Model autonomous learning methods applied to predictive control in process industries
The model autonomous learning method addresses nonlinear and multiconditional modeling issues by dividing systems into multiple-input, single-output subsystems, using lognormal distribution and Bayesian optimization, ensuring reliable and stable predictive control in process industries.
Patent Information
- Application Number
- JP2024529805
- Authority / Receiving Office
- JP · JP
- Patent Type
- Applications
- Current Assignee / Owner
- Priority Date
- 2022-07-22
- Filing Date
- 2022-12-01
- Publication Date
- 2024-11-13
- Estimated Expiration
- 2042-12-01
AI Technical Summary
Existing model predictive control methods in process industries face challenges with nonlinear and multiconditional modeling, difficulty in training models online, and inadequate handling of dynamic characteristic changes due to factors like production variations and equipment aging, leading to mismatch between predictive models and actual process dynamics.
A model autonomous learning method that divides systems into multiple-input, single-output subsystems, uses lognormal distribution for parameter modeling, integrates process knowledge, and employs Bayesian optimization with differential evolution algorithms for parameter learning, incorporating parameter deadbands to ensure smoothness and reliability.
Enables online learning of model parameters with physical significance, improving prediction accuracy and control stability by integrating prior knowledge and verifying parameters, reducing fluctuations, and enhancing the reliability of predictive models.
Smart Images

Figure 2024542229000001_ABST
Abstract
Description
[Technical field]
[0001] The present invention relates to the technical field of process industry forecasting, and in particular to a model autonomous learning method applied to predictive control in the process industry. [Background technology]
[0002] Since the end of the 20th century, the process industry has undergone rapid development, including the successful application of model predictive control in the process industry. Compared with traditional PID control, model predictive control can not only be applied to difficult-to-control objects such as multi-coupling and large time lag, but also model constraints such as process conditions and economic indicators, and use optimization techniques to complete control tasks. Therefore, model predictive control is expected to have a wide range of applications in the process industry. However, in actual production, the dynamic characteristics of the process change due to various factors such as changes in production targets, changes in the ratio of production raw materials, aging equipment, and insufficient catalyst activity. This change causes the problem of mismatch between the predictive model of predictive control and the dynamic characteristics of the actual process. Therefore, if the degree of matching between the predictive model and the dynamic characteristics of the actual process can be increased, the controller can enable more accurate prediction of the process output, and thus improve the control effect. At present, there are several ideas and methods in the industry to deal with the problem of changing dynamic characteristics, such as nonlinear modeling, multi-condition modeling, online identification, and other methods. There are the following disadvantages when implementing existing methods:
[0003] The first drawback is in the case of nonlinear modeling and multi-condition modeling. In the process industry, there are many processes of "three transfers and one reaction" (momentum transfer, heat transfer, mass transfer and chemical reaction), which makes dynamic modeling difficult. In addition, it is difficult to cover all operating conditions with on-site operation data. Therefore, data-driven modeling methods can only build local models or models for some operating conditions. When operating under unknown conditions during actual operation, a method for learning the model online is still required.
[0004] The second shortcoming is the method for online identification. On the one hand, most of the existing methods adopt linear regression to model the dynamic characteristics, and the parameters have no real physical meaning, and there are no reasonable constraints on the parameters during identification, so the degree of freedom of the identification parameters is too high; on the other hand, the dynamic characteristics of the process tend to change slowly, and the currently operating prediction model should be a reference for new dynamic characteristic parameters. However, a technical method that can effectively balance these two aspects has not yet been found. Summary of the Invention
[0005] To overcome the above technical shortcomings, the present invention provides a model autonomous learning method applied to predictive control in process industries.
[0006] The features of this method are as follows. First, the dynamic characteristics of the actual process can be learned online autonomously, and the method of dividing into a multi-input, single-output subsystem can be applied to a multi-input, multi-output system, and of course, can also be applied to other systems that are simpler than a multi-input, multi-output system. Second, the type of the parameter to be learned can be automatically selected according to the preset model parameters so that the parameter to be learned has a real physical meaning, and in the learning process, the preset model has a constraint effect on the learning result of the model to improve the reliability of the model. Third, the learning result of the parameters is verified and the prediction result of the model is evaluated to further improve the reliability of the model. Fourth, the smoothness of the model parameters is improved by setting a parameter deadband, and the fluctuation of the control effect caused by frequent changes of the model parameters is reduced. The method of the present invention is an online learning method that fuses model knowledge and process data, and is applied to model autonomous learning of predictive control in the process industry.
[0007] Explanation of terms: 1. Gain: Magnification factor. 2. Natural angular frequency: A physical quantity that represents the vibration speed of an object and is related to the inherent characteristics of the vibration system. 3. Damping coefficient: A parameter that represents the characteristic of the system's energy decreasing during motion in second-order or higher systems. 4. Log-normal distribution: If the logarithm of a random variable follows normal distribution, then the random variable follows log-normal distribution. 5. Predictive Model: A quantitative relationship between things described in mathematical language or formulas used in predictive control. 6. Parameter Deadband: This means that when a parameter changes, if the change does not exceed the value, the parameter is considered as unchanged.
[0008] The technical solutions adopted in the present invention to overcome the technical problems are as follows:
[0009] A model autonomous learning method applied to predictive control in the process industry, which is realized by using a multi-input, single-output system as a basic subsystem, and includes the following steps:
[0010] Step 1, setting of models and automatic selection of learning parameters: All transfer function models are set, each transfer function model is calculated, and the parameters to be learned are automatically selected.
[0011] Here, the transfer function model is specifically set as follows: It involves setting up a transfer function model between each input variable and each output variable under a template of a fixed second-order system + first-order integral + pure lag transfer function model.
[0012] JPEG2024542229000002.jpg24170JPEG2024542229000003.jpg15170
[0013] JPEG2024542229000004.jpg12170
[0014] JPEG2024542229000005.jpg10170First, we perform calculations for each transfer function model, and then analyze it as follows: JPEG2024542229000006.jpg12170JPEG2024542229000007.jpg37170
[0015] JPEG2024542229000008.jpg24170JPEG2024542229000009.jpg8170
[0016] Step 2, parameter distribution modeling: For each transfer function model involved in the training, establish a log-normal distribution model of the training target parameters using the log-normal distribution, which specifically includes:
[0017] JPEG2024542229000010.jpg48170JPEG2024542229000011.jpg16170
[0018] Step 3, data collection and processing: Collect the data of input variables and output variables, and the quality codes corresponding to the input variables and output variables online, and process the cases of poor quality codes. Specifically, the following is included:
[0019] JPEG2024542229000012.jpg24170JPEG2024542229000013.jpg12170JPEG2024542229000014.jpg30170
[0020] Step 4, model parameter learning: Based on the Bayesian optimization theory, establish a mathematical optimization model of transfer function model autonomous learning, and use the differential evolution algorithm to optimize the transfer function model parameters, which specifically includes:
[0021] JPEG2024542229000015.jpg18170
[0022] An impulse response sequence of the system is calculated according to the transfer function model, and a predicted sequence of the system output variables is calculated using a convolution of the impulse response sequence and the system input variables.
[0023] JPEG2024542229000016.jpg93170JPEG2024542229000017.jpg12170JPEG2024542229000018.jpg14170
[0024] JPEG2024542229000019.jpg41170
[0025] Step 5, model parameter validation: For each parameter of each transfer function model, parameter validation including normality validation, parameter deadband judgment and mean value validation is performed.
[0026] JPEG2024542229000020.jpg7170
[0027] JPEG2024542229000021.jpg26170JPEG2024542229000022.jpg9170
[0028] JPEG2024542229000023.jpg20170
[0029] JPEG2024542229000024.jpg28170
[0030] Step 6, compare model prediction errors: use the original transfer function model to calculate the predicted root mean square error, and use the learned transfer function model to calculate the predicted root mean square error, and then determine whether to update the transfer function model of the multi-input single-output system according to the magnitude relationship between the two, and then return to step 1 to continue execution. Specifically, this includes the following:
[0031] JPEG2024542229000025.jpg27170JPEG2024542229000026.jpg18170
[0032] The beneficial effects of the present invention are as follows: 1. The present invention provides an autonomous learning technology for model parameters that supports online operation, and can autonomously learn model parameters using online data. 2. In the present invention, the general form of the second-order transfer function is adopted as the transfer function template, and not only the stable second-order real eigenvalue system model but also the stable second-order conjugate complex root system model can be learned, so that the learning parameters have a clear physical meaning. 3. In the present invention, the reliability of the online learning result of the transfer function model is improved by two methods. The first method is to integrate the a priori knowledge of the process into the learning process of the transfer function model, and the a priori knowledge includes the structure of the transfer function model of the operating controller, the parameter value, the sign of the gain, and the type of the parameters to be learned of the transfer function model that are automatically calculated from the coefficients and have physical meaning. The second method includes two steps of parameter verification and comparison of the prediction error result of the transfer function model, and only when the parameters pass the hypothesis verification and the transfer function model passes the prediction error result comparison, the transfer function model parameters are automatically updated to the learned result. 4. In the present invention, the method of parameter deadband is adopted to improve the smoothness of the transfer function model parameters. When the change of the transfer function model parameters is within the deadband set by the user, the transfer function model parameters will not change, so that the smoothness of the transfer function model parameters is improved and the stability of the control effect is improved. JPEG2024542229000027.jpg18170On the other hand, if the gain of the transfer function model is large, control will be slow, but if it is small, the system may become unstable, so the gain must be reduced more carefully.In addition, the probability density function curve of the log-normal distribution has a steeper shape to the left of the maximum value, and the probability density decays faster as the gain becomes smaller, so this type of a priori model is suitable for modeling risks when the gain is small. [Brief description of the drawings]
[0033] [Figure 1]FIG. 1 is a block diagram showing a flow of a model autonomous learning method applied to predictive control in the process industry according to the present invention. [Diagram 2] This is a flowchart showing the three phases of the model parameter validation step: normality validation, parameter deadband judgment, and mean value validation. [Diagram 3] The learning results of the transfer function model at different η2 and the step response curves of the original transfer function model are shown. [Figure 4] This shows a part of the training result of a transfer function model without using the parameter deadband method and the step response curve of the original transfer function model. DETAILED DESCRIPTION OF THE PREFERRED EMBODIMENTS
[0034] In order to facilitate the understanding of the present invention by those skilled in the art, the present invention is described in more detail below in conjunction with drawings and specific examples, which are merely illustrative and do not limit the protection scope of the present invention.
[0035] The model autonomous learning method applied to predictive control in the process industry according to the present invention is applicable to a multi-input multi-output system, and specifically, is based on a multi-input single-output subsystem division. Therefore, in a specific embodiment, a detailed description will be given using a multi-input single-output subsystem.
[0036] The following will be described using a two-input, one-output subsystem as an example, but the present invention is not limited to the two-input, one-output subsystem described in this embodiment, and can also be applied to other multiple-input, one-output subsystems.
[0037] The current transfer function model of the two-input, one-output system is expressed by the following equation (7). JPEG2024542229000028.jpg20170
[0038] Step 1, setting of models and automatic selection of learning parameters: All transfer function models are set, each transfer function model is calculated, and the parameters to be learned are automatically selected.
[0039] According to the formulas (1) and (7), in this embodiment, the parameters of the two transfer function models are set as follows: JPEG2024542229000029.jpg6170
[0040] JPEG2024542229000030.jpg22170
[0041] JPEG2024542229000031.jpg22170
[0042] Step 2, Parameter Distribution Modeling: For each transfer function model involved in training, a log-normal distribution model of the training target parameters is established using the log-normal distribution.
[0043] Since neither of the two transfer function models has a zero gain, nor does the real part of the characteristic root have a positive, pure imaginary root, both of the transfer function models need to be trained.
[0044] JPEG2024542229000032.jpg42170
[0045] Step 3, data collection and processing: Collect the data of input variables and output variables, and the quality codes corresponding to the input variables and output variables online, and process the cases where the quality codes have bad values.
[0046] JPEG2024542229000033.jpg19170JPEG2024542229000034.jpg12170JPEG2024542229000035.jpg59170
[0047] Step 4, model parameter learning: Based on the Bayesian optimization theory, a mathematical optimization model of transfer function model autonomous learning is established, and the transfer function model parameters are optimized using the differential evolution algorithm.
[0048] JPEG2024542229000036.jpg18170
[0049] JPEG2024542229000037.jpg67170JPEG2024542229000038.jpg13170
[0050] JPEG2024542229000039.jpg30170JPEG2024542229000040.jpg13170JPEG2024542229000041.jpg14170
[0051] JPEG2024542229000042.jpg27170
[0052] Step 5, model parameter validation: For each parameter of each transfer function model, parameter validation including normality validation, parameter deadband judgment and mean value validation is performed.
[0053] JPEG2024542229000043.jpg12170
[0054] JPEG2024542229000044.jpg20170
[0055] JPEG2024542229000045.jpg28170
[0056] JPEG2024542229000046.jpg42170
[0057] Step 6, compare model prediction errors: calculate the predicted root mean square error using the original transfer function model, calculate the predicted root mean square error using the trained transfer function model, and determine whether to update the transfer function model of the multi-input single-output system according to the relationship between the two, and then return to step 1 to continue execution.
[0058] JPEG2024542229000047.jpg27170JPEG2024542229000048.jpg15170
[0059] JPEG2024542229000049.jpg21170
[0060] Then, go back to step 1 and continue the loop.
[0061] The following description will be continued with another specific embodiment, which is a multi-input, single-output system shown in formula (12), where the input variables are the water temperature set values of the two branch circuits, the output variable is the water temperature of the dry circuit, and the control method is to use the two branch circuits to control the water temperature of the dry circuit. As a multi-input, single-output system, the specific embodiment is the same as that described above, so the description of the implementation process is omitted and the emphasis is on the effects of the present invention. JPEG2024542229000050.jpg8170
[0062] If the transfer function model of the controller during actual operation is expressed as equation (13), JPEG2024542229000051.jpg20170
[0063] JPEG2024542229000052.jpg70170
[0064] JPEG2024542229000053.jpg61170
[0065] The above only describes the basic principles and preferred embodiments of the present invention, but those skilled in the art can make many modifications and improvements according to the above description, and all of these modifications and improvements shall fall within the protection scope of the present invention.
Claims
1. This is realized by using a multi-input, single-output system as the basic subsystem. Step 1 of setting all transfer function models and automatically selecting parameters to be trained by performing calculations for each transfer function model; Step 2 of establishing, for each transfer function model involved in the training, a log-normal distribution model of the parameters to be trained using a log-normal distribution; Step 3: Collecting input and output variable data and quality codes corresponding to the input and output variables online, and processing the cases with poor quality codes; Step 4: establishing a mathematical optimization model of transfer function model autonomous learning based on Bayesian optimization theory, and optimizing transfer function model parameters using a differential evolution algorithm; Step 5: performing parameter validation for each parameter of each transfer function model, including normality validation, parameter deadband judgment and mean value validation; and a step 6 of calculating, for each multi-input single-output system, a prediction root-mean-square error using an original transfer function model, calculating a prediction root-mean-square error using the learned transfer function model, determining whether to update the transfer function model of the multi-input single-output system depending on the magnitude relationship between the two, and then returning to step 1 to continue execution.
2. In step 1, the transfer function model is specifically set as follows: The model autonomous learning method applied to predictive control in the process industry as described in claim 1, characterized in that it includes setting a transfer function model between each input variable and each output variable under a template of a fixed second-order system + first-order integral + pure lag transfer function model.
3.
4. In step 1, calculating each transfer function model and automatically selecting the parameters to be learned is specifically performed as follows: First, perform calculations for each transfer function model, then analyze as follows: If the gain of a transfer function model is 0, the transfer function model does not participate in the model online learning. otherwise, determine the pole positions of the second-order system; (1) if there is a root with a positive real part or a pure imaginary root, the transfer function model does not participate in the model online learning; and (2) if there are two conjugate complex roots with negative real parts, calculate the natural angular frequency and damping factor of the transfer function model, and use the model gain, natural angular frequency and damping factor as parameters to be learned; and except for (1) and (2) above, factorize the denominator polynomial of the transfer function model in the real domain, and use all non-zero first-order term coefficients and model gains after the decomposition as parameters to be learned.
5. Specifically, step 2 is 2. The model autonomous learning method applied to predictive control in the process industry according to claim 1, characterized in that it includes calculating log-normal distribution parameters for each transfer function model involved in the learning according to coefficients related to a distribution shape set by a user, and modeling parameters to be learned using the log-normal distribution.
6. Specifically, step 3 is Collect input and output variable data and the quality codes corresponding to the input and output variables online, and judge and process them as follows: If bad quality code data occurs for an input variable, clear all collected variable history data and collect it again.
7. Step 4 is specifically: When the amount of data of the input variables or output variables is accumulated to a predetermined number, for each of the multi-input single-output subsystems, a mathematical optimization model of the transfer function model autonomous learning is established based on the Bayesian optimization theory; Calculating an impulse response sequence of the system according to the transfer function model, and calculating a predicted sequence of the system output variables using a convolution of the impulse response sequence and the system input variables; The method for model autonomous learning applied to predictive control in the process industry according to claim 1, comprising optimizing transfer function model parameters using a differential evolution algorithm.
8. The model autonomous learning method applied to predictive control in the process industry as described in claim 7, characterized in that the differential evolution algorithm includes three steps: population evolution, individual mutation and individual selection.
9. Step 5 is specifically: The model autonomous learning method applied to predictive control in the process industry according to claim 1, characterized in that for each parameter of each transfer function model, three phases, namely, normality verification, parameter deadband judgment, and average value verification, are sequentially performed, and if all three phases are passed, the parameter learning result is valid; if not, the parameter learning result is invalid, and the original parameters are maintained.
10. Step 6 is specifically: The model autonomous learning method applied to predictive control in the process industry according to claim 1, further comprising returning to step 1 and continuing execution thereafter.
Citation Information
Patent Citations
PID parameter setting optimization method based on improved Bayesian model
CN112000001A
Tuning aid device for dynamic characteristics model parameter
JP1997006407A
Data processor, control system, data processing method, and program
JP2019021186A
