Multi-condition industrial control method and system based on integrated space-time model prediction

Abstract

The application discloses a kind of multi-working condition industrial control method and system based on integrated space-time model prediction, method includes: the typical working condition feature extraction method based on orthogonal test, constructs working condition identifier;The time dynamic model and spatial distribution model under each working condition are constructed, and the optimal parameters of the space-time model corresponding to each working condition are obtained using data-driven and integrated training method;The current working condition of multi-working condition industrial system is identified in real time using working condition identifier, and the space-time model under the current working condition is used as prediction model, and the current optimal control input of system is obtained by rolling optimization method.The application learns the space-time correlation of observable point and unobservable point, establishes the prediction model of unobservable point, and integrates it into the prediction control framework, to realize the accurate control of unobservable point.

Description

Technical Field

[0001] This invention belongs to the field of control technology, specifically relating to a multi-condition industrial control method and system for unobservable states of distributed parameter systems. Background Technology

[0002] With the development and progress of modern industrial technology, industrial processes are becoming increasingly complex, such as 600KA aluminum electrolysis cells and 152m... 2 Calcining furnaces. These systems involve multiphase, multi-field reaction systems that are coupled and interact, often exhibiting characteristics such as spatiotemporal coupling, infinite dimensions, and nonlinearity. This makes them difficult to model and control accurately, hindering long-term stable operation. Therefore, precise control of large-scale industrial processes is crucial.

[0003] Existing control methods are mainly divided into two categories: lumped parameter system methods and distributed parameter system methods. Lumped parameter system methods assume that the state indices within the system do not change spatially, and control of the entire system can be achieved through observation and control of specific locations. While lumped parameter system methods can control the system to a certain extent, they neglect the spatial distribution characteristics of large-scale industrial processes, making it difficult to accurately model and control the system. Distributed parameter system methods, on the other hand, use partial differential equations combined with specific initial and boundary conditions to describe the system. Because partial differential equations contain time and spatial partial derivatives, distributed parameter system methods can describe the system from both temporal dynamics and spatial distribution perspectives. However, traditional distributed parameter system modeling methods typically require known PDE equations, which is often lacking in clear mechanistic knowledge in industrial settings, making it difficult to meet this requirement.

[0004] With the development of information technologies such as big data and sensor technology, a sufficient number of various sensors have been deployed in industrial sites, accumulating a large amount of data. This data provides new paths for system control, leading to the proposal of data-driven distributed parameter system modeling and control methods. These methods utilize the large amount of data accumulated in industrial sites to learn a spatiotemporal reduced-order model of the system and design a controller based on this model. However, traditional data-driven distributed parameter system modeling and control methods typically assume that all observed states are known and neglect the model mismatch problem caused by changes in operating conditions. Summary of the Invention

[0005] To address the problem of precise control of large-scale industrial processes with unobservable states in existing technologies, this invention provides a multi-condition industrial control method and system based on an integrated spatiotemporal model prediction. By learning the spatiotemporal correlation between observable and unobservable points, a predictive model for unobservable points is established and integrated into the predictive control framework to achieve precise control of unobservable points.

[0006] To achieve the above technical objectives, the present invention adopts the following technical solution:

[0007] A multi-condition industrial control method based on integrated spatiotemporal model prediction includes:

[0008] S1, Based on the typical working condition feature extraction method of orthogonal experiment, a working condition identifier is constructed;

[0009] S2, construct the time dynamic model and spatial distribution model under each working condition, and use data-driven and integrated training methods to obtain the optimal parameters of the spatiotemporal model corresponding to each working condition;

[0010] S3 uses a condition identifier to identify the current operating conditions of a multi-condition industrial system in real time, and uses the spatiotemporal model under the current operating conditions as a prediction model to obtain the current optimal control input of the system through rolling optimization.

[0011] Furthermore, step S1 specifically includes:

[0012] First, identify the key influencing factors of the multi-condition industrial system state, and denote the number of key influencing factors as follows: ;

[0013] Then, for each key influencing factor, the level value is determined based on its range of variation, and denoted as the [missing value]. The number of levels for each key influencing factor is ;

[0014] Then based on the number of factors and level number The typical orthogonal array is obtained by matching and correspondence, and the operating parameters of the system under each typical working condition are obtained according to the orthogonal array; wherein, each element in the typical orthogonal array corresponds to a typical working condition;

[0015] Finally, training samples for condition recognition are constructed using the operating parameters and corresponding condition labels for each typical working condition, and a neural network is trained to obtain a condition recognizer.

[0016] Furthermore, the operating parameters for each typical working condition are obtained by collecting data for a preset duration using sensors distributed at various locations throughout the system; The operating parameters under typical working conditions are expressed as follows: :

[0017]

[0018] In the formula, the right subscript Representing the A typical working condition, Indicates the first time Sensor readings at the location It is the number of sensors. It is the length of the training data for each working condition.

[0019] Furthermore, a time-dynamic model is established for each sensor, the first... The input-output relationship of the time dynamic model of each sensor is as follows:

[0020]

[0021]

[0022] In the formula, It is the first One sensor Predicted output at time step; Indicates the first One sensor The time-dynamic model is constructed using a neural network. This represents the model parameters of the corresponding time-based dynamic model; It is the first One sensor The input to the time-dynamic model; It is a multi-condition industrial system sequentially sampling at time points. Input, Input time delay label; It is the first Each sensor at the sampling time point The monitoring value, To output the time delay label.

[0023] Furthermore, for unobservable points in multi-condition industrial systems where no sensors are installed, a spatial distribution model is established, represented as:

[0024]

[0025]

[0026] In the formula, These are the spatial coordinates of any unobservable point. yes Location Predicted output at time step; It is a spatial distribution model, constructed using a neural network. Parameters representing the spatial distribution model; Representing the time dynamic model of all sensors The vector formed by the predicted outputs at time points.

[0027] Furthermore, a data-driven and integrated training method is used to obtain the optimal parameters of the spatiotemporal model corresponding to each working condition. The objective function used is:

[0028]

[0029] In the formula, This represents the model parameters of the time-dynamic model corresponding to n sensors. Indicates the first The loss function term of the time-dynamic model corresponding to each sensor. The loss function term represents the spatial distribution model. This is to control the relative contribution of the loss function terms of each time-dynamic model and the loss function terms of the spatial distribution model.

[0030] Furthermore, the first The loss function for the time dynamic model corresponding to each sensor is:

[0031]

[0032] In the formula, Indicates the first One sensor Monitoring values ​​at any given time;

[0033] The loss function for the spatial distribution model is:

[0034]

[0035] In the formula, express Location The true value of a moment.

[0036] A multi-condition industrial control system based on integrated spatiotemporal model prediction includes a condition identifier and several spatiotemporal models that correspond one-to-one with different conditions.

[0037] The operating condition identifier is constructed based on the typical operating condition feature extraction method of orthogonal experiment and is used to identify the current operating condition of multi-operating condition industrial systems in real time.

[0038] The spatiotemporal model consists of a time dynamic model and a spatial distribution model. Its optimal parameters are obtained by using a data-driven and integrated training method, and are used as a prediction model under the current operating conditions. The optimal control input of the system is obtained through rolling optimization.

[0039] Beneficial effects

[0040] This invention proposes a multi-condition industrial control method and system based on an integrated spatiotemporal model prediction. By learning the spatiotemporal correlation between observable and unobservable points, a prediction model for unobservable points is established and integrated into the predictive control framework, thereby achieving precise control of unobservable points under multi-condition operation and providing a guarantee for the stable and safe operation of industrial processes. Attached Figure Description

[0041] Figure 1 This is a general framework diagram of the method described in the embodiments of this application;

[0042] Figure 2 This is the modeling process of the integrated spatiotemporal model described in the embodiments of this application;

[0043] Figure 3 This is the control strategy framework of the embodiments of this application;

[0044] Figure 4 This describes the predictive control process and data visualization effect of the catalyst rod described in the embodiments of this application;

[0045] Figure 5 These are the predicted effects of using different test methods on the catalyst rod described in the embodiments of this application;

[0046] Figure 6 This is the experimental effect of the single-condition and multi-condition control experiments described in the embodiments of this application. Detailed Implementation

[0047] The embodiments of the present invention will be described in detail below. These embodiments are based on the technical solutions of the present invention and provide detailed implementation methods and specific operation processes to further explain the technical solutions of the present invention.

[0048] This embodiment provides a multi-condition industrial control method based on an integrated spatiotemporal model prediction, applicable to multi-condition industrial control of distributed parameter systems with unobservable points. (Refer to...) Figure 1 As shown, it includes the following steps:

[0049] S1, Based on the typical working condition feature extraction method of orthogonal experiment, a working condition identifier is constructed;

[0050] S2, construct the time dynamic model and spatial distribution model under each working condition, and use data-driven and integrated training methods to obtain the optimal parameters of the spatiotemporal model corresponding to each working condition;

[0051] S3 uses a condition identifier to identify the current operating conditions of a multi-condition industrial system in real time, and uses the spatiotemporal model under the current operating conditions as a prediction model to obtain the current optimal control input of the system through rolling optimization.

[0052] The following sections will introduce the construction of the working condition identifier and the integrated spatiotemporal model respectively.

[0053] I. Operating Condition Identifier Based on Orthogonal Experiment

[0054] The operating condition acquisition method based on orthogonal experiments mainly includes three steps: selection of key factors, selection of levels, and orthogonal array matching. First, the influence of factors such as heat transfer, chemical reaction, and heat source terms on the system state is analyzed to extract the key influencing factors of the system. The number of factors obtained is [number missing]. Then, for each key influencing factor, based on the range of variation and combined with the operational mechanism analysis, the level value of each key influencing factor was obtained, and the number of levels obtained was... Finally, based on the number of factors and level number Match the corresponding typical orthogonal array and obtain the operating parameters of the system under typical operating conditions based on the orthogonal array. After the above steps, the typical operating conditions of the system can be obtained. The formula for calculating the required number of experiments is as follows:

[0055]

[0056] consider Each level and A system with several factors is obtained by designing an orthogonal array based on existing technology. Under each typical working condition, the system is fully stimulated to obtain data for that condition, forming a multi-working-condition training dataset. ,in and Representing the first Operating data and corresponding operating condition labels under various working conditions:

[0057]

[0058]

[0059] in, Indicates the first time Sensor readings at the location It is the number of sensors. This refers to the length of the training data for each operating condition. Therefore, The identification result of the system operating condition at any given time can be expressed as:

[0060]

[0061]

[0062] in, This indicates a working condition identifier built based on orthogonal experiments. It is constructed using a feedforward neural network, and the model parameters are optimized using the gradient descent method.

[0063] II. Integrated Spatiotemporal Model

[0064] To accurately describe the spatiotemporal distribution characteristics between observable and unobservable points and achieve precise prediction of unobservable points, this invention proposes an integrated spatiotemporal model (Joint Spatial-Temporal Model, JSTM) modeling method. This method uses a temporal dynamic model to describe the temporal dynamic characteristics of observable points and a spatial distribution model to describe the spatial distribution characteristics between observable points (with sensor monitoring) and unobservable points (without sensor monitoring). It combines the processes of these two models to construct an integrated spatiotemporal model. Specifically, firstly, the network structure and parameters of the temporal dynamic model and the spatial distribution model are initialized; then, the loss functions of each temporal dynamic model and the spatial distribution model are calculated, and these loss functions are weighted and summed to form the overall loss function of the spatiotemporal model; finally, based on the overall loss function of the spatiotemporal model, the parameters of the temporal dynamic model and the spatial distribution model are updated in an integrated manner using the gradient descent algorithm to obtain the overall optimal model parameters. The specific modeling steps are as follows:

[0065] 1. Temporal Model

[0066] Considering the system has One sensor Used for data acquisition to obtain a set of training data under a certain working condition. ,in It is system input. It is located in a spatial position The output, Indicates the first Each sampling time, and These represent the number of sensors and the sampling time range, respectively. Predicted output of each sensor It can be represented in the following form:

[0067] (6)

[0068]

[0069] in Indicates the first One sensor The time-dynamic model at that location, and The time delay labels represent the input and output, respectively. Indicates the first One sensor The model input at location. A neural network model is used to describe the time dynamic model, at location. The time-dynamic model at this point can be represented in the following form:

[0070]

[0071] in, It is the first The connection weights between the input layer and the hidden layer in a time-dynamic model Indicates the hidden layer bias. This represents the connection weight matrix between the hidden layer and the output layer. This indicates the output layer bias. It is the activation function. Therefore, formula (8) can be simplified to the following form:

[0072]

[0073] in, This represents the model parameters of the corresponding time-based dynamic model. Therefore, the first... The loss function of a time-dynamic model can be expressed in the following form:

[0074]

[0075] The above objective function is only related to the model parameters The optimal parameters of the model can be obtained by using the gradient descent method.

[0076] 2. Spatial Model

[0077] To reconstruct unobservable points, this paper uses a neural network model to describe the spatial distribution function, which can be expressed in the following form:

[0078]

[0079]

[0080] in It is a spatial distribution model. Represents the spatial coordinates of unobservable points. This represents the vector formed by the predicted outputs of the time-dynamic model of the observable points. Considering the parameters in the spatial distribution model, (11) can be transformed into the following form:

[0081]

[0082] in, This represents the connection weights between the input layer and the hidden layer in a spatial distribution model. Indicates the hidden layer bias. This represents the connection weight matrix between the hidden layer and the output layer. This indicates the output layer bias. Therefore, equation (13) can be further simplified to the following form:

[0083]

[0084] in, The parameters represent the spatial distribution model. Therefore, the loss function of the spatial distribution model can be expressed as follows:

[0085]

[0086] Similarly, the objective function (15) is only related to the parameters of the spatial distribution model, so the parameters of the optimal spatial distribution model can be obtained by gradient descent.

[0087] 3. Integrated spatiotemporal model training

[0088] To achieve overall optimal parameters for both the time-dynamic model and the spatial distribution model, this invention proposes an integrated optimization modeling method, with the following specific objective function:

[0089]

[0090] in, Indicates the first The loss function term of a time-dynamic model, The loss function term represents the spatial distribution model. This is to control the relative contributions of the loss function terms of each temporal dynamic model and the spatial distribution model. The integrated spatiotemporal model can theoretically obtain more accurate reconstruction predictions of unobservable points from a global perspective. The specific modeling process is as follows: Figure 2 As shown.

[0091] After training to obtain the integrated spatiotemporal model, it is used as the predictive model for model predictive control. The optimal control input for the system is then obtained through rolling optimization. Figure 3 As shown.

[0092] To verify the effectiveness of the method proposed in this invention, a predictive control experiment of the catalyst rod was designed. The process and data visualization of the catalyst rod are shown below. Figure 4 As shown; the method of this invention is compared with the traditional method to test the prediction effect of different methods for unobservable points. The experimental results are as follows. Figure 5 As shown in the figure, and as can be seen from the figure, the method of the present invention has the best prediction effect.

[0093] In addition, this experiment also designed single-condition and multi-condition control experiments, and the experimental results are as follows: Figure 6 As shown in the figure, and as can be seen from the figure, the method of the present invention can quickly and stably track the control target.

[0094] The above experiments verified the superiority of the method of the present invention in predicting and controlling unobservable points compared with traditional methods.

[0095] The multi-condition industrial control method based on an integrated spatiotemporal model proposed in this invention can be applied to predictive control of unobservable points in multi-condition distributed parameter systems. This method overcomes the unreasonable assumptions about spatial distribution made by traditional methods and extends predictive control to multi-condition operation. This method enables precise control of large-scale industrial systems, ensuring the stable and safe operation of industrial processes.

[0096] The above embodiments are preferred embodiments of this application. Those skilled in the art can make various changes or improvements based on them. Without departing from the overall concept of this application, these changes or improvements should fall within the scope of protection claimed in this application.

Claims

1. A multi-condition industrial control method based on integrated spatiotemporal model prediction, characterized in that, include: S1, Based on the typical working condition feature extraction method of orthogonal experiment, a working condition identifier is constructed; S2, construct the time dynamic model and spatial distribution model under each working condition, and use data-driven and integrated training methods to obtain the optimal parameters of the spatiotemporal model corresponding to each working condition; In this process, a time dynamic model is established for each sensor; and a spatial distribution model is established for unobservable points in multi-condition industrial systems where no sensors are installed. S3 uses a condition identifier to identify the current operating conditions of a multi-condition industrial system in real time, and uses the spatiotemporal model under the current operating conditions as a prediction model to obtain the current optimal control input of the system through rolling optimization.

2. The multi-condition industrial control method based on integrated spatiotemporal model prediction according to claim 1, characterized in that, Step S1 is as follows: First, identify the key influencing factors of the multi-condition industrial system state, and denote the number of key influencing factors as follows: ; Then, for each key influencing factor, the level value is determined based on its range of variation, and denoted as the [missing value]. The number of levels for each key influencing factor is ; Then based on the number of factors and level number The typical orthogonal array is obtained by matching and correspondence, and the operating parameters of the system under each typical working condition are obtained according to the orthogonal array; wherein, each element in the typical orthogonal array corresponds to a typical working condition; Finally, training samples for condition recognition are constructed using the operating parameters and corresponding condition labels for each typical working condition, and a neural network is trained to obtain a condition recognizer.

3. The multi-condition industrial control method based on integrated spatiotemporal model prediction according to claim 2, characterized in that, The operating parameters for each typical working condition are obtained by collecting data for a preset duration using sensors distributed at various locations throughout the system; The operating parameters under typical working conditions are expressed as follows: : ； In the formula, the right subscript Representing the A typical working condition, Indicates the first time Sensor readings at the location It is the number of sensors. It is the length of the training data for each working condition.

4. The multi-condition industrial control method based on integrated spatiotemporal model prediction according to claim 1, characterized in that, No. The input-output relationship of the time dynamic model of each sensor is as follows: ； ； In the formula, It is the first One sensor Predicted output at time step; Indicates the first One sensor The time-dynamic model is constructed using a neural network. This represents the model parameters of the corresponding time-based dynamic model; It is the first One sensor The input to the time-dynamic model; It is a multi-condition industrial system sequentially sampling at time points. Input, Input time delay label; It is the first Each sensor at the sampling time point The monitoring value, To output the time delay label.

5. The multi-condition industrial control method based on integrated spatiotemporal model prediction according to claim 4, characterized in that, The spatial distribution model is expressed as: ； ； In the formula, These are the spatial coordinates of any unobservable point. yes Location Predicted output at time step; It is a spatial distribution model, constructed using a neural network. Parameters representing the spatial distribution model; Representing the time dynamic model of all sensors The vector formed by the predicted outputs at time points.

6. The multi-condition industrial control method based on integrated spatiotemporal model prediction according to claim 5, characterized in that, The optimal parameters of the spatiotemporal model for each working condition are obtained using a data-driven and integrated training method. The objective function used is: ； In the formula, This represents the model parameters of the time-dynamic model corresponding to n sensors. Indicates the first The loss function term of the time-dynamic model corresponding to each sensor. The loss function term represents the spatial distribution model. This is to control the relative contribution of the loss function terms of each time-dynamic model and the loss function terms of the spatial distribution model.

7. The multi-condition industrial control method based on integrated spatiotemporal model prediction according to claim 6, characterized in that, No. The loss function for the time dynamic model corresponding to each sensor is: ； In the formula, Indicates the first One sensor Monitoring values ​​at any given time; The loss function for the spatial distribution model is: ； In the formula, express Location The true value of a moment.

8. A multi-condition industrial control system based on an integrated spatiotemporal model prediction, characterized in that, This includes a working condition identifier and several spatiotemporal models that correspond one-to-one with different working conditions; The operating condition identifier is constructed based on the typical operating condition feature extraction method of orthogonal experiment, and is used to identify the current operating condition of multi-operating condition industrial systems in real time. The spatiotemporal model consists of a time dynamic model and a spatial distribution model. Its optimal parameters are obtained by using a data-driven and integrated training method, and are used as a prediction model under the current working conditions. The optimal control input of the system at the current time is obtained through rolling optimization. In this process, a time-dynamic model is established for each sensor; and a spatial distribution model is established for unobservable points in multi-condition industrial systems where no sensors are set up.

Images