Blast furnace parameter optimization method based on multi-scale LOWESS regression
By using the multi-scale LOWESS regression method, the problem of capturing the nonlinear relationship of blast furnace parameters was solved, and efficient, low-consumption, and stable blast furnace production was achieved. Through multi-objective fusion optimization and visualization output, the stability of blast furnace production and ease of operation were improved.
Patent Information
- Application Number
- CN202511602920.5
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2025-11-04
- Publication Date
- 2026-03-03
AI Technical Summary
Existing technologies cannot effectively capture the nonlinear relationships of blast furnace parameters, have poor data processing adaptability, are easily affected by abnormal data in optimization results, have high operational thresholds, and suffer from imbalances in multi-objective optimization, making it difficult to achieve efficient, low-consumption, and stable blast furnace production.
The multi-scale LOWESS regression method is adopted. Through multi-time-scale data preprocessing and adaptive LOWESS parameter calculation, combined with multi-objective fusion optimization, a nonlinear model is constructed, and a visual fitting curve is output, supporting multi-objective optimization and parameter visualization.
It improves the accuracy and stability of blast furnace parameter optimization, reduces the difficulty of operation, achieves multi-objective balance optimization, and enhances the efficiency and stability of blast furnace production.
Smart Images
Figure FT_1 
Figure FT_2 
Figure FT_3
Abstract
Description
Technical Field
[0001] This invention belongs to the field of parameter optimization and control technology in iron and steel smelting process. Specifically, it relates to a blast furnace parameter optimization method based on multi-scale LOWESS regression, which is particularly suitable for intelligent optimization of blast furnace control parameters and furnace charge parameters. It can be integrated into the intelligent production control system of blast furnaces in iron and steel enterprises to realize closed-loop linkage between parameter optimization and production control. Background Technology
[0002] Blast furnace ironmaking is a core process in the steel industry, requiring the coordinated control of multiple parameters to achieve the production goals of "high efficiency, low consumption, and stability." Core blast furnace parameters include control parameters (blast volume, oxygen content, coal ratio, etc.), furnace condition parameters (molten iron temperature, silicon content, etc.), and burden parameters (raw material and fuel ratio). These parameters exhibit strong nonlinear coupling relationships, and blast furnace data is collected frequently and in large quantities. Existing analysis methods struggle to effectively uncover the patterns in these parameters. Blast furnace parameter optimization methods (such as linear regression) and conventional data analysis tools are ill-suited to the functional requirements of blast furnace parameter characteristics. The main problems are as follows: (1) Lack of nonlinear modeling capability: Existing linear analysis methods cannot capture the nonlinear relationship of parameters of large blast furnaces, resulting in large optimization errors and failing to meet the accuracy requirements of the system's "local weighted regression statistical analysis"; (2) Poor data processing adaptability: Blast furnace data has short-term fluctuations and long-term trends. The existing linear regression method has not been processed on multiple scales and is not well adapted to the system's "data cleaning and transformation" function. Abnormal data interferes with the optimization results. (3) Insufficient practicality: The optimization process relies on manual parameter tuning (such as the number of hidden layer nodes in the neural network and the crossover probability of the genetic algorithm), which has a high operational threshold; the optimization results are mostly output in numerical form, lacking the support of visual curves, making it difficult to assist furnace operators in making quick decisions; (4) Multi-objective balance failure: Blast furnaces need to simultaneously achieve "increased output, reduced fuel ratio, and smooth furnace operation". Existing methods often focus on a single indicator (such as only optimizing output), which can easily lead to fluctuations in furnace conditions. They have not formed a multi-objective optimization logic that is compatible with the system's "generating target parameters (multi-weight configuration)" function. Summary of the Invention
[0003] The purpose of this invention is to provide a blast furnace parameter optimization method based on multi-scale LOWESS regression, thereby solving the problems mentioned in the background art.
[0004] To achieve the above objectives, the present invention adopts the following technical solution: A method for optimizing blast furnace parameters based on multi-scale Lowes regression is described below: S1: Multi-timescale data preprocessing: Addressing the time-dimensional differences in blast furnace data, and leveraging the system's data processing capabilities to improve data quality. S1.1. Generation of multi-timescale data For the historical / real-time data imported into the system, moving averages at three time scales are calculated to adapt to the short-term fluctuations and long-term trends of large blast furnace parameters. Blast furnace parameters include control parameters (blast volume, oxygen content, coal ratio), burden parameters (raw material ratio, batch weight), and furnace condition parameters (molten iron temperature, silicon content, furnace top pressure). (1) Short-term scale: 3-8 hour moving average; (2) Medium-term scale: 8-15 hour moving average; (3) Long-term scale: 15-25 hour moving average; S1.2. Dual Outlier Handling Using the system's "outlier filtering" function, outlier data is removed in two steps: (1) IQR outlier removal: Calculate the 25th percentile (Q1) and 75th percentile (Q3) of parameter x to obtain IQR=Q3-Q1. Keep the data where Q1-r×IQR≤x≤Q3+r×IQR, and r takes the value of 1.5-5.0; (2) Quantile range filtering: Configure the lower quantile q_l (values range from 0.25% to 2.5%) and the upper quantile q_h (values range from 97.5% to 99.75%), and retain x within [x q_l ,x q_h Data within the specified range is used to further reduce the impact of extreme values; Dual outlier processing filters out more than 95% of abnormal data (such as sensor failures and instantaneous fluctuations).
[0005] S2: Adaptive LOWESS parameter calculation S2.1. LOWESS parameter calculation principle LOWESS (Locally Weighted Scatter Smoothing Regression) is a nonparametric regression method whose core logic is "local weighting, low-order fitting, and iterative correction," and it is suitable for handling nonlinear data. (1) Local neighborhood determination: For each target data point x_i, a local neighborhood is determined with it as the center (the size of the neighborhood is controlled by the parameter frac). (2) Weight calculation: Gaussian weight function (or triangular weight function) is used to weight the data in the neighborhood. The closer the data is to x_i, the greater the weight, and vice versa, to reduce the interference of far data on the fitting. (3) Low-order fitting: Perform first-order or second-order polynomial fitting on the weighted data to obtain the fitted value y_hat_i at x_i; (4) Iterative correction: Adjust the weights based on the residual (the difference between the actual value y_i and the fitted value y_hat_i) to reduce the weights of outliers (2-4 iterations) and further improve the fitting accuracy.
[0006] This principle can accurately capture the nonlinear relationship of blast furnace parameters, avoiding the modeling bias of traditional linear methods.
[0007] S2.2. Adaptive rule for smoothing parameter delta The core hyperparameters of the LOWESS model include delta (controlling the smoothness of the fitted curve), frac (the proportion of locally weighted data), and it (the number of iterations). Among them, delta directly affects the "smoothness-accuracy" balance of the fitted curve. Based on the numerical characteristics of the blast furnace parameters, delta is automatically adjusted to avoid errors from manual parameter tuning. The specific rules are as follows: (1) If x_median>5000: delta=x_median / 200; (2) If 500 <x_median≤5000:delta=x_median / 100; (3) If 50 <x_median≤500:delta=x_median / 50; (4) If x_median≤50: delta=x_median / 20; (5) The final delta = calculated value × r_delta, r_delta is configured through the system interface to adapt to the analysis needs of different parameters.
[0008] S3: Multi-scale Lowes Fitting (based on the system's core analysis functions) The `statsmodels.nonparametric.lowess` regression function is called to perform Lowess regression on the original parameters and the three time-scale parameters respectively, constructing a nonlinear model: S3.1. Fitting Parameter Configuration Refer to the system's default analysis parameters and set frac and iteration count it; S3.2. Multi-scale fitting execution Fit the target parameter y to the four sets of data (x, x_6, x_12, x_18) of the influence factor: (1) Fitting 1: y_hat1=LOWESS(y,x,frac,it,delta) (original air volume data); (2) Fitting 2: y_hat2=LOWESS(y,x_6,frac,it,delta) (3-8 hour average); (3) Fitting 3: y_hat3=LOWESS(y,x_12,frac,it,delta) (8-15 hour average); (4) Fitting 4: y_hat4=LOWESS(y,x_18,frac,it,delta) (average value of 15-25 hours); S3.3. Output of Fitting Results Four fitting curves were plotted to visually demonstrate the nonlinear relationship between the parameters.
[0009] S4: Multi-objective fusion optimization Based on the blast furnace production requirements (output, cost, furnace condition score), the optimization objective and weights are determined, and the final optimal parameters are obtained by integrating the multi-scale fitting results. Specific steps include: S4.1. Optimize target and weight configuration Referring to the "Target Parameter Mode Selection" function in the system interface, two types of target settings and weight configurations are supported: (1) Single-objective optimization If the objective y is "the larger the value, the better" (e.g., output): the optimal x(k) = argmax(y_hat(k)) (k = 1, 2, 3, 4, corresponding to 4 scales); If the objective y is "the smaller the value, the better" (e.g., fuel ratio): the optimal x(k) = argmin(y_hat(k)); (2) Multi-objective optimization Based on the characteristics of the furnace condition parameters, the furnace condition parameters are evaluated and scored according to three modes: Mode 1: The lower the parameter index value X, the better. If it is below the lower limit L1, the score is 100. If it is above the upper limit U1, the score is 0. The evaluation scoring formula between L1 and U1 is: Single item score R = (U1-X) / (U1-L1) × 100; Mode 2: The higher the parameter index value X, the better. If it is below the lower limit L1, the score is 0 points. If it is above the upper limit U1, the score is 100 points. The evaluation scoring formula between L1 and U1 is: Single item score R = (X-U1) / (U1-L1)×100; Mode 3: A good score is when the parameter index X is in the middle, that is, between the lower limit L2 and the upper limit U2. The further away from this range, the lower the score. If it exceeds the lower limit L1 or the upper limit U1, the score is 0. The evaluation scoring formula for values between L1 and L2 or U2 and U1 is: Single item score R = (X - L1) / (L2 - L1) × 100 or R = (U1 - X) / (U1 - U2) × 100. Based on the characteristics of blast furnace production, the furnace condition indicators are divided into five aspects: pressure-volume relationship, airflow distribution, material feeding, furnace heat level and technical indicators. These are scored and comprehensively evaluated in modules, with multiple objective weights configured. The objectives include output, fuel ratio and furnace condition stability, and the total weight is 1. S4.2. Multi-scale result fusion S4.3. Results Output: Visualized output, displaying multi-scale fitting curves, optimal values at each scale, final optimal value, and expected target value through the system interface.
[0010] Preferably, the value of r_delta in step S2.2 is in the range of 0.2-1.2, and is manually configured through the interface of the blast furnace intelligent control system.
[0011] Preferably, in step S3.2, the LOWESS regression in the multi-scale fitting process uses a Gaussian weighting function to weight the data in the local neighborhood and then perform second-order polynomial fitting.
[0012] Preferably, the multi-scale result fusion operation in step S4.2 is as follows: The final optimal x is calculated using the equal-weighted averaging method (to meet the stability requirements of large blast furnace parameters and avoid the influence of single-scale fluctuations). The formula is as follows: x_best = [x(1) + x(2) + x(3) + x(4)] / 4.
[0013] Preferably, the specific operation for outputting the result of step S4.3 is as follows: (1) The optimal values x(1)-x(4) at each scale and the corresponding y_hat values; (2) The final optimal value x_best and the expected target value y_best after fusion; (3) Comparison of multi-scale fitting curves with actual data to help engineers verify the rationality of optimization.
[0014] Compared with the prior art, the beneficial effects achieved by the present invention are as follows: (1) There is no need to re-adjust the model for specific parameters (air volume, coal ratio, silicon content, etc.). The system analysis function can be directly reused, adapting to all core parameters of blast furnaces. The parameter adaptability is wide. It is also applicable to blast furnaces of different volumes and has strong versatility. (2) Through multi-objective fusion optimization, it supports weighted optimization of output, cost and furnace condition stability, improves multi-objective balance performance, and avoids furnace condition imbalance caused by a single objective; (3) The multi-scale LOWESS fitting results output a visual fitting curve, which reduces the operation threshold, improves practicality and interpretability, and is easy to integrate and promote. Attached Figure Description
[0015] Figure 1 This is a locally weighted regression analysis graph of [Si] and comprehensive score in Embodiment 1 of the present invention; Figure 2 This is a local weighted regression analysis diagram of the four-burned SiO2 and the comprehensive score in Example 2 of the present invention; Figure 3This is a locally weighted regression analysis diagram of the total heat load and comprehensive score in Embodiment 3 of the present invention; Figure 4 This is a local weighted regression analysis diagram of the heat load and comprehensive score of floors 1-13 in Embodiment 4 of the present invention; Figure 5 This is a locally weighted regression analysis diagram of coal ratio and comprehensive score in Example 5 of the present invention. Detailed Implementation
[0016] The technical solution of the present invention will be described in detail below with reference to the embodiments.
[0017] For a 4350m³ large blast furnace, the optimization parameters are as follows: the independent variable x represents the silicon content [Si] in molten iron, the SiO2 content of four layers, the total heat load, the heat load of layers 1-13, and the coal ratio; the dependent variable y represents the overall score of the blast furnace (1-100 points, with higher scores indicating better overall blast furnace operation). The optimization objective is to accurately determine the optimal control targets for molten iron [Si], SiO2 content of four layers, total heat load, heat load of layers 1-13, and coal ratio, thereby improving the overall score of the blast furnace. The blast furnace parameter optimization method based on multi-scale Lowes regression of this invention is adopted, and the specific steps are as follows: Example 1
[0018] S1: Data Acquisition and Preprocessing S1.1. Data Acquisition: Export historical data of molten iron [Si] (x) and blast furnace comprehensive score (y) from the secondary system; S1.2. Multi-timescale data generation: Calculating the moving average of molten iron [Si] over three timescales: (1) Short-term scale: 6-hour moving average; (2) Medium-term scale: 12-hour moving average; (3) Long-term scale: 18-hour moving average; S1.3. Dual outlier handling (for optimizing the process characteristics of molten iron [Si]): (1) IQR outlier removal: Calculate the 25th percentile Q1=0.38 and the 75th percentile Q3=0.78 of molten iron [Si], and get IQR=0.40. r takes the value of 2.5, that is, retain "0.38-2.5×0.40≤x≤0.78+2.5×0.40" which is -0.62≤x≤1.78; (2) Quantile range filtering: Configure the lower quantile q_l=1% and the upper quantile q_h=99% to retain data in the range of [0.31,0.99].
[0019] S2: Adaptive LOWESS parameter (delta) calculation S2.1. Calculation of median of parameters: The median of molten iron [Si] after statistical preprocessing is x_median=0.55, which belongs to the value range of "x_median≤50"; S2.2. Basic delta calculation: According to the rule "when x_median≤50, delta_base=x_median / 20", we get delta_base=0.55 / 20=0.0275; S2.3. Final delta calculation: Set r_delta=0.3, final delta=0.0275×0.3=0.00825.
[0020] S3: Multi-scale Lowes Fitting and Optimization S3.1. Fitting Parameter Configuration: Set the core parameters of the LOWESS model: frac=0.4, iteration count it=3. S3.2. Multi-scale fitting execution: The system's built-in LOWESS function is called to fit the target parameter y (blast furnace comprehensive score) and the four sets of data (x, x_6, x_12, x_18) of the influence factor x ([Si]) respectively: (1) Fitting 1: y_hat1=LOWESS(y,x,frac,it,delta) (original air volume data); (2) Fitting 2: y_hat2=LOWESS(y,x_6,frac,it,delta) (6-hour average); (3) Fitting 3: y_hat3=LOWESS(y,x_12,frac,it,delta) (12-hour average); (4) Fitting 4: y_hat4=LOWESS(y,x_18,frac,it,delta) (18-hour average); S3.3. Fitting Result Output: Plot 4 fitting curves, as shown below. Figure 1 As shown, the nonlinear relationship between parameters is intuitively displayed, and the optimal target value of molten iron [Si] is found to be 0.42. Example 2
[0021] S1: Data Acquisition and Preprocessing S1.1. Data Acquisition: Export historical data of four-stage SiO2 (x) and blast furnace comprehensive score (y) from the secondary system; S1.2. Multi-timescale data generation: Calculating the moving average of SiO2 across 3 timescales: (1) Short-term scale: 6-hour moving average; (2) Medium-term scale: 12-hour moving average; (3) Long-term scale: 18-hour moving average; S1.3. Dual Outlier Handling: (1) IQR outlier removal: Calculate the 25th percentile Q1=5.26 and the 75th percentile Q3=5.38 of the four-dimensional SiO2, and get IQR=0.12. r takes the value of 2.5, and retain "5.26-2.5×0.12≤x≤5.38+2.5×0.12" that is, 4.96≤x≤5.68; (2) Quantile range filtering: Configure the lower quantile q_l=1% and the upper quantile q_h=99%, and retain data in the range of [5.18, 5.51].
[0022] S2: Adaptive LOWESS parameter (delta) calculation S2.1. Calculation of median of parameters: After statistical preprocessing, the median of the four-burned SiO2 is x_median=5.34, which falls within the value range of "x_median≤50"; S2.2. Basic delta calculation: According to the rule "when x_median≤50, delta_base=x_median / 20", we get delta_base=5.34 / 20=0.267; S2.3. Final delta calculation: Set r_delta=0.3, the final delta=0.267×0.3=0.080.
[0023] S3: Multi-scale Lowes Fitting and Optimization S3.1. Fitting parameter configuration: Set the core parameters of the LOWESS model: frac=0.4, iteration number it=3; S3.2. Multi-scale fitting execution: Call the system's built-in LOWESS function to fit the target parameter y (blast furnace comprehensive score) and the four sets of data (x, x_6, x_12, x_18) of the influence factor x (four-burned SiO2) respectively; S3.3. Fitting Result Output: Plot 4 fitting curves, as shown below. Figure 2 As shown, the nonlinear relationship between parameters is intuitively displayed, and the optimal target value of four-burned SiO2 is found to be 5.18. Example 3
[0024] S1: Data Acquisition and Preprocessing S1.1. Data Acquisition: Export historical data of total heat load (x) and blast furnace comprehensive score (y) from the secondary system; S1.2. Multi-timescale data generation: Calculate the moving average of the total heat load over three time scales: (1) Short-term scale: 6-hour moving average; (2) Medium-term scale: 12-hour moving average; (3) Long-term scale: 18-hour moving average; S1.3. Dual outlier handling (process characteristic optimization for total heat load): (1) IQR outlier removal: Calculate the 25th percentile Q1=10592 and the 75th percentile Q3=14283 of the total heat load, and get IQR=3691. r takes the value of 2.5, and retain "10592-2.5×3691≤x≤14283+2.5×3691", that is, 1364.5≤x≤23510.5; (2) Quantile range filtering: Configure the lower quantile q_l=1% and the upper quantile q_h=99%, and retain data in the range of [8112,21526].
[0025] S2: Adaptive LOWESS parameter (delta) calculation S2.1. Calculation of median parameter: The median of the total heat load after statistical preprocessing is x_median=14203, which falls within the value range of "x_median>5000"; S2.2. Basic delta calculation: According to the rule "when x_median>5000, delta_base=x_median / 200", we get delta_base=14203 / 200=71.015; S2.3. Final delta calculation: Set r_delta=0.3, final delta=71.015×0.3=21.305.
[0026] S3: Multi-scale Lowes Fitting and Optimization S3.1. Fitting parameter configuration: Set the core parameters of the LOWESS model: frac=0.4, iteration number it=3; S3.2. Multi-scale fitting execution: Call the system's built-in LOWESS function to fit the target parameter y (blast furnace comprehensive score) and the four sets of data (x, x_6, x_12, x_18) of the influence factor x (total heat load) respectively; S3.3. Fitting Result Output: Plot 4 fitting curves, as shown below. Figure 3 As shown, the nonlinear relationship between parameters is intuitively displayed, and the optimal target for total heat load is found to be 10544*10MJ / h. Example 4
[0027] S1: Data Acquisition and Preprocessing S1.1. Data Acquisition: Export historical data of heat load (x) and blast furnace comprehensive score (y) for floors 1-13 from the secondary system; S1.2. Multi - time - scale data generation: Calculate the moving averages of the heat loads of 1 - 13 layers at three time scales: (1) Short - term scale: 6 - hour moving average; (2) Medium - term scale: 12 - hour moving average; (3) Long - term scale: 18 - hour moving average; S1.3. Dual outlier handling (optimization for the process characteristics of the heat loads of 1 - 13 layers): (1) IQR outlier rejection: Calculate the 25% quantile Q1 = 1669 and 75% quantile Q3 = 2546 of the heat loads of 1 - 13 layers, getting IQR = 877, r takes the value of 2.5, and retain the data where "1669 - 2.5×877 ≤ x ≤ 2546 + 2.5×877", that is, - 523.5 ≤ x ≤ 4738.5; (2) Quantile range screening: Configure the lower quantile q_l = 1% and the upper quantile q_h = 99%, and retain the data where x is within the range of [1223, 3721].
[0028] S2: Calculation of adaptive LOWESS parameter (delta) S2.1. Median calculation of parameters: Statistically calculate the median x_median = 2678 of the heat loads of 1 - 13 layers after pre - processing, which belongs to the numerical range of "500 < x_median ≤ 5000"; S2.2. Basic delta calculation: According to the rule "when 500 < x_median ≤ 5000, delta_base = x_median / 100", get delta_base = 2678 / 100 = 26.78; S2.3. Final delta calculation: Set r_delta = 0.3, and the final delta = 26.78×0.3 = 8.034.
[0029] S3: Multi - scale LOWESS fitting and optimal value seeking S3.1. Fitting parameter configuration: Set the core parameters of the LOWESS model: frac = 0.4, number of iterations it = 3; S3.2. Multi - scale fitting execution: Call the built - in LOWESS function of the system to respectively fit the four groups of data (x, x_6, x_12, x_18) of the target parameter y (comprehensive score of the blast furnace) and the influencing factor x (heat loads of 1 - 13 layers); S3.3. Output of fitting results: Plot four fitting curves, as Figure 4 shown, visually display the non - linear relationship between parameters, and obtain the optimal target of the heat loads of 1 - 13 layers as 3126.0 * 10 MJ / h. Example 5
[0030] S1: Data collection and preprocessing S1.1. Data collection: Export the historical data of coal ratio (x) and blast furnace comprehensive score (y) from the secondary system; S1.2. Generation of multi-time scale data: Calculate the moving average of the coal ratio at 3 time scales: (1) Short-term scale: 6-hour moving average; (2) Medium-term scale: 12-hour moving average; (3) Long-term scale: 18-hour moving average; S1.3. Dual outlier processing (optimization for the process characteristics of coal ratio): (1) IQR outlier removal: Calculate the 25% quantile Q1 = 132.7 and 75% quantile Q3 = 153.8 of the coal ratio, obtaining IQR = 21.1. Take r as 2.5, and retain the data where "132.7 - 2.5×21.1 ≤ x ≤ 153.8 + 2.5×21.1", that is, 79.95 ≤ x ≤ 206.55; (2) Quantile range screening: Configure the lower quantile q_l = 1% and the upper quantile q_h = 99%, and retain the data where x is within the range of [99.6, 192.5].
[0031] S2: Calculation of adaptive LOWESS parameter (delta) S2.1. Calculation of parameter median: Statistically calculate the median x_median = 155 of the coal ratio after preprocessing, which belongs to the numerical interval of "50 < x_median ≤ 500"; S2.2. Calculation of basic delta: According to the rule "when 50 < x_median ≤ 500, delta_base = x_median / 50", obtain delta_base = 155 / 50 = 3.1; S2.3. Calculation of final delta: Set r_delta = 0.3, and the final delta = 3.1×0.3 = 0.93.
[0032] S3: Multi-scale LOWESS fitting and optimal value search S3.1. Configuration of fitting parameters: Set the core parameters of the LOWESS model: frac = 0.4, number of iterations it = 3; S3.2. Execution of multi-scale fitting: Call the built-in LOWESS function of the system to fit the 4 groups of data (x, x_6, x_12, x_18) of the target parameter y (blast furnace comprehensive score) and the influencing factor x (coal ratio) respectively; S3.3. Output of fitting results: Plot 4 fitting curves, as Figure 5As shown, the nonlinear relationship between parameters is intuitively displayed, and the optimal target for the coal ratio is found to be 150.6.
[0033] The five embodiments of this invention target a 4350m³ blast furnace. Through multi-timescale preprocessing, adaptive Lowes fitting, and multi-scale fusion, the optimal control targets for molten iron [Si] are accurately positioned as 0.42, for quaternary SiO2 as 5.18, for total heat load as 10544*10MJ / h, for heat load of layers 1-13 as 3216*10MJ / h, and for coal ratio as 150.6. This improves the overall score of the blast furnace and enhances its overall operating condition.
Claims
1. A method for optimizing blast furnace parameters based on multi-scale Lowes regression, characterized in that, The specific steps are as follows: S1: Multi-timescale data preprocessing: Addressing the temporal differences in blast furnace data and leveraging the system's data processing capabilities to improve data quality. S1.
1. Generation of multi-timescale data For the historical / real-time data imported into the system, moving averages are calculated at three time scales to adapt to the short-term fluctuations and long-term trends of large blast furnace parameters, including control parameters, burden parameters, and furnace condition parameters. (1) Short-term scale: 3-8 hour moving average; (2) Medium-term scale: 8-15 hour moving average; (3) Long-term scale: 15-25 hour moving average; S1.
2. Dual Outlier Handling Using the system's "outlier filtering" function, outlier data is removed in two steps: (1) IQR outlier removal: Calculate the 25th percentile (Q1) and 75th percentile (Q3) of parameter x to obtain IQR=Q3-Q1. Keep the data where Q1-r×IQR≤x≤Q3+r×IQR, and r takes the value of 1.5-5.0; (2) Quantile range filtering: Configure the lower quantile q_l, with a value of 0.25%-2.5%; the upper quantile q_h, with a value of 97.5%-99.75%, and retain x within [x q_l ,x q_h Data within the specified range; S2: Adaptive LOWESS parameter calculation S2.
1. LOWESS parameter calculation principle LOWESS, or locally weighted scatter smoothing regression, is a nonparametric regression method. Its core logic is "local weighting, low-order fitting, and iterative correction," and it is suitable for processing nonlinear data. (1) Local neighborhood determination: For each target data point x_i, a local neighborhood is determined with it as the center. The size of the neighborhood is controlled by the parameter frac. (2) Weight calculation: Gaussian weight function or triangular weight function is used to weight the data in the neighborhood. The closer the data is to x_i, the greater the weight, and vice versa, to reduce the interference of far data on the fitting. (3) Low-order fitting: Perform first-order or second-order polynomial fitting on the weighted data to obtain the fitted value y_hat_i at x_i; (4) Iterative correction: Adjust the weights based on the residuals, where the residuals are the difference between the actual value y_i and the fitted value y_hat_i, reduce the weights of outliers, and iterate 2-4 times; S2.
2. Adaptive rule for smoothing parameter delta The core hyperparameters of the LOWESS model include delta, frac, and it. delta controls the smoothness of the fitted curve, frac is the proportion of locally weighted data, and it is the number of iterations. Among them, delta directly affects the "smoothness-accuracy" balance of the fitted curve. Based on the numerical characteristics of the blast furnace parameters, delta is automatically adjusted according to the following rules: (1) If x_median>5000: delta=x_median / 200; (2) If 500 <x_median≤5000:delta=x_median / 100; (3) If 50 <x_median≤500:delta=x_median / 50; (4) If x_median≤50: delta=x_median / 20; (5) The final delta = calculated value × r_delta, where r_delta is configured through the system interface to adapt to the analysis needs of different parameters; S3: Multiscale Lowes Fit The `statsmodels.nonparametric.lowess` regression function is called to perform Lowess regression on the original parameters and the three time-scale parameters respectively, constructing a nonlinear model: S3.
1. Fitting Parameter Configuration Refer to the system's default analysis parameters and set frac and iteration count it; S3.
2. Multi-scale fitting execution Fit the target parameter y to the four sets of data (x, x_6, x_12, x_18) of the influence factor: (1) Fitting 1: y_hat1=LOWESS(y,x,frac,it,delta) (original air volume data); (2) Fitting 2: y_hat2=LOWESS(y,x_6,frac,it,delta) (3-8 hour average); (3) Fitting 3: y_hat3=LOWESS(y,x_12,frac,it,delta) (8-15 hour average); (4) Fitting 4: y_hat4=LOWESS(y,x_18,frac,it,delta) (average value of 15-25 hours); S3.
3. Output of Fitting Results Four fitting curves were plotted to visually demonstrate the nonlinear relationship between the parameters; S4: Multi-objective fusion optimization Based on the blast furnace production requirements (output, cost, furnace condition score), the optimization objective and weights are determined, and the final optimal parameters are obtained by integrating the multi-scale fitting results. Specific steps include: S4.
1. Optimize target and weight configuration Referring to the "Target Parameter Mode Selection" function in the system interface, two types of target settings and weight configurations are supported: (1) Single-objective optimization If the objective y is "the larger the value, the better" (e.g., output): the optimal x(k) = argmax(y_hat(k)) (k = 1, 2, 3, 4, corresponding to 4 scales); If the objective y is "the smaller the value, the better" (e.g., fuel ratio): the optimal x(k) = argmin(y_hat(k)); (2) Multi-objective optimization Based on the characteristics of the furnace condition parameters, the furnace condition parameters are evaluated and scored according to three modes: Mode 1: The lower the parameter index value X, the better. If it is below the lower limit L1, the score is 100. If it is above the upper limit U1, the score is 0. The evaluation scoring formula between L1 and U1 is: Single item score R = (U1-X) / (U1-L1) × 100; Mode 2: The higher the parameter index value X, the better. If it is below the lower limit L1, the score is 0 points. If it is above the upper limit U1, the score is 100 points. The evaluation scoring formula between L1 and U1 is: Single item score R = (X-U1) / (U1-L1)×100; Mode 3: A good score is when the parameter index X is in the middle, that is, between the lower limit L2 and the upper limit U2. The further away from this range, the lower the score. If it exceeds the lower limit L1 or the upper limit U1, the score is 0. The evaluation scoring formula for values between L1 and L2 or U2 and U1 is: Single item score R = (X - L1) / (L2 - L1) × 100 or R = (U1 - X) / (U1 - U2) × 100. Based on the characteristics of blast furnace production, the furnace condition indicators are divided into five aspects: pressure-volume relationship, airflow distribution, material feeding, furnace heat level and technical indicators. These are scored and comprehensively evaluated in modules, with multiple objective weights configured. The objectives include output, fuel ratio and furnace condition stability, and the total weight is 1. S4.
2. Multi-scale result fusion S4.
3. Results Output: Visualized output, displaying multi-scale fitting curves, optimal values at each scale, final optimal value, and expected target value through the system interface.
2. The blast furnace parameter optimization method based on multi-scale Lowes regression according to claim 1, characterized in that, The value of r_delta in step S2.2 is in the range of 0.2-1.2 and can be manually configured through the interface of the blast furnace intelligent control system.
3. The blast furnace parameter optimization method based on multi-scale Lowes regression according to claim 1, characterized in that, In step S3.2, the LOWESS regression uses a Gaussian weighting function to weight the data in the local neighborhood and then perform a second-order polynomial fitting.
4. The blast furnace parameter optimization method based on multi-scale Lowes regression according to claim 1, characterized in that, The multi-scale result fusion operation in step S4.2 is as follows: The final optimal x is calculated using the equal-weighted averaging method, and the formula is as follows: x_best = [x(1) + x(2) + x(3) + x(4)] / 4.
5. The blast furnace parameter optimization method based on multi-scale Lowes regression according to claim 1, characterized in that, The specific steps for outputting the results in step S4.3 are as follows: (1) The optimal values x(1)-x(4) at each scale and the corresponding y_hat values; (2) The final optimal value x_best and the expected target value y_best after fusion; (3) Comparison of multi-scale fitting curves with actual data to help engineers verify the rationality of optimization.