Rolling temperature control method

Abstract

The application relates to the technical field of steel hot rolling process control and industrial intelligent optimization, and discloses a rolling temperature control method, which comprises the following steps: acquiring PDI information, process point target temperature, and rolling speed and measured temperature collected online; determining genetic characteristic parameters by using a regression chain model, obtaining a first predicted temperature through an instant learning model and an RNN time sequence model; taking the measured temperature at the rough rolling outlet as an initial temperature, calculating a second predicted temperature by using a mechanism temperature model, and positively regulating and controlling cooling and speed according to the deviation after the first predicted temperature is corrected; if the coiling temperature and the finishing rolling temperature cannot both meet the standards under the current capacity, the coiling temperature is preferentially corrected in a reverse optimization mode, and a control instruction is generated for execution. Through genetic characteristic extraction and mechanism-data fusion modeling driven by full-process data, the positive online regulation and control and the reverse process optimization are cooperatively and circularly controlled, so that the process temperature hit rate and the product performance consistency are improved.

Description

Technical Field

[0001] This application relates to the field of hot rolling process control and industrial intelligent optimization technology, and in particular to a rolling temperature control method. Background Technology

[0002] In the production of hot-rolled strip steel, temperature is the core process variable determining the product's microstructure, dimensional accuracy, and surface quality. After being heated in the furnace, the strip steel undergoes multiple processes, including rough rolling deformation, intermediate slab cooling, finish rolling deformation, and post-rolling cooling. The temperature field distribution and evolution of each process directly affect the microstructure transformation and final mechanical properties of the metal. With downstream users increasingly demanding consistent strip steel performance, and with hot-rolling production lines moving towards higher speeds and greater automation, higher requirements are being placed on the precision and coordination of temperature control throughout the entire process.

[0003] Currently, the temperature control of hot-rolled strip steel mainly adopts a single-process local control method, such as independently adjusting the outlet temperature of the heating furnace, the outlet temperature of the finishing mill, or the coiling temperature. The temperature setting and control of each process are relatively fragmented, lacking a systematic consideration of the temperature inheritance and transmission relationship of the entire process of "heating-roughing-finishing-cooling". Summary of the Invention

[0004] In view of this, embodiments of this application provide a rolling temperature control method, including: Obtain the PDI information of the strip steel to be rolled and the target temperature of each preset process point in the entire rolling process, as well as the rolling speed and the measured temperature of each temperature node of the rolled part collected online. Using a pre-built regression chain model, the genetic feature parameters corresponding to the current temperature node are determined based on the measured temperature, target temperature, control setpoint, equipment status parameters, and steel plate operating parameters of each preceding temperature node upstream of the current temperature node. A pre-built real-time learning model is used to obtain self-learning values ​​with the genetic feature parameters as input, and a pre-built RNN time series model is used to determine the first predicted temperature of each preset process point based on the self-learning values. Using the measured temperature at the roughing mill exit as the initial temperature, the pre-constructed mechanism temperature model is used to calculate the second predicted temperature of each preset process point according to the rolling direction, based on the PDI information and the target temperature of each preset process point. The second predicted temperature is then corrected based on the first predicted temperature of each preset process point to obtain the final predicted temperature. Based on the deviation between the final predicted temperature and the corresponding target temperature, the parameters of the cooling device and the running speed of the steel plate are positively controlled. Using the target coiling temperature and target final rolling temperature as fixed process target constraints, it is determined whether the final rolling temperature and coiling temperature can simultaneously reach the corresponding fixed process target constraints under the current cooling capacity and current rolling speed regime. If they can be reached simultaneously, the positive control result is executed; otherwise, the target coiling temperature is given the primary priority and the target final rolling temperature is given the secondary priority, and reverse optimization correction is performed in the direction opposite to the rolling direction. Control commands are generated based on the revised control variables for each process, and the speed actuator and cooling device are controlled based on the control commands.

[0005] By employing the above technical solutions, the rolling temperature control method provided in this application extracts genetic feature parameters from the entire process data using a regression chain model, and jointly predicts the first predicted temperature for each preset process point using an instant learning model and an RNN time series model, achieving rapid and accurate prediction of the entire process temperature evolution using a data-driven method. The second predicted temperature for each preset process point is recursively calculated according to the rolling direction using a mechanistic temperature model, and the second predicted temperature is corrected using the first predicted temperature to obtain the final predicted temperature, achieving an organic integration of the mechanistic model and the data-driven model. This ensures both the physical interpretability of the prediction results and improves adaptability to complex working conditions. Real-time online temperature adjustment is achieved by positively controlling the deviation between the final predicted temperature and the target temperature. Reverse optimization correction with priority given to the coiling temperature when the final rolling temperature and coiling temperature cannot be simultaneously met ensures the pass rate of the final product performance under extreme working conditions. Overall, this embodiment can overcome the problems of insufficient full-process coordination, unstable prediction accuracy and weak online optimization capability in the prior art, realize integrated temperature control of heating-rolling-cooling and multi-process collaborative optimization, and improve the final cooling temperature hit rate and product quality stability.

[0006] The above description is only an overview of the technical solution of this application. In order to better understand the technical means of this application and to implement it in accordance with the contents of the specification, and to make the above and other objects, features and advantages of this application more obvious and understandable, the following are specific embodiments of this application. Attached Figure Description

[0007] The accompanying drawings, which are included to provide a further understanding of this application and form part of this application, illustrate exemplary embodiments and are used to explain this application, but do not constitute an undue limitation of this application. In the drawings: Figure 1 A schematic flowchart of a rolling temperature control method provided in an embodiment of this application is shown; Figure 2 A schematic flowchart of another rolling temperature control method provided in an embodiment of this application is shown; Figure 3 A schematic diagram of the overall architecture of a temperature control system provided in an embodiment of this application is shown. Detailed Implementation

[0008] The present application will be described in detail below with reference to the accompanying drawings and embodiments. It should be noted that, unless otherwise specified, the embodiments and features described in the embodiments of the present application can be combined with each other.

[0009] This embodiment provides a rolling temperature control method, such as Figure 1 As shown, the method includes: Step 101: Obtain the PDI information of the strip steel to be rolled and the target temperature of each preset process point in the entire rolling process, as well as the rolling speed collected online and the measured temperature of each temperature node of the rolled part.

[0010] Among them, PDI information refers to the product design information of strip steel, including basic data such as steel composition and finished product specifications. These data determine the thermophysical parameters and phase transformation characteristics of strip steel during heating and deformation. Preset process points refer to key temperature detection locations pre-set along the rolling direction, including locations such as after intermediate billet cooling, at the finish mill inlet, finish mill outlet, and coiler inlet. High-temperature gauges are installed at these locations for temperature measurement. Target temperatures refer to the desired temperature settings at each preset process point, including the target intermediate billet cooling temperature, target finish mill inlet temperature, target final rolling temperature, and target coiling temperature. Measured temperatures refer to the temperature values ​​measured by the high-temperature gauges during actual production, including the measured temperatures at the roughing mill outlet, finish mill inlet, finish mill outlet, and coiling.

[0011] Specifically, the production line control system first acquires the PDI (Pressure Intake) information of the strip steel to be rolled. This information is generated and stored in a database by the host computer when the production plan is issued. It also reads the target temperature values ​​for each preset process point, which are pre-set according to the steel grade, specifications, and process procedures. Simultaneously, pyrometers located at key positions on the production line collect rolling speed signals and measured temperatures at each temperature node online. For example, pyrometers are installed at the roughing mill exit, finishing mill entrance, finishing mill exit, and coiler entrance to monitor the surface temperature of the rolled strip steel in real time.

[0012] Step 102: Using a pre-built regression chain model, determine the genetic feature parameters corresponding to the current temperature node based on the measured temperature, target temperature, control setpoint, equipment status parameters, and steel plate operating parameters of each preceding temperature node upstream of the current temperature node; use a pre-built instant learning model with the genetic feature parameters as input to obtain self-learning values, and use a pre-built RNN time series model to determine the first predicted temperature of each preset process point based on the self-learning values.

[0013] Among them, the regression chain model refers to a sequence of regression models constructed sequentially along the rolling direction according to adjacent temperature nodes, with each regression sub-model representing the temperature transfer relationship between two adjacent temperature nodes. Genetic feature parameters refer to feature vectors representing the comprehensive influence of upstream process temperature state, speed regime, cooling conditions, and equipment conditions on downstream temperature nodes. These parameters are obtained by dimensionless transformation and fusion of one or more of the following: cross-process temperature transfer influence, temperature drop sensitivity influence, thickness influence, composition influence, equipment state drift influence, and operating condition fluctuation influence. The instant learning model refers to a fast prediction model based on a local neighborhood search strategy, including one or a combination of OPC, GRP, and KNN. This model achieves rapid prediction by retrieving the most similar historical samples to the current operating condition from a historical database. The self-learning value refers to the correction bias output by the instant learning model, used to compensate for the prediction bias of the mechanistic model under the current operating condition. The RNN time series model refers to a recurrent neural network model used to capture the time series features of temperature evolution along the rolling direction.

[0014] Specifically, the measured temperatures, target temperatures, control setpoints, equipment status parameters, and steel plate operating parameters of each preceding temperature node upstream of the current temperature node are input into a pre-constructed regression chain model. The input data organization method for the regression chain model is as follows: for the i-th temperature node, all preceding data from the roughing mill exit to the (i-1)-th temperature node are used as input features. After calculation by the regression chain model, the corresponding genetic feature parameters for the current node are output. These genetic feature parameters are then input into an instant learning model. The instant learning model retrieves the K most similar historical samples to the current genetic feature parameters from the historical database and calculates the self-learning value under the current operating condition based on the measured deviation values ​​of these similar samples. Subsequently, the self-learning value is input into an RNN time series model. The RNN model utilizes its sequence memory capability, combined with the time series characteristics of the self-learning value, to output the first predicted temperature for each preset process point.

[0015] By combining regression chain model, instant learning model and RNN time series model for prediction, rapid and accurate prediction of temperature at each preset process point in the whole process is achieved. The data-driven method can effectively capture complex nonlinear relationships that are difficult to describe by mechanistic models. At the same time, the introduction of instant learning mechanism enables the model to quickly adapt to steel grade switching and operating condition changes.

[0016] Step 103: Using the measured temperature at the roughing mill exit as the initial temperature, the second predicted temperature of each preset process point is calculated sequentially according to the rolling direction based on the PDI information and the target temperature of each preset process point using a pre-constructed mechanism temperature model. The second predicted temperature is then corrected based on the first predicted temperature of each preset process point to obtain the final predicted temperature. Based on the deviation between the final predicted temperature and the corresponding target temperature, the parameters of the cooling device and the running speed of the steel plate are positively adjusted.

[0017] The mechanistic temperature model refers to a physical mechanism model constructed based on heat transfer, thermodynamics, and plastic deformation theories. This model covers the roughing rolling process, descaling process, intermediate billet cooling process, finishing rolling process, and post-rolling cooling process. It includes models for roughing descaling, intermediate billet conveying temperature drop, intermediate billet cooling, finishing rolling deformation temperature rise, finishing rolling friction temperature rise, finishing mill inter-stand temperature drop, finishing mill exit temperature prediction, and post-rolling cooling temperature drop. This model performs discrete temperature calculations along the length and thickness of the strip. The initial temperature refers to the starting temperature value calculated in a forward recursive manner, using the actual temperature measured by a pyrometer at the roughing mill exit as the initial temperature. The second predicted temperature refers to the temperature prediction values ​​for each preset process point calculated sequentially by the mechanistic temperature model according to the rolling direction. The final predicted temperature is the temperature value obtained after correcting the second predicted temperature with the first predicted temperature. Forward control refers to the real-time adjustment of the cooling device parameters and operating speed according to the deviation between the predicted temperature and the target temperature, based on the strip's running direction.

[0018] For example, the step of correcting the second predicted temperature based on the first predicted temperature of each preset process point to obtain the final predicted temperature includes: using a weighted fusion or residual correction fusion method, taking the first predicted temperature of each preset process point as the correction benchmark for the corresponding process point, and correcting the temperature value of the corresponding node in the second predicted temperature.

[0019] Specifically, the measured temperature at the roughing mill exit pyrometer is used as the initial temperature for forward recursive calculation. PDI information and the target temperatures for each preset process point are input into a pre-constructed mechanistic temperature model. The mechanistic temperature model, following the strip steel running direction, sequentially calls each sub-model for recursive calculation: starting from the roughing mill exit, it sequentially performs temperature drop calculations for descaling, intermediate billet conveying, intermediate billet cooling, finishing mill deformation, finishing mill friction, finishing mill inter-stand temperature drop, finishing mill exit temperature prediction, and post-rolling cooling temperature drop calculations, ultimately obtaining the second predicted temperature for each preset process point. After obtaining the second predicted temperature, the first predicted temperature is used as the correction benchmark for the corresponding process point. A weighted fusion method or a residual correction fusion method is used to correct the second predicted temperature to obtain the final predicted temperature. In the weighted fusion method, the sum of the weight coefficients of the mechanistic model and the data-driven model is 1, and their weights can be dynamically adjusted according to the confidence level of each model under the current operating conditions. In the residual correction fusion method, the deviation between the first predicted temperature and the second predicted temperature is used as the residual, and this residual is superimposed on the second predicted temperature of subsequent nodes. Then, the deviation between the final predicted temperature and the corresponding target temperature is calculated, and based on the sign and magnitude of this deviation, the parameters of the intermediate billet cooling device, the cooling device between the finishing mill stands, and the cooling device in the post-rolling cooling zone, as well as the running speed of the steel plate, are positively controlled.

[0020] Step 104: Using the target coiling temperature and target final rolling temperature as fixed process target constraints, determine whether the final rolling temperature and coiling temperature can simultaneously reach the corresponding fixed process target constraints under the current cooling capacity and current rolling speed regime; if they can be reached simultaneously, execute according to the positive control result; otherwise, take the target coiling temperature as the primary priority and the target final rolling temperature as the secondary priority, and perform reverse optimization correction in the direction opposite to the rolling direction.

[0021] Among these, fixed process target constraints refer to the target coiling temperature and target final rolling temperature as rigid process targets that must be met first, with higher priority than other dynamic targets. Cooling capacity refers to the maximum cooling intensity that the current cooling device can provide, which is limited by factors such as cooling water flow rate, water pressure, water temperature, and valve opening. Rolling speed regime refers to the rolling speed setting scheme for each stand of the finishing mill. Reverse optimization correction refers to the process of calculating the correction amount of the control variables of each upstream process in the opposite direction of rolling, starting from the coiling temperature target, when forward control cannot simultaneously achieve the target final rolling temperature.

[0022] Specifically, the target coiling temperature and target final rolling temperature are used as fixed process target constraints. It is determined whether, under the current cooling capacity and rolling speed regime, the final rolling temperature and coiling temperature can simultaneously reach the corresponding temperature values ​​within the fixed process target constraint range. This determination is based on the predicted values ​​of the target coiling temperature and target final rolling temperature in the final predicted temperature. If the determination result indicates that they can be reached simultaneously, the positive control result of step 103 is executed without triggering additional reverse correction. If the determination result indicates that they are difficult to reach simultaneously, the target coiling temperature is given the primary priority, and the target final rolling temperature the secondary priority, and the reverse optimization correction process is initiated. The specific process of reverse optimization correction is as follows: using at least one of the following as the control variable to be optimized: post-rolling cooling parameters, finishing rolling speed regime, finishing mill inter-stand cooling parameters, intermediate billet cooling parameters, intermediate roller speed, and conveying rhythm, under the conditions of satisfying the cooling device capacity constraint, speed adjustment range constraint, and rolling stability constraint, the correction amount of the control variable to be optimized is calculated in the direction opposite to the rolling direction. When a combination of control variables exists that simultaneously satisfies the allowable deviation ranges of both the target coiling temperature and the target final rolling temperature, this combination is output as a reverse correction value. If no such combination exists, the combination of control variables that prioritizes satisfying the allowable deviation range of the target coiling temperature and minimizes the deviation of the target final rolling temperature is output as the reverse correction value. During the reverse optimization process, if the iterative calculation exceeds the preset maximum number of iterations and still fails to converge, the system automatically outputs the current combination of control variables that minimizes the objective function value and issues a warning message to prompt operator intervention.

[0023] By prioritizing and correcting through reverse optimization, the winding temperature is adjusted with the highest priority when the equipment capacity is insufficient, thus maximizing the mechanical performance qualification rate of the final product. At the same time, the reverse optimization mechanism provides operators with the optimal control scheme under extreme working conditions.

[0024] In an optional embodiment, during the reverse optimization correction process, the method further includes: When there exists a first combination of control variables to be optimized that simultaneously satisfies the allowable deviation ranges of the target coiling temperature and the target final rolling temperature, the first combination of control variables to be optimized is output as a reverse correction amount. When there is no first combination of control variables that simultaneously satisfies the allowable deviation ranges of both the target coiling temperature and the target final rolling temperature, the output of the second combination of control variables that prioritizes satisfying the allowable deviation range of the target coiling temperature and minimizes the deviation of the target final rolling temperature is used as a reverse correction.

[0025] In this embodiment, the target coiling temperature and the target final rolling temperature are used as the reverse optimization targets. At least one of the following parameters is used as the control variables to be optimized: post-rolling cooling parameters, finishing rolling speed regime, finishing mill inter-stand cooling parameters, intermediate billet cooling parameters, intermediate roller speed, and conveying rhythm. Under the conditions of satisfying the cooling device capacity constraints, speed adjustment range constraints, and rolling stability constraints, the mechanistic temperature model is used to calculate the predicted coiling temperature and the predicted final rolling temperature corresponding to each combination of control variables to be optimized in the opposite direction to the rolling direction.

[0026] During the traversal calculation, the system first determines whether there exists a first combination of control variables that simultaneously satisfies the allowable deviation ranges of both the target coiling temperature and the target final rolling temperature. If such a combination exists, it is output as a back-correction quantity for subsequent control command generation. If no first combination of control variables simultaneously satisfies both temperature allowable deviation ranges, the constraint on the target final rolling temperature is relaxed. The allowable deviation range of the target coiling temperature is used as the primary screening condition. Among all combinations of control variables that can ensure the predicted coiling temperature falls within the allowable deviation range of the target coiling temperature, the second combination of control variables that minimizes the deviation between the predicted final rolling temperature and the target final rolling temperature is selected and output as a back-correction quantity.

[0027] By using the above reverse optimization correction method, it is possible to achieve both the final rolling temperature and the coiling temperature when the equipment capacity is sufficient, and to ensure the mechanical properties of the final product by prioritizing the coiling temperature when the equipment capacity is insufficient.

[0028] In an optional embodiment, the objective function of the reverse optimization is: , In the formula, To encode temperature weights, Weighted by the final rolling temperature. > , and These are the calculation of the winding temperature and the target winding temperature, respectively. and The values ​​are calculated final rolling temperature and target final rolling temperature, respectively. U is the correction amount of the control variables to be optimized, including correction amount of ultra-fast cooling water volume, laminar cooling water volume, finishing rolling speed, inter-stand cooling correction amount, intermediate billet cooling correction amount, and intermediate roller table speed correction amount. λ is a preset coefficient.

[0029] In this embodiment, the aforementioned reverse optimization correction is achieved by constructing an objective function and solving for its minimum value. Specifically, the coiling temperature weight and the final rolling temperature weight are assigned to the coiling temperature deviation term and the final rolling temperature deviation term, respectively, with the coiling temperature weight being greater than the final rolling temperature weight. The coiling temperature deviation term is represented as the square of the difference between the calculated coiling temperature and the target coiling temperature, and the final rolling temperature deviation term is represented as the square of the difference between the calculated final rolling temperature and the target final rolling temperature. Simultaneously, the correction amount of the control variable to be optimized is added to the objective function in norm form and assigned a preset penalty coefficient to suppress large fluctuations in the control variable and prevent the correction amount from exceeding the equipment's execution capacity. The correction amounts of the control variables to be optimized include the correction amount for ultra-fast cooling water volume, laminar cooling water volume, finishing mill speed, inter-stand cooling, intermediate billet cooling, and intermediate roller table speed. By finding the minimum value of the objective function—that is, under the constraint of prioritizing the coiling temperature—the optimal solution is found that minimizes both the final rolling temperature deviation and the control variable correction. The corresponding control variable corrections for this optimal solution are then output as reverse correction values. By introducing norm terms for the control variable corrections into the objective function and setting penalty coefficients, large jumps in the control variables during the reverse optimization correction process are prevented, allowing the correction values ​​to change smoothly. This is beneficial for the smooth operation of the on-site actuators and the long-term stable operation of the equipment.

[0030] Step 105: Generate control instructions based on the corrected control variables of each process, and control the speed actuator and cooling device based on the control instructions.

[0031] Among them, control commands refer to electrical or digital signals generated based on the modified control variables of each process to drive the field actuators. Speed ​​actuators refer to drive devices used to adjust the mill speed and roller speed, including the main drive motor and its speed regulating device. Cooling devices include intermediate billet cooling devices located between the roughing and finishing mill areas, inter-stand cooling devices located in the finishing mill area, and ultra-fast cooling devices and laminar flow cooling devices located in the post-rolling cooling area.

[0032] For example, each preset process point includes a roughing mill exit temperature node, an intermediate billet cooling temperature node, a finishing mill inlet temperature node, a finishing mill exit temperature node, and a coiling temperature node; the measured temperatures include the roughing mill exit measured temperature, the finishing mill inlet measured temperature, the finishing mill exit measured temperature, and the coiling measured temperature; the cooling device includes an intermediate billet cooling device disposed between the roughing mill area and the finishing mill area, a finishing mill stand cooling device disposed in the finishing mill area, and an ultra-fast cooling device and / or a laminar flow cooling device disposed in the post-rolling cooling area.

[0033] Specifically, based on the control variables of each process obtained after forward regulation in step 103 or reverse optimization in step 104, corresponding control commands are generated. For example, the control commands are generated by converting the values ​​of the control variables of each process into signal formats that can be recognized by the corresponding actuators. For the cooling device, the control commands include valve opening signals, flow setting signals, and pressure setting signals; for the speed actuator, the control commands include speed setting signals and acceleration setting signals.

[0034] By applying the technical solution of this embodiment, the genetic feature parameters of the entire process data are extracted through a regression chain model, and the first predicted temperature of each preset process point is jointly predicted by an instant learning model and an RNN time series model, realizing rapid and accurate prediction of the temperature evolution of the entire process using a data-driven method. The second predicted temperature of each preset process point is recursively calculated according to the rolling direction using a mechanistic temperature model, and the second predicted temperature is corrected using the first predicted temperature to obtain the final predicted temperature. This achieves the organic integration of the mechanistic model and the data-driven model, ensuring both the physical interpretability of the prediction results and improving adaptability to complex working conditions. Real-time online temperature adjustment is achieved by positively controlling the deviation between the final predicted temperature and the target temperature. Reverse optimization correction with priority given to the coiling temperature when the final rolling temperature and coiling temperature cannot be simultaneously met ensures the pass rate of the final product performance under extreme working conditions. Overall, this embodiment overcomes the problems of insufficient full-process coordination, unstable prediction accuracy, and weak online optimization capability in existing technologies, achieving integrated temperature control of heating-rolling-cooling and multi-process collaborative optimization, improving the final cooling temperature hit rate and product quality stability.

[0035] In an optional embodiment, the target coiling temperature and target final rolling temperature are used as fixed process target constraints. It is determined whether the final rolling temperature and coiling temperature can simultaneously reach the corresponding fixed process target constraints under the current cooling capacity and current rolling speed regime. If they can be reached simultaneously, the positive control result is executed; otherwise, the target coiling temperature is given primary priority and the target final rolling temperature secondary priority, and reverse optimization correction is performed in the direction opposite to the rolling direction, including: Using the target coiling temperature, target final rolling temperature, and target finishing mill inlet temperature as fixed process target constraints, it is determined whether the final rolling temperature, coiling temperature, and finishing mill inlet temperature can simultaneously reach the corresponding fixed process target constraints under the current cooling capacity and current rolling speed regime. If they can be reached simultaneously, the positive control result is executed; otherwise, the target coiling temperature is given the first priority, the target final rolling temperature the second priority, and the target finishing mill inlet temperature the third priority, and reverse optimization correction is performed in the direction opposite to the rolling direction.

[0036] In this embodiment, the target coiling temperature, target final rolling temperature, and target finishing mill entry temperature are first used as fixed process target constraints. It is then determined whether the predicted values ​​of these three temperatures, under the current cooling capacity and rolling speed regime, can simultaneously fall within the allowable deviation range defined by their respective fixed process target constraints. This determination is based on the predicted values ​​corresponding to the coiling temperature, final rolling temperature, and finishing mill entry temperature from the final predicted temperatures obtained after fusion correction. When all three predicted values ​​are within their respective allowable deviation ranges, it is determined that the target can be met simultaneously, and the forward control result is executed without triggering the reverse optimization correction process. When it is determined that the above three fixed process target constraints cannot be met simultaneously, reverse optimization correction is initiated. In reverse optimization correction, optimization is performed stepwise in the order of coiling temperature as the primary priority, final rolling temperature as the secondary priority, and finishing mill entry temperature as the third priority. By incorporating the finishing mill entry temperature into the fixed process target constraint system and assigning it a third priority, the finishing mill entry temperature receives progressive protection in reverse optimization correction, which helps improve the rolling stability of the finishing rolling process. The three-level priority screening strategy takes into account the primary and secondary impacts of each temperature node on product quality, while ensuring that executable reverse corrections can be output under various operating conditions.

[0037] In an optional embodiment, during the reverse optimization correction process, the method further includes: When there exists a third combination of control variables to be optimized that simultaneously satisfies the allowable deviation ranges of the target coiling temperature, the target final rolling temperature, and the target finishing rolling inlet temperature, the third combination of control variables to be optimized is output as a reverse correction amount. When there is no third combination of control variables that simultaneously satisfies the allowable deviation ranges of the target coiling temperature, target final rolling temperature, and target finishing mill inlet temperature, the fourth combination of control variables that prioritizes satisfying the allowable deviation ranges of the target coiling temperature and target final rolling temperature and minimizes the deviation of the target finishing mill inlet temperature is output as a reverse correction amount. When there is no fourth combination of control variables that simultaneously satisfies the allowable deviation ranges of the target coiling temperature and the target final rolling temperature, the fifth combination of control variables that prioritizes satisfying the allowable deviation range of the target coiling temperature and minimizes the deviation of the target final rolling temperature is output as a reverse correction amount. The objective function of the reverse optimization is: , In the formula, To encode temperature weights, Weighted by the final rolling temperature. Weighted by the inlet temperature of the finishing mill. , and These are the calculation of the winding temperature and the target winding temperature, respectively. and These are the calculation of the final rolling temperature and the target final rolling temperature, respectively. and The values ​​are calculated for the finishing mill inlet temperature and the target finishing mill inlet temperature, respectively. U represents the correction amount of the control variables to be optimized, including the correction amount for ultra-fast cooling water volume, laminar cooling water volume, finishing mill speed, inter-stand cooling, intermediate billet cooling, and intermediate roller speed. λ is a preset coefficient.

[0038] In this embodiment, at least one of the following parameters—post-rolling cooling parameters, finishing mill speed regime, finishing mill inter-stand cooling parameters, intermediate slab cooling parameters, intermediate roller speed, and conveying rhythm—is used as the control variable to be optimized. First, it is determined whether, under the conditions of satisfying the cooling device capacity constraint, speed adjustment range constraint, and rolling stability constraint, there exists a third combination of control variables to be optimized that allows the predicted values ​​of coiling temperature, final rolling temperature, and finishing mill inlet temperature to simultaneously fall within their respective allowable deviation ranges. Specifically, the determination method is as follows: a traversal search is performed within the feasible value range of all control variables to be optimized. Each combination of control variables to be optimized is substituted into the mechanistic temperature model and recursively calculated in the direction opposite to the rolling direction to obtain the corresponding predicted values ​​of coiling temperature, final rolling temperature, and finishing mill inlet temperature. Each of these is then checked to see if all three fall within their respective allowable deviation ranges. If at least one combination satisfies the above conditions, the combination that minimizes the weighted sum of the squares of the three temperature deviations is output as the third combination of control variables to be optimized as a reverse correction.

[0039] If it is determined that there is no third combination of control variables that can simultaneously satisfy the allowable deviation ranges of the three temperatures, then the second level of screening is performed: priority is given to satisfying the allowable deviation ranges of the target coiling temperature and the target final rolling temperature. Among all combinations of control variables that can simultaneously satisfy the allowable deviation ranges of the coiling temperature and the final rolling temperature, the combination that minimizes the deviation between the predicted value of the finishing mill inlet temperature and the target finishing mill inlet temperature is selected, and this combination is output as the fourth combination of control variables to be optimized as the reverse correction amount.

[0040] If it is determined that there is no fourth combination of control variables that can simultaneously satisfy the allowable deviation ranges of both coiling temperature and final rolling temperature, then the third level of screening is performed: taking the allowable deviation range of the target coiling temperature as the first constraint, among all combinations of control variables that can satisfy the allowable deviation range of the coiling temperature, the combination that minimizes the deviation between the predicted value of the final rolling temperature and the target final rolling temperature is selected, and this combination is output as the fifth combination of control variables to be optimized as the reverse correction amount.

[0041] In an optional embodiment, the pre-construction process of the regression chain model includes: acquiring historical full-process batch data, which includes the historical measured temperature of each batch of strip steel at each temperature node, the historical target temperature of each preset process point, the historical control setpoint, the historical equipment status parameters, and the historical steel plate operating parameters; using the temperature change between the (i-1)th temperature node and the ith temperature node as the output target, and using the historical measured temperature, historical target temperature, historical control setpoint, historical equipment status parameters, and historical steel plate operating parameters of each preceding temperature node before the (i-1)th temperature node as input features, training the ith regression sub-model to obtain the regression relationship between the (i-1)th temperature node and the ith temperature node; and traversing all adjacent temperature node pairs to obtain a regression chain model covering the regression relationship between adjacent temperature nodes throughout the entire process.

[0042] The regression chain model refers to a sequence of regression models constructed sequentially along the rolling direction at adjacent temperature nodes. Each regression sub-model characterizes the temperature transfer regression relationship between two adjacent temperature nodes. The temperature change refers to the temperature difference that occurs as the strip steel moves from the (i-1)th temperature node to the ith temperature node. This difference reflects the combined effect of various process factors on temperature drop or rise within that section. The regression sub-models are regression models trained separately for each pair of adjacent temperature nodes, and these sub-models are connected in series along the rolling direction to form a complete regression chain.

[0043] In this embodiment, historical batch data of the entire process is first obtained from the production line historical database. This historical data includes the measured temperature values ​​of each batch of strip steel at each temperature node, the target temperature setpoints of each preset process point, the control setpoints of each process control device, equipment status monitoring parameters, and operating parameters such as speed and reduction during the rolling process. After obtaining the historical data, the data is preprocessed: outliers that are significantly outside the reasonable range are removed, and missing values ​​are imputed using data from adjacent batches or adjacent time points to ensure the completeness and effectiveness of the training data. Then, the i-th regression sub-model is trained. For the i-th temperature node, the historical measured temperature, historical target temperature, historical control setpoints, historical equipment status parameters, and historical steel plate operating parameters of all preceding temperature nodes before the (i-1)-th temperature node are used as input features, and the temperature change between the (i-1)-th temperature node and the i-th temperature node is used as the output target. A regression algorithm is used to fit the parameters of the sub-model to obtain the regression relationship between the (i-1)-th temperature node and the i-th temperature node. For example, the regression algorithm used during training includes any one of multiple linear regression, partial least squares regression, support vector regression, or neural network regression. Following the above method, the corresponding regression sub-model is trained sequentially for each pair of adjacent temperature nodes: the first sub-model is trained for the first temperature node to the second temperature node, the second sub-model is trained for the second temperature node to the third temperature node, and so on, until all pairs of adjacent temperature nodes are traversed. All sub-models are then concatenated and combined along the rolling direction to obtain a regression chain model covering the regression relationships between adjacent temperature nodes throughout the entire process.

[0044] The regression chain model constructed through the above optional embodiments enables explicit modeling of the temperature transfer relationship between adjacent temperature nodes. Each sub-model is trained independently and used in series, which reduces the complexity of model training and allows the temperature deviation to be gradually transferred along the rolling direction, providing a data-driven foundation with process interpretability for accurate prediction of the temperature of the entire process.

[0045] In one optional embodiment, the pre-construction process of the instant learning model includes: Acquire historical full-process batch data, extract the historical genetic characteristic parameter vectors of each batch of plate and strip steel at each temperature node and the corresponding historical measured temperature deviation values, and construct a historical sample database; The step of obtaining self-learning values ​​using a pre-built instant learning model with the genetic feature parameters as input includes: The instant learning model employs a local neighborhood search strategy. When the genetic characteristic parameters of the current strip steel are received, the model retrieves the K most similar historical samples to the genetic characteristic parameters from the historical sample database. Based on the historical measured temperature deviation values ​​of the K historical samples, the model calculates the self-learning value corresponding to the genetic characteristic parameters using local weighted regression or local weighted averaging.

[0046] The instant learning model refers to a fast prediction model based on a local neighborhood search strategy. This model does not perform global parameter fitting but dynamically selects historically similar samples for local modeling upon receiving a query sample. The historical genetic feature parameter vector is a feature vector composed of genetic feature parameters extracted from historical batch data at each temperature node. This vector includes one or more feature components obtained through dimensionless and fusion processing, including cross-process temperature transfer effects, temperature drop sensitivity effects, thickness effects, composition effects, equipment state drift effects, and operating condition fluctuation effects. The historical measured temperature deviation value refers to the difference between the measured temperature of the same strip steel in a historical batch at the same temperature node and the temperature predicted by the mechanism model. The local neighborhood search strategy refers to a search method that retrieves the K most similar historical samples in the feature space to the current query sample from the historical sample database. Local weighted regression refers to weighted regression fitting based on the selected K historical samples, assigning different weights according to the similarity between each sample and the current query sample. The self-learning value refers to the correction deviation calculated by the instant learning model based on locally similar samples, used to compensate for the prediction deviation of the mechanism model under the current operating conditions.

[0047] In this embodiment, historical batch data is first acquired. This historical data includes the measured temperature values ​​of each batch of strip steel at each temperature node, the mechanism-predicted temperature values ​​calculated by the mechanism temperature model, and the genetic feature parameter vectors at each temperature node. The genetic feature parameter vectors and corresponding measured temperature deviation values ​​for each batch of strip steel at each temperature node are extracted from the acquired historical data. The measured temperature deviation value is calculated by subtracting the measured temperature value at the temperature node from the mechanism-predicted temperature value calculated by the mechanism temperature model. The extracted genetic feature parameter vectors and corresponding measured temperature deviation values ​​are associated and stored to construct a historical sample database. When constructing the historical sample database, historical samples are classified and stored according to steel type and specification, and an independent index is created for each type of sample to accelerate the subsequent retrieval process.

[0048] The instant learning model obtains self-learning values ​​using the constructed historical sample database as follows: When the genetic characteristic parameters of the current strip steel are received, these parameters are used as query vectors to retrieve the K most similar historical samples from the historical sample database. Similarity can be measured using Euclidean distance, Mahalanobis distance, or cosine similarity. During the retrieval process, if the distance between the most similar historical sample and the current query sample in the historical sample database is greater than a preset threshold, it is determined that the current working condition lacks effective similar samples. The system automatically uses a default self-learning value as the output and simultaneously marks the current sample as a new type of sample. Once this type of sample accumulates to a preset number, it is re-included in the historical sample database. After retrieving the K historical samples, the self-learning value corresponding to the current genetic characteristic parameter is calculated based on the historical measured temperature deviation values ​​of these K historical samples. The calculation method includes local weighted regression or local weighted averaging. In the locally weighted averaging method, the weight of each sample is determined according to the similarity between each historical sample and the current query sample; the higher the similarity, the greater the weight. The measured temperature deviation values ​​of all historical samples are then weighted and averaged to obtain the self-learning value corresponding to the current genetic feature parameter. In the locally weighted regression method, a local regression model is established based on the correspondence between the genetic feature parameters of K historical samples and the measured temperature deviation values. The current genetic feature parameter is then substituted into this local regression model to obtain the corresponding self-learning value.

[0049] Through the above-described construction of an instant learning model and inference process, rapid self-learning value prediction based on locally similar samples is achieved. This model can adapt to new working conditions without global retraining and has a rapid response capability to steel grade switching and specification changes. At the same time, when similar historical samples are lacking, it can automatically switch to the default output and label new types of samples, providing a data accumulation mechanism for continuous model optimization.

[0050] In an optional embodiment, the pre-construction process of the RNN time series model includes: Acquire historical full-process batch data and extract the historical process parameter sequences of each batch of plate and strip steel at multiple historical moments as time-series input feature samples; The RNN network is trained by using the sequence of historical process parameters from N times prior to any given time and the historical self-learning value at any given time as input features, and the measured temperature values ​​of the corresponding batch of strip steel at each preset process point as output targets. The RNN network adopts a long short-term memory network or a gated recurrent unit network.

[0051] Specifically, the RNN time-series model can be a recurrent neural network model used to capture the time-series characteristics of temperature evolution along the rolling direction. This model has a memory unit structure and can learn the temporal dependence and variation patterns of temperature. The historical process parameter sequence refers to the numerical sequence of multiple process parameters recorded sequentially along time during the rolling process of the same batch of strip steel. Process parameters include at least one of temperature, rolling speed, and cooling water volume. N time points refer to the length of the time window before the current time point, where N is a positive integer. This time window length is determined based on the production line speed and temperature response time. Long Short-Term Memory (LSTM) networks are recurrent neural network variants containing input gates, forget gates, and output gates. Through gating mechanisms, they control the long-term memory and short-term updates of information, effectively solving the gradient vanishing problem in the training process of traditional recurrent neural networks. Gated Recurrent Unit (GRU) networks are recurrent neural network variants containing update gates and reset gates. Their structural complexity is lower than LSM networks, resulting in faster training speeds, making them suitable for applications with relatively small training data volumes. The measured temperature value refers to the preset process point temperature value actually collected and recorded by a pyrometer during historical production, serving as the ground truth label for the RNN network's output target during the training phase.

[0052] In this embodiment, historical full-process batch data is first obtained. This historical data includes the values ​​of multiple process parameters collected along the time sequence during the rolling process of each batch of strip steel. The process parameters include the measured temperature values ​​of each temperature node, the rolling speed setpoint and feedback value, the water volume and water pressure signal of the cooling device, and the measured temperature values ​​of each preset process point.

[0053] After acquiring historical data, the historical process parameter sequences of each batch of strip steel at multiple historical moments are extracted from this data as time-series input feature samples. The extraction method is as follows: the entire rolling process of each batch of strip steel from the roughing mill exit to the coiler entrance is divided into multiple moments in chronological order. At each moment, the values ​​of process parameters such as temperature, rolling speed, and cooling water volume are recorded, forming a complete time-series sequence of process parameters reflecting the temperature evolution of that batch of strip steel. Subsequently, a training sample set for the RNN network is constructed. For any given moment, the historical process parameter sequences of the previous N moments and the historical self-learning value at that moment are used as input features, and the measured temperature value of the preset process point at the corresponding position of that batch of strip steel at that moment is used as the output target, forming a set of training samples. A sliding window is used to traverse all moments according to the time step, resulting in multiple sets of training samples, constituting a complete training sample set. After the training sample set is constructed, it is divided into a training set and a validation set according to a preset ratio. The training set is used for model training, and the validation set is used for model performance monitoring and early stop judgment during the training process. The RNN network is then trained using the above training sample set. The RNN network employs either a Long Short-Term Memory (LSTM) network or a gated recurrent unit (GRU) network structure. During training, the input feature sequence is sequentially fed into the RNN network. The network calculates the predicted temperature value through forward propagation and then calculates the gradient of the loss function between the predicted and measured temperatures through backpropagation. The network weights are adjusted based on this gradient until the deviation between the predicted and measured temperatures meets the preset accuracy requirements. During training, if the prediction deviation on the validation set no longer decreases after a preset number of iterations, training is terminated using an early stopping strategy. The network weights of the current iteration are used as the final RNN temporal model parameters to prevent overfitting. If the loss function value oscillates and fails to converge during training, the learning rate is automatically reduced, and training continues until the loss function value stabilizes or the preset maximum number of training iterations is reached.

[0054] The RNN time series model constructed through the above optional embodiments uses a long short-term memory network or a gated recurrent unit network to model the long-term dependencies in the temperature time series data, enabling the model to capture the dynamic characteristics of temperature evolution along the rolling direction and provide data-driven support with time series memory capability for temperature prediction at each preset process point.

[0055] Furthermore, as a refinement and extension of the specific implementation of the above embodiments, and to fully illustrate the specific implementation process of this embodiment, another rolling temperature control method is provided, such as... Figure 2 As shown, the method includes: Step 201: When the measured temperature at the finishing mill inlet is obtained, the setting parameters of the cooling devices between the finishing mill stands and the cooling devices in the post-rolling cooling zone are corrected based on the deviation between the measured temperature at the finishing mill inlet and the final predicted temperature at the finishing mill inlet. Step 202: When the final rolling measured temperature is obtained, the setting parameters of the cooling device in the post-rolling cooling zone and the conveyor roller speed are corrected according to the deviation between the final rolling measured temperature and the final rolling predicted temperature. Step 203: When the actual winding temperature is obtained, adjust the setting parameters of the final cooling device in the post-rolling cooling zone according to the deviation between the actual winding temperature and the final predicted winding temperature.

[0056] The measured temperature at the finishing mill inlet refers to the actual temperature value of the strip steel detected by a pyrometer located at the inlet of the finishing mill. The final predicted temperature at the finishing mill inlet refers to the final predicted temperature at the finishing mill inlet location obtained after the fusion correction of the mechanistic model and the data-driven model in step 103. The cooling device between the finishing mill stands refers to the cooling equipment installed between adjacent stands of the finishing mill, used to regulate the temperature of the strip steel between each stand. The post-rolling cooling zone refers to the cooling section between the finishing mill exit and the coiler, equipped with at least one of an ultra-fast cooling device and a laminar flow cooling device. The measured temperature at the final rolling mill outlet refers to the actual temperature value of the strip steel detected by a pyrometer located at the finishing mill exit. The final predicted temperature at the final rolling mill outlet location obtained after the fusion correction in step 103. The conveyor roller speed refers to the speed setting value of the roller conveyor device between the finishing mill exit and the coiler. The measured temperature at the coiler inlet refers to the actual temperature value of the strip steel detected by a pyrometer located at the coiler inlet. The final predicted temperature for coiling refers to the final predicted temperature at the coiler inlet obtained after fusion correction in step 103. The final cooling device refers to the cooling equipment in the post-rolling cooling area near the coiler, including at least one of the final ultra-fast cooling manifold and the final laminar flow cooling manifold.

[0057] In this embodiment, when the measured temperature at the finishing mill inlet is obtained, the deviation between the measured temperature and the final predicted temperature at the finishing mill inlet obtained in step 103 is calculated. Based on the sign and magnitude of the deviation, the setting parameters of the cooling device between the finishing mill stands and the setting parameters of the cooling device in the post-rolling cooling zone are corrected. Specifically, the correction method is as follows: when the measured temperature at the finishing mill inlet is higher than the final predicted temperature at the finishing mill inlet, it indicates that the actual temperature of the strip steel entering the finishing mill is too high. At this time, the cooling water volume or water pressure of the cooling device between the finishing mill stands is increased, and the setting parameters of the cooling device in the post-rolling cooling zone are adjusted accordingly to compensate for the impact of the high finishing mill inlet temperature on subsequent temperature nodes; when the measured temperature at the finishing mill inlet is lower than the final predicted temperature at the finishing mill inlet, the cooling water volume or water pressure of the cooling device between the finishing mill stands is decreased, and the setting parameters of the cooling device in the post-rolling cooling zone are adjusted accordingly.

[0058] When the final rolling measured temperature is obtained, the deviation between the measured temperature and the final rolling predicted temperature obtained in step 103 is calculated. Based on this deviation, the setting parameters of the cooling device in the post-rolling cooling zone and the conveyor roller speed are corrected. Specifically, the correction method is as follows: when the final rolling measured temperature is higher than the final rolling predicted temperature, it indicates that the temperature of the strip steel at the finishing mill exit is too high. In this case, the cooling water volume or water pressure of the cooling device in the post-rolling cooling zone is increased, and the conveyor roller speed is appropriately reduced to prolong the residence time of the strip steel in the cooling zone, so that the coiling temperature can drop back to the target range; when the final rolling measured temperature is lower than the final rolling predicted temperature, the cooling water volume or water pressure of the cooling device in the post-rolling cooling zone is reduced, and the conveyor roller speed is appropriately increased to shorten the residence time of the strip steel in the cooling zone, preventing the coiling temperature from being too low.

[0059] When the actual winding temperature is obtained, the deviation between the actual temperature and the final predicted winding temperature obtained in step 103 is calculated. Based on this deviation, the setting parameters of the final cooling device in the post-rolling cooling zone are adjusted. Specifically, when the actual winding temperature is higher than the final predicted winding temperature, the cooling water volume or water pressure of the final cooling device is increased to improve the cooling intensity before winding; when the actual winding temperature is lower than the final predicted winding temperature, the cooling water volume or water pressure of the final cooling device is decreased to reduce the cooling intensity before winding.

[0060] Through the online correction in this embodiment, the cooling device parameters and speed parameters of the corresponding sections are corrected in real time by using the measured temperatures at the three locations of the finishing mill inlet, the final mill outlet, and the coiler inlet. This ensures that the control accuracy of the forward control process and the reverse optimization correction process is maintained continuously, forming a closed-loop control of the entire process of "prediction → control → measurement → correction". This eliminates the influence of model prediction deviation and external disturbances on the final coiling temperature hit rate.

[0061] In a specific application scenario, such as Figure 3As shown in the figure, this is a schematic diagram of the overall architecture of a temperature control system provided in this application. It fully demonstrates the closed-loop workflow of the system from three levels: data acquisition, model processing, and control execution. The bottom of the figure, starting from the furnace outlet, sequentially passes through the roughing mill, roughing mill exit, intermediate billet cooling, finishing mill, final rolling mill exit, ultra-fast cooling, and finally coiling, forming a complete temperature node sequence. The strip temperature transfer relationship between these nodes reflects the "temperature inheritance" characteristic; the temperature state, speed regime, and cooling conditions of upstream processes continuously affect downstream temperature nodes along the rolling direction. The data flow in the figure is divided into three paths: process data (rolling speed, cooling water volume, etc.) is directly input into the mechanistic temperature model; process data (measured temperature, target temperature, etc.) is input into the real-time learning model; and environmental data participates in system calculations as auxiliary information. The mechanistic model provides physical baseline predictions, the real-time learning model rapidly outputs self-learned values ​​based on locally similar samples, and the genetic coefficient is transferred between the mechanistic model and the real-time learning model to characterize the strength of genetic influence between adjacent temperature nodes. The recurrent neural network model receives the self-learned values ​​and, combined with its temporal memory capability, outputs predicted values ​​for key temperature nodes. These multiple models work together to form a hybrid prediction architecture of "mechanistic temperature model + real-time learning model + recurrent neural network model." In the diagram, the forward control arrows indicate that the system, based on the predicted values, issues control commands along the rolling direction to the cooling device and speed actuator; the reverse control arrows indicate that when the final quality target is difficult to achieve, the system calculates the upstream temperature setting and control variable correction in the opposite direction to the rolling direction. Finally, all control commands are executed in segments by two controllers: controller one is responsible for roughing and front-end cooling control, and controller two is responsible for finishing and rear-end cooling control, forming a closed-loop control system of "prediction-control-execution-feedback."

[0062] The technical features of the above embodiments can be combined in any way. For the sake of brevity, not all possible combinations of the technical features in the above embodiments are described. However, as long as there is no contradiction in the combination of these technical features, they should be considered to be within the scope of this specification.

[0063] The embodiments described above are merely illustrative of several implementation methods of this application, and while the descriptions are specific and detailed, they should not be construed as limiting the scope of this patent application. It should be noted that those skilled in the art can make various modifications and improvements without departing from the concept of this application, and these all fall within the protection scope of this application. Therefore, the protection scope of this application should be determined by the appended claims.

Claims

1. A rolling temperature control method, characterized in that, include: Obtain the PDI information of the strip steel to be rolled and the target temperature of each preset process point in the entire rolling process, as well as the rolling speed and the measured temperature of each temperature node of the rolled part collected online. Using a pre-built regression chain model, the genetic feature parameters corresponding to the current temperature node are determined based on the measured temperature, target temperature, control setpoint, equipment status parameters, and steel plate operating parameters of each preceding temperature node upstream of the current temperature node. A pre-built real-time learning model is used to obtain self-learning values ​​with the genetic feature parameters as input, and a pre-built RNN time series model is used to determine the first predicted temperature of each preset process point based on the self-learning values. Using the measured temperature at the roughing mill exit as the initial temperature, the pre-constructed mechanism temperature model is used to calculate the second predicted temperature of each preset process point according to the rolling direction, based on the PDI information and the target temperature of each preset process point. The second predicted temperature is then corrected based on the first predicted temperature of each preset process point to obtain the final predicted temperature. Based on the deviation between the final predicted temperature and the corresponding target temperature, the parameters of the cooling device and the running speed of the steel plate are positively controlled. Using the target coiling temperature and target final rolling temperature as fixed process target constraints, it is determined whether the final rolling temperature and coiling temperature can simultaneously reach the corresponding fixed process target constraints under the current cooling capacity and current rolling speed regime; if they can be reached simultaneously, then the positive control result is executed. Otherwise, with the target coiling temperature as the primary priority and the target final rolling temperature as the secondary priority, reverse optimization correction is performed in the opposite direction to the rolling direction. Control commands are generated based on the revised control variables for each process, and the speed actuator and cooling device are controlled based on the control commands.

2. The rolling temperature control method according to claim 1, characterized in that, The preset process points include the roughing mill exit temperature node, the intermediate billet cooling temperature node, the finishing mill inlet temperature node, the finishing mill exit temperature node, and the coiling temperature node; the measured temperatures include the roughing mill exit temperature, the finishing mill inlet temperature, the finishing mill exit temperature, and the coiling temperature; the cooling devices include an intermediate billet cooling device located between the roughing mill area and the finishing mill area, a finishing mill stand cooling device located in the finishing mill area, and an ultra-fast cooling device and / or a laminar flow cooling device located in the post-rolling cooling area.

3. The rolling temperature control method according to claim 1, characterized in that, The step of correcting the second predicted temperature based on the first predicted temperature at each preset process point to obtain the final predicted temperature includes: By employing a weighted fusion or residual correction fusion method, the first predicted temperature of each preset process point is used as the correction benchmark for the corresponding process point, and the temperature value of the corresponding node in the second predicted temperature is corrected.

4. The rolling temperature control method according to claim 1, characterized in that, The reverse optimization correction includes: Using at least one of the following as the control variables to be optimized: post-rolling cooling parameters, finishing mill speed regime, finishing mill inter-stand cooling parameters, intermediate billet cooling parameters, intermediate roller speed, and conveying rhythm, the correction amount of the control variables to be optimized is calculated in the direction opposite to the rolling direction, under the conditions of satisfying the cooling device capacity constraints, speed adjustment range constraints, and rolling stability constraints.

5. The rolling temperature control method according to claim 4, characterized in that, In the process of reverse optimization correction, the method further includes: When there exists a first combination of control variables to be optimized that simultaneously satisfies the allowable deviation ranges of the target coiling temperature and the target final rolling temperature, the first combination of control variables to be optimized is output as a reverse correction amount. When there is no first combination of control variables that simultaneously satisfies the allowable deviation ranges of the target coiling temperature and the target final rolling temperature, the output of the second combination of control variables that prioritizes satisfying the allowable deviation range of the target coiling temperature and minimizes the deviation of the target final rolling temperature is used as a reverse correction amount. The objective function of the reverse optimization is: ， In the formula, For winding temperature weighting, Weighted by the final rolling temperature. > , and These are the calculation of the winding temperature and the target winding temperature, respectively. and The values ​​are calculated final rolling temperature and target final rolling temperature, respectively. U is the correction amount of the control variables to be optimized, including correction amount of ultra-fast cooling water volume, laminar cooling water volume, finishing rolling speed, inter-stand cooling correction amount, intermediate billet cooling correction amount, and intermediate roller table speed correction amount. λ is a preset coefficient.

6. The rolling temperature control method according to claim 1 or 4, characterized in that, Using the target coiling temperature and target final rolling temperature as fixed process target constraints, it is determined whether the final rolling temperature and coiling temperature can simultaneously reach the corresponding fixed process target constraints under the current cooling capacity and current rolling speed regime; if they can be reached simultaneously, then the positive control result is executed. Otherwise, prioritizing the target coiling temperature and secondarily prioritizing the target final rolling temperature, a reverse optimization correction is performed in the opposite direction to the rolling direction, including: Using the target coiling temperature, target final rolling temperature, and target finishing mill inlet temperature as fixed process target constraints, it is determined whether the final rolling temperature, coiling temperature, and finishing mill inlet temperature can simultaneously reach the corresponding fixed process target constraints under the current cooling capacity and current rolling speed regime; if they can be reached simultaneously, then the positive control result is executed. Otherwise, with the target coiling temperature as the primary priority, the target final rolling temperature as the secondary priority, and the target finishing rolling inlet temperature as the next priority, reverse optimization correction is performed in the direction opposite to the rolling direction.

7. The rolling temperature control method according to claim 6, characterized in that, In the process of reverse optimization correction, the method further includes: When there exists a third combination of control variables to be optimized that simultaneously satisfies the allowable deviation ranges of the target coiling temperature, the target final rolling temperature, and the target finishing rolling inlet temperature, the third combination of control variables to be optimized is output as a reverse correction amount. When there is no third combination of control variables that simultaneously satisfies the allowable deviation ranges of the target coiling temperature, target final rolling temperature, and target finishing mill inlet temperature, the fourth combination of control variables that prioritizes satisfying the allowable deviation ranges of the target coiling temperature and target final rolling temperature and minimizes the deviation of the target finishing mill inlet temperature is output as a reverse correction amount. When there is no fourth combination of control variables that simultaneously satisfies the allowable deviation ranges of the target coiling temperature and the target final rolling temperature, the fifth combination of control variables that prioritizes satisfying the allowable deviation range of the target coiling temperature and minimizes the deviation of the target final rolling temperature is output as a reverse correction amount. The objective function of the reverse optimization is: ， In the formula, For winding temperature weighting, Weighted by the final rolling temperature. Weighted by the inlet temperature of the finishing mill. , and These are the calculation of the winding temperature and the target winding temperature, respectively. and These are the calculation of the final rolling temperature and the target final rolling temperature, respectively. and The values ​​are calculated for the finishing mill inlet temperature and the target finishing mill inlet temperature, respectively. U represents the correction amount of the control variables to be optimized, including the correction amount for ultra-fast cooling water volume, laminar cooling water volume, finishing mill speed, inter-stand cooling, intermediate billet cooling, and intermediate roller speed. λ is a preset coefficient.

8. The rolling temperature control method according to claim 1, characterized in that, The method further includes: When the measured temperature at the finishing mill inlet is obtained, the setting parameters of the cooling devices between the finishing mill stands and the cooling devices in the post-rolling cooling zone are corrected based on the deviation between the measured temperature at the finishing mill inlet and the final predicted temperature at the finishing mill inlet. When the final rolling temperature is obtained, the setting parameters of the cooling device in the post-rolling cooling zone and the conveyor roller speed are corrected based on the deviation between the final rolling temperature and the final rolling predicted temperature. When the actual winding temperature is obtained, the setting parameters of the final cooling device in the post-rolling cooling zone are corrected based on the deviation between the actual winding temperature and the final predicted winding temperature.

9. The rolling temperature control method according to claim 1, characterized in that, The pre-construction process of the regression chain model includes: Acquire historical full-process batch data, which includes the historical measured temperature of each batch of plate and strip steel at each temperature node, the historical target temperature of each preset process point, historical control setpoints, historical equipment status parameters, and historical steel plate operating parameters. The temperature change between the (i-1)th temperature node and the ith temperature node is taken as the output target, and the historical measured temperature, historical target temperature, historical control setpoint, historical equipment status parameters, and historical steel plate operating parameters of each preceding temperature node before the (i-1)th temperature node are taken as input features. The ith regression sub-model is trained to obtain the regression relationship between the (i-1)th temperature node and the ith temperature node. By traversing all adjacent temperature node pairs, a regression chain model covering the regression relationship between adjacent temperature nodes throughout the entire process is obtained. And / or, The pre-construction process of the RNN time series model includes: Acquire historical full-process batch data and extract the historical process parameter sequences of each batch of plate and strip steel at multiple historical moments as time-series input feature samples; The RNN network is trained by using the sequence of historical process parameters from N times prior to any given time and the historical self-learning value at any given time as input features, and the measured temperature values ​​of the corresponding batch of strip steel at each preset process point as output targets. The RNN network adopts a long short-term memory network or a gated recurrent unit network.

10. The rolling temperature control method according to claim 1, characterized in that, The pre-construction process of the instant learning model includes: Acquire historical full-process batch data, extract the historical genetic characteristic parameter vectors of each batch of plate and strip steel at each temperature node and the corresponding historical measured temperature deviation values, and construct a historical sample database. The step of obtaining self-learning values ​​using a pre-built instant learning model with the genetic feature parameters as input includes: The instant learning model employs a local neighborhood search strategy. When the genetic characteristic parameters of the current strip steel are received, the model retrieves the K most similar historical samples to the genetic characteristic parameters from the historical sample database. Based on the historical measured temperature deviation values ​​of the K historical samples, the model calculates the self-learning value corresponding to the genetic characteristic parameters using local weighted regression or local weighted averaging.

Images