Open channel predictive control method based on SCN deep Koopman model

Abstract

The disclosed embodiments provide an open channel predictive control method based on the SCN deep Koopman model, belonging to the field of control technology. The method specifically includes: Step 1: collecting state data and control sequences corresponding to the open channel system; Step 2: training an encoder and decoder based on a random configuration network based on the state data and control sequences; Step 3: establishing a state update equation based on the trained encoder and decoder; Step 4: constructing and solving an optimization problem based on the trained encoder and decoder and the state update equation to obtain an optimized controller for predictive control of the open channel system. The disclosed solution improves the efficiency, accuracy, and robustness of water level control in the open channel system.

Description

Technical Field

[0001] The embodiments of the present disclosure relate to the field of control technology, and in particular to an open channel predictive control method based on an SCN deep Koopman model. Background Art

[0002] The open channel system is a typical channel system. Its mechanism model is a hyperbolic quasi-linear partial differential equation called the Saint-Venant equation, which is established based on the conservation of mass and momentum of open channel water flow. Because it is impossible to obtain an analytical solution, it is difficult to directly apply general control strategies.

[0003] Existing open channel system modeling methods primarily discretize the Saint-Venant equations and then linearize them at equilibrium points to obtain ordinary differential equations or algebraic equations that meet the application requirements of some classical control algorithms. Common discretization methods include Preissmann implicit difference and finite volume method. However, these methods inevitably lead to the following problems:

[0004] Using a simplified model as the control object will inevitably lead to a loss of model accuracy. When the error reaches a certain level, the controller will fail and will be unable to cope with the impact of changes in environmental factors.

[0005] Since the linearization method is performed near the system equilibrium point, as the system gradually moves away from the equilibrium point, the model accuracy will become lower and lower, gradually exceeding the stabilization range of the controller, resulting in control failure.

[0006] It can be seen that there is an urgent need for an open channel predictive control method based on the SCN deep Koopman model with high control efficiency and accuracy. Summary of the Invention

[0007] In view of this, an embodiment of the present disclosure provides an open channel predictive control method based on the SCN deep Koopman model, which at least partially solves the problems of poor control efficiency and accuracy in the prior art.

[0008] The present disclosure provides an open channel predictive control method based on the SCN deep Koopman model, including:

[0009] Step 1: Collect the state data and control sequence corresponding to the open channel system;

[0010] Step 2: Train the encoder and decoder based on the random configuration network based on the state data and control sequence;

[0011] Step 3: Establish a state update equation based on the trained encoder and decoder;

[0012] Step 4: Construct and solve the optimization problem based on the trained encoder and decoder and the state update equation to obtain the optimized controller for predictive control of the open channel system.

[0013] According to a specific implementation of the embodiment of the present disclosure, step 1 specifically includes:

[0014] Step 1.1, use the Saint-Venant equation to describe the dynamic model corresponding to the open channel system;

[0015] In step 1.2, the Preissmann implicit difference method is used to discretize it, a set of gate control sequences with a suitable range is randomly given as the excitation signal, and the values ​​of the water level in front of the gate within N time steps are calculated as the state data.

[0016] According to a specific implementation of the embodiment of the present disclosure, step 2 specifically includes:

[0017] Step 2.1: construct an encoder and decoder based on a random configuration network, wherein the encoder is used to map the original state space to the feature state space, and the decoder is used to reconstruct the feature space back to the original space;

[0018] In step 2.2, the encoder and decoder are trained respectively using the state data and the control sequence according to the preset training method.

[0019] According to a specific implementation of the embodiment of the present disclosure, the preset training method is

[0020] Step 2.2.1, calculating the network parameters corresponding to the nodes in the last hidden layer of the encoder based on the inequality error, wherein the network parameters include weight parameters and bias;

[0021] Step 2.2.2, calculate the optimal output weight of the encoder based on the weight parameters and bias;

[0022] Step 2.2.3, calculating the current output of the encoder based on the optimal output weight of the encoder, the state data, and the control sequence, and calculating the first error accordingly;

[0023] Step 2.2.4: Determine whether the first error is less than a first threshold. If so, terminate encoder training. If not, increase the number of nodes in the last hidden layer and repeat steps 2.2.1 to 2.2.3 until the first error is less than the first threshold.

[0024] Step 2.2.5, calculate the optimal output weight of the decoder based on the weight parameters and bias

[0025] Step 2.2.6, using the state data and the encoder output as input to the decoder, calculating the current output of the decoder according to the optimal output weight of the decoder and specifically calculating the second error;

[0026] Step 2.2.7, determine whether the second error is less than the second threshold. If so, end the decoder training. If not, increase the number of nodes in the last hidden layer and re-execute steps 2.2.5 to 2.2.6 until the second error is less than the second threshold.

[0027] According to a specific implementation of the embodiment of the present disclosure, step 3 specifically includes:

[0028] Step 3.1: Convert the output of the trained encoder and decoder into mathematical expressions.

[0029] In step 3.1, the Koopman operator matrix is ​​estimated using the least squares method and the state update equation is constructed by combining the mathematical expressions of the encoder and decoder.

[0030] According to a specific implementation of the embodiment of the present disclosure, step 4 specifically includes:

[0031] Step 4.1: Use the trained encoder and decoder and the state update equation as equality constraints to construct the optimization problem.

[0032] Step 4.2, solve the optimization problem at the current time step and obtain a set of length N c The optimization control sequence is applied to the open channel system, and the optimization problem is solved again at the next time step. The above steps are repeated to obtain a forward rolling optimization controller and achieve the control target accordingly. c To control the time domain.

[0033] The open channel predictive control scheme based on the SCN deep Koopman model in the embodiment of the present disclosure includes: step 1, collecting state data and control sequences corresponding to the open channel system; step 2, training an encoder and decoder based on a random configuration network based on the state data and control sequence; step 3, establishing a state update equation based on the trained encoder and decoder; step 4, constructing and solving an optimization problem based on the trained encoder and decoder and the state update equation to obtain an optimized controller for predictive control of the open channel system.

[0034] The beneficial effects of the embodiments of the present disclosure are as follows: through the scheme of the present disclosure, first, in order to ensure the accuracy of modeling and improve the robustness of the control system, a network framework based on a stochastic configuration network (Stochastic Configuration Network) is applied to the modeling of the open channel system. It is a data-driven randomized learning method, which has a faster training speed than the traditional neural network based on gradient descent. When learning the dynamic evolution of the system, a global linearization modeling method based on the Koopman operator is used to linearly evolve the high-dimensional features learned by the stochastic configuration network to predict the state of the system at the next moment. Finally, the system models obtained by combining the above two methods are used to design a model predictive controller to achieve efficient and accurate control of the water level of the open channel system. BRIEF DESCRIPTION OF THE DRAWINGS

[0035] In order to more clearly illustrate the technical solutions of the embodiments of the present disclosure, the following briefly introduces the drawings required for use in the embodiments. Obviously, the drawings described below are only some embodiments of the present disclosure. For ordinary technicians in this field, other drawings can be obtained based on these drawings without any creative work.

[0036] Figure 1 A flow chart of an open channel predictive control method based on the SCN deep Koopman model provided in an embodiment of the present disclosure;

[0037] Figure 2 A structural diagram of an open channel system provided in an embodiment of the present disclosure;

[0038] Figure 3 A structural diagram of a random configuration network autoencoder provided in an embodiment of the present disclosure;

[0039] Figure 4 A schematic diagram of water level prediction results under white noise excitation provided by an embodiment of the present disclosure;

[0040] Figure 5 A schematic diagram of gate opening under full excitation provided in an embodiment of the present disclosure, wherein (a) represents the opening of gate 1, and (b) represents the opening of gate 2;

[0041] Figure 6 A schematic diagram of a water level tracking condition according to an embodiment of the present disclosure;

[0042] Figure 7 A schematic diagram of gate openings in a working condition provided by an embodiment of the present disclosure, wherein (a) represents the opening of gate 1, and (b) represents the opening of gate 2;

[0043] Figure 8A schematic diagram of water level tracking in working condition 2 provided by an embodiment of the present disclosure;

[0044] Figure 9 A schematic diagram of gate openings in two working conditions provided in an embodiment of the present disclosure, wherein (a) represents the opening of gate 1 and (b) represents the opening of gate 2. DETAILED DESCRIPTION

[0045] The embodiments of the present disclosure are described in detail below with reference to the accompanying drawings.

[0046] The following describes the embodiments of the present disclosure through specific examples, and those skilled in the art can easily understand other advantages and effects of the present disclosure from the contents disclosed in this specification. Obviously, the described embodiments are only a part of the embodiments of the present disclosure, rather than all of the embodiments. The present disclosure can also be implemented or applied through other different specific embodiments, and the details in this specification can also be modified or changed in various ways based on different viewpoints and applications without departing from the spirit of the present disclosure. It should be noted that, in the absence of conflict, the following embodiments and features in the embodiments can be combined with each other. Based on the embodiments in the present disclosure, all other embodiments obtained by ordinary technicians in this field without making creative work are within the scope of protection of the present disclosure.

[0047] It should be noted that various aspects of the embodiments within the scope of the appended claims are described below. It should be apparent that the aspects described herein can be embodied in a wide variety of forms, and any specific structure and / or function described herein is merely illustrative. Based on this disclosure, it should be understood by those skilled in the art that an aspect described herein can be implemented independently of any other aspect, and two or more of these aspects can be combined in various ways. For example, any number of aspects described herein can be used to implement an apparatus and / or practice a method. In addition, other structures and / or functionalities other than one or more of the aspects described herein can be used to implement this apparatus and / or practice this method.

[0048] It should also be noted that the illustrations provided in the following embodiments are only schematic illustrations of the basic concept of the present disclosure. The illustrations only show components related to the present disclosure and are not drawn according to the number, shape and size of components in actual implementation. In actual implementation, the type, quantity and proportion of each component can be changed at will, and the component layout type may also be more complicated.

[0049] Additionally, in the following description, specific details are provided to provide a thorough understanding of the examples. However, one skilled in the art will appreciate that the aspects described can be practiced without these specific details.

[0050] An embodiment of the present disclosure provides an open channel predictive control method based on the SCN deep Koopman model, which can be applied to the open channel control process.

[0051] See also Figure 1 , is a flow chart of an open channel predictive control method based on the SCN deep Koopman model provided by an embodiment of the present disclosure. Figure 1 As shown, the method mainly includes the following steps:

[0052] Step 1: Collect the state data and control sequence corresponding to the open channel system;

[0053] Furthermore, the step 1 specifically includes:

[0054] Step 1.1, use the Saint-Venant equation to describe the dynamic model corresponding to the open channel system;

[0055] In step 1.2, the Preissmann implicit difference method is used to discretize it, a set of gate control sequences with a suitable range is randomly given as the excitation signal, and the values ​​of the water level in front of the gate within N time steps are calculated as the state data.

[0056] In specific implementation, the present invention addresses the modeling and control issues of open channel systems. The key lies in building a randomly configured network framework to learn the system's characteristic states, representing the evolution of the characteristic states through the Koopman operator, and finally reconstructing the high-dimensional state back to the original state. In this way, a data-driven model of the open channel system is established, and a model predictive controller is designed to achieve water level control of the open channel system. The specific implementation plan is as follows:

[0057] Open channel system is a typical channel system, its structure is as follows Figure 2 As shown in the figure, a single channel pond is used as an example, which consists of a channel pond and two upstream and downstream gates. Its dynamic model is usually described by the Saint-Venant equation:

[0058]

[0059] Among them, t is the time coordinate, l represents the spatial coordinate along the river from upstream to downstream, Q is the flow rate, A is the cross-sectional area of ​​the water flow, S0 and S f They represent the riverbed slope and hydraulic slope, H is the water depth in the river, g is the acceleration of gravity, S f It can be calculated using the Manning-Strickler empirical formula, where n is the Manning coefficient and R is the hydraulic radius.

[0060]

[0061] Since Equation (1) cannot be solved analytically, in order to obtain the channel water level data for modeling, the Preissmann implicit difference method is used to discretize it, and a set of gate control sequences U = [u1, u2, ..., u N ] 2×N As the excitation signal, calculate the internal value of the water level before the gate in N time steps X = [x1, x2, ..., x N ] d×N As a system status, in actual scenarios, water level data can be collected directly through sensors in front of the gate.

[0062] Step 2: Train the encoder and decoder based on the random configuration network based on the state data and control sequence;

[0063] Based on the above embodiment, step 2 specifically includes:

[0064] Step 2.1: construct an encoder and decoder based on a random configuration network, wherein the encoder is used to map the original state space to the feature state space, and the decoder is used to reconstruct the feature space back to the original space;

[0065] In step 2.2, the encoder and decoder are trained respectively using the state data and the control sequence according to the preset training method.

[0066] Furthermore, the preset training method is

[0067] Step 2.2.1, calculating the network parameters corresponding to the nodes in the last hidden layer of the encoder based on the inequality error, wherein the network parameters include weight parameters and bias;

[0068] Step 2.2.2, calculate the optimal output weight of the encoder based on the weight parameters and bias;

[0069] Step 2.2.3, calculating the current output of the encoder based on the optimal output weight of the encoder, the state data, and the control sequence, and calculating the first error accordingly;

[0070] Step 2.2.4: Determine whether the first error is less than a first threshold. If so, terminate encoder training. If not, increase the number of nodes in the last hidden layer and repeat steps 2.2.1 to 2.2.3 until the first error is less than the first threshold.

[0071] Step 2.2.5, calculate the optimal output weight of the decoder based on the weight parameters and bias

[0072] Step 2.2.6, using the state data and the encoder output as input to the decoder, calculating the current output of the decoder according to the optimal output weight of the decoder and specifically calculating the second error;

[0073] Step 2.2.7, determine whether the second error is less than the second threshold. If so, end the decoder training. If not, increase the number of nodes in the last hidden layer and re-execute steps 2.2.5 to 2.2.6 until the second error is less than the second threshold.

[0074] In specific implementation, the method disclosed herein constructs an autoencoder based on a random configuration network for learning the feature mapping of the system, and can also map the high-dimensional state back to the original state. The specific network structure is as follows: Figure 2 As shown in the figure, unlike traditional neural networks, it incrementally constructs the network by adding nodes one by one and randomly assigns parameters to each node, which has faster training speed and better generalization performance.

[0075] According to the state data X and control sequence U collected in step 1, assuming that the hidden layer L-1 node of the random configuration network has been generated, the output f of the network is L-1 It can be expressed as:

[0076]

[0077] Among them, g j is the activation function, which is the standard sigmoid function selected in this invention, β j is the output weight of the jth node, w j and b j are the weight parameters and biases randomly assigned to the j-th node, respectively, and are calculated as follows:

[0078] w j =λ×(2×rand(d,1)-1)

[0079] b j =λ×(2×rand(1,1)-1) (4)

[0080] Where λ is a control factor used to determine the range of random sampling of weights and biases. The nodes to be assigned are selected from the candidate node pool, and the selection is made to make the inequality constraint The maximum parameter is given to the pending node:

[0081]

[0082] Where q is the label dimension, e L-1 To add L-1 nodes to the network output residual, h L is the output of the current hidden layer, f is the target output of the network, r∈(0,1), μ L is a non-negative real number sequence, and both need to satisfy μ L ≤1-r, Ensure that the network can converge, eL-1 and h L The expression is as follows:

[0083] e L-1 =e L-1 (X) = ff L-1 (X) (6)

[0084] h L =h L (X) = [g L (x1w L +b L ),…,g L (x N w L +b L )] T (7)

[0085] When selecting nodes from the candidate node pool, the larger the value of (5) is, the better. This is to ensure that the randomly configured network has universal approximation properties and accelerate network convergence. When evaluating the output weight β, it is necessary to fix the node parameters. The output weight is calculated as follows:

[0086]

[0087] Among them, α is the weight factor of the regularization term. Compared with the general least squares method, the introduction of the l2 regularization term is conducive to obtaining network parameters with stronger generalization performance and reducing the risk of overfitting.

[0088] like Figure 2 As shown in the figure, the encoder DSCN-E and the decoder DSCN-D are trained using the above method, where the role of DSCN-E is to map the original state space X to the feature state space X lift , and the role of DSCN-D is to reconstruct the feature space back to the original space. The two parts of training are carried out separately. When training DSCN-D, the original state X and the feature state X are f The combination is used as the input of the decoder, which is beneficial to the convergence of the decoder.

[0089] Step 3: Establish a state update equation based on the trained encoder and decoder;

[0090] Furthermore, the step 3 specifically includes:

[0091] Step 3.1: Convert the output of the trained encoder and decoder into mathematical expressions.

[0092] In step 3.1, the Koopman operator matrix is ​​estimated using the least squares method and the state update equation is constructed by combining the mathematical expressions of the encoder and decoder.

[0093] In specific implementation, in step 2, we obtain the encoder DSCN-E and decoder DSCN-D respectively, and the mathematical expression of their corresponding outputs can be simplified as follows:

[0094]

[0095] The fusion state X com =[XX f ], Ψ and Ψ -1 The mapping functions of the encoder and decoder are simplified expressions. In order to capture the dynamic characteristics of the open channel system and reduce the complexity of the model, the Koopman operator method is used to represent the dynamic evolution process of the system.

[0096] The Koopman operator is a global linear modeling method. According to the algorithm, in a higher-dimensional feature space, the dynamic process of the system can be expressed as a linear change. Therefore, the state update equation of the following form is established:

[0097] X com (k+1)=AX com (k)+BU(k) (10)

[0098] The present invention uses the least squares method to estimate the Koopman operator matrix

[0099]

[0100] in therefore:

[0101]

[0102] in The Penrose pseudoinverse is a generalization of the matrix inverse for non-square matrices.

[0103] Step 4: Construct and solve the optimization problem based on the trained encoder and decoder and the state update equation to obtain the optimized controller for predictive control of the open channel system.

[0104] Based on the above embodiment, step 4 specifically includes:

[0105] Step 4.1: Use the trained encoder and decoder and the state update equation as equality constraints to construct the optimization problem.

[0106] Step 4.2, solve the optimization problem at the current time step and obtain a set of length N cThe optimization control sequence is applied to the open channel system, and the optimization problem is solved again at the next time step. The above steps are repeated to obtain a forward rolling optimization controller and achieve the control target accordingly. c To control the time domain.

[0107] In specific implementation, model predictive control is an advanced control strategy widely used in industrial automation, process control and other fields. It is based on a dynamic model and uses the predicted future state to solve an optimization problem at each time step to optimize the control input to achieve the desired system performance.

[0108]

[0109] Formula (13) describes the optimization problem to be solved when the model predictive control is applied in the present invention, where Y is the state output of the open channel system, which is the water level in the present invention, and Y ref is the given water level target value, N p For the prediction time domain, Q and R are the weight matrices of state and control quantities respectively. and are the constraint sets that the state variables and control variables must satisfy respectively. The open channel system data-driven prediction model established in steps 2 and 3 exists here as an equality constraint. By solving equation (13), we can obtain a set of N c The control sequence, where N c To control the time domain, N is usually satisfied c ≤N p Under the condition of , we apply the first component of the obtained control sequence to the system, solve equation (13) again at the next time step, repeat the above steps, and obtain a forward rolling optimization controller to achieve the control target.

[0110] Regarding the modeling and control issues of open channel systems, the method disclosed in this disclosure has the following three beneficial effects:

[0111] The proposed modeling method based on random configuration network autoencoders does not need to rely on gradient information, so it has greater advantages in training speed and generalization performance;

[0112] The deep Koopman model is used to capture the dynamic behavior of the system. Due to its global linearization rather than linearization only at the equilibrium point, the model has higher accuracy. At the same time, combined with the random configuration network, it can also reduce the complexity of the network and speed up the solution of the control algorithm.

[0113] The use of model predictive control can effectively handle the nonlinear optimization problem of neural networks, and it also has robustness to changes in the external environment.

[0114] The open channel predictive control method based on the SCN deep Koopman model provided in this embodiment is applied to the modeling of the open channel system through a network framework based on a random configuration network. It is a data-driven randomized learning method, which has a faster training speed than the traditional gradient descent-based neural network. When learning the dynamic evolution of the system, a global linearization modeling method based on the Koopman operator is used to linearly evolve the high-dimensional features learned by the random configuration network to predict the state of the system at the next moment. Finally, the system models obtained by the above two methods are combined to design a model predictive controller to achieve efficient and accurate control of the water level of the open channel system; the method of combining the random configuration network and the Koopman operator establishes a data-driven model of the open channel system. When the amount of data is sufficient, this method can more accurately capture the nonlinear characteristics of the system than the traditional equilibrium point linearization modeling method. The model predictive controller designed based on this data-driven model shows higher robustness in complex environments.

[0115] The method of the present disclosure will be further described below with reference to a specific embodiment. Figure 2 Shows the schematic diagram of the single-channel system, where H up and H do are the upstream and downstream water levels of the channel, L is the channel length, u0 and u L They are the gate openings upstream and downstream of the channel respectively, and the channel water level is controlled by optimizing the scheduling of the gates.

[0116] Figure 3 The network structure of an autoencoder based on a random configuration network is shown. The encoder (DSCN-E) is on the left, and the decoder (DSCN-D) is on the right. Nodes are generated incrementally. Each hidden layer is directly connected to the output to evaluate the quality of the newly added nodes.

[0117] Figure 4 and Figure 5 The water level prediction model based on the collected water level data is shown. The excitation signal is in the form of white noise. The final prediction root mean square error (RMSE) is 0.1123. The RMSE is calculated as follows:

[0118]

[0119] Figure 6 and Figure 7 The water level tracking situation is shown when the initial water level is 1.5m and the target setting value is 2m. It can be seen that the water level tends to stabilize after 50 minutes, and the final steady-state error is 0.0191m. The gate opening remains within the set range.

[0120] Figure 8 and Figure 9 The water level tracking situation is shown when the initial water level is 1.5m, the target setting value before 225 minutes is 2m, and the target setting value after 225 minutes is 1m. It can be seen that the water level can stabilize within 50 minutes, and the steady-state error in the first half is 0.0191m, and the steady-state error in the second half is 0.0488m. The gate opening can be maintained within the set range.

[0121] In another embodiment, for data-driven modeling and control of an open channel system, specific open channel system structural parameters are shown in Table 1:

[0122]

[0123] Table 1

[0124] The present invention collects a training data set consisting of 1000 samples, wherein the sampling time is 60s. The openings of the upstream and downstream gates are randomly sampled in the range of 0 to 1. The same method is used to obtain a test dataset consisting of 300 samples. The detailed information of the hyperparameter values ​​and network structure required for network training is shown in Table 2:

[0125]

[0126] Table 2

[0127] It should be understood that various parts of the present disclosure can be implemented in hardware, software, firmware, or a combination thereof.

[0128] The above description is merely a specific embodiment of the present disclosure, but the scope of protection of the present disclosure is not limited thereto. Any changes or substitutions that can be easily conceived by a person skilled in the art within the technical scope disclosed in this disclosure should be included in the scope of protection of the present disclosure. Therefore, the scope of protection of the present disclosure should be based on the scope of protection of the claims.

Claims

1. An open channel predictive control method based on the SCN deep Koopman model, characterized in that: include: Step 1: Collect the state data and control sequence corresponding to the open channel system; Step 2: Train the encoder and decoder based on the random configuration network based on the state data and control sequence; The step 2 specifically includes: Step 2.1: construct an encoder and decoder based on a random configuration network, wherein the encoder is used to map the original state space to the feature state space, and the decoder is used to reconstruct the feature space back to the original space; Step 2.2: According to the preset training method, the encoder and decoder are trained using the state data and control sequence respectively; The preset training method is Step 2.2.1, calculating the network parameters corresponding to the nodes in the last hidden layer of the encoder based on the inequality error, wherein the network parameters include weight parameters and bias; Step 2.2.2, calculate the optimal output weight of the encoder based on the weight parameters and bias; Step 2.2.3, calculating the current output of the encoder based on the optimal output weight of the encoder, the state data, and the control sequence, and calculating the first error accordingly; Step 2.2.4: Determine whether the first error is less than a first threshold. If so, terminate encoder training. If not, increase the number of nodes in the last hidden layer and repeat steps 2.2.1 to 2.2.3 until the first error is less than the first threshold. Step 2.2.5, calculate the optimal output weight of the decoder based on the weight parameters and bias Step 2.2.6, using the state data and the encoder output as input to the decoder, calculating the current output of the decoder according to the optimal output weight of the decoder and specifically calculating the second error; Step 2.2.7: Determine whether the second error is less than the second threshold. If so, terminate decoder training. If not, increase the number of nodes in the last hidden layer and repeat steps 2.2.5 to 2.2.6 until the second error is less than the second threshold. Step 3: Establish a state update equation based on the trained encoder and decoder; Step 4: Construct and solve the optimization problem based on the trained encoder and decoder and the state update equation to obtain the optimized controller for predictive control of the open channel system.

2. The method according to claim 1, characterized in that The step 1 specifically includes: Step 1.1, use the Saint-Venant equation to describe the dynamic model corresponding to the open channel system; In step 1.2, the Preissmann implicit difference method is used to discretize it, a set of gate control sequences with a suitable range is randomly given as the excitation signal, and the values ​​of the water level in front of the gate within N time steps are calculated as the state data.

3. The method according to claim 2, characterized in that The step 3 specifically includes: Step 3.1: Convert the output of the trained encoder and decoder into mathematical expressions. In step 3.1, the Koopman operator matrix is ​​estimated using the least squares method and the state update equation is constructed by combining the mathematical expressions of the encoder and decoder.

4. The method according to claim 3, characterized in that The step 4 specifically includes: Step 4.1: Use the trained encoder and decoder and the state update equation as equality constraints to construct the optimization problem. Step 4.2, solve the optimization problem at the current time step and obtain a set of lengths The optimization control sequence is applied to the open channel system, and the optimization problem is solved again at the next time step. The above steps are repeated to obtain a forward rolling optimization controller and achieve the control target accordingly. To control the time domain.

Images