Distributed intelligent autonomous control method under cascade coupling
Through a distributed autonomous control method based on state self-learning, using neural networks to perform multi-step collaborative prediction and cascade control between devices, the complexity of coupling and nonlinear characteristics between devices in the multi-device cascade process is solved, and the complexity of the coupling and nonlinear characteristics between devices is realized Higher control accuracy and refined control.
Patent Information
- Application Number
- CN202410951768.0
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Priority Date
- 2024-01-17
- Filing Date
- 2024-07-16
- Publication Date
- 2024-11-08
AI Technical Summary
In the process industry of multi-device cascade, existing control methods are difficult to achieve precise multivariate coordinated control, especially due to the complexity of cascade coupling between devices and nonlinear characteristics, resulting in poor control effects.
Using a distributed autonomous control method based on state self-learning, a nonlinear dynamic model is established through a neural network, multi-step collaborative prediction between devices is carried out, and distributed autonomous control of cascading devices is realized based on the prediction results, and online learning is used by RBF neural network. Autoregressive model parameters capture the nonlinear dynamic characteristics of the device, realize real-time information interaction between devices and intelligent autonomous control of key parameters.
It significantly improves control accuracy and provides a refined management and control solution for multi-device cascade process, which significantly improves the control effect compared to traditional methods.
Smart Images

Figure SMS_2 
Figure SMS_4 
Figure SMS_6
Abstract
Description
Technical Field
[0001] The present invention relates to the field of industrial autonomous control methods, and in particular to a distributed intelligent autonomous control method under cascade coupling. Background Art
[0002] Automatic control is an important means to achieve stable and efficient operation of industrial processes. Multi-device cascade is a common phenomenon in process industries. A series of physical and chemical reactions occur in the cascade device, resulting in complex nonlinear reactions. In addition, since the outlet of the previous device is connected to the inlet of the next device, the operating status of the previous device will affect the status of the subsequent device, which poses a great challenge to the precise control of the entire process.
[0003] PID control and fuzzy control are commonly used control methods. These methods are usually used for single variable and single loop control processes and are not suitable for multivariable coordinated control under multi-device cascade. Although emerging control methods, such as reinforcement learning, have been widely studied and applied, the challenges in establishing accurate environmental models make it difficult to train suitable controllers, which hinders their application in process industry processes. In recent years, model predictive control has been widely studied. At each time point, model predictive control uses a prediction model to predict the future value of the controlled variable; then the future control sequence is obtained by optimization so that the actual value of the controlled variable is close to the expected value. Finally, the first control quantity of the control sequence is applied to the controlled object, and real-time control is achieved through rolling optimization. Model predictive control has many advantages, such as its ability to handle complex constraints and control multiple variables. Model predictive control can be divided into three forms: centralized, decentralized and distributed. Among them, centralized model predictive control is the most common. However, for long process and nonlinear multi-device cascade processes, on the one hand, it is difficult to establish an accurate global prediction model, and on the other hand, the centralized model predictive control method is difficult to solve this complex large-scale non-convex optimization problem, which limits its application in real-time control of the entire process. Although the distributed model predictive control method reduces the computational requirements and has a flexible control mode, it is difficult to achieve satisfactory results due to the lack of consideration of cascade coupling between devices. In recent years, the distributed model predictive control method has been gradually applied to process industry processes. Compared with the centralized model predictive control, it significantly reduces the computational overhead and has outstanding control performance compared with the distributed model predictive control method. The accuracy of the model is the key to ensuring the effectiveness of predictive control. At present, the distributed model predictive control method in the process industry mainly establishes the prediction model of the subsystem through the mechanism, and converts the nonlinear model into a linear model through local linearization. There is a certain degree of error in the mechanism model itself, which is aggravated during the linearization process, resulting in inaccurate prediction model and affecting the control effect. With the advent of the new generation of information technology and the widespread deployment of automatic control systems in production sites, more and more field data are accumulated. It provides a basis for the research and application of learning-based prediction models. Nonlinear autoregressive models can be directly established by using neural networks or other data models. When a single device is under different operating conditions, the input-output relationship of the model will change, resulting in the difficulty of the directly established nonlinear model to accurately describe the variable operating state of the device. The state-dependent model predictive control method uses a neural network to learn the coefficients of the prediction model at each state point. It can describe time-varying nonlinearities and has shown good results in single system or device control.However, for the current process industry with multiple cascaded devices, a method with self-learning of device status, real-time information interaction between devices, and intelligent and autonomous control of key parameters of equipment is needed to achieve refined management and control of multiple cascaded devices in the process industry.
[0004] To this end, this patent proposes a distributed autonomous control method based on state self-learning to solve the control problem of the cascade process equipment under the influence of the equipment's front and rear and the changing equipment operation status. It includes nonlinear dynamic modeling based on state self-learning, multi-step collaborative prediction of the front and rear cascade equipment, and distributed autonomous control of the cascade equipment. Summary of the invention
[0005] Based on this, the purpose of the present invention is to provide a distributed intelligent autonomous control method under cascade coupling. First, a neural network is used to self-learn the time-varying working state of each device to obtain the linear autoregressive model parameters under different states, wherein the nonlinear dynamic characteristics of the device are captured by the local linearization of the prediction model. Then, the models of different devices are used to perform multi-step prediction of the future state, and the prediction information of the previous device outlet is used as the future entry information of the next device to achieve collaborative prediction of cascade devices. Subsequently, distributed autonomous control of each device is achieved in a cascade manner based on the prediction results.
[0006] A distributed intelligent autonomous control method under cascade coupling comprises the following steps:
[0007] Step 1: For each device, a nonlinear dynamic model based on state self-learning is established to obtain L distributed nonlinear prediction models;
[0008] Step 2: Each device learns the autoregressive model parameters online according to its own state;
[0009] Step 3: For online control, obtain a multi-step prediction of the future output of the lth device and then optimize the control sequence;
[0010] Step 4: Apply the first value of the control sequence to device l. If l = L, jump to step 2 to continue rolling optimization, otherwise continue to step 5.
[0011] Step 5: The prediction information is transmitted along the cascade direction to calculate the future entry condition of the device l+1;
[0012] Step 6: Set l=l+1 and jump to step 3 to implement cascade distributed control.
[0013] Furthermore, in step 1, L distributed nonlinear prediction models are obtained by formula (1),
[0014]
[0015] Furthermore, in the second step, let t = t + 1, and each device learns the autoregressive model parameters online according to its own state through formula (2).
[0016]
[0017] Among them, W(t-1) represents the state variable of the device at the current time t-1, is the hidden layer coefficient of the RBF neural network,
[0018]
[0019] Represents the output layer coefficient of the RBF neural network.
[0020] Furthermore, the third step obtains a multi-step prediction of the future output of the lth device through formula (6),
[0021]
[0022] in
[0023]
[0024]
[0025] E{X} represents the expectation of X.
[0026] Furthermore, the third step is optimized by formula (17):
[0027]
[0028] Furthermore, the step 5 calculates the future entry condition of the l+1th device by formula (14):
[0029]
[0030] Among them, E l (t) represents the prediction error of the lth device at time t, and the function g is determined by material balance and other mechanism knowledge.
[0031] The distributed intelligent autonomous control method under cascade coupling described in the present invention first uses a neural network to self-learn the time-varying working state of each device to obtain the linear autoregressive model parameters under different states, wherein the nonlinear dynamic characteristics of the device are captured by the local linearization of the prediction model. Then, the future state is predicted in multiple steps using the models of different devices, and the prediction information of the previous device outlet is used as the future entry information of the next device to achieve collaborative prediction of cascade devices. Subsequently, distributed autonomous control of each device is achieved in a cascade manner based on the prediction results. Compared with the control accuracy, it is greatly improved, providing a new solution for refined control under multi-device cascade.
[0032] For better understanding and implementation, the present invention is described in detail below with reference to the accompanying drawings. BRIEF DESCRIPTION OF THE DRAWINGS
[0033] Figure 1 A process schematic diagram of an application example of this application;
[0034] Figure 2 The predictive modeling effect based on state self-learning in the application example of this application;
[0035] Figure 3 This is a comparison diagram of prediction errors of different methods in the application example of this application;
[0036] Figure 4 This is a comparison diagram of the control effects of different methods in the application example of this application. DETAILED DESCRIPTION
[0037] The following will be combined with the drawings in the embodiments of the present invention to clearly and completely describe the technical solutions in the embodiments of the present invention. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without creative work are within the scope of protection of the present invention.
[0038] In the description of the present invention, it should be noted that the terms "vertical direction", "up", "down", "horizontal", etc., indicating the orientation or positional relationship, are based on the orientation or positional relationship shown in the drawings, and are only for the convenience of describing the present invention and simplifying the description, rather than indicating or implying that the device or element referred to must have a specific orientation, be constructed and operated in a specific orientation, and therefore cannot be understood as limiting the present invention. In addition, "first", "second", "third", and "fourth" are only used for descriptive purposes and cannot be understood as indicating or implying relative importance.
[0039] In the description of the present invention, it is also necessary to explain that, unless otherwise clearly specified and limited, the terms "set", "install", "connect", and "connect" should be understood in a broad sense, for example, it can be a heat-conducting connection, a detachable connection, or an integral connection, a mechanical connection, an electrical connection, a direct connection, or a connection through an intermediate medium, or the internal communication of two components. For ordinary technicians in this field, the specific meanings of the above terms in the present invention can be understood according to specific circumstances.
[0040] Example
[0041] The present invention provides a distributed intelligent autonomous control method under cascade coupling. First, a neural network is used to self-learn the time-varying working state of each device to obtain the linear autoregressive model parameters under different states, wherein the nonlinear dynamic characteristics of the device are captured by the local linearization of the prediction model. Then, the models of different devices are used to perform multi-step prediction of the future state, and the prediction information of the previous device outlet is used as the future entry information of the next device to achieve collaborative prediction of cascade devices. Subsequently, distributed autonomous control of each device is achieved in a cascade manner based on the prediction results.
[0042] The specific steps include:
[0043] Step 1: For each device, a nonlinear dynamic model based on state self-learning is established to obtain L distributed nonlinear prediction models, as shown in formula (1):
[0044] And set t=0;
[0045] Among them, y(t), u(t) and ξ(t) represent the controlled variable, control variable and noise at time t respectively, and the noise obeys the Gaussian distribution with a mean value of 0; a i,t-1 、b i,t-1 and φ(t-1) are the autoregressive model coefficients, which vary with time and depend on the operating state of the device at time t-1; r(ti)∈R d It consists of d input condition variables at time ti, and d(ti)∈R d It also consists of the autoregressive model coefficients;
[0046] Step 2: Let t = t + 1, each device learns the autoregressive model parameters online according to its own state through formula (2), and then sets l = 1;
[0047]
[0048] Among them, W(t-1) represents the state variable of the device at the current time t-1, is the hidden layer coefficient of the RBF neural network,
[0049]
[0050] Represents the output layer coefficient of the RBF neural network.
[0051] Step 3: For online control, use formula (6) to obtain the multi-step prediction of the future output of the lth device, and optimize the control sequence through formula (17);
[0052]
[0053] in,
[0054] E{X} represents the expectation of X,
[0055]
[0056]
[0057]
[0058] Step 4: Set the control sequence U l The first value u of (t) l (t) Apply to device l, if l = L, jump to step 2 to continue rolling optimization, otherwise continue to step 5;
[0059] Step 5: The prediction information is transmitted along the cascade direction, and the future entry condition of the device l+1 is calculated by formula (14);
[0060]
[0061] Step 6: Set l=l+1 and jump to step 3 to implement cascade distributed control.
[0062] Specifically, for the modeling of a single device in the cascade process, linearization is performed at different state points. Since the input of a single device includes not only the control quantity but also the input condition variable, the linearized autoregressive model with multiple external inputs is specifically expressed as follows:
[0063]
[0064] Among them, y(t), u(t) and ξ(t) represent the controlled variable, control variable and noise at time t respectively, and the noise obeys the Gaussian distribution with a mean value of 0; a i,t-1 、b i,t-1and φ(t-1) are the autoregressive model coefficients, which vary with time and depend on the operating state of the device at time t-1; r(ti)∈R d It consists of d input condition variables at time ti, and d(ti)∈R d It also consists of the autoregressive model coefficients.
[0065] The above autoregressive model coefficients can be learned using RBF neural network based on the current state variables of the device. The specific expression is as shown in formula (2):
[0066]
[0067] Among them, W(t-1) represents the state variable of the device at the current time t-1, is the hidden layer coefficient of the RBF neural network,
[0068]
[0069] Represents the output layer coefficient of the RBF neural network.
[0070] Multi-step collaborative prediction of cascaded devices
[0071] Formula (1) can be converted into state space. First, the state vector is defined as formula (3):
[0072]
[0073] The state vector X(t) can be derived from the current time output y(t) and the input and output of the past time. In addition, the state space model can be expressed as (4):
[0074]
[0075] in
[0076]
[0077] Formula (4) can be used to predict the output. If N represents the prediction domain, N u represents the control domain, and the control amount is in the Nth u step remains unchanged, that is, u(t+j)=u(t+N u -1)(j≥N u ). In addition, the future output of the device is also affected by the input conditions. If it is possible to obtain the time from step t+1 to step t+N r The future information of the entry condition of the step
[0078] r(t+j)=r(t+N r -1)(j≥Nr ). The future output of a single device can be predicted in multiple steps. The calculation formula is as follows:
[0079]
[0080] in
[0081]
[0082]
[0083] E{X} represents the expectation of X. The input and output vectors in equation (6) can be found in equation (7), and the system matrix can be found in equations (9) to (12). It is worth noting that when making predictions, the state variable w(t+i) (i=1,2,…,N-1) at the future time t has no detection value. Therefore, we assume that w(t+i)=w(t)(i=1,…,N-1).
[0084] Going further, we can get the formula for multi-step collaborative prediction
[0085]
[0086] The superscript l represents the lth device. Since the outlet of the lth device is connected to the inlet of the l+1th device, the multi-step prediction information of the lth device must be used as the future input information of the l+1th device. This relationship can be expressed as follows:
[0087]
[0088] E l (t) represents the prediction error of the lth device at time t. Function g can be determined through material balance and other mechanism knowledge.
[0089] For each cascade device, the proposed distributed autonomous control method is able to set a separate control target and prediction step size, and has error self-correction capabilities. Therefore, the following objective function can be derived.
[0090]
[0091] in, Represents X T ΩX, is the identity matrix;
[0092] are the control amount and penalty term corresponding to the control increment of the lth device respectively. is the error self-correction term, which can be estimated using the previous prediction error. Formula (15) can be further converted to
[0093]
[0094] The above optimization problem can be solved using quadratic programming. In practical applications, when the output is not limited, the following formula can be used directly to obtain the control sequence solution:
[0095]
[0096] Application examples:
[0097] like Figure 1 , and the application verification was carried out in the zinc smelting leaching process. The front end of the zinc smelting leaching process includes three cascade reactors. The input of the leaching process includes zinc roasted sand, spent electrolyte, mixed liquor and acid leaching supernatant, etc. The first three reactors are mainly used to dissolve zinc compounds into zinc sulfate solution under appropriate acidity. The stability of pH is the key to ensure the leaching rate, so the pH of each reactor needs to be controlled to ensure the best reaction atmosphere and ensure that the zinc element in the zinc roasted sand enters the solution as much as possible to achieve the production goal of high leaching rate.
[0098] First, the nonlinear dynamic modeling based on state self-learning is used to perform autoregressive modeling on each device to predict the pH value of each device. The effect of pH prediction is as follows: Figure 2 As shown in the figure, the first 1300 groups of samples are training sets, and the last 300 groups of samples are test sets. The results of single-step prediction show that the prediction model of cascade equipment under complex nonlinear conditions can be accurately constructed through equipment state self-learning; in addition, the autoregressive prediction model based on state self-learning is compared with the commonly used nonlinear autoregressive model based on neural networks, such as Figure 3 As shown in the figure, it can be seen that better modeling effect can be achieved through state self-learning. Then, the pH value of each device is controlled through multi-step collaborative prediction and autonomous control methods. The control effect is shown in the figure below. Figure 4 As shown, compared with the commonly used PID control method, the control method proposed in this patent is much better than the conventional control method in terms of control effect. The mean square error of the control is calculated using formula (18), as shown in Table 1, which also shows that the control method proposed in this patent has a smaller control error.
[0099]
[0100] Among them, Num is the number of control samples, y(i) is the actual controlled variable value corresponding to the i-th control sample, and y r (i) is the corresponding setting value.
[0101] Table 1 Comparison of control error MSE of different methods
[0102]
[0103] The above-mentioned embodiments only express several implementation methods of the present invention, and the description is relatively specific and detailed, but it cannot be understood as limiting the scope of the invention patent. It should be pointed out that for ordinary technicians in this field, several modifications and improvements can be made without departing from the concept of the present invention, and the present invention is also intended to include these modifications and modifications.
Claims
1. A distributed intelligent autonomous control method under cascade coupling, characterized in that: The steps include: Step 1: For each device, a nonlinear dynamic model based on state self-learning is established to obtain L distributed nonlinear prediction models; Step 2: Each device learns the autoregressive model parameters online according to its own state; Step 3: For online control, obtain a multi-step prediction of the future output of the lth device and then optimize the control sequence; Step 4: Apply the first value of the control sequence to device l. If l = L, jump to step 2 to continue rolling optimization, otherwise continue to step 5. Step 5: The prediction information is transmitted along the cascade direction to calculate the future entry condition of the device l+1; Step 6: Set l=l+1 and jump to step 3 to implement cascade distributed control.
2. The distributed intelligent autonomous control method under cascade coupling according to claim 1 is characterized in that: In the first step, L distributed nonlinear prediction models are obtained by formula (1):
3. The distributed intelligent autonomous control method under cascade coupling according to claim 2 is characterized in that: In the second step, let t = t + 1, and each device learns the autoregressive model parameters online according to its own state through formula (2). Among them, W(t-1) represents the state variable of the device at the current time t-1, is the hidden layer coefficient of the RBF neural network, Represents the output layer coefficient of the RBF neural network.
4. The distributed intelligent autonomous control method under cascade coupling according to claim 3 is characterized in that: The third step obtains the multi-step prediction of the future output of the lth device through formula (6): in E{X} represents the expectation of X.
5. The distributed intelligent autonomous control method under cascade coupling according to claim 4 is characterized in that: The third step is to optimize through formula (17):
6. The distributed intelligent autonomous control method under cascade coupling according to claim 5 is characterized in that: The step 5 calculates the future entry condition of the l+1th device by formula (14): Among them, E l (t) represents the prediction error of the lth device at time t, and the function g is determined by material balance and other mechanism knowledge.
