the non-steady-state system;
[0031] The determining unit is configured to determine an optimal state feedback matrix corresponding to the non-steady-state system using a state vector of the non-steady-state system as a system model input; the system model is obtained by training the initial system model based on a forward propagation function, a backward propagation function, and a weight update function using a random state vector as an initial system model input and an initial state feedback matrix as an initial parameter of the initial system model; the initial state feedback matrix is obtained by calculating a discrete-time algebraic Riccati equation solution function using a system matrix, an input matrix, a state weight matrix, and a control weight matrix as inputs;
[0032] The control unit is used to control the non-steady-state system based on the optimal state feedback matrix.
[0033] In a possible implementation, the apparatus includes a training unit;
[0034] The training unit is specifically used to:
[0035] Inputting a random state vector into an initial system model, using an initial state feedback matrix as an initial parameter of the initial system model, and generating a first matrix based on a forward propagation function;
[0036] Obtaining gradient data based on the first matrix and the back propagation function;
[0037] Based on the gradient data, the weight update function value and the initial state feedback matrix are updated to generate the trained system model.
[0038] In a possible implementation, the training unit is specifically configured to:
[0039] Based on the gradient data, the weight update function value and the initial state feedback matrix are updated to generate an updated state feedback matrix, an updated weight update function value, and an updated system model;
[0040] The random state vector, the updated state feedback matrix, the system matrix, the input matrix, the state weight matrix and the control weight matrix are used as inputs of the forward propagation function to generate the second matrix;
[0041] Obtaining updated gradient data based on the second matrix and the back propagation function;
[0042] Based on the updated gradient data, the updated weight update function value and the updated state feedback matrix are updated to generate the trained system model.
[0043] In a possible implementation, the training unit is specifically configured to:
[0044] Determining a loss function based on the first matrix;
[0045] The loss function is used as the input of the back propagation function to obtain gradient data.
[0046] In a possible implementation, the training unit is specifically configured to:
[0047] Obtaining a minimized quadratic cost function based on the sum of the squares of the norms of each element in the first matrix, wherein each element is represented based on a state weight matrix and a control weight matrix;
[0048] The minimized quadratic cost function multiplied by 1 / (N+1) is determined as a loss function; wherein N is an integer greater than or equal to 0.
[0049] In a possible implementation, the training unit is specifically configured to:
[0050] The unsteady system is modeled to generate an initial system model; the initial system model is represented by a z-dimensional state vector of the unsteady system at time k, an m-dimensional input vector of the unsteady system at time k, a system matrix, an input matrix, and an uncertainty vector; wherein k is an integer greater than or equal to; z and m are positive integers.
[0051] In a possible implementation, the control unit is specifically configured to:
[0052] Based on the optimal state feedback matrix, feedback adjustment is performed on the non-steady-state system to adjust the non-steady-state system to a steady-state system.
[0053] A third aspect: An embodiment of the present application provides a computer device, the computer device comprising: a processor and a memory;
[0054] The memory is used to store program code and transmit the program code to the processor;
[0055] The processor is configured to execute the steps of the control method for a non-steady-state system according to the instructions in the program code.
[0056] A fourth aspect: An embodiment of the present application provides a computer-readable storage medium, on which a computer program is stored. When the computer program is executed by a processor, the steps of the control method of a non-steady-state system as described above are implemented.
[0057] Fifth aspect: An embodiment of the present application provides a program product. When the program product runs on a computer, the computer can execute the steps of the control method of a non-steady-state system as described above. BRIEF DESCRIPTION OF THE DRAWINGS
[0058] Figure 1 A schematic diagram of an application scenario of a control method for a non-steady-state system provided in an embodiment of the present application;
[0059] Figure 2 A flow chart of a control method for an unsteady-state system provided in an embodiment of the present application;
[0060] Figure 3 A network architecture diagram of a system model provided in an embodiment of the present application;
[0061] Figure 4 A schematic structural diagram of a control device for an unstable system is provided for an embodiment of the present application. DETAILED DESCRIPTION
[0062] The terms "first", "second" and "third" in the specification, claims and drawings of this application are used to distinguish different objects rather than to limit a specific order.
[0063] In the embodiments of this application, words such as "exemplary" or "for example" are used to indicate examples, illustrations, or descriptions. Any embodiment or design described as "exemplary" or "for example" in the embodiments of this application should not be interpreted as being preferred or advantageous over other embodiments or designs. Rather, the use of words such as "exemplary" or "for example" is intended to present the relevant concepts in a concrete manner.
[0064] Currently, to regulate an unsteady system to a steady state, the optimal state feedback matrix can be obtained by explicitly solving the Riccati equation. However, this method requires sufficient explicit model knowledge of the unsteady system being controlled. For unsteady systems with additive noise interference, explicit model knowledge is often lacking. In this case, it is difficult to control unsteady systems with additive noise interference by explicitly solving the Riccati equation.
[0065] For non-steady-state systems with a lack of explicit model cognition, an optimal state feedback matrix can be obtained based on learning-based reinforcement learning, so as to adjust the non-steady-state system to a stable state based on the optimal state feedback matrix.
[0066] However, the method of obtaining the optimal state feedback matrix based on learning-based reinforcement learning has the following three main problems: First, since the model is obtained through training under the condition of known additive noise interference distribution, this method is only applicable to the case where there is additive noise interference with a specific distribution in the non-steady-state system; Second, since this method requires a large amount of sample data to conduct a large-scale search of the problem space to find the optimal state feedback matrix, as the model dimension increases, the scope of the search in the problem space will increase, the depth of the reinforcement learning network and the volume of parameters will increase exponentially, the amount of calculation will increase, and the control efficiency of the non-steady-state system will be reduced; Third, the reinforcement learning process requires a large amount of sample data, the learning efficiency is low, and this can also lead to a decrease in the control efficiency of the non-steady-state system.
[0067] Based on this, an embodiment of the present application provides a control method and related device for an unsteady-state system, which extracts features from the data of the unsteady-state system to obtain a state vector of the unsteady-state system; uses the state vector of the unsteady-state system as the input of the system model to determine the optimal state feedback matrix corresponding to the unsteady-state system; and controls the unsteady-state system based on the optimal state feedback matrix. The system model is obtained by training the initial system model based on a forward propagation function, a backward propagation function, and a weight update function using a random state vector as the input of the initial system model and an initial state feedback matrix as the initial parameters of the initial system model; the initial state feedback matrix is obtained by calculating a discrete-time algebraic Riccati equation solution function using a system matrix, an input matrix, a state weight matrix, and a control weight matrix as inputs.
[0068] In the method provided in the embodiment of the present application, based on the system model with the initial state feedback matrix as the initial parameter, the process of controlling the non-steady-state system only requires a small number of state vectors to quickly obtain the optimal state feedback matrix corresponding to the non-steady-state system, which effectively reduces the amount of calculation and improves the control efficiency of the non-steady-state system.
[0069] The following combination Figure 1 The application scenario of a control method for an unsteady-state system provided in an embodiment of the present application is introduced. Figure 1 As shown, the figure is a schematic diagram of an application scenario of a control method for a non-steady-state system provided in an embodiment of the present application.
[0070] In the embodiment of the present application, the client 1101 of the non-steady-state system can interact with the server 1102 equipped with the system model.
[0071] Exemplarily, the client 1101 may send an adjustment instruction to the server 1102 , where the adjustment instruction is used to instruct the server to solve the optimal state feedback matrix of the non-steady-state system.
[0072] In response to the adjustment instruction, the server 1102 extracts features from the data of the non-steady-state system to obtain a state vector of the non-steady-state system; uses the state vector of the non-steady-state system as a system model input to determine an optimal state feedback matrix corresponding to the non-steady-state system; and based on the optimal state feedback matrix, sends a control instruction to the client 1101 to control the non-steady-state system.
[0073] In which, the system model is obtained by training the initial system model with a random state vector as the input of the initial system model and an initial state feedback matrix as the initial parameter of the initial system model based on a forward propagation function, a backpropagation function, and a weight update function; the initial state feedback matrix is obtained by calculating the discrete-time algebraic Riccati equation solution function with the system matrix, input matrix, state weight matrix and control weight matrix as input.
[0074] The following combination Figure 2 A control method for an unsteady-state system provided in an embodiment of the present application is introduced. Figure 2 As shown, the figure is a flow chart of a control method for a non-steady-state system provided in an embodiment of the present application, including S201-S203.
[0075] S201 , extracting features from data of an unsteady-state system to obtain a state vector of the unsteady-state system.
[0076] In the embodiment of the present application, feature extraction may be performed on data that can reflect the operating state of the non-steady-state system to obtain a state vector of the non-steady-state system.
[0077] S202: Using the state vector of the non-steady-state system as a system model input, determining an optimal state feedback matrix corresponding to the non-steady-state system.
[0078] The system model is obtained by training the initial system model based on a forward propagation function, a backward propagation function, and a weight update function, using a random state vector as the input of the initial system model and an initial state feedback matrix as the initial parameter of the initial system model.
[0079] In one possible implementation, the present application provides a joint model-based recurrent neural network (J-MoRNN) and learning algorithm based on a recurrent neural network (RNN). The J-MoRNN utilizes the similarity between the architecture mechanisms of a linear time-invariant system and RNN to design a network architecture of the system model. Figure 3 As shown, this figure is a network architecture diagram of a system model provided in an embodiment of the present application.
[0080] Since the network architecture of the system model in this application is similar to the architecture of RNN, it does not require specific assumptions about the introduced additive noise interference. At the same time, the system model can be trained based on fewer parameters, which reduces the amount of calculation and can effectively provide control efficiency for non-steady-state systems.
[0081] The following describes the training steps of the system model. The training steps of the system model include steps 21 to 23:
[0082] Step 21: Inputting a random state vector into an initial system model, using an initial state feedback matrix as an initial parameter of the initial system model, and generating a first matrix based on a forward propagation function.
[0083] In an embodiment of the present application, a random state vector is input into an initial system model, and an initial state feedback matrix is used as an initial parameter of the initial system model. Based on the forward propagation function, before generating the first matrix, the non-steady-state system can be modeled to generate an initial system model. The initial system model can be represented by a z-dimensional state vector of the non-steady-state system at time k, an m-dimensional input vector of the non-steady-state system at time k, a system matrix, an input matrix, and an uncertainty vector; wherein k is an integer greater than or equal to; z and m are positive integers.
[0084] For example, the initial system model can be expressed as follows by equation (1):
[0085] x(k+1)=Ax(k)+Bu(k)+δ(k) (1)
[0086] Where x(k) is the z-dimensional state vector of the unsteady system at time k, u(k) is the m-dimensional input vector of the unsteady system at time k, A and B are the known system matrix and input matrix respectively, δ(k) is an uncertainty vector.
[0087] In the embodiment of the present application, δ(k) can be additive noise interference of any distribution, which improves the applicability of the model compared to the method of obtaining the optimal state feedback matrix based on reinforcement learning.
[0088] In one possible implementation, the random state vector can be expressed as X(0)=[x 0 (0), x 1 (0),……,x i (0),……,x l-1 (0)] T , l is a positive integer.
[0089] The initial state feedback matrix can be calculated by the discrete-time algebraic Riccati equation solver, the system matrix, the input matrix, the state weight matrix, and the control weight matrix.
[0090] For example, the discrete linear time-invariant system can be expressed as follows by equation (2):
[0091] x(k+1)=Ax(k)+Bu(k) (2)
[0092] In order to make the system run at a minimum cost, u(k) can be controlled by the initial state feedback matrix, which is expressed as follows by formula (3):
[0093] u(k)=-Kx(k) (3)
[0094] Among them, K represents the initial state feedback matrix,
[0095] Solving K requires minimizing the quadratic cost function, as shown in formula (4):
[0096]
[0097] Among them, Q and R are the state weight matrix and control weight matrix respectively,
[0098] Based on minimizing the quadratic cost function, K can be obtained as shown in formula (5):
[0099] K=(R+B T PB) -1 (B T PA) (5)
[0100] Among them, P can be obtained by solving the discrete-time algebraic Riccati equation (DARE) function, as shown in formula (6):
[0101] P=A T PA-(A T PB)(R+B T PB) T (B T PA)+Q (6)
[0102] In the embodiment of the present application, after determining the random state vector and the initial state feedback matrix, the random state vector, the initial state feedback matrix, the system matrix, the input matrix, the state weight matrix, and the control weight matrix can be used as parameters of the forward propagation function to generate a first matrix, as shown in formula (7):
[0103] Forward(x i (0),K,A,B,Q,R)→O(k)
[0104] O(k)=[o(0), o(1),…,o(k),…,o(N)] T (7)
[0105] Among them, Forward() is the forward propagation function; O(k) is the first matrix,
[0106] In the embodiment of the present application, using the initial state feedback matrix as the initial parameter of the initial system model can effectively reduce the number of state vectors required for training.
[0107] For example, in the embodiment of the present application, the number of random state vectors may be less than or equal to 10. Compared with the sample data required for reinforcement learning, the method provided in the embodiment of the present application can effectively reduce the amount of calculation and improve the control efficiency of the non-steady-state system.
[0108] Step 22: Obtain gradient data based on the first matrix and the back propagation function.
[0109] In the embodiment of the present application, a loss function can be determined based on the first matrix. After the loss function is determined, the loss function is used as an input of the back propagation function to obtain gradient data.
[0110] Exemplarily, based on the sum of the squares of the norms of each element in the first matrix, a minimized quadratic cost function can be obtained, and the minimized quadratic cost function multiplied by 1 / (N+1) is determined as the loss function, as shown in formula (8):
[0111]
[0112] Where L represents the loss function and N is an integer greater than or equal to 0.
[0113] After determining the loss function, the loss function can be used as the input of the back propagation function to obtain gradient data, as shown in formula (9):
[0114] Backward(L(O i ))→▽ K L (9)
[0115] Among them, Backward() represents the back propagation function, ▽ K L represents the gradient data.
[0116] In the embodiment of the present application, a loss function is determined based on the first matrix, which accelerates the iteration of model parameters and improves learning efficiency. Furthermore, since the initial parameters of the system model are the initial state feedback matrix, which is obtained based on a noiseless system, using this initial state feedback matrix as a priori indication allows the optimal state feedback matrix to be quickly obtained using only a small number of state vectors.
[0117] Step 23: Based on the gradient data, the weight update function value and the initial state feedback matrix are updated to generate the trained system model.
[0118] Among them, the weight update function can be expressed as Update(K, ▽ K L).
[0119] In an embodiment of the present application, the weight update function value can be updated based on the gradient data, thereby updating the initial state feedback matrix and generating the trained system model.
[0120] In one possible implementation, based on gradient data, the weight update function value and the initial state feedback matrix are updated to generate an updated state feedback matrix, an updated weight update function value, and an updated system model. A random state vector, an updated state feedback matrix, a system matrix, an input matrix, a state weight matrix, and a control weight matrix are used as inputs to a forward propagation function to generate a second matrix. Based on the second matrix and a backward propagation function, updated gradient data is obtained. Based on the updated gradient data, the updated weight update function value and the updated state feedback matrix are updated to generate the trained system model.
[0121] In the embodiment of the present application, each x i (0) Input the initial system model, perform forward propagation, backpropagation, and weight update in sequence to update the initial state feedback matrix, obtain the updated state feedback matrix and the updated system model. On this basis, continue to train the updated system model and update the updated state feedback matrix.
[0122] For example, the training process can be repeated multiple times, and the state feedback matrix will be updated in each round of training until each x i (0) After multiple trainings, the state feedback matrix is updated and the final state feedback matrix is obtained.
[0123] It is understandable that the embodiments of the present application do not specifically limit the number of training times, and it can be set according to actual needs.
[0124] In the embodiment of the present application, after the system model is obtained through training, the state vector of the non-steady-state system may be used as the system model input to determine the optimal state feedback matrix corresponding to the non-steady-state system so as to execute S203.
[0125] S203: Control the non-steady-state system based on the optimal state feedback matrix.
[0126] In the embodiment of the present application, feedback adjustment is performed on the non-steady-state system based on the optimal state feedback matrix, so as to adjust the non-steady-state system to a steady-state system.
[0127] In summary, in the method provided in the embodiment of the present application, based on the system model with the initial state feedback matrix as the initial parameter, the process of controlling the non-steady-state system only requires a small number of state vectors to quickly obtain the optimal state feedback matrix corresponding to the non-steady-state system, which effectively reduces the amount of calculation and improves the control efficiency of the non-steady-state system.
[0128] At the same time, in the embodiment of the present application, during the training of the system model, there is no need to interfere with the introduced additive noise with known distribution. The initial system model can be trained based on arbitrary additive noise interference, thereby improving the universal applicability of the system model.
[0129] The present application provides a control device for an unsteady-state system, such as Figure 4 As shown in the figure, this figure is a structural schematic diagram of a control device for a non-steady-state system provided in an embodiment of the present application. Its specific implementation method is consistent with the implementation method and the technical effect achieved in the embodiment of the above method, and some contents will not be repeated here.
[0130] The embodiment of the present application provides a control device 4100 for an unsteady-state system, including:
[0131] Extraction unit 4101, determination unit 4102, and control unit 4103;
[0132] The extraction unit 4101 is used to extract features from the data of the non-steady-state system to obtain a state vector of the non-steady-state system;
[0133] The determining unit 4102 is configured to determine an optimal state feedback matrix corresponding to the non-steady-state system using the state vector of the non-steady-state system as a system model input; the system model is obtained by training the initial system model based on a forward propagation function, a backward propagation function, and a weight update function using a random state vector as an initial system model input and an initial state feedback matrix as an initial parameter of the initial system model; the initial state feedback matrix is obtained by calculating a discrete-time algebraic Riccati equation solution function using a system matrix, an input matrix, a state weight matrix, and a control weight matrix as inputs;
[0134] The control unit 4103 is used to control the non-steady-state system based on the optimal state feedback matrix.
[0135] In a possible implementation, the apparatus includes a training unit;
[0136] The training unit is specifically used to:
[0137] Inputting a random state vector into an initial system model, using an initial state feedback matrix as an initial parameter of the initial system model, and generating a first matrix based on a forward propagation function;
[0138] Obtaining gradient data based on the first matrix and the back propagation function;
[0139] Based on the gradient data, the weight update function value and the initial state feedback matrix are updated to generate the trained system model.
[0140] In a possible implementation, the training unit is specifically configured to:
[0141] Based on the gradient data, the weight update function value and the initial state feedback matrix are updated to generate an updated state feedback matrix, an updated weight update function value, and an updated system model;
[0142] The random state vector, the updated state feedback matrix, the system matrix, the input matrix, the state weight matrix and the control weight matrix are used as inputs of the forward propagation function to generate the second matrix;
[0143] Obtaining updated gradient data based on the second matrix and the back propagation function;
[0144] Based on the updated gradient data, the updated weight update function value and the updated state feedback matrix are updated to generate the trained system model.
[0145] In a possible implementation, the training unit is specifically configured to:
[0146] Determining a loss function based on the first matrix;
[0147] The loss function is used as the input of the back propagation function to obtain gradient data.
[0148] In a possible implementation, the training unit is specifically configured to:
[0149] Obtaining a minimized quadratic cost function based on the sum of the squares of the norms of each element in the first matrix, wherein each element is represented based on a state weight matrix and a control weight matrix;
[0150] The minimized quadratic cost function multiplied by 1 / (N+1) is determined as a loss function; wherein N is an integer greater than or equal to 0.
[0151] In a possible implementation, the training unit is specifically configured to:
[0152] The unsteady system is modeled to generate an initial system model; the initial system model is represented by a z-dimensional state vector of the unsteady system at time k, an m-dimensional input vector of the unsteady system at time k, a system matrix, an input matrix, and an uncertainty vector; wherein k is an integer greater than or equal to; z and m are positive integers.
[0153] In a possible implementation, the control unit is specifically configured to:
[0154] Based on the optimal state feedback matrix, feedback adjustment is performed on the non-steady-state system to adjust the non-steady-state system to a steady-state system.
[0155] In summary, the device provided in the embodiment of the present application controls the non-steady-state system based on a system model with an initial state feedback matrix as the initial parameter. Only a small number of state vectors are required to quickly obtain the optimal state feedback matrix corresponding to the non-steady-state system, which effectively reduces the amount of calculation and improves the control efficiency of the non-steady-state system.
[0156] At the same time, in the embodiment of the present application, during the training of the system model, there is no need to interfere with the introduced additive noise with known distribution. The initial system model can be trained based on arbitrary additive noise interference, thereby improving the universal applicability of the system model.
[0157] An embodiment of the present application further provides a computer device, the computer device comprising: a processor and a memory;
[0158] The memory is used to store program code and transmit the program code to the processor;
[0159] The processor is configured to execute the steps of the control method for a non-steady-state system according to the instructions in the program code.
[0160] An embodiment of the present application further provides a computer-readable storage medium having a computer program stored thereon. When the computer program is executed by a processor, the steps of the control method for a non-steady-state system as described above are implemented.
[0161] An embodiment of the present application further provides a program product. When the program product is run on a computer, the computer can execute the steps of the control method for a non-steady-state system as described above.
[0162] The above is only a specific embodiment of the present application, but the scope of protection of this application is not limited to this. Any changes or substitutions within the technical scope disclosed in this application should be included in the scope of protection of this application. Therefore, the scope of protection of this application should be based on the scope of protection of the claims.
Claims
1. A control method for an unsteady system, characterized in that: include: Extract features from the data of the unsteady-state system to obtain the state vector of the unsteady-state system; The state vector of the non-steady-state system is used as a system model input to determine the optimal state feedback matrix corresponding to the non-steady-state system; the system model is obtained by training the initial system model based on a forward propagation function, a backward propagation function, and a weight update function using a random state vector as an initial system model input and an initial state feedback matrix as an initial parameter of the initial system model; the initial state feedback matrix is obtained by calculating a discrete-time algebraic Riccati equation solution function using a system matrix, an input matrix, a state weight matrix, and a control weight matrix as inputs; The non-steady-state system is controlled based on the optimal state feedback matrix.
2. The method according to claim 1, characterized in that The training steps of the system model include: Inputting a random state vector into an initial system model, using an initial state feedback matrix as an initial parameter of the initial system model, and generating a first matrix based on a forward propagation function; Obtaining gradient data based on the first matrix and the back propagation function; Based on the gradient data, the weight update function value and the initial state feedback matrix are updated to generate the trained system model.
3. The method according to claim 2, characterized in that The updating of the network weights and the initial state feedback matrix based on the gradient data to generate the trained system model includes: Based on the gradient data, the weight update function value and the initial state feedback matrix are updated to generate an updated state feedback matrix, an updated weight update function value, and an updated system model; The random state vector, the updated state feedback matrix, the system matrix, the input matrix, the state weight matrix and the control weight matrix are used as inputs of the forward propagation function to generate the second matrix; Obtaining updated gradient data based on the second matrix and the back propagation function; Based on the updated gradient data, the updated weight update function value and the updated state feedback matrix are updated to generate the trained system model.
4. The method according to claim 2, characterized in that The step of obtaining gradient data based on the first matrix and the back propagation function includes: Determining a loss function based on the first matrix; The loss function is used as the input of the back propagation function to obtain gradient data.
5. The method according to claim 4, characterized in that The determining of the loss function based on the first matrix includes: Obtaining a minimized quadratic cost function based on the sum of the squares of the norms of each element in the first matrix, wherein each element is represented based on a state weight matrix and a control weight matrix; The minimized quadratic cost function multiplied by 1 / (N+1) is determined as a loss function; wherein N is an integer greater than or equal to 0.
6. The method according to claim 2, characterized in that Before inputting the random state vector into the initial system model and using the initial state feedback matrix as the initial parameter of the initial system model and generating the first matrix based on the forward propagation function, the method further includes: The unsteady system is modeled to generate an initial system model; the initial system model is represented by a z-dimensional state vector of the unsteady system at time k, an m-dimensional input vector of the unsteady system at time k, a system matrix, an input matrix, and an uncertainty vector; wherein k is an integer greater than or equal to; z and m are positive integers.
7. The method according to any one of claims 1 to 6, characterized in that The controlling of the non-steady-state system based on the optimal state feedback matrix includes: Based on the optimal state feedback matrix, feedback adjustment is performed on the non-steady-state system to adjust the non-steady-state system to a steady-state system.
8. A control device for an unsteady-state system, characterized in that: include: an extracting unit, a determining unit, and a controlling unit; The extraction unit is used to extract features from the data of the non-steady-state system to obtain a state vector of the non-steady-state system; The determining unit is configured to determine an optimal state feedback matrix corresponding to the non-steady-state system using a state vector of the non-steady-state system as a system model input; the system model is obtained by training the initial system model based on a forward propagation function, a backward propagation function, and a weight update function using a random state vector as an initial system model input and an initial state feedback matrix as an initial parameter of the initial system model; the initial state feedback matrix is obtained by calculating a discrete-time algebraic Riccati equation solution function using a system matrix, an input matrix, a state weight matrix, and a control weight matrix as inputs; The control unit is used to control the non-steady-state system based on the optimal state feedback matrix.
9. A computer device, characterized in that: The computer device includes: a processor and a memory; The memory is used to store program code and transmit the program code to the processor; The processor is configured to execute the steps of a control method for a non-steady-state system according to any one of claims 1 to 7 according to the instructions in the program code.
10. A computer-readable storage medium, characterized in that The computer-readable storage medium stores a computer program, and when the computer program is executed by a processor, the steps of the control method for a non-steady-state system according to any one of claims 1 to 7 are implemented.