Accurate parking control method for polymerization reaction heating temperature based on generative AI

Abstract

The invention discloses a polymerization reaction heating temperature accurate parking control method based on generative AI, and the method comprises the steps: confirming a control target in a technological process containing a polymerization reaction; analyzing and determining state characteristics influencing a control target by utilizing a process principle; performing time sequence characteristic correlation analysis by using a sliding window, and determining time sequence characteristics and control variables influencing a control target; constructing a discriminator and a training data set thereof, and training the discriminator through the training data set to obtain a trained discriminator; constructing a generator and a sample set thereof, and training the generator through the sample set and the trained discriminator to obtain a trained generator; the trained generator is applied to the technological process containing polymerization reaction, prediction time sequence data of control variables are generated, and accurate parking control over the temperature rise temperature is achieved. Accurate parking control over the temperature rise temperature of the polymerization reaction is achieved on the basis of the generation type AI, namely setting of the discriminator and the generator.

Description

Technical Field

[0001] This invention relates to the field of industrial data detection and processing technology, and in particular to a method for precise temperature control of polymerization reaction heating based on generative AI. Background Technology

[0002] Polymerization, a typical and complex process in chemical production, always presents a critical challenge for process optimization and a significant technical bottleneck in its heating phase. The precision of temperature control during the heating process directly determines the effectiveness of reaction initiation and the consistency of product quality. 1. Process risks associated with insufficient heating: If the reaction temperature fails to accurately reach the process set value (e.g., within ±0.5℃), the polymerization reaction will not be effectively initiated, resulting in insufficient reactivity and low conversion rate. In this case, emergency process measures must be taken—including depressurizing the system, draining the condensate, and then reintroducing steam to raise the temperature. This not only extends the production cycle but also leads to energy waste and raw material loss.

[0003] 2. Risk of uncontrolled temperature rise: When the temperature exceeds the upper limit of the process window, it will trigger a violent exothermic reaction, accompanied by a sudden increase in reaction pressure, imbalance in the feed rate, and other chain effects. In severe cases, it may trigger a safety interlock shutdown. Under such conditions, key quality indicators such as the molecular weight distribution and melt index of the entire batch of materials will exceed the control range, leading to product downgrading or even scrapping.

[0004] The main challenges of heating up the polymerization reaction are: 1. Lag and long duration of impact: There is a time delay in the temperature measurement of the reaction system (temperature measuring point), the adjustment of the heating / cooling medium (actuator), and the temperature homogenization of the material in the vessel. Furthermore, the heating stage is mainly heated by steam condensate, but the condensate does not fill the vessel. The cooling water added in the early stage, after mixing with the condensate, still does not fill the vessel, thus having a long-term effect on the vessel. The cooling water added in the early stage has a continuous impact time, which may be as long as ten minutes. 2. Positive Feedback Risk: Increased temperature further accelerates the reaction rate, which in turn releases more heat, leading to even higher temperatures. If not properly controlled, this positive feedback loop of "accelerated reaction → increased heat release → increased temperature" can easily cause "runaway temperatures," resulting in either defective products or, in severe cases, explosive polymerization, equipment overpressure, and other safety accidents. Traditional solutions for increasing the temperature during polymerization reactions include: 1. Conventional PID control exhibits significant limitations in adaptability during the heating process of polymerization reactions. Due to its linear control characteristics, it struggles to effectively handle the nonlinear dynamic changes in the reaction system, easily leading to temperature overshoot or response lag during the reaction triggering phase. Fixed parameter settings cannot adapt to actual production conditions such as catalyst activity fluctuations and changes in raw material composition, resulting in decreased control accuracy. Furthermore, PID control struggles to compensate for the typical large hysteresis characteristic of polymerization reactions in a timely manner, often causing oscillations in the transient process and excessively long settling times.

[0005] 2. While MPC technology has theoretical advantages, it faces significant challenges in practical applications. It heavily relies on accurate process mechanism models, and the complex thermodynamic and kinetic characteristics of polymerization reactions make modeling extremely difficult. The strong coupling of multiple variables also increases the complexity of control strategy design. Furthermore, MPC's high computational resource requirements limit its real-time application capabilities in industrial DCS systems. Different reaction stages require different control models, but existing switching strategies often lead to instability in the transition process.

[0006] To address the shortcomings of PID and MPC, neural network-based predictive control methods also exist. However, these are essentially equivalent to PID advance compensation, but due to the uncertainty of the lag in the initial temperature rise, the predictive compensation still cannot ensure that the final target temperature reaches the final stable process control target range. Summary of the Invention

[0007] Technical objective: To address the deficiencies in existing technologies, this invention discloses a method for precise temperature control during the heating of polymerization reactions based on generative AI. This method achieves precise temperature control during the heating of polymerization reactions by using generative AI, specifically by setting up a discriminator and a generator.

[0008] Technical solution: To achieve the above technical objectives, the present invention adopts the following technical solution.

[0009] A method for precise temperature control during the heating of a polymerization reaction based on generative AI, the method comprising: In a process flow involving polymerization reactions, the process start time is taken as the starting point, the time interval for heating preparation is identified as the first time interval, the time interval for unstable temperature after heating is identified as the second time interval, and the temperature at several time points in the second time interval is selected as the control target, that is, the controlled object of the heating stop control. By analyzing the process principles, the state characteristics affecting the control objective are determined; by using a sliding window to perform time-series characteristic correlation analysis, the time-series characteristics affecting the control objective and the control variables are determined. Construct a discriminator and its training dataset, and train the discriminator using the training dataset to obtain the trained discriminator; Construct a generator and its sample set, and train the generator using the sample set and the trained discriminator to obtain the trained generator; The trained generator is applied to processes involving polymerization reactions to generate predicted time-series data of control variables, enabling precise temperature control during the heating process.

[0010] Preferably, a sliding window is used to perform time-series feature correlation analysis to determine the time-series features and control variables affecting the control objective, including: In processes involving polymerization reactions, several first-order characteristics are obtained through process experience; Construct the time-series data window of the control target and the time-series data windows of all first time-series features; After introducing a lag time into the time series data window of the control target, for each first time series feature, the lag time is continuously adjusted and the correlation between the time series data window of the control target and the time series data window of the first time series feature is calculated by using a sliding window method. The maximum correlation value of each first time series feature is selected, and the first time series feature whose maximum correlation value exceeds the preset correlation threshold is taken as the second time series feature. All second time series features are the time series features that affect the control target. The second time series feature corresponding to the maximum correlation value is taken as the control variable.

[0011] Preferably, the training dataset includes historical data on temperature, temporal characteristics, and state characteristics affecting the control target during the period from the start of the process to the end of the second time interval.

[0012] Preferably, the discriminator is a spliced ​​model of several linear neural networks and one BP neural network. The number of linear neural networks is the number of temporal features affecting the control target. The input data of one linear neural network is the historical data of a temporal feature affecting the control target, and the input data of the BP neural network is the historical data of the state feature.

[0013] Preferably, the formula for the discriminator includes: , in, The control target for the next time point predicted by the discriminator, i.e., the (M+1)th time point, is... The control objective at the current time point M is... Let m be the data from the m-th historical time point in the time series characteristics that affect the control objective for the q-th objective. M represents the current time point, and the value of M ranges from the start time of the process to the end of the second time interval. The state features are processed by a BP neural network. The resulting impact coefficient, Let be the weight of the q-th linear neural network, that is, the weight of the neural network to which the q-th temporal feature affecting the control objective belongs. For the data at the nth historical time point in the control variables, N is the end point of the second time interval.

[0014] Preferably, the process of training the discriminator using the training dataset includes: training the discriminator using the training error E, wherein the training error E includes the prediction error. and correction error Prediction error Used to train the coefficients of the discriminator, wherein the coefficients of the discriminator include and Correction error Used to introduce empirical error.

[0015] Preferably, the formula for calculating the training error E includes: , in, For prediction error, To correct the error, The control objective predicted by the discriminator at time point n is... To train the historical data of the control target at time point n in the dataset; N is the end point of the second time interval; This indicates the empirical error after adding the valve. This represents the empirical error after reducing the valve; sign represents the sign function; Y is the predicted output of the discriminator. + To improve the discriminator's prediction output by adding a small positive number to the control variable, I is a preset positive integer, representing the discriminator's predicted output after the control variable is reduced by a small positive number.

[0016] Preferably, I = 2S + 1; S is the number of control targets.

[0017] Preferably, the sample set includes historical data on temperature, temporal characteristics and state characteristics affecting the control target during the period from the start time of the process to the end time of the second time interval; The generator takes temperature, time-series characteristics affecting the control objective, and state characteristics as input data at the same time point as input data, and outputs predicted time-series data of the control variables as output data. The generator takes historical data of state characteristics, historical data of time-series characteristics affecting the control objective, and predicted time-series data of the control variables as input data, and outputs the predicted control objective through a trained discriminator.

[0018] Preferably, the generator training error formula includes: , in, For generator training error, This is the setpoint corresponding to the s-th control target. The s-th control target predicted by the trained discriminator. S represents the number of control targets. This is the mean value of the s-th control target predicted by the trained discriminator.

[0019] Beneficial effects: This invention is based on generative AI, that is, the setting of discriminator and generator to achieve precise control of the heating temperature of polymerization reaction, so that the temperature of the heating stage of polymerization reaction, which is most prone to overheating, as well as after the heating stage and in the early stage of polymerization reaction, can be controlled within the allowable range of process indicators, which greatly reduces the overheating phenomenon that is prone to occur in this stage. Attached Figure Description

[0020] Figure 1 This is a flowchart of a method according to an embodiment of the present invention; Figure 2 This is a schematic diagram of historical data from an embodiment of the present invention; Figure 3 This is a schematic diagram illustrating the effect of polymerization reaction temperature control in an embodiment of the present invention without implementing the method of the present invention; Figure 4 This is a schematic diagram illustrating the control effect of implementing the method of the present invention in an embodiment of the present invention. Detailed Implementation

[0021] To enable those skilled in the art to better understand the present application, the technical solutions in the embodiments of the present application will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present application, and not all embodiments. Based on the embodiments in the present application, all other embodiments obtained by those of ordinary skill in the art without creative effort are within the scope of protection of the present application. Example

[0022] As attached Figure 1 As shown in the figure, this embodiment takes a PVDF polymerization process as an example to explain the present invention's method for precise temperature control of polymerization reaction based on generative AI, including the following steps: S1. In a process flow containing polymerization reaction, the process start time is taken as the starting point, the time interval for heating preparation is identified as the first time interval, the time interval for unstable temperature after heating is taken as the second time interval, and the temperature at several time points in the second time interval is selected as the control target, that is, the controlled object of heating stop control. Based on actual production conditions, process experience, or historical data statistics, the first and second time intervals in the polymerization reaction process are obtained. The second time interval is generally the period within which the temperature is prone to exceed the limit after heating. The time start point, i.e., the process start time, is defined as... The first time interval is The second time interval is When obtaining the second time interval through historical data statistics, the earliest and latest occurrence times of overheating during the polymerization process can be statistically analyzed in batches to obtain the final time interval range, which serves as the second time interval. In this embodiment, taking a PVDF polymerization process as an example, after setting the process start time as the starting point, the first time interval is from the first minute to the tenth minute, i.e. The second time interval is from the fifteenth minute to the twentieth minute, that is... It allows for precise temperature control of the reactor.

[0023] The selection of control targets is based on preset rules or process experience. Preset rules can be interval selection, such as selecting several time points within the second time interval, i.e., the unstable period after heating. In this embodiment, if control targets are selected according to preset rules, the reactor temperatures at the 15th, 17th, and 19th minutes of the second time interval are selected, for a total of three control targets. Alternatively, if control targets are selected based on process experience, the reactor temperatures at the time the pressure reaches 4.0, and then at 5 minutes and 10 minutes thereafter, for a total of three time points, are selected as evaluation points for the heating halt, i.e., control targets. In the method of this invention, a successful control result is achieved when all control targets do not exceed the preset controlled temperature range.

[0024] S2. Utilize process principles to analyze and determine the state characteristics affecting the control objective; use a sliding window to perform time-series characteristic correlation analysis to determine the time-series characteristics and control variables affecting the control objective; time-series characteristics affecting the control objective. for ,in, Let q be the temporal characteristic affecting the control objective. Q represents the number of temporal characteristics affecting the control objective. It is also a control variable; state characteristic for ,in, For the p-th state feature, P represents the number of state features.

[0025] State characteristics are constant features that do not change during changes in the control target. For polymerization reactions, if heating is achieved through steam or a heater, and temperature control is achieved through cooling water, then the state characteristics affecting the entire system are mainly the cooling water temperature and the cumulative feed rate. Higher cooling water temperatures lead to higher generation valves and higher predicted temperatures, while lower cooling water temperatures lead to lower generation valves and lower predicted temperatures. Lower cumulative feed rates result in more drastic temperature fluctuations due to the heat generated by the same feed reaction. In this embodiment, the state characteristics are the cooling water temperature and the cumulative feed rate.

[0026] The process of using a sliding window to perform time-series feature correlation analysis to determine the time-series features and control variables affecting the control objective includes the following steps: S21. In a process flow containing polymerization reaction, several first time-series characteristics are obtained through process experience; in this embodiment, the first time-series characteristics are jacket temperature, cooling water valve, cooling water flow rate, instantaneous feed rate, pressure, and reactor liquid level.

[0027] S22. Construct the time series data window of the control target and the time series data windows of all first time series features; S23. After introducing a lag time into the time-series data window of the control target, for each first time-series feature, the lag time is continuously adjusted, and the correlation between the time-series data window of the control target and the time-series data window of the first time-series feature is calculated using a sliding window method. The maximum correlation value of each first time-series feature is selected, and the first time-series feature whose maximum correlation value exceeds a preset correlation threshold is taken as the second time-series feature. All second time-series features are the time-series features that affect the control target. The second time-series feature corresponding to the maximum correlation value is taken as the control variable.

[0028] In S23, the formula for the time-series data window of the control target after introducing the lag time is: , in, To introduce a time-series data window for the control target after lag time, Time series data window The starting time series data, i.e. The temperature of the reactor at a given time point Time series data window The endpoint time series data, i.e. The temperature of the reactor at a given time point The length of the time series data window. The time lag is the time delay. The formula for the time series data window of the first time series feature is: , in, The time series data window representing the first time series characteristic. Time series data window Starting point time series data, Time series data window The endpoint time series data; During the continuous adjustment of the lag time, the time series data window is calculated. With time series data window The correlation between the features is calculated; the maximum correlation value of each first time-series feature is selected, and if the maximum correlation value is greater than a preset correlation threshold, the first time-series feature is used as the second time-series feature. In this embodiment, the correlation threshold is a fixed value set based on experience. The purpose of introducing a lag time is that different first time-series features have different impact times on the control target, or different action times. In this invention, the lag time is continuously adjusted during the correlation calculation process to obtain the maximum correlation value, thereby enabling the selection of all time-series features that affect the control target. In this invention, the correlation calculation is existing technology and will not be described in detail here.

[0029] In this embodiment, by calculating the correlation through a sliding window and continuously adjusting the lag time, the maximum correlation value corresponding to the reactor liquid level is always less than the preset correlation threshold. Therefore, the second time-series features after screening are jacket temperature, cooling water valve, cooling water flow rate, instantaneous feed rate, and pressure, which are used as time-series features affecting the control target. Simultaneously, the second time-series feature corresponding to the maximum correlation value is the cooling water valve, which is used as a control variable.

[0030] S3. Construct the discriminator and its training dataset. Train the discriminator using the training dataset to obtain the trained discriminator.

[0031] The training dataset includes the time from the start of the process to the end of the second time interval, i.e. During this period, historical data on temperature, temporal characteristics, and state characteristics affecting the control target were collected.

[0032] The discriminator is structured as a concatenation of several linear neural networks and one backpropagation (BP) neural network. The number of linear neural networks corresponds to the number of temporal features influencing the control objective. Both linear and BP neural networks are existing network architectures and will not be elaborated upon here. The input data for a linear neural network is historical data of a temporal feature influencing the control objective, while the input data for a BP neural network is historical data of a state feature. It is important to emphasize that the input data for both linear and BP neural networks are time series of the same length at the same point in time. If there are inconsistencies in length, a mask is used to expand the temporal features influencing the control objective; for state features, a fixed value is used for expansion. Since state features are constant features that do not change during changes in the control objective, they can be directly expanded using the same fixed value.

[0033] The formula for the discriminator is: , in, The control target for the next time point predicted by the discriminator, i.e., the (M+1)th time point, is... The control objective at the current time point M is... Let m be the data from the m-th historical time point in the time series characteristics that affect the control objective for the q-th objective. M represents the current time point, and the value of M ranges from the start time of the process to the end of the second time interval, i.e. , The state features are processed by a BP neural network. The resulting impact coefficient, Let be the weight of the q-th linear neural network, that is, the weight of the neural network to which the q-th temporal feature affecting the control objective belongs. For the data at the nth historical time point in the control variables, N is the end point of the second time interval, i.e., the time point. .

[0034] As attached Figure 2 As shown in the schematic diagram of some input data in this embodiment, X1 is control variable data, and its values ​​after the current time are all masked using MASK. X2 is a temporal feature data that affects the control objective, and S is a state feature data.

[0035] In this invention, the discriminator, as a multivariable proportional-integral controlled model, predicts the temperature at multiple heating stop points. Simultaneously, the discriminator's structure distinguishes between the time-series characteristics of the control variable and other factors influencing the control objective. Specifically, for the time-series characteristics of other factors influencing the control objective, only historical data before the current time point M is used as input parameters. For the control variable, the time from the start of the process to the end of the second time interval is used. All historical data are used as input parameters. The discriminator's structure is designed to match the generator. Since the generator predicts the time series data of the control variables, the "future time series data" of the control variables are also used as input data during the discriminator's training process.

[0036] The process of training the discriminator using the training dataset includes: training the discriminator using the training error E, where the training error E includes the prediction error. and correction error Prediction error Used to train the coefficients of the discriminator, wherein the coefficients of the discriminator include and Correction error Used to introduce empirical error.

[0037] Prediction error The calculation formula is: , in, The control objective predicted by the discriminator at time point n is... To train the historical data of the control target at time point n in the dataset; N is the end point of the second time interval, i.e., the time point. When training with historical data, the prediction error is used to correct the network parameters.

[0038] However, during actual prediction, it was found that when the control variable, namely the cooling water valve, was increased, the predicted temperature might also rise. This is because the model is too complex, and the effects of the cooling water valve and the control target, i.e., temperature, may have been learned in reverse. Furthermore, the effects of the cooling water valve and temperature cannot both have positive and negative effects simultaneously. Therefore, in this invention, prediction error is addressed. Based on this, add an error term for both positive and negative actions as a correction error. During the discriminator's prediction process, a small random valve is added between time points M and N of the control variable, i.e., at the valve generating the future. Theoretically, all three stationary temperature points should decrease. Conversely, all three stationary temperature points should increase. The prediction error correction value is the adjustment error. The calculation formula is: , in, This indicates the empirical error after adding the valve. This represents the empirical error after reducing the valve; sign represents the sign function; Y is the predicted output of the discriminator. + To explain the discriminator's predicted output after adding a small positive number to the control variable, in this embodiment, it means that if a very small valve is added to the historical valve, the discriminator's predicted output will... To determine the discriminator's predicted output after the control variable is reduced by a small positive number, I is a preset positive integer, I = 2S + 1; S is the number of control targets. The final value is a positive number. In this embodiment, S is 3 and I is 7. The final value is in the interval [1, 13]. The more empirical errors there are in the stationary valve, the larger the value. The formula for calculating the final training error E is: , Ultimately, the effectiveness of network training is judged by ensuring that historical experience errors are minimized while also preventing the occurrence of experience-based errors.

[0039] S4. Construct the generator and its sample set. Train the generator using the sample set and the trained discriminator to obtain the trained generator. The generator is a TFT model. In this embodiment, the generator directly adopts the existing TFT (Temporal Fusion Transformer) model, which will not be elaborated here. In this invention, the discriminator and generator are used as a generative AI network.

[0040] The sample set includes the time from the start of the process to the end of the second time interval, i.e. During this period, historical data on temperature, temporal characteristics, and state characteristics affecting the control target were collected.

[0041] The generator's input data consists of time-series data of temperature, temporal characteristics affecting the control objective, and state characteristics at the same time point. The output data is the predicted time-series data of the control variables. Using historical data of state characteristics, historical data of temporal characteristics affecting the control objective, and the predicted time-series data of the control variables as input, the trained discriminator outputs the predicted control objective. The generator's training error formula is then constructed as follows: , in, For generator training error, This is the setpoint corresponding to the s-th control target. The s-th control target predicted by the trained discriminator. S represents the number of control targets. Let be the mean of the s-th control target predicted by the trained discriminator. For predicting the bias term, To control the mean square error of the target, the generator training error is designed to make the control target approach the set value while minimizing fluctuations between control targets.

[0042] In this embodiment, the timing sequence lengths of the timing data affecting the timing characteristics and state characteristics of the generator's input data are the same. If there are inconsistent lengths, the timing characteristics affecting the control target are expanded using a mask method; the state characteristics are expanded using a fixed value method. Since the state characteristics are constant characteristics that do not change during the change of the control target, they can be directly expanded using the same fixed value.

[0043] S5. The trained generator is applied to a process involving polymerization reactions to generate predicted time-series data for control variables, achieving precise temperature control during the heating and stabilization phase. In this embodiment, the generator is used to generate future valves for controlling the final temperature stabilization during the heating phase of the polymerization reaction.

[0044] In this embodiment, a PVDF polymerization reaction project is used as an example. After implementing the method of this invention, the probability of overall ultra-high or ultra-low temperatures in the early stage of the reaction is significantly reduced. Three points are selected as evaluation points for temperature rise and halt: the pressure reaches 4.0, and then for 5 minutes and 10 minutes thereafter. Using correlation analysis, the reactor temperature, jacket temperature, cooling water valve, cooling water flow rate, instantaneous feed rate, and pressure are determined to be highly correlated with the future reactor temperature. According to process analysis, cooling water temperature and cumulative feed rate are considered as state characteristics affecting future temperature prediction and valve generation.

[0045] The discriminator is trained using time-series data such as reactor temperature, jacket temperature, cooling water valve position, cooling water flow rate, instantaneous feed rate, and pressure. Cooling water temperature and cumulative feed rate are used as input data for the BP neural network. The training iterations are up to 1000 times, and the process terminates after 20 iterations if no better error is found. Once the discriminator is trained, the generator is trained using it.

[0046] Figure 4 This refers to the situation where, after implementation, the temperature rises and then stagnates. (Compared to...) Figure 3 A comparison was made between the heating conditions in the reactor and those not using the method of this invention. In the heating process without the method of this invention, the temperature was either extremely high or extremely low. However, after using the method of this invention, the temperature of the reactor was stable and remained within the set range.

[0047] The above are merely examples and do not limit the application areas provided by this invention.

[0048] In the embodiments of this application, the terms "first" and "second" (if they exist) are used only as name identifiers and do not represent the order of first and second.

[0049] The above description is only a preferred embodiment of the present invention. It should be noted that for those skilled in the art, several improvements and modifications can be made without departing from the principle of the present invention, and these improvements and modifications should also be considered within the scope of protection of the present invention.

Claims

1. A method for precise stopping control of the temperature rise in a polymerization reaction based on generative AI, characterized by, The method comprises: In the process containing polymerization reaction, the temperature preparation time interval is determined as the first time interval, the temperature unstable time interval after temperature rising is determined as the second time interval, and the temperature at several time points in the second time interval is selected as the control target, that is, the controlled object of temperature rising stop control, with the process start time as the time starting point; The state characteristics influencing the control target are determined by process principle analysis; The time sequence characteristics influencing the control target and the control variable are determined by time sequence characteristic correlation analysis using a sliding window; A discriminator and its training data set are constructed, the discriminator is trained by the training data set, and the trained discriminator is obtained; A generator and its sample set are constructed, the generator is trained by the sample set and the trained discriminator, and the trained generator is obtained; The trained generator is applied to the process containing polymerization reaction to generate predicted time sequence data of the control variable, and precise temperature rising temperature stop control is realized.

2. The method of claim 1, wherein the method is based on generative AI. The time sequence characteristics influencing the control target and the control variable are determined by time sequence characteristic correlation analysis using a sliding window, comprising: In the process containing polymerization reaction, a plurality of first time sequence characteristics are obtained through process experience; The time sequence data window of the control target and the time sequence data window of all first time sequence characteristics are constructed; After introducing a lag time into the time sequence data window of the control target, for each first time sequence characteristic, the lag time is adjusted and the correlation between the time sequence data window of the control target and the time sequence data window of the first time sequence characteristic is calculated by means of a sliding window, the maximum correlation value of each first time sequence characteristic is screened out, the first time sequence characteristic corresponding to the maximum correlation value exceeding a preset correlation threshold is selected as a second time sequence characteristic, and all second time sequence characteristics are the time sequence characteristics influencing the control target; the second time sequence characteristic corresponding to the maximum maximum correlation value is selected as the control variable.

3. The method of claim 1, wherein the method is based on generative AI. The training data set includes historical data of temperature, time sequence characteristics influencing the control target and state characteristics during the period from the time starting point of process start to the end of the second time interval.

4. The method of claim 1, wherein the method is based on generative AI. The structure of the discriminator is a splicing model of a plurality of linear neural networks and one BP neural network, and the number of linear neural networks is equal to the number of time sequence characteristics influencing the control target. The input data of one linear neural network is the historical data of one time sequence characteristic influencing the control target, and the input data of the BP neural network is the historical data of the state characteristic.

5. The method of claim 1, wherein the method is based on generative AI. The formula of the discriminator includes: ， wherein, is the control target at the next time point, i.e. the M+1th time point, predicted by the discriminator, is the control target at the current time point M, is the data of the mth historical time point in the timing feature of the qth influencing control target, , M is the current time point, and the value range of M is from the starting time point of the process to the ending time point of the second time interval, is the influence coefficient generated by the state feature via the BP neural network, is the influence coefficient generated by the state feature via the BP neural network, is the weight of the qth linear neural network, i.e. the weight of the neural network to which the timing feature of the qth influencing control target belongs, is the data of the nth historical time point in the control variable, , N is the ending time point of the second time interval.

6. The method of claim 1, wherein the method is based on generative AI. The process of training the discriminator by the training data set comprises: training the discriminator by a training error E, wherein the training error E comprises a prediction error and a correction error , the prediction error is used for training the coefficients of the discriminator, wherein the coefficients of the discriminator comprise and ; the correction error is used for introducing an experience error.

7. The method of claim 1, wherein the method is based on generative AI. The calculation formula of the training error E includes: ， wherein, is a prediction error, is a correction error, is a control target at the nth time point predicted by the discriminator, is a control target history data at the nth time point in the training data set; N is an end point of the second time interval; represents an empirical error after increasing the valve, represents an empirical error after decreasing the valve, and sign represents a sign function; Y is a prediction output of the discriminator, and Y + is a prediction output of the discriminator after increasing a small positive number to the control variable, is a prediction output of the discriminator after decreasing a small positive number to the control variable, and I is a preset positive integer.

8. The method of claim 1, wherein the method is based on generative AI. I=2S+1; S is the number of control targets.

9. The method of claim 1, wherein the method is based on generative AI. The sample set includes historical data of temperature, time sequence characteristics influencing the control target and state characteristics during the period from the time starting point of process start to the end of the second time interval. The input data of the generator is the time sequence data of temperature, time sequence characteristics influencing the control target and state characteristics at the same time point, and the output data is the predicted time sequence data of the control variable; the historical data of the state characteristic at the same time point, the historical data of the time sequence characteristic influencing the control target and the predicted time sequence data of the control variable are input, and the predicted control target is output by the trained discriminator.

10. The method of claim 9, wherein the method is based on generative AI. The generator training error formula includes: ， wherein, is the generator training error, is the setpoint corresponding to the s-th control target, is the s-th control target predicted by the trained discriminator, S is the number of control targets, is the mean of the s-th control target predicted by the trained discriminator.

Images