Model prediction control method fused with incremental learning

By integrating incremental learning and model prediction control, the problem of model mismatch in industrial scenarios is solved, accurate prediction and adaptive adjustment of the impact of disturbances are achieved, and the robustness and performance of the control system are improved.

CN118938814AInactive Publication Date: 2024-11-12ANHUI SHANGGAO DATA TECH CO LTD
View PDF 0 Cites 2 Cited by

Patent Information

Application Number
CN202410963836.5
Authority / Receiving Office
CN · China
Patent Type
Applications(China)
Current Assignee / Owner
Filing Date
2024-07-18
Publication Date
2024-11-12
Estimated Expiration
Not applicable · inactive patent

AI Technical Summary

Technical Problem

Complex time-varying characteristics and unpredictable perturbations in modern industrial scenarios lead to model mismatch in model prediction control systems, resulting in degradation of control performance and potential system divergence.

Method used

The model prediction control method with fusion incremental learning is adopted, and the transfer function model and incremental learning model is constructed, the online sequence prediction model is established, the model is updated using incremental data, the model error is corrected, and the optimal control input is achieved through iterative optimization.

Benefits of technology

Effectively reflect the dynamic changes of the actual system, improve the robustness of the system control, reduce the deviation caused by disturbances, and significantly improve the control effect.

✦ Generated by Eureka AI based on patent content.

Smart Images

  • Figure CN118938814A_ABST
    Figure CN118938814A_ABST
Patent Text Reader

Abstract

The invention is applicable to the technical field of model predictive control, and provides a model predictive control method fused with incremental learning, which comprises the following steps of: constructing an incremental learning model, supplementing a predictive deviation when a state space model deals with disturbance, feeding back a predictive result of the incremental learning model to a predictive time domain of the state space model, and correcting a model error. And the model is updated through incremental data, so that the influence caused by disturbance is effectively predicted. Finally, through feedback correction and iterative optimization, optimal control input is achieved, and the purpose of restraining unmeasurable disturbance is achieved. According to the fusion method, the model can reflect the dynamic change of an actual system more accurately, adaptive adjustment can be carried out when the working condition changes, and the control robustness of the system is improved. Simulation experiment results show that the method provided by the invention has the advantage of overcoming deviation caused by disturbance, and has remarkable superiority compared with other methods.
Need to check novelty before this filing date? Find Prior Art

Description

Technical Field

[0001] The present invention belongs to the technical field of model predictive control, and in particular to a model predictive control method integrating incremental learning. Background Art

[0002] Model predictive control is widely used in multiple industrial scenarios. However, modern industrial scenarios are extremely complex, and various unpredictable disturbances with time-varying characteristics seriously affect the high-quality operation of the control system. Predictive control technology based on state-space models has been widely used in the industrial field. However, due to the existence of various disturbances and uncertainties in the actual production process, the model mismatch problem will gradually emerge and have an adverse effect on the control system. In the early stage of operation, the model predictive control effect can generally achieve the designed performance, but as time goes by and the production process is accompanied by various disturbances, the transfer function model cannot capture the error effects caused by these changes, resulting in a large deviation between the model output and the true value, and the model mismatch occurs slowly, and the control performance will gradually decrease. When the degree of model mismatch is too large, it may cause the control system to diverge, causing huge losses in the production process. Summary of the invention

[0003] The present invention provides a model predictive control method integrating incremental learning, aiming to solve the above-mentioned problems.

[0004] The present invention is implemented in this way: a model predictive control method integrating incremental learning comprises the following steps:

[0005] Step 1: Construct a transfer function model;

[0006] Step 2: Build an incremental learning model;

[0007] Step 3: Control based on incremental learning prediction model.

[0008] Preferably, the step three further comprises: sub-step one, establishing an augmented state space equation;

[0009] Sub-step 2: iterative optimization control process;

[0010] Sub-step three: System status update.

[0011] Preferably, the transfer function in step 1 is as follows:

[0012]

[0013] Among them, K p is the gain factor, T p is the time constant, T d For time delay, an optimization algorithm is usually used to identify the system and determine the values ​​of the three parameters.

[0014] Assume that the manipulated process input is u and the measurable process output is y, with steady-state values ​​u0 and y0 respectively. Without loss of generality, the response of a linear first-order system with time delay can be expressed as a differential equation:

[0015]

[0016] Usually, the least square method can be used to solve the problem and find the best K p 、T p and T d , so that the sum of squares of the difference between the predicted value of y and the actual value of y is minimized. The transfer function model of the system is obtained based on the obtained parameters, and the following state space equation is obtained based on the transfer function model:

[0017]

[0018] Where k represents the discrete time series, x represents the system state variable, y represents the system predicted output, u represents the control input variable, and A, B, and C are coefficient matrices of corresponding dimensions.

[0019] Preferably, the sequence prediction model under the online incremental learning framework in step 2 can be defined as follows: for the real-time data block stream, slices are made every T time to form data sets D0, D1, ... D t , there will be a corresponding series of new models generated M0, M1, ...M t Dataset D t ={x i ,y i},x i ∈R p ,y i ∈R q It represents a set of sequences with dimension +, where represents the dimension of the input sequence I and represents the dimension of the output sequence O.

[0020] Preferably, in the incremental learning framework, each data block is used to update the model only once. Once the model is updated, this set of data blocks is no longer used. The core of this framework is the model M t-1 To Model M t Update, M t The update of parameters mainly depends on two aspects. One is the model M at the previous moment. t-1 and the previous moment data set D t-1 , and the second is the current moment data set D t , where the dataset D t Split into training set D tT and validation set D tV Model M t It is directly in the old model M t-1Based on the dataset D t-1 and D tT Train and use the validation set D tV For model M t-1 Perform verification and calculate the loss of the new model on the verification set. The loss function is as follows:

[0021] L = MSE(rea_D tV ,pre_D tV )

[0022] where rea_D tV It is the dataset D tV The actual value of pre_D tV is the predicted value of the model. Iterate the training until the loss is within the set range and obtain the new model M t .

[0023] In the real-time time series data stream, each sub-dataset is generated in sequence. Assume that the historical data set is D0 and the new data set is D1. The model trained in the historical data set is M0, and the new data set is divided into the training set D1t and the validation set D1v. The task is transformed into training a new model M1 on all the training data D0+D1, so that its loss on the validation set D1v is minimized.

[0024] Preferably, the incremental learning algorithm is described as follows: Input: historical data set D0, new data D1, model M0 obtained by historical data training

[0025] Output: New Model M1

[0026] Step 1: Initialize the weight W of the weight initialization algorithm on the data set i (i=1,2,…,m), where m is the number of samples;

[0027] Step 2: Divide the new data set D1 into training set D 1t and validation set D 1v

[0028] For iteration from 1 to N:

[0029] 1. D0+D 1t , bring in the existing model M0 for training and generate a new model M1_pre;

[0030] 2. Use the validation set D 1v Verify the new model M1_pre and calculate the loss of the new model on the validation set: L = MSE (rea_D tV ,pre_D tV )

[0031] 3. Adjust model parameters and weights, and update sample weights:

[0032]

[0033] Iterate training until L <= ε

[0034] End for

[0035] Step 3: Output the new model M1

[0036] Preferably, a discrete state space model is also included:

[0037]

[0038] Where x represents the system state variable, y represents the system predicted output, u represents the control input variable, and A, B, and C are all coefficient matrices. Since there are many constraints in the actual system, it is specifically expressed in the following mathematical form: u min <=u(k)<=u max , Δu min <=Δu(k)<=Δu max ,y min <=y(k)<=y max Assuming that the set value is r and the model prediction value is y, when a large disturbance occurs, y can still be close to r. According to the basic principle of predictive control, under the constraint of time k, the optimal solution u that meets the constraint conditions is obtained.

[0039] Preferably, in step 3, in sub-step 1, since the loss function needs to optimize the change of the input vector, we first make some transformations to replace the input vector in the state equation with the input change. Let Δx(k+1)=x(k+1)-x(k), we have

[0040]

[0041] Written in compact form

[0042]

[0043] Obviously, we can define a new state space whose input vector is the change of the original input vector. Define:

[0044]

[0045] The new state space equation is:

[0046]

[0047] Preferably, in the sub-step 2, at each time k, assuming N c To control the step size, N p is the prediction step size. Set N at time ki c The continuous control increment is ΔU, then ΔU can be expressed as:

[0048] ΔU=[Δu(k i ) T Δu(k i +1) T ...Δu(k i +N c -1)] T

[0049] According to the state space equation of the formula, we can get x(k i +1|k i ),…x(k i +N p |k i ), where x(k i +N p |k i ) can be expressed as:

[0050]

[0051] According to the above state space equation, we can still get y(k i +1|k i ),…y(k i +N p |k i ), where y(k i +N p |k i ) can be expressed as:

[0052]

[0053] Considering the sequence state in the prediction time domain, the controller will predict a total of Np states as the basis for optimization. Define x(k+N|k) to represent the value of variable x at time k+N predicted at time k. Then we have:

[0054]

[0055] The above formula can be simplified as

[0056] Y=Fx(k i )+ΦΔU

[0057] Where Y = [y(k i +1|k i ) T y(ki +2|k i ) T ...y(k i +N p |k i ) T ]

[0058]

[0059] At the same time, the fusion incremental learning model built in the previous section is used to predict the future Np Y values: Y L =[y l (k i +1|k i ) T y l (k i +2|k i ) T ...y l (k i +N p |k i ) T ], by using Y L If Y is corrected, the corrected value for the future forecast is r = Y L -Y, that is:

[0060] r=[(y l (k i +1|k i )-y(k i +1|k i )) T (y l (k i +2|k i )-y(k i +2|k i )) T ...(y l (k i +N p |k i )-y(k i +N p |k i )) T ].

[0061] By adding this correction value to the loss function, we can easily construct the loss function J(ΔU)

[0062]

[0063] where r p =[r(k),r(k+1),...,r(k+Np-1)]T , r(k), r(k+1) and r(k+Np-1) are reference input signals, Q and R are weight matrix coefficients, which are semi-positive matrices. Their function is to set the weights of the two optimization items. When Q is significantly large, the system can track the expected vector r p Tighter, which can make the system input change smaller when R is significantly large.

[0064] Q = diag(q 1 ,q 2 ,q 3 , ..., q n ,), R = diag(r 1 , r 2 , r 3 , ..., r n , ),

[0065] The loss function can be transformed into:

[0066]

[0067] Taking the partial derivative of the above loss function, we can get

[0068]

[0069] ΔU*=(Φ T QΦ+R) -1 Φ T Q T (r p +r-Fx(k i ))

[0070] Preferably, in the sub-step 3, in order to prevent the error caused by the open-loop optimization relying only on the model, the MPC first calculates the error between the actual output and the predicted output before calculating the control optimization at time K+1:

[0071]

[0072] In the state space model, the Kalman gain is calculated by using the Kalman filter function. The Kalman gain represents the ratio of the model prediction error to the measurement error in the state optimal estimation process. It is updated to the system state, and further based on the system state space equation, the purpose of correcting the predicted output is achieved. The state optimal estimate X(k|k)' is calculated based on the state prediction value and the X(k|k) Kalman gain:

[0073] X(k|k)'=X(k|k)+Lε(k)

[0074] Based on the optimal estimate at the previous moment, calculate the state prediction value at the next moment:

[0075] X(k+1|k)=AX(k|k)′+BΔU

[0076] And update the model prediction value according to the state space direction

[0077] Y(k+1|k)=CX(k+1|k).

[0078] Compared with the prior art, the embodiments of the present application have the following beneficial effects:

[0079] By constructing an incremental learning model, the prediction deviation of the state-space model in response to disturbances is supplemented, the prediction results of the incremental learning model are fed back to the prediction domain of the state-space model, the model error is corrected, and the model is updated through incremental data to effectively predict the impact of disturbances. Finally, through feedback correction and iterative optimization, the optimal control input is achieved to achieve the purpose of suppressing unmeasurable disturbances. This fusion method enables the model to more accurately reflect the dynamic changes of the actual system, and can make adaptive adjustments when the working conditions change, thereby improving the robustness of system control. The simulation experimental results show that the method proposed in the present invention has the ability to overcome the deviation caused by disturbances and has significant advantages over other methods. BRIEF DESCRIPTION OF THE DRAWINGS

[0080] Figure 1 It is the overall block diagram of the model predictive control method integrating incremental learning of the present invention;

[0081] Figure 2 It is a schematic diagram of the training process structure of the present invention. DETAILED DESCRIPTION

[0082] Unless otherwise defined, all technical and scientific terms used herein have the same meaning as those commonly understood by technicians in the technical field of this application; the terms used in the specification of the application herein are only for the purpose of describing specific embodiments and are not intended to limit this application; the terms "including" and "having" and any variations thereof in the specification and claims of this application and the above-mentioned drawings are intended to cover non-exclusive inclusions. The terms "first", "second", etc. in the specification and claims of this application or the above-mentioned drawings are used to distinguish different objects, not to describe a specific order.

[0083] Reference to "embodiments" herein means that a particular feature, structure, or characteristic described in conjunction with the embodiments may be included in at least one embodiment of the present application. The appearance of the phrase in various locations in the specification does not necessarily refer to the same embodiment, nor is it an independent or alternative embodiment that is mutually exclusive with other embodiments. It is explicitly and implicitly understood by those skilled in the art that the embodiments described herein may be combined with other embodiments.

[0084] The embodiment of the present invention provides a model predictive control method integrating incremental learning, as shown in the figure, including the following steps:

[0085] Step 1: Construct a transfer function model;

[0086] Step 2: Build an incremental learning model;

[0087] Step 3: Control based on incremental learning prediction model.

[0088] Step three also includes: sub-step one, establishing the augmented state space equation;

[0089] Sub-step 2: iterative optimization control process;

[0090] Sub-step three: System status update.

[0091] The transfer function in step 1 is as follows:

[0092]

[0093] Among them, K p is the gain factor, T p is the time constant, T d For time delay, an optimization algorithm is usually used to identify the system and determine the values ​​of the three parameters.

[0094] Assume that the manipulated process input is u and the measurable process output is y, with steady-state values ​​u0 and y0 respectively. Without loss of generality, the response of a linear first-order system with time delay can be expressed as a differential equation:

[0095]

[0096] Usually, the least square method can be used to solve the problem and find the best K p , T p and T d , so that the sum of squares of the difference between the predicted value of y and the actual value of y is minimized. The transfer function model of the system is obtained based on the obtained parameters, and the following state space equation is obtained based on the transfer function model:

[0097]

[0098] Where k represents the discrete time series, x represents the system state variable, y represents the system predicted output, u represents the control input variable, and A, B, and C are coefficient matrices of corresponding dimensions.

[0099] The sequence prediction model under the online incremental learning framework in step 2 can be defined as follows: for the real-time data block stream, slice it every T time to form data sets D0, D1, ... D t, there will be a corresponding series of new models generated M0, M1, ...M t Dataset D t ={x i ,y i},x i ∈R p ,y i ∈R q It represents a set of sequences with dimension +, where represents the dimension of the input sequence I and represents the dimension of the output sequence O.

[0100] In the incremental learning framework, each data block is used to update the model only once. Once the model is updated, this set of data blocks is no longer used. The core of this framework is the model M t-1 To Model M t Update, M t The update of parameters mainly depends on two aspects. One is the model M at the previous moment. t-1 and the previous moment data set D t-1 , and the second is the current moment data set D t , where the dataset D t Split into training set D tT and validation set D tV Model M t It is directly in the old model M t-1 Based on the dataset D t-1 and D tT Train and use the validation set D tV For model M t-1 Perform verification and calculate the loss of the new model on the verification set. The loss function is as follows:

[0101] L = MSE(rea_D tV ,pre_D tV )

[0102] where rea_D tV It is the dataset D tV The actual value of pre_D tV is the predicted value of the model. Iterate the training until the loss is within the set range and obtain the new model M t .

[0103] In the real-time time series data stream, each sub-dataset is generated in sequence. Assume that the historical data set is D0 and the new data set is D1. The model trained in the historical data set is M0, and the new data set is divided into the training set D1t and the validation set D1v. The task is transformed into training a new model M1 on all the training data D0+D1, so that its loss on the validation set D1v is minimized.

[0104] Incremental learning algorithm description: Input: historical data set D0, new data D1, historical data training to get model M0

[0105] Output: New Model M1

[0106] Step 1: Initialize the weight W of the weight initialization algorithm on the data set i (i=1,2,…,m), where m is the number of samples;

[0107] Step 2: Divide the new data set D1 into training set D 1t and validation set D 1v

[0108] For iteration from 1 to N:

[0109] 1. D0+D 1t , bring in the existing model M0 for training and generate a new model M1_pre;

[0110] 2. Use the validation set D 1v Verify the new model M1_pre and calculate the loss of the new model on the validation set: L = MSE (rea_D tV ,pre_D tV )

[0111] 3. Adjust model parameters and weights, and update sample weights:

[0112]

[0113] Iterate training until L <= ε

[0114] End for

[0115] Step 3: Output the new model M1

[0116] The predictive control method based on the state space model can improve the prediction accuracy and control effect, but when unknown disturbances occur in complex industrial sites, it is easy to cause model distortion and poor control effect. In this case, timely adjustment of identification parameters can alleviate model distortion, but frequent adjustment of identification parameters will cause the entire control system to diverge due to identification problems, causing huge losses to the production process. Therefore, it is necessary to find a method that does not require frequent changes in system parameters and effectively considers the impact of disturbances, so that stable control can still be performed when the working conditions change. The control method is a discrete state space model.

[0117] Discrete state space models are also included:

[0118]

[0119] Where x represents the system state variable, y represents the system predicted output, u represents the control input variable, and A, B, and C are all coefficient matrices. Since there are many constraints in the actual system, it is specifically expressed in the following mathematical form: u min <=u(k)<=u max , Δu min <=Δu(k)<=Δu max ,y min <=y(k)<=y max Assuming that the set value is r and the model prediction value is y, when a large disturbance occurs, y can still be close to r. According to the basic principle of predictive control, under the constraint of time k, the optimal solution u that meets the constraint conditions is obtained.

[0120] Model predictive control is a control algorithm based on a predictive model. The predictive model is a model established based on the historical information of the controlled object, and predicts the process output in the future. The predictive model has a wide range of extensions. The model does not emphasize the structural form, but only emphasizes the functional role, as long as it can predict the future output of the controlled object. The predictive model can be obtained through the mathematical models of traditional control such as state equations and transfer functions. For linear stable systems, the predictive model can be established through the finite step response or finite impulse response method.

[0121] In step 3, since the loss function needs to optimize the change of the input vector in sub-step 1, we first make some transformations and replace the input vector in the state equation with the input change. Let Δx(k+1)=x(k+1)-x(k), we have

[0122]

[0123] Written in compact form

[0124]

[0125] Obviously, we can define a new state space whose input vector is the change of the original input vector. Define:

[0126]

[0127] The new state space equation is:

[0128]

[0129] Iterative optimization is a very important feature of predictive control. Predictive control does not adopt an unchanging global optimization target, but an iterative finite-time domain optimization control strategy, that is, at each sampling moment, the performance index ranges from that moment to a finite time domain in the future. When the next sampling moment is executed, the optimization performance index is pushed forward one moment.

[0130] The iterative finite time domain optimization method can compensate for the uncertainty caused by factors such as model mismatch, external interference, and time variation, and can always establish the optimization based on the latest actual working conditions. In reality, since complex industrial processes always have uncertain factors such as model mismatch, interference, and time variation, it is more effective to establish an iterative finite time domain optimization method. MPC is an algorithm that determines the control sequence through optimization. In sub-step 2, at each time k, assume that N c To control the step size, N p is the prediction step size. Set N at time ki c The continuous control increment is ΔU, then ΔU can be expressed as:

[0131] ΔU=[Δu(k i ) T Δu(k i +1) T ...Δu(k i +N c -1)] T

[0132] According to the state space equation of the formula, we can get x(k i +1|k i ),…x(k i +N p |k i ), where x(k i +N p |k i ) can be expressed as:

[0133]

[0134] According to the above state space equation, we can still get y(k i +1|k i ),…y(k i +N p |k i ), where y(k i +N p |k i ) can be expressed as:

[0135]

[0136] Considering the sequence state in the prediction time domain, the controller will predict a total of Np states as the basis for optimization. Define x(k+N|k) to represent the value of variable x at time k+N predicted at time k. Then we have:

[0137]

[0138] The above formula can be simplified as

[0139] Y=Fx(k i )+ΦΔU

[0140] Where Y = [y(k i +1|k i ) T y(k i +2|k i ) T ...y(k i +N p |k i ) T ]

[0141]

[0142] At the same time, the fusion incremental learning model built in the previous section is used to predict the future Np Y values: Y L =[y l (k i +1|k i ) T y l (k i +2|k i ) T ...y l (k i +N p |k i ) T ], by using Y L If Y is corrected, the corrected value for the future forecast is r = Y L -Y, that is:

[0143] r=[(y l (k i +1|k i )-y(k i +1|k i )) T (y l (k i +2|k i )-y(k i +2|k i )) T ...(y l(k i +N p |k i )-y(k i +N p |k i )) T ].

[0144] The present invention modifies the reference value by using the correction value instead of directly correcting the predicted value. By adding the correction value to the loss function, the loss function J(ΔU) can be easily constructed.

[0145]

[0146] where r p =[r(k),r(k+1),...,r(k+Np-1)] T , r(k), r(k+1) and r(k+Np-1) are reference input signals, Q and R are weight matrix coefficients, which are semi-positive matrices. Their function is to set the weights of the two optimization items. When Q is significantly large, the system can track the expected vector r p Tighter, which can make the system input change smaller when R is significantly large.

[0147] Q = diag (q 1 ,q 2 ,q 3 , ..., q n ,), R = diag(r 1 , r 2 , r 3 , ..., r n , ),

[0148] The loss function can be transformed into:

[0149]

[0150] Taking the partial derivative of the above loss function, we can get

[0151]

[0152] ΔU*=(Φ T QΦ+R) -1 Φ T Q T (r p +r-Fx(k i ))

[0153] Since the model prediction value may deviate from the actual value, it is necessary to use real-time information for feedback correction in a timely manner. The next step of optimization will be based on inaccurate model prediction. As the process proceeds, the predicted output may deviate more and more from the actual output. In sub-step 3, to prevent the error caused by open-loop optimization that only relies on the model, MPC first calculates the error between the actual output and the predicted output before calculating the control optimization at time K+1:

[0154]

[0155] In the state space model, the Kalman gain is calculated by using the Kalman filter function. The Kalman gain represents the ratio of the model prediction error to the measurement error in the state optimal estimation process. It is updated to the system state, and further based on the system state space equation, the purpose of correcting the predicted output is achieved. The state optimal estimate X(k|k)' is calculated based on the state prediction value and the X(k|k) Kalman gain:

[0156] X(k|k)'=X(k|k)+Lε(k)

[0157] Based on the optimal estimate at the previous moment, calculate the state prediction value at the next moment:

[0158] X(k+1|k)=AX(k|k)′+BΔU

[0159] And update the model prediction value according to the state space direction

[0160] Y(k+1|k)=CX(k+1|k).

[0161] By constructing an incremental learning model, the prediction deviation of the state-space model in response to disturbances is supplemented, the prediction results of the incremental learning model are fed back to the prediction domain of the state-space model, the model error is corrected, and the model is updated through incremental data to effectively predict the impact of disturbances. Finally, through feedback correction and iterative optimization, the optimal control input is achieved to achieve the purpose of suppressing unmeasurable disturbances. This fusion method enables the model to more accurately reflect the dynamic changes of the actual system, and can make adaptive adjustments when the working conditions change, thereby improving the robustness of system control. The simulation experimental results show that the method proposed in the present invention has the ability to overcome the deviation caused by disturbances and has significant advantages over other methods.

[0162] It should be noted that, for the above-mentioned embodiments, for the sake of simplicity, they are all described as a series of action combinations, but those skilled in the art should know that the present invention is not limited by the described order of actions, because according to the present invention, some steps may be performed in other orders or simultaneously. Secondly, those skilled in the art should also know that the embodiments described in the specification are all preferred embodiments, and the actions and modules involved are not necessarily required by the present invention.

[0163] In the several embodiments provided in the present application, it should be understood that the disclosed devices can be implemented in other ways. For example, the device embodiments described above are merely schematic, such as the division of the above-mentioned units. There may be other division methods in actual implementation, such as multiple units or components can be combined or integrated into another system, or some features can be ignored or not executed. Another point is that the mutual coupling or communication connection shown or discussed can be through some interfaces, and the indirect coupling or communication connection between devices or units can be in the form of telecommunication or other forms.

[0164] The units described above as separate components may or may not be physically separated, and the components shown as units may or may not be physical units, that is, they may be located in one place or distributed on multiple network units. Some or all of the units may be selected according to actual needs to achieve the purpose of the solution of this embodiment.

[0165] The above embodiments are only used to illustrate the technical solutions of the present invention, rather than to limit the scope of protection of the invention. Obviously, the described embodiments are only some embodiments of the present invention, rather than all embodiments. Based on these embodiments, all other embodiments obtained by ordinary technicians in this field without making creative work are within the scope of protection of the present invention. Although the present invention has been described in detail with reference to the above embodiments, ordinary technicians in this field can still combine, add, delete or make other adjustments to the features in the various embodiments of the present invention according to the circumstances without conflict, without making creative work, so as to obtain different other technical solutions that do not deviate from the concept of the present invention in essence, and these technical solutions also belong to the scope of protection of the present invention.

Claims

1. A model predictive control method integrating incremental learning, characterized in that: Includes the following steps: Step 1: Construct a transfer function model; Step 2: Build an incremental learning model; Step 3: Control based on incremental learning prediction model.

2. The model predictive control method integrating incremental learning according to claim 1, characterized in that: The step three also includes: sub-step one, establishing the augmented state space equation; Sub-step 2: iterative optimization control process; Sub-step three: System status update.

3. The model predictive control method integrating incremental learning as claimed in claim 2, characterized in that: The transfer function in step 1 is as follows: Among them, K p is the gain factor, T p is the time constant, T d For time delay, an optimization algorithm is usually used to identify the system and determine the values ​​of the three parameters. Assume that the manipulated process input is u and the measurable process output is y, with steady-state values ​​u0 and y0 respectively. Without loss of generality, the response of a linear first-order system with time delay can be expressed as a differential equation: Usually, the least square method can be used to solve the problem and find the best K p , T p and T d , so that the sum of squares of the difference between the predicted value of y and the actual value of y is minimized. The transfer function model of the system is obtained based on the obtained parameters, and the following state space equation is obtained based on the transfer function model: Where k represents the discrete time series, x represents the system state variable, y represents the system predicted output, u represents the control input variable, and A, B, and C are coefficient matrices of corresponding dimensions.

4. The model predictive control method integrating incremental learning according to claim 1, characterized in that: The sequence prediction model under the online incremental learning framework in step 2 can be defined as follows: For the real-time data block stream, slices are made every T time to form data sets D0, D1, ... D t , there will be a corresponding series of new models generated M0, M1, ...M t Dataset D t ={x i ,y i },x i ∈R p ,y i ∈R q It represents a set of sequences with dimension +, where represents the dimension of the input sequence I and represents the dimension of the output sequence O.

5. The model predictive control method integrating incremental learning as claimed in claim 4, characterized in that: In the incremental learning framework, each data block is used to update the model only once. Once the model is updated, this set of data blocks is no longer used. The core of this framework is the model M t-1 To Model M t Update, M t The update of parameters mainly depends on two aspects. One is the model M at the previous moment. t-1 and the previous moment data set D t-1 , and the second is the current moment data set D t , where the dataset D t is split into training set D tT and validation set D tV Model M t It is directly in the old model M t-1 Based on the dataset D t-1 and D tT Train and use the validation set D tV For model M t-1 Perform verification and calculate the loss of the new model on the verification set. The loss function is as follows: L=MSE(rea_D tV ,pre_D tV ) where rea_D tV It is the dataset D tV The actual value of pre_D tV is the predicted value of the model. Iterate the training until the loss is within the set range and obtain the new model M t . In the real-time time series data stream, each sub-dataset is generated in sequence. Assume that the historical data set is D0 and the new data set is D1. The model trained in the historical data set is M0, and the new data set is divided into the training set D1t and the validation set D1v. The task is transformed into training a new model M1 on all the training data D0+D1, so that its loss on the validation set D1v is minimized.

6. The model predictive control method integrating incremental learning as claimed in claim 5, characterized in that: The incremental learning algorithm is described as follows: Input: historical data set D0, new data D1, model M0 obtained by historical data training Output: New Model M1 Step 1: Initialize the weight W of the weight initialization algorithm on the data set i (i=1,2,…,m), where m is the number of samples; Step 2: Divide the new data set D1 into training set D 1t and validation set D 1v For iteration from 1 to N:

1. D0+D 1t , bring in the existing model M0 for training and generate a new model M1_pre; 2. Use the validation set D 1v Verify the new model M1_pre and calculate the loss of the new model on the validation set: L = MSE (rea_D tV ,pre_D tV ) 3. Adjust model parameters and weights, and update sample weights: Iterate training until L <= ε End for Step 3: Output the new model M1.

7. The model predictive control method integrating incremental learning as claimed in claim 2, characterized in that: Discrete state space models are also included: Where x represents the system state variable, y represents the system predicted output, u represents the control input variable, and A, B, and C are all coefficient matrices. Since there are many constraints in the actual system, it is specifically expressed in the following mathematical form: u min <=u(k)<=u max , Δu min <=Δu(k)<=Δu max ,y min <=y(k)<=y max Assuming that the set value is r and the model prediction value is y, when a large disturbance occurs, y can still be close to r. According to the basic principle of predictive control, under the constraint of time k, the optimal solution u that meets the constraint conditions is obtained.

8. The model predictive control method integrating incremental learning as claimed in claim 2, characterized in that: In step 3, in sub-step 1, since the loss function needs to optimize the change of the input vector, we first make some transformations to replace the input vector in the state equation with the input change. Let Δx(k+1)=x(k+1)-x(k), we have Written in compact form Obviously, we can define a new state space whose input vector is the change of the original input vector. Define: The new state space equation is:

9. The model predictive control method integrating incremental learning as claimed in claim 2, characterized in that: In the sub-step 2, at each time k, assume that N c To control the step size, N p is the prediction step size. Set N at time ki c The continuous control increment is ΔU, then ΔU can be expressed as: ΔU=[Δu(k i ) T Δu(k i +1) T ...Δu(k i +N c -1)] T According to the state space equation of the formula, we can get x(k i +1|k i ),…x(k i +N p |k i ), where x(k i +N p |k i ) can be expressed as: According to the above state space equation, we can still get y(k i +1|k i ),…y(k i +N p |k i ), where y(k i +N p |k i ) can be expressed as: Considering the sequence state in the prediction time domain, the controller will predict a total of Np states as the basis for optimization. Define x(k+N|k) to represent the value of variable x at time k+N predicted at time k. Then we have: The above formula can be simplified as Y=Fx(k i )+FDU where Y = [y(k i +1|k i ) T y(k i +2|k i ) T ...y(k i +N p |k i ) T ​ At the same time, the fusion incremental learning model built in the previous section is used to predict the future Np Y values: Y L =[y l (k i +1|k i ) T y l (k i +2|k i ) T ...y l (k i +N p |k i ) T ], by using Y L If Y is corrected, the corrected value for the future forecast is r = Y L -Y, that is: r=[(y l (k i +1|k i )-y(k i +1|k i )) T (y l (k i +2|k i )-y(k i +2|k i )) T ...(y l (k i +N p |k i )-y(k i +N p |k i )) T ]。 By adding this correction value to the loss function, we can easily construct the loss function J(ΔU) where r p =[r(k),r(k+1),...,r(k+Np-1)] T , r(k), r(k+1) and r(k+Np-1) are reference input signals, Q and R are weight matrix coefficients, which are semi-positive matrices. Their function is to set the weights of the two optimization items. When Q is significantly large, the system can track the expected vector r p Tighter, which can make the system input change smaller when R is significantly large. Q=diag(q 1 ,q 2 ,q 3 ,...,q n ,),R=diag(r 1 ,r 2 ,r 3 ,...,r n ,), The loss function can be transformed into: Taking the partial derivative of the above loss function, we can find ΔU*=(Φ T QF+R) -1 F T Q T (r p +r-Fx(k i ))。 10. The model predictive control method integrating incremental learning as claimed in claim 2, characterized in that: In the sub-step 3, in order to prevent the error caused by open-loop optimization that only relies on the model, before calculating the control optimization at time K+1, MPC first calculates the error between the actual output and the predicted output: In the state space model, the Kalman gain is calculated by using the Kalman filter function. The Kalman gain represents the ratio of the model prediction error to the measurement error in the state optimal estimation process. It is updated to the system state, and further based on the system state space equation, the purpose of correcting the predicted output is achieved. The state optimal estimate X(k|k)' is calculated based on the state prediction value and the X(k|k) Kalman gain: X(k|k)'=X(k|k)+Lε(k) Based on the optimal estimate at the previous moment, calculate the state prediction value at the next moment: X(k+1|k)=AX(k|k)′+BΔU And update the model prediction value according to the state space direction Y(k+1|k)=CX(k+1|k).

Citation Information

Cited By

  • Double-layer data driving model predictive control method of coal slurry concentration control system in coal gasification process

    CN120802589A

  • Multi-source sensing data augmentation method and system

    CN121010852A