Loose moisture regaining machine outlet material temperature prediction control method and system

Abstract

The application discloses a kind of loose rehumidifier outlet material temperature prediction control method and system.The method includes the steps of establishing thermodynamic equation model, data collection and preprocessing, model parameter optimization and verification, real-time temperature prediction and control, etc., the correlation analysis and time lag evaluation of system input variable are carried out in the application, the prediction model is split into quantifiable linear part and difficult to directly measure nonlinear residual term, the interpretability of physical model and the nonlinear fitting ability of machine learning are fully utilized, so as to realize high-precision, low-complexity temperature prediction control.

Description

Technical Field

[0001] This invention relates to the field of tobacco processing equipment technology, specifically to a method and system for predicting and controlling the temperature of the material exiting a loosening and rehydration machine. Background Technology

[0002] In the tobacco processing technology, the rehumidifier is used to adjust the moisture content and temperature of the tobacco shreds to meet the requirements of subsequent processes. Currently, the temperature control of the material exiting the rehumidifier mainly adopts the traditional PID control method. However, PID control has significant shortcomings in practical applications due to the following reasons:

[0003] (1) System time delay and overshoot: The loose rehumidifier has a large thermal inertia, which leads to a significant time delay in the system. The PID controller is difficult to respond in time and is prone to overshoot and lag, which affects the accuracy and stability of temperature control.

[0004] (2) Complex environmental impact: The temperature of the outlet material is affected by a variety of environmental factors that cannot be directly measured, such as heat radiation loss and material heat absorption differences. Traditional PID control is difficult to accurately model and compensate for these nonlinear factors.

[0005] (3) Limitations of pure data-driven methods: Some studies have attempted to use machine learning methods (such as neural networks) for temperature prediction, but due to a lack of understanding of the physical mechanisms, the interpretability and generalization ability of the models are insufficient, and model training requires a large amount of data, resulting in high computational complexity. Summary of the Invention

[0006] To overcome the problems in existing technologies, this invention provides a method and system for predictive control of material outlet temperature in a loose rehumidifier that combines thermodynamic equations with machine learning. This method, through correlation analysis and time delay evaluation of system input variables, decomposes the prediction model into a quantifiable linear component and a nonlinear residual term that is difficult to measure directly. It fully utilizes the interpretability of the physical model and the nonlinear fitting capability of machine learning, thereby achieving high-precision, low-computational-complexity temperature prediction control.

[0007] In this invention, correlation analysis is a key step in predictive control. Because the loose rehumidifier system has numerous input features, and the influence of each feature on the outlet material temperature is complex, in-depth correlation analysis of each input variable is necessary to ensure that the relationship between the input features and the system output (outlet material temperature) is fully revealed. Specifically, Spearman rank correlation analysis is used to quantitatively evaluate the correlation between the input variables and the outlet temperature. This method can capture nonlinear correlations and identify which input features have the greatest impact on temperature control.

[0008] Correlation analysis not only effectively eliminates input features with little impact on temperature changes, simplifying the model and reducing computational complexity, but also identifies time-delay effects between different input features. Time-delay effects refer to the fact that the influence of certain input features on system temperature is not instantaneous but has a certain delay. Through time-delay evaluation, input data can be precisely aligned, allowing the model to capture the critical moments that best reflect system dynamics. This time-delay processing significantly improves the accuracy of temperature prediction, thereby ensuring the timeliness and reliability of control strategies.

[0009] Therefore, correlation analysis plays a crucial role in control not only by optimizing characteristic inputs but also by laying the foundation for improving the overall system response and prediction accuracy. It provides a more streamlined and efficient set of input variables for subsequent temperature control, resulting in more precise and stable temperature regulation.

[0010] In real-world production environments, many input characteristics cannot be consistently and accurately obtained due to technological limitations, measurement conditions, or cost constraints. For example, factors such as material heat radiation loss, energy loss in the duct system, and differences in heat absorption by the cylinder wall are difficult to measure directly and are dynamically changing. Furthermore, the production environment of a loosening and rehumidifying machine is complex, influenced by multiple factors such as ambient temperature, humidity, and equipment aging, making it difficult to accurately predict and control the system using only linear physical models.

[0011] To address these complex, unquantifiable, or difficult-to-obtain factors, this invention introduces the modeling of the residual term A. Residual A represents the system energy loss and impact that physical models cannot accurately describe. In residual modeling, we use the XGBoost regression model, which possesses powerful nonlinear fitting capabilities and can effectively capture the potentially complex relationships in the production process. XGBoost's parallel processing and optimization algorithms enable it not only to accurately model the nonlinear factors in the residual term but also to maintain high computational efficiency and rapidly respond to system changes.

[0012] In actual production processes, uncontrollable factors such as equipment status and environmental fluctuations frequently occur. The existence of residual term A can compensate for the shortcomings of the physical model and ensure the stability of the system in complex production environments. Residual modeling incorporates these difficult-to-quantify external factors into the predictive control process through machine learning, enabling the model to maintain high prediction accuracy when dealing with changing actual production conditions. In addition, the introduction of residual A reduces the system's dependence on massive amounts of data, enhances the model's robustness, and can provide relatively accurate predictions even when some data is incomplete or inaccurate.

[0013] In summary, modeling the residual term A not only effectively supplements existing physical models but is also a key measure to address the uncertainties and complexities of real-world production environments. By accurately modeling the residual A using machine learning algorithms, this invention significantly improves the system's predictive and control capabilities in complex environments, making temperature regulation more flexible and reliable.

[0014] The technical solution adopted in this invention is as follows:

[0015] A method for predicting and controlling the temperature of the material exiting a loose rehumidifier includes the following steps:

[0016] S1 establishes a thermodynamic equation model

[0017] Based on the physical characteristics of the loose rehumidifier, a thermodynamic equation describing the outlet material temperature is established as the linear part of the model. This equation captures the main energy exchange processes in the system, including the following quantifiable influencing factors:

[0018] (1) Feed water flow rate (Q_water): used to adjust the moisture content of the material and affect the thermal balance of the system.

[0019] (2) Steam flow rate (Q_steam): Heat is transferred through steam to increase the temperature of the material.

[0020] (3) Hot air volume (Q_hot): provides heat energy to the system and affects the heating effect of materials.

[0021] Since the circulating air circulates within the system, its impact on the total system energy is small and can be ignored. Energy losses that cannot be directly measured, such as energy carried away by the exhaust air, thermal radiation loss from the cylinder wall, and differences in material heat absorption, are summarized as residual term A.

[0022] Its thermodynamic differential equation is expressed as follows:

[0023]

[0024] Parameter description:

[0025] T_sys: System temperature, i.e., the temperature of the material after absorbing heat.

[0026] t: time variable

[0027] T_hot: Hot air temperature

[0028] T_steam: Saturated steam temperature, which can be calculated from the atomized steam pressure.

[0029] T_water: Feed water temperature

[0030] C_air: Specific heat capacity of air

[0031] C_steam: Specific heat capacity of saturated steam

[0032] C_water: Specific heat capacity of water

[0033] Q_hot: Hot air volume

[0034] Q_steam: Steam flow

[0035] R: Steam moisture gain coefficient

[0036] L_steam: Latent heat of vaporization coefficient

[0037] C: Unknown term, representing the total heat capacity of the system.

[0038] A: Residual term (compensation term) is used to represent the unquantifiable heat effects in the system.

[0039] The differential equation is structured as follows: the left side of the equation is the total heat capacity of the system multiplied by the system rate of change, in kJ / s; the right side of the equation is the rate of change of hot air energy, the rate of change of sensible heat energy of steam and water in steam, the rate of change of latent heat energy of steam, the rate of change of feed water flow energy, and the unmeasurable residual A, in kJ / s.

[0040] By using thermodynamic equations, the main physical mechanisms of a system can be effectively described, providing an accurate and interpretable description of the system's behavior. Experimental results show that the greatest advantage of this physical model lies in reducing reliance on large amounts of data, thus improving the model's reliability and demonstrating significant advantages over purely data-driven methods.

[0041] Modeling of S2 residual term A

[0042] For nonlinear effects that cannot be directly quantified, they are summarized as residual terms A and modeled using machine learning methods. Specifically, the XGBoost regression model is chosen to predict A.

[0043] XGBoost excels at capturing complex nonlinear relationships, effectively handling the nonlinear effects in residual terms, and boasts high computational efficiency and speed. This algorithm employs parallel processing and an optimized loss function, improving both training speed and prediction accuracy. Furthermore, XGBoost's built-in regularization mechanism provides strong resistance to overfitting, enhancing the model's generalization performance. These characteristics make XGBoost an ideal choice for residual term modeling, balancing high predictive performance with computational efficiency.

[0044] S3 data collection and preprocessing includes:

[0045] S3.1 Data Collection

[0046] Collect operational data from the loose rehumidifier, including input characteristics and output targets;

[0047] The input features include: inlet material moisture content, scale instantaneous flow rate, feed water flow rate, inlet hot air temperature and volume, steam flow rate, atomized steam pressure, exhaust damper opening, circulation damper opening, drum rotation speed, etc.

[0048] The output target includes: the actual value of the outlet material temperature.

[0049] S3.2 Data preprocessing includes:

[0050] (1) Moving average filtering: smooths the time series data, reduces noise interference, and ensures the accuracy and reliability of the model input.

[0051] (2) Spearman rank correlation analysis: Determine the correlation between each input feature and the outlet temperature, calculate the optimal time delay, and align the data. This helps to extract effective features and enhance the predictive performance of the model.

[0052] S4 model parameter optimization and validation, including:

[0053] S4.1 Determine the total heat capacity C of the system

[0054] The total heat capacity C of the system is a key parameter in the thermodynamic equations, representing the system's ability to store heat. By constructing an optimization problem with the difference between the predicted and actual temperatures as the loss function, and using the BFGS algorithm to minimize the loss function, the optimal value of C is found, ensuring the reliability of the thermodynamic model.

[0055] S4.2 Calculate and model residual A

[0056] After determining C, the actual residual A_calculated is calculated based on the known data. Then, the XGBoost model is used, with the input features as independent variables, to model A. XGBoost's high predictive performance and computational efficiency enable it to effectively capture the nonlinear effects in the residuals, improving overall prediction accuracy.

[0057] S5 Real-time Predictive Control

[0058] In practical applications, a pre-trained model is deployed for real-time temperature prediction. Based on the currently collected input features, the residual A and the outlet temperature T_sys_predicted are predicted. The predicted temperature is compared with the target temperature, and the control variables that need to be adjusted (such as steam flow rate) are calculated and input into the PID controller in advance.

[0059] A computer-readable storage medium having a computer program stored thereon, the computer program being executable by a processor to implement the steps of the method for predicting and controlling the temperature of the material exiting a loose rehumidifier as described in this invention.

[0060] A predictive control system for the temperature of the material exiting a loose rehumidifier, the system comprising:

[0061] Sensors used to acquire information such as inlet material moisture content, instantaneous flow rate of the scale, feed water flow rate, inlet hot air temperature and volume, steam flow rate, atomized steam pressure, exhaust damper opening, circulating damper opening, and drum speed;

[0062] A computer-readable storage medium as described in this invention;

[0063] The computer acquires data such as the moisture content of the inlet material, the instantaneous flow rate of the scale, the feed water flow rate, the inlet hot air temperature and volume, the steam flow rate, the atomized steam pressure, the opening of the exhaust damper, the opening of the circulating damper, and the drum rotation speed by communicating with the aforementioned sensors; and executes the program on the computer-readable storage medium.

[0064] The beneficial effects of this invention include:

[0065] (1) Improved prediction accuracy

[0066] By combining thermodynamic and XGBoost models, the linear and nonlinear effects of the system were effectively captured. Experimental results show that the proportion of samples with prediction errors less than 0.3℃ reaches over 80%, significantly improving the accuracy of temperature prediction.

[0067] (2) Reduced time delay and overshoot:

[0068] Correlation analysis reveals the time it takes for the model to predict the outlet temperature in advance. The PID controller can adjust the control input in a timely manner, reducing system time delay and overshoot, and improving control performance.

[0069] (3) Reduced the computational complexity of the model

[0070] By splitting the model into linear and nonlinear components, the model structure is simplified. Compared to purely data-driven machine learning models, this reduces computational cost and model complexity while maintaining high prediction accuracy.

[0071] (4) Enhanced model interpretability

[0072] Thermodynamic models describe the main energy exchange processes of a system and have clear physical meaning. Machine learning models handle nonlinear residuals, and the combination of the two makes the model easier to understand and maintain.

[0073] (5) Improved product quality and production efficiency

[0074] Precise temperature control helps improve the quality of tobacco processing products, reduce energy consumption, and increase production efficiency, resulting in significant economic and social benefits. Attached Figure Description

[0075] Figure 1 : Flowchart of the method of the present invention.

[0076] Figure 2 Model training flowchart.

[0077] Figure 3 Real-time predictive control flowchart.

[0078] Figure 4 Input correlation heatmap.

[0079] Figure 5 : Comparison of actual and predicted values ​​of test set residual A.

[0080] Figure 6 : Comparison chart of actual and predicted values ​​of outlet material temperature during the steady-state phase of the test set. Detailed Implementation

[0081] Example 1

[0082] 1. Data Collection and Preprocessing

[0083] 1.1 Data Collection

[0084] Thirty batches of data were collected from the Körber ultra-loose rehumidifier at the YX cigarette factory as the training set, and five batches as the test set. The data included input characteristics such as inlet material moisture content, instantaneous flow rate of the weigher, feed water flow rate, inlet hot air temperature and volume, steam flow rate, and atomized steam pressure, as well as the actual value of the outlet material temperature.

[0085] 1.2 Data Preprocessing

[0086] (1) Use moving average filtering to smooth input features and output targets, reduce noise and improve data quality.

[0087] (2) Spearman rank correlation analysis was performed. 21 batches were used to analyze the correlation coefficients. 300s was used as the maximum time delay range for analysis. The correlation between each input feature and the outlet temperature was determined, and the optimal time delay was calculated to ensure that the model can accurately capture the relationship between the input and the output.

[0088] (3) Time alignment based on analysis results. The quantity with the smallest time delay among the strongly correlated inputs in the differential equation is taken as the advance prediction control quantity. The output of the training set is uniformly advanced by the time delay of this input point. Then, the time delay difference that other strongly correlated points in the equation need to be shifted later is calculated and aligned. Timestamp alignment means that the specific time delay of each input point relative to the material flow at the inlet will be obtained. That is, when the material micro-element enters the equipment, the predicted output value can be obtained at the latest input point. This directly determines how far in advance the model can make predictions and the effectiveness of the model's prediction value.

[0089] 1.3 Model Parameter Optimization

[0090] (1) Determine the total heat capacity C of the system

[0091] Construct an optimization problem with the difference between the predicted and actual temperatures as the loss function:

[0092]

[0093] The BFGS algorithm is used to find the optimal C value by minimizing the loss function, thus ensuring the reliability of the thermodynamic model.

[0094] (2) Calculate and model residual A

[0095] After determining C, the actual residuals A of the training set are calculated. An XGBoost regression model is used, with the input features as independent variables, to model A. XGBoost excels at handling nonlinear relationships, efficiently capturing complex patterns in the residuals and improving the model's prediction accuracy.

[0096] 2. Model Validation and Evaluation

[0097] 2.1 Residual Prediction

[0098] On the test set, the trained XGBoost model is used to predict the residual A_predicted, and then substituted into the thermodynamic equation to calculate the predicted value T_sys_predicted of the outlet material temperature.

[0099] 2.2 Performance Evaluation

[0100] like Figure 4 As shown, correlation analysis revealed the three points with the strongest correlation: hot air flow rate, steam flow rate, and feed water flow rate. Their time delays were 100s, 220s, and 150s, respectively. Therefore, we can conclude that the model can predict the outlet temperature 100s in advance. Furthermore, the PID controller can adjust the control input in a timely manner, reducing system time delay and overshoot, thus improving control performance.

[0101] The mean squared error (MSE) and absolute error were calculated to evaluate the model's predictive performance. The results show that, in the steady-state phase, the proportion of samples with prediction errors less than 0.3℃ exceeds 80%, indicating that the model has high accuracy.

[0102] 2.3 Visualization

[0103] Plot a comparison chart of the actual and predicted residual values, as well as a comparison chart of the actual and predicted outlet temperatures, to visually demonstrate the model's predictive capabilities.

[0104] 3. Implementation of temperature prediction control

[0105] 3.1 Real-time Prediction and Control

[0106] In actual production, a pre-trained model is deployed for real-time prediction. Based on the real-time collected input characteristics, the residual A and the outlet temperature T_sys_predicted are predicted. The prediction results are used as the PV value of the PID controller to adjust control variables (such as steam flow rate), and are input into the PID controller in advance.

[0107] 3.2 Improve control effect

[0108] By adjusting the control input in advance, the system's time delay and overshoot were reduced, improving the control system's response speed and stability, and enhancing the performance of temperature control.

[0109] 4. Experimental Results

[0110] like Figure 5 As shown, the vertical axis represents the value of the residual A, in units of 10^5 kJ, and the horizontal axis represents the time index, with each point spaced one second apart. The blue curve represents the actual value of A, and the yellow dashed line represents the model's predicted value of A. From the graph, we can clearly see that the predicted value of A has a very good fit.

[0111] like Figure 6 As shown, the vertical axis represents the outlet material temperature in °C, and the horizontal axis represents the time index, with each point spaced one second apart. The blue curve represents the actual outlet material temperature, and the yellow dashed line represents the corrected outlet material temperature after compensation by the residual A prediction value. From the graph, we can see that the predicted and actual values ​​have a good fit, both in terms of magnitude and trend.

[0112] 5. Model Performance

[0113] On the test set, the model demonstrated high-precision predictive ability. The XGBoost model effectively predicted the residual A, and the overall predicted outlet temperature closely matched the actual value. In practical applications, using this model to optimize the PID controller can reduce the fluctuation of the outlet material temperature, accelerate the response speed of the control system, and improve its stability, thus verifying the effectiveness of the method of this invention.

[0114] 6. Conclusion

[0115] This invention innovatively splits the temperature prediction model for the outlet material of a loose rehumidifier into a linear (quantifiable) part and a nonlinear (residual) part by combining thermodynamic equations and machine learning methods. The physical model captures the main energy exchange processes of the system, while the machine learning model handles complex nonlinear effects, improving prediction accuracy, reducing computational complexity, and enhancing model interpretability. In practical applications, it effectively solves the time delay and overshoot problems in PID control, improving the accuracy and stability of temperature control, and has significant potential for widespread application.

Claims

1. A method for predicting and controlling the temperature of the material exiting a loose rehumidifier, characterized in that, Includes the following steps: S1 establishes a thermodynamic equation model Based on the physical characteristics of the loose rehumidifier, a thermodynamic equation describing the temperature of the outlet material is established as the linear part of the model; For nonlinear effects that cannot be directly quantified, they are summarized as residual term A and modeled using machine learning methods; The thermodynamic equation model includes the following quantifiable influencing factors: (1) Feed water flow rate Used to adjust the moisture content of materials, affecting the thermal balance of the system. (2) Steam flow rate Heat is transferred through steam to raise the temperature of the material. (3) Hot air volume It provides heat energy to the system, affecting the heating effect of the materials; For energies that cannot be directly measured, they are summarized as residual term A; Its thermodynamic differential equation is expressed as follows: ； in: Temperature of the material after absorbing heat, t: time variable. Hot air temperature : Saturated steam temperature Feed water temperature, Specific heat capacity of air Specific heat capacity of saturated steam R: Specific heat capacity of water; Specific heat capacity of steam; Moisture gain coefficient of steam. Latent heat coefficient of steam; S2 data collection and preprocessing; S3 model parameter optimization and validation, including: S3.1 Determine the total heat capacity C of the system; S3.2 Modeling the residual A After determining the total heat capacity C of the system, the actual residual is calculated based on the known data. Then, the XGBoost model is used to model the residual A with the input features as independent variables. S4 Real-time Temperature Prediction and Control Based on the currently collected input characteristics, predict the residual A and the outlet temperature. The predicted temperature is compared with the target temperature to calculate the control variables that need to be adjusted, and then input into the PID controller in advance for control.

2. The method according to claim 1, characterized in that, The energy that cannot be directly measured includes the energy carried away by the exhaust air, the heat radiation loss from the cylinder wall, and the difference in heat absorption of the materials.

3. The method according to claim 1, characterized in that, In step S1, the XGBoost regression model is selected to predict A.

4. The method according to claim 1, characterized in that, Step S2 also includes: S2.1 Data Collection Collect operational data from the loose rehumidifier, the operational data including input characteristics and output targets; The input features include the moisture content of the inlet material, the instantaneous flow rate of the weighing scale, the feed water flow rate, the inlet hot air temperature and volume, the steam flow rate, the atomized steam pressure, the opening of the exhaust damper, the opening of the circulating damper, and the drum rotation speed. The output target includes the actual temperature of the outlet material; S2.2 Data preprocessing includes: (1) Moving average filtering; (2) Spearman rank correlation analysis: determine the correlation between each input feature and the outlet temperature, calculate the optimal time delay, and align the data.

5. The method according to claim 4, characterized in that, The moving average filtering includes smoothing the time-series data to reduce noise interference.

6. The method according to any one of claims 1-5, characterized in that, Determining the total heat capacity C of the system in step S4 also includes: By constructing an optimization problem with the difference between predicted and actual temperatures as the loss function, and using the BFGS algorithm, the problem is solved by minimizing the loss function. Find the optimal total heat capacity C of the system.

7. The method according to claim 6, characterized in that, The minimized loss function include: 。 8. A computer-readable storage medium having a computer program stored thereon, characterized in that, The computer program can be executed by a processor to implement the steps of the method for predicting and controlling the temperature of the material outlet of a loose rehumidifier as described in any one of claims 1-7.

9. A predictive control system for the outlet material temperature of a loose rehumidifier, the system comprising: Sensors used to acquire inlet material moisture content, scale instantaneous flow rate, feed water flow rate, inlet hot air temperature and volume, steam flow rate, atomized steam pressure, exhaust damper opening, circulating damper opening, and drum speed. A computer-readable storage medium as described in claim 8; The computer acquires the inlet material moisture content, instantaneous flow rate of the scale, feed water flow rate, inlet hot air temperature and volume, steam flow rate, atomized steam pressure, exhaust damper opening, circulation damper opening, and drum rotation speed by communicating with the sensors; and executes the computer program stored on the computer-readable storage medium.

Images