A rotary lime kiln temperature control method based on neural network prediction

Abstract

The application discloses a rotary lime kiln temperature control method based on neural network prediction and belongs to the technical field of industrial process control. The method acquires historical process parameter data in the running process of a rotary lime kiln and pre-processes the process parameter data; a double-channel neural network prediction model is constructed and trained, the time sequence evolution law of system state and the influence law of control quantity on state are respectively learned through independent channels, both are explicitly superimposed to predict the running state parameter of the rotary lime kiln in the future control period; a proxy air volume is constructed based on the running parameter of a fan, and a double-cross limiting control mechanism between the proxy air volume and the coal feed quantity is established; on this basis, a prediction control optimization model introducing the double-cross limiting control is constructed, the optimal control sequence of the control quantity is solved through a model prediction control solver, and the first control quantity in the optimal control sequence is written into a controller for execution.

Description

Technical Field

[0001] This invention relates to the field of industrial process control technology, and in particular to a method for temperature control of a rotary lime kiln based on neural network prediction. Background Technology

[0002] The rotary lime kiln is a key piece of equipment in the lime production process. Its main function is to decompose limestone through high-temperature calcination to produce active lime. During the production process of the rotary lime kiln, the kiln temperature is a crucial parameter affecting the quality of the lime product and energy utilization efficiency. Therefore, stable control of the kiln temperature is of great significance for improving production efficiency and product quality.

[0003] However, due to the characteristics of rotary lime kilns, such as large hysteresis, strong coupling, and nonlinearity, their temperature control process is quite complex. Furthermore, the operation of a rotary lime kiln involves fan operating parameters, feeding parameters, and various pressure and temperature parameters, which have complex dynamic relationships, making it difficult to achieve precise temperature regulation within the kiln using traditional control methods.

[0004] In existing technologies, proportional-integral-derivative (PID) control is commonly used to regulate the operating state of rotary lime kilns. However, due to the significant time-varying and nonlinear characteristics of rotary lime kiln systems, traditional PID control methods struggle to adequately adapt to changes in system operating conditions, easily leading to large temperature fluctuations and thus affecting production stability.

[0005] In recent years, with the development of artificial intelligence technology, neural networks have been widely used in industrial process modeling and prediction. By using neural networks for data-driven modeling of industrial processes, the dynamic characteristics of complex systems can be better described. Meanwhile, Model Predictive Control (MPC), as an advanced control method, can obtain a better control strategy by predicting the future state of the system and optimizing the solution under constraints.

[0006] Meanwhile, when the variation range of the control variables in the training data is limited (i.e., operators tend to maintain stable operating conditions and rarely make significant adjustments to the control variables in actual production), the end-to-end recurrent neural network exhibits a failure mode of control signal inundation. Specifically, the gating mechanism of LSTM tends to rely primarily on historical state time-series features for extrapolation, and is insensitive to changes in the input of control variables (fan frequency, coal feed rate, etc.). This causes the system dynamics model implicitly learned by the neural network to essentially degenerate into a free-evolutionary model rather than a controlled system model. The consequence is that when the model predictive control solver attempts to change the control variable value during optimization, the neural network's predictive output hardly reflects this change, causing the MPC to fail to find an effective optimization direction, resulting in a severe decrease in temperature control accuracy. Summary of the Invention

[0007] The purpose of this invention is to provide a rotary lime kiln temperature control method based on neural network prediction, in order to solve the problem that the air-coal ratio is difficult to control stably in the existing rotary lime kiln temperature control process due to the strong nonlinearity of the system, the complex dynamic coupling, and the difficulty in directly measuring the air intake, which in turn affects the temperature stability inside the kiln, thereby improving the production efficiency and product quality of the rotary lime kiln.

[0008] To achieve the above objectives, the present invention adopts the following technical solution: a rotary lime kiln temperature control method based on neural network prediction, characterized in that the method is applied to a rotary lime kiln control system, extracting time-series features from process parameter data during the operation of the rotary lime kiln using a neural network model, and predicting the operating state parameters of the rotary lime kiln within future control cycles, wherein the operating state parameters include at least the kiln temperature parameters, and constructing a proxy air volume based on the fan operating parameters, establishing a double-cross-limiting control mechanism to limit the air-coal ratio within a preset range, and solving for the optimal control sequence of the control quantity under constraints using a model predictive control algorithm, thereby achieving stable control of the kiln temperature, the method comprising: S1. Obtain historical process parameter data during the operation of the rotary lime kiln; S2. Construct and train a neural network model based on the process parameter data; S3. Establish a predictive control optimization model for the air-coal ratio double cross-limit control mechanism between the agent air volume and the coal feed rate, and configure the corresponding model predictive control solver. S4. Obtain the process parameter data under the current operating conditions of the rotary lime kiln, and input the process parameter data into the model predictive control solver to obtain the optimal control sequence of the control quantity. S5. Write the optimal control sequence obtained from the solution into the controller of the rotary lime kiln control system.

[0009] Furthermore, the process parameter data obtained in step S1 during the operation of the rotary lime kiln includes rotary lime kiln operating parameters, fan operating parameters, and material conveying parameters; the rotary lime kiln operating parameters include the operating frequency and operating current of the rotary lime kiln drive motor, and the pressure and temperature parameters of one or more measuring points inside the rotary lime kiln; the fan operating parameters include the operating frequency and current parameters of the primary fan, secondary fan, and pulverized coal injection fan; the material conveying parameters are the pushing time interval of the raw material conveying push rod and the flow rate of the pulverized coal rotor scale.

[0010] Furthermore, the method for obtaining process parameter data during the operation of the rotary lime kiln described in step S1 includes, but is not limited to, collecting relevant process parameters through field sensors in the rotary lime kiln control system, and transmitting, recording, and storing the collected process parameter data through a programmable logic controller (PLC) or a distributed control system (DCS).

[0011] Furthermore, the neural network model described in step S2 adopts a state-control decoupled dual-path prediction architecture (also known as a dual-path structure). Its core design principle is to forcibly decouple the free evolution of the system state from the causal effect of the control input on the state, enabling changes in the control quantity to explicitly influence the predicted output through independent paths, reducing the risk of the control signal being overwhelmed by the state history. This architecture consists of three independent and complementary functional modules: the first module is the system free dynamics encoding path, i.e., the state evolution path, which uses a deep long short-term memory network (LSTM) to encode the time series of historical state parameters; the second module is the external stimulus-response encoding path, i.e., the control response path, independent of the state path, which uses small gated recurrent units (GRUs) to encode the historical control quantity sequence to extract the time-cumulative effect of the control action, and explicitly concatenates the time-series features encoded by the GRU with the instantaneous value of the control quantity at the current moment, mapping it to the control effect increment through a fully connected network; the third module is the residual superposition output module, calculated as follows: ; In the above formula, For predicted values, This is the current state value. For state evolution increments, The increment of the control effect represents the first... The predicted value of the first control cycle is determined by the first... The final prediction result is obtained by superimposing the current state observation, state self-evolution increment, and control effect increment for each control cycle.

[0012] Furthermore, the dual-cross-limit control mechanism described in S3 is implemented by constructing a proxy airflow based on the fan operating parameters. In the actual industrial scenario of a rotary lime kiln, the total intake air volume of primary and secondary air required for combustion is difficult to measure accurately directly using a flow meter. This is because the kiln tail and kiln head are under slight negative pressure conditions, the duct cross-sectional area is large and the dust content is high, and the measurement accuracy of conventional differential pressure or thermal flow meters is severely affected by dust deposition and flow field inhomogeneity, resulting in high maintenance costs over the long term. However, the combustion airflow is essentially determined by the fan operating parameters—the fan is driven by a frequency converter at a given frequency, and its output airflow has an empirical mapping relationship with the fan frequency, fan load rate, and duct back pressure that can be calibrated using historical operating data. Based on this physical fact, this invention utilizes the frequency feedback, current feedback, and duct pressure measurement points of the fan frequency converter already configured in the rotary lime kiln control system to construct a proxy airflow that can be calculated in real time, replacing direct flow measurement for air-coal ratio constraint control. For each control cycle, the proxy airflow is constructed using the following formula: ; In the above formula, A collection of fans that participate in combustion air supply. For wind turbine index; For the first Typhoon machine in The frequency feedback value of the inverter for each control cycle; This is the inverter current feedback value for the fan. Its rated current; The corresponding air duct for this fan is in the [number]th [section]. Pressure measurement values ​​for each control cycle Local atmospheric pressure; For the first The air volume weighting coefficient for typhoon fans is determined based on the proportion of the rated air volume of each fan, and meets the following requirements. ; The pressure attenuation index of the fan can be obtained by calibration based on historical operating data, reflecting the nonlinear suppression effect of increased pipeline back pressure on actual volumetric flow rate.

[0013] The physical basis for the above proxy variable construction form is that the fan output air volume is directly proportional to the frequency. Used to characterize changes in fan load rate, and to empirically correct frequency-estimated air volume, its coefficients and exponents can be obtained through calibration using historical operating data. The invention introduces a compression correction for volumetric flow rate based on pipeline back pressure. The weighted sum of the three products yields a comprehensive proxy airflow index, which can be used as the airflow representation quantity in dual-cross-limit control constraints. Compared to existing methods that directly use a single fan frequency as an approximation of airflow, the proxy airflow constructed in this invention offers substantial improvements in three aspects: multi-fan joint representation, load rate correction, and back pressure compensation. It tracks the changing trend of the actual comprehensive intake airflow with engineering-acceptable accuracy without requiring additional on-site flow measurement hardware.

[0014] Furthermore, the dual-cross-limit control mechanism described in step S3 is used to limit the air-coal ratio of the rotary lime kiln to within a preset range, thereby reducing the risk of oxygen-deficient combustion. To achieve the dual-cross-limit control constraint, the air-side control quantity and the coal feed control quantity satisfy the following constraint relationship: ; ; In the above formula, This is the final setpoint for the pulverized coal rotor scale. The coal feed rate requirement given by the solver. and This is a proportionality coefficient used to establish the lower and upper limits of the proportional constraint relationship between the coal feed rate and the proxy air volume. For the currently acquired agent air volume, and These are the minimum and maximum allowable coal feed rates, respectively. This formula indicates that the final coal feed rate must not only follow the coal feed demand value given by the solver, but must also be limited to the allowable range determined by the agent air volume and the upper and lower limits of the equipment itself. The final given value for the agent's airflow. The agent's air volume requirement value is given by the solver. and This is a proportionality coefficient used to establish the lower and upper limits of the proportional constraint relationship between the agent air volume and the coal feed volume. This represents the current coal feed rate. and These are the minimum and maximum allowable proxy air volumes, respectively. This formula indicates that the final proxy air volume value must not only follow the proxy air volume requirement value given by the solver, but must also be limited to the allowable range jointly determined by the coal feed rate and the upper and lower limits of the equipment itself.

[0015] Furthermore, step S3, constructing the predictive control optimization model, includes: S31: Construct an objective function based on the control target of the kiln temperature to minimize the deviation between the predicted value of the kiln temperature and the corresponding desired temperature. The formula is as follows: ; In the above formula, To optimize the objective function for model predictive control, To predict the length of the time domain, The number of temperature parameters, For the first The weighting coefficients corresponding to each temperature parameter For the first During the first control cycle, the first Predicted values ​​for each temperature parameter For the first During the first control cycle, the first The expected value of each temperature parameter is obtained by minimizing the objective function to obtain the optimal control sequence of the control quantity in the model predictive control solution process. S32: Determine the control quantities that can be executed by the rotary lime kiln control system as model predictive control optimization variables. The control quantities include the setpoint values ​​of the primary fan frequency, secondary fan frequency, pulverized coal fan frequency, pulverized coal rotor scale flow rate, rotary lime kiln drive frequency, and / or raw material conveying push rod push time interval. Input the kiln temperature parameters, kiln pressure parameters, motor current parameters, fan current parameters, and feedback values ​​of each actuator as state variables, feedback variables, or constraint variables into the data-driven time-series prediction model. S33: Simultaneously, constraints are set on the control quantity, the control quantity is normalized, and the change range of the control quantity within adjacent control cycles is limited to not exceeding a preset threshold. The corresponding formula is shown below: ; ; In the above formula, Let be the normalized value for the i-th control cycle. For the first The control quantity in the first The actual value for each control cycle. For the first The minimum value of each control variable For the first The maximum value of each control quantity; Indicates the first in adjacent control cycles The maximum value of the normalized change of each control variable.

[0016] Further, the process parameter data input to the model predictive control solver in step S4 to obtain the optimal control sequence for future control cycles includes: acquiring process parameter data for the current control cycle and several previous historical control cycles, and constructing time series input data; constructing an initial candidate control sequence based on the optimal control sequence translation result of the previous control cycle, the current control quantity, or a preset benchmark control strategy; inputting the time series input data and the initial candidate control sequence into a model predictive control solver constructed based on a neural network model; and during the solution process, performing rolling predictions on the rotary lime kiln operating state parameters for future control cycles, and iteratively optimizing the candidate control sequence according to the objective function and constraints to obtain the optimal control sequence for the control quantity.

[0017] Further, in step S2, before constructing and training the neural network model, the process parameter data is preprocessed. The preprocessing includes data cleaning, outlier handling, and data normalization. Specifically, data cleaning performs integrity checks and consistency verification on the collected historical process parameter data, removing missing data, duplicate records, and obviously erroneous data to improve data reliability. Outlier handling identifies process parameter data that deviates from the normal operating range and removes or corrects these outliers to reduce interference with the neural network model training process. Data normalization unifies the scale of process parameter data with different dimensions and value ranges, mapping each process parameter data to a unified numerical range, thereby reducing the impact of differences in parameter dimensions on the model training process and improving the numerical stability and prediction accuracy of the neural network model training process.

[0018] Furthermore, in each control cycle, the first control quantity in the optimal control sequence is executed by the programmable logic controller (PLC) or distributed control system (DCS), and in the next control cycle, the process parameter data under the operating conditions of the rotary lime kiln is re-acquired, and steps S4 and S5 are re-executed to update the control quantity. Attached Figure Description

[0019] Figure 1 This is a flowchart of a rotary lime kiln temperature control method based on neural network prediction.

[0020] Figure 2 This is a diagram of the dual-pathway neural network structure in a rotary lime kiln temperature control method based on neural network prediction.

[0021] Figure 3 The diagram shows the neural network prediction and MPC closed-loop control block diagram in the temperature control method of rotary lime kiln based on neural network prediction.

[0022] Figure 4This is a schematic diagram of the double-cross amplitude limiting control relationship in the rotary lime kiln temperature control method based on neural network prediction. Detailed Implementation

[0023] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. It should be noted that the following embodiments are only used to illustrate the technical solutions of the present invention and do not constitute a limitation on the scope of protection of the present invention. However, the implementation of the present invention is not limited thereto. Therefore, based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0024] Please refer to Figures 1 to 4 The present invention provides a technical solution: This embodiment provides a rotary lime kiln temperature control method based on neural network prediction. The method is applied to a rotary lime kiln control system, extracting time-series features from process parameter data during rotary lime kiln operation using a neural network model, and predicting the rotary lime kiln's operating state parameters within future control cycles. These operating state parameters include at least the kiln temperature parameter. A double-cross-limit control mechanism is established based on the fan operating parameters to construct a proxy airflow, limiting the air-coal ratio to a preset range. Under these constraints, a model predictive control algorithm is used to solve for the optimal control sequence of the control variables, thereby achieving stable control of the kiln temperature. The method includes: S1. Obtain historical process parameter data during the operation of the rotary lime kiln; S2. Construct and train a neural network model based on the process parameter data; S3. Establish a predictive control optimization model for the air-coal ratio double cross-limit control mechanism between the agent air volume and the coal feed rate, and configure the corresponding model predictive control solver. S4. Obtain the process parameter data under the current operating conditions of the rotary lime kiln, and input the process parameter data into the model predictive control solver to obtain the optimal control sequence of the control quantity. S5. Write the optimal control sequence into the controller of the rotary lime kiln control system.

[0025] Furthermore, the method for obtaining process parameter data during the operation of the rotary lime kiln described in step S1 includes, but is not limited to, collecting relevant process parameters through field sensors in the rotary lime kiln control system, and transmitting, recording, and storing the collected process parameter data through a programmable logic controller (PLC) or a distributed control system (DCS).

[0026] The neural network model described in step S2 adopts a state-control decoupled dual-path prediction architecture, i.e., a dual-path structure. Its core design principle is to forcibly decouple the free evolution law of the system state from the causal effect law of the control input on the state, so that the change of the control quantity can explicitly affect the prediction output through an independent path, reducing the risk of the control signal being submerged by the state history. The architecture consists of three independent yet complementary functional modules: the first module is the system free dynamics encoding path, i.e., the state evolution path, which uses a deep long short-term memory network to encode the time series of historical state parameters. Based on the historical state series, it extracts the system's thermal inertia, hysteresis, and autoregressive evolution trend, outputting the state residual increment. The second module is the external stimulus-response encoding path, i.e., the control response path, which is independent of the state path. It uses small gated recurrent units (GRUs) to encode the historical control quantity sequence to extract the time cumulative effect of the control action. The temporal features encoded by the GRUs are explicitly concatenated with the instantaneous value of the control quantity at the current moment and mapped to the control effect increment through a fully connected network. The third module is the residual superposition output module, which explicitly superimposes the current state observation value, the state self-evolution increment, and the control effect increment to form the final prediction result, i.e.: Predicted value = Current state value + State evolution increment + Control effect increment. This structure enhances the explicit influence of the control quantity on the predicted output by setting an independent control response path in the network topology, reducing the risk of the control signal being overwhelmed by historical state features. The effectiveness of the control response pathway can be constrained and verified through methods such as training with samples containing control disturbances, weighting samples with control changes, and verifying the sensitivity of the predicted output to the control.

[0027] Training employed the Adam optimizer with an initial learning rate of 0.0001, a batch size of 64, and the mean squared error (MSE) loss function. To prevent overfitting, an early stopping strategy was implemented: training was terminated and the model parameters that minimized the validation set loss were rolled back to the minimum value within 20 consecutive epochs. The input sequence length was selected as 10 control cycles based on the thermal response time constant of the rotary lime kiln; that is, the model uses historical data from the past 10 control cycles to predict the state for the next control cycle. State variables and control inputs were normalized before being input into the network, mapping them to the [0,1] interval.

[0028] Furthermore, the dual-cross-limit control mechanism described in S3 is implemented by constructing a proxy airflow based on the fan operating parameters. In the actual industrial scenario of a rotary lime kiln, the total intake air volume of primary and secondary air required for combustion is difficult to measure accurately directly using a flow meter. This is because the kiln tail and kiln head are under slight negative pressure conditions, the duct cross-sectional area is large and the dust content is high, and the measurement accuracy of conventional differential pressure or thermal flow meters is severely affected by dust deposition and flow field inhomogeneity, resulting in high maintenance costs over the long term. However, the combustion airflow is essentially determined by the fan operating parameters—the fan is driven by a frequency converter at a given frequency, and its output airflow has an empirical mapping relationship with the fan frequency, fan load rate, and duct back pressure that can be calibrated using historical operating data. Based on this physical fact, this invention utilizes the frequency feedback, current feedback, and duct pressure measurement points of the fan frequency converter already configured in the rotary lime kiln control system to construct a proxy airflow that can be calculated in real time, replacing direct flow measurement for air-coal ratio constraint control. For each control cycle, the proxy airflow is constructed using the following formula: ; In the above formula, A collection of fans that participate in combustion air supply. For wind turbine index; For the first Typhoon machine in The frequency feedback value of the inverter for each control cycle; This is the inverter current feedback value for the fan. Its rated current; The corresponding air duct for this fan is in the [number]th [section]. Pressure measurement values ​​for each control cycle Local atmospheric pressure; For the first The air volume weighting coefficient for typhoon fans is determined based on the proportion of the rated air volume of each fan, and meets the following requirements. ; The pressure attenuation index of the fan can be obtained by calibration based on historical operating data, reflecting the nonlinear suppression effect of increased pipeline back pressure on actual volumetric flow rate.

[0029] The physical basis for the above proxy variable construction form is that the fan output air volume is directly proportional to the frequency. Used to characterize changes in fan load rate, and to empirically correct frequency-estimated air volume, its coefficients and exponents can be obtained through calibration using historical operating data. The invention introduces a compression correction for volumetric flow rate based on pipeline back pressure. The weighted sum of the three products yields a comprehensive proxy airflow index, which can be used as the airflow representation quantity in dual-cross-limit control constraints. Compared to existing methods that directly use a single fan frequency as an approximation of airflow, the proxy airflow constructed in this invention offers substantial improvements in three aspects: multi-fan joint representation, load rate correction, and back pressure compensation. It tracks the changing trend of the actual comprehensive intake airflow with engineering-acceptable accuracy without requiring additional on-site flow measurement hardware.

[0030] Furthermore, the dual-cross-limiting control mechanism described in step S3 is used to limit the air-coal ratio in the rotary lime kiln to within a preset range, reducing the risk of oxygen-deficient combustion. The control relationship is as follows: Figure 4 As shown. The limiting control module satisfies the following constraints on the wind-side control quantity and the coal feeding control quantity: ; ; In the above formula, This is the final setpoint for the pulverized coal rotor scale. The coal feed rate requirement given by the solver. and This is a proportionality coefficient used to establish the lower and upper limits of the proportional constraint relationship between the coal feed rate and the proxy air volume. For the currently acquired agent air volume, and These are the minimum and maximum allowable coal feed rates, respectively. This formula indicates that the final coal feed rate must not only follow the coal feed demand value given by the solver, but must also be limited to the allowable range determined by the agent air volume and the upper and lower limits of the equipment itself. The final given value for the agent's airflow. The agent's air volume requirement value is given by the solver. and This is a proportionality coefficient used to establish the lower and upper limits of the proportional constraint relationship between the agent air volume and the coal feed volume. This represents the current coal feed rate. and These are the minimum and maximum allowable proxy air volumes, respectively. This formula indicates that the final proxy air volume value must not only follow the proxy air volume requirement value given by the solver, but must also be limited to the allowable range jointly determined by the coal feed rate and the upper and lower limits of the equipment itself.

[0031] Furthermore, step S3, constructing the predictive control optimization model, includes: S31: Construct an objective function based on the control target of the kiln temperature to minimize the deviation between the predicted value of the kiln temperature and the corresponding desired temperature. The formula is as follows: ; In the above formula, To optimize the objective function for model predictive control, To predict the length of the time domain, The number of temperature parameters, For the first The weighting coefficients corresponding to each temperature parameter For the first During the first control cycle, the first Predicted values ​​for each temperature parameter For the first During the first control cycle, the first The expected value of each temperature parameter is obtained by minimizing the objective function to obtain the optimal control sequence of the control quantity in the model predictive control solution process. S32: Determine the control quantities that can be executed by the rotary lime kiln control system as model predictive control optimization variables. The control quantities include the setpoint values ​​of the primary fan frequency, secondary fan frequency, pulverized coal fan frequency, pulverized coal rotor scale flow rate, rotary lime kiln drive frequency, and / or raw material conveying push rod push time interval. Input the kiln temperature parameters, kiln pressure parameters, motor current parameters, fan current parameters, and feedback values ​​of each actuator as state variables, feedback variables, or constraint variables into the data-driven time-series prediction model. S33: Simultaneously, constraints are set on the control quantity, the control quantity is normalized, and the change range of the control quantity within adjacent control cycles is limited to not exceeding a preset threshold. The corresponding formula is shown below: ; ; In the above formula, Let be the normalized value for the i-th control cycle. For the first The control quantity in the first The actual value for each control cycle. For the first The minimum value of each control variable For the first The maximum value of each control quantity; Indicates the first in adjacent control cycles The maximum value of the normalized change of each control variable.

[0032] Step S4, which involves inputting process parameter data into a model predictive control solver to obtain the optimal control sequence for future control cycles, includes: acquiring process parameter data for the current control cycle and several previous historical control cycles, and constructing time series input data; inputting the time series input data into a model predictive control solver built on a neural network model; during the solving process of the model predictive control solver, constructing an initial control sequence based on the optimal control sequence translation result of the previous control cycle, the current control quantity, or a preset baseline control strategy; inputting the initial control sequence and historical process parameters into the predictive model for rolling prediction, and iteratively optimizing based on the prediction results to obtain the optimal control sequence.

[0033] Further, in step S2, before constructing and training the neural network model, the process parameter data is preprocessed. The preprocessing includes data cleaning, outlier handling, and data normalization. Specifically, data cleaning performs integrity checks and consistency verification on the collected historical process parameter data, removing missing data, duplicate records, and obviously erroneous data to improve data reliability. Outlier handling identifies process parameter data that deviates from the normal operating range and removes or corrects these outliers to reduce interference with the neural network model training process. Data normalization unifies the scale of process parameter data with different dimensions and value ranges, mapping each process parameter data to a unified numerical range, thereby reducing the impact of differences in parameter dimensions on the model training process and improving the numerical stability and prediction accuracy of the neural network model training process.

[0034] Furthermore, in each control cycle, the first control quantity in the optimal control sequence is executed by the programmable logic controller (PLC) or distributed control system (DCS), and in the next control cycle, the process parameter data under the operating conditions of the rotary lime kiln is re-acquired, and steps S4 and S5 are re-executed to update the control quantity.

[0035] The present invention provides a method for controlling the temperature of a rotary lime kiln based on neural network prediction, which has the following substantial advantages over the prior art: At the neural network modeling level, this study reveals and solves the failure mode of control signal overwhelmance in industrial time series prediction. By forcibly decoupling the state evolution law and the control response law at the network topology level, and independently encoding them using heterogeneous recurrent units (deep LSTM and small GRU), the predicted output is constructed by explicitly superimposing three terms: state baseline, self-evolution increment, and control effect increment. This strengthens the explicit influence channel of control quantity changes on the predicted output from the model structure, reducing the risk of control signals being overwhelmed by historical state characteristics. This decoupling design is a structural improvement on existing end-to-end neural network prediction methods, rather than a simple adjustment of network depth or width parameters, and is therefore non-obvious. Secondly, regarding the air-coal ratio control, addressing the engineering challenge of directly and accurately measuring the air intake of rotary lime kilns, a method is proposed that uses the combined operating parameters (frequency, current, and pipeline pressure) of multiple fans to construct a proxy air volume. A dual cross-limiting mechanism is established to constrain the proxy variable and the coal feed rate. This solution achieves a safe range constraint on the air-coal ratio using software algorithms without requiring additional on-site flow measurement hardware, representing a substantial improvement over existing cross-limiting control methods in terms of operational adaptability and measurement robustness. Third, at the system integration level, the aforementioned decoupled predictive model and dual-cross-limiting mechanism are uniformly embedded into the optimization framework of Model Predictive Control (MPC). This allows the optimizer to simultaneously consider the temperature control objective and the air-coal ratio safety constraint when searching for the optimal control quantity, avoiding coupling mismatch and constraint violation problems that may occur when neural network prediction and constraint control are designed independently. The synergistic operation of these three elements enables this invention to achieve temperature control stability and air-coal ratio safety coordination, which are difficult to simultaneously achieve with existing PID control or conventional end-to-end neural network plus independent constraint control methods, in rotary lime kiln industrial applications characterized by strong nonlinearity, large hysteresis, and complex coupling.

[0036] As a supplementary explanation to the above embodiments, the process parameters involved in the agent air volume in this invention and their corresponding OPC Tag identifiers in the rotary lime kiln control system can be collected from the corresponding variable addresses in the PLC / DCS. The specific variable names can be replaced according to the configuration of different control systems. The identifier is only a variable naming convention in a specific embodiment. In different manufacturers' PLC / DCS systems, it can be replaced with equivalent hardware addresses or variable names, and its physical meaning and function in this invention remain unchanged.

[0037] It should be understood that the above embodiments are only used to illustrate the technical solutions of the present invention, and are not intended to limit the scope of protection of the present invention. Those skilled in the art can make equivalent substitutions or modifications to the neural network structure, model prediction control parameter setting methods, prediction time domain length, constraint condition forms, and process parameter selection methods without departing from the spirit and substance of the present invention, and all such equivalent substitutions or modifications should fall within the scope of protection of the present invention.

Claims

1. A method for temperature control of a rotary lime kiln based on neural network prediction, characterized in that, The method is applied to the control system of a rotary lime kiln. It uses a data-driven time-series prediction model to extract time-series features from process parameter data during the operation of the rotary lime kiln and predicts the operating state parameters of the kiln within future control cycles. These operating state parameters include at least the kiln temperature parameter. A proxy airflow is constructed based on the fan operating parameters, and a double-cross-limiting mechanism is established between the proxy airflow and the coal feed rate to restrict the air-coal ratio within a preset range. Under the constraint of the double-cross-limiting mechanism, the optimal control sequence of the control variables is solved using a model predictive control algorithm, thereby achieving stable control of the kiln temperature. The method includes: S1. Obtain historical process parameter data during the operation of the rotary lime kiln; S2. Construct and train a data-driven time-series prediction model based on the process parameter data; S3. Establish a predictive control optimization model for the air-coal ratio dual-cross-limit control mechanism between the agent air volume and the coal feed rate, and configure the corresponding model predictive control solver. The steps for constructing the predictive control optimization model include: S31: Construct an objective function based on the control target of the kiln temperature parameters to minimize the deviation between the predicted values ​​and the corresponding expected values ​​of the kiln temperature parameters. The formula is as follows: ； In the above formula, To optimize the objective function for model predictive control, To predict the length of the time domain, The number of temperature parameters, For the first The weighting coefficients corresponding to each temperature parameter For the first During the first control cycle, the first Predicted values ​​for each temperature parameter For the first During the first control cycle, the first The expected value of each temperature parameter is obtained by minimizing the objective function to obtain the optimal control sequence of the control quantity in the model predictive control solution process. S32: Determine the control quantities that can be executed by the rotary lime kiln control system as model predictive control optimization variables. The control quantities include the setpoint values ​​of the primary fan frequency, secondary fan frequency, pulverized coal fan frequency, pulverized coal rotor scale flow rate, rotary lime kiln drive frequency, and / or raw material conveying push rod push time interval. Input the kiln temperature parameters, kiln pressure parameters, motor current parameters, fan current parameters, and feedback values ​​of each actuator as state variables, feedback variables, or constraint variables into the data-driven time-series prediction model. S33: Simultaneously, constraints are set on the control quantity, the control quantity is normalized, and the change range of the control quantity within adjacent control cycles is limited to not exceeding a preset threshold. The corresponding formula is shown below: ； ； In the above formula, Let be the normalized value for the i-th control cycle. For the first The control quantity in the first The actual value for each control cycle. For the first The minimum value of each control variable For the first The maximum value of each control quantity; Indicates the first in adjacent control cycles The maximum value of the normalized change of each control quantity; S4. Obtain the process parameter data under the current operating conditions of the rotary lime kiln, and input the process parameter data into the model predictive control solver to obtain the optimal control sequence of the control quantity. S5. Write the first control quantity in the optimal control sequence into the controller of the rotary lime kiln control system for execution, and use the remaining control sequence as a candidate initial sequence or a re-optimized reference sequence for the next control cycle.

2. The method for controlling the temperature of a rotary lime kiln based on neural network prediction according to claim 1, characterized in that: The process parameter data obtained in step S1 during the operation of the rotary lime kiln include the rotary lime kiln operating parameters, the blower operating parameters, and the material conveying parameters. The rotary lime kiln operating parameters include the operating frequency and current of the rotary lime kiln drive motor, and the pressure and temperature parameters of one or more measuring points inside the rotary lime kiln. The operating parameters of the blowers mentioned above include the operating frequency and current parameters of the primary blower, secondary blower, and pulverized coal blower; The material conveying parameters mentioned above are the pushing time interval of the raw material conveying push rod and the flow rate of the pulverized coal rotor scale.

3. The method for controlling the temperature of a rotary lime kiln based on neural network prediction according to claim 1, characterized in that: The neural network model described in step S2 adopts a state-control decoupled time-series prediction architecture, namely a dual-path structure: the first path is the system free dynamics encoding path, which learns the evolution law of the system state itself based on the time series data of historical process parameters; the second path is the external stimulus-response encoding path, which receives the historical control quantity sequence and the candidate control quantity sequence generated in the prediction time domain during the model predictive control solution process, and learns the time cumulative influence law of the control quantity on the system state; the outputs of the two paths are superimposed to form the final prediction result. This structure separates the system free evolution and control action into model models, ensuring that the changes in control quantity can be effectively reflected in the prediction output, and provides an explicit control response modeling foundation for the subsequent optimization solution of model predictive control.

4. The method for controlling the temperature of a rotary lime kiln based on neural network prediction according to claim 1, characterized in that: The dual-cross-limit control mechanism described in S3 is implemented by constructing a proxy air volume based on the fan operating parameters. In the actual industrial scenario of a rotary lime kiln, the total intake air volume of primary and secondary air required for combustion is difficult to measure accurately directly using a flow meter. Utilizing the frequency feedback, current feedback, and duct pressure measurement points already configured in the rotary lime kiln control system, a proxy air volume that can be calculated in real time is constructed to replace direct flow measurement for air-coal ratio constraint control. For each control cycle, the proxy air volume is constructed using the following formula: ； In the above formula, A collection of fans that participate in combustion air supply. For wind turbine index; For the first Typhoon machine in The frequency feedback value of the frequency converter for each control cycle; This is the inverter current feedback value for the fan. Its rated current; The corresponding air duct for this fan is in the [number]th [section]. Pressure measurement values ​​for each control cycle Local atmospheric pressure; For the first The air volume weighting coefficient for typhoon fans is determined based on the proportion of the rated air volume of each fan, and meets the following requirements. ; The pressure attenuation index of the fan can be obtained by calibration based on historical operating data, reflecting the nonlinear suppression effect of increased pipeline back pressure on actual volumetric flow rate.

5. The method for controlling the temperature of a rotary lime kiln based on neural network prediction according to claim 1, characterized in that: The predictive control optimization model of the air-coal ratio dual-cross-limit control mechanism described in step S3 is used to limit the air-coal ratio of the rotary lime kiln to a preset range, thereby reducing the risk of oxygen-deficient combustion. To achieve the air-coal ratio dual-cross-limit control, the following constraints must be satisfied on the air-side control quantity and the coal feed control quantity: ； ； In the above formula, This is the final setpoint for the pulverized coal rotor scale. The coal feed rate requirement given by the solver. and This is a proportionality coefficient used to establish the lower and upper limits of the proportional constraint relationship between the coal feed rate and the proxy air volume. For the currently acquired agent air volume, and These are the minimum and maximum allowable coal feed rates, respectively. This formula indicates that the final coal feed rate must not only follow the coal feed demand value given by the solver, but must also be limited to the allowable range determined by the agent air volume and the upper and lower limits of the equipment itself. The final given value for the agent's airflow. The agent's air volume requirement value is given by the solver. and This is a proportionality coefficient used to establish the lower and upper limits of the proportional constraint relationship between the agent air volume and the coal feed volume. This represents the current coal feed rate. and These are the minimum and maximum allowable proxy air volumes, respectively. This formula indicates that the final proxy air volume value must not only follow the proxy air volume requirement value given by the solver, but must also be limited to the allowable range jointly determined by the coal feed rate and the upper and lower limits of the equipment itself.

6. The method for controlling the temperature of a rotary lime kiln based on neural network prediction according to claim 1, characterized in that: In step S4, the process of inputting process parameter data into the model predictive control solver to obtain the optimal control sequence for future control cycles includes: acquiring process parameter data for the current control cycle and several previous historical control cycles, and constructing time series input data; constructing an initial candidate control sequence based on the translation result of the optimal control sequence of the previous control cycle, the current control quantity, or a preset benchmark control strategy; inputting the time series input data and the initial candidate control sequence into the model predictive control solver built on a neural network model; during the solving process of the model predictive control solver, rolling predictions are made for the kiln operating state parameters for future control cycles, and the initial candidate control sequence is iteratively optimized according to the objective function and constraints to obtain the optimal control sequence for the control variables; executing the first control quantity in the optimal control sequence in each control cycle, and re-acquiring the process parameter data under the operating conditions of the rotary lime kiln in the next control cycle, and re-executing steps S4 and S5 to update the control variables.

7. The method for controlling the temperature of a rotary lime kiln based on neural network prediction according to claim 1, characterized in that: In step S2, before constructing and training the data-driven time series prediction model, the process parameter data is preprocessed. The preprocessing includes data cleaning, outlier handling, and data normalization.

Images