Prediction method of silicon content in molten iron based on TD-LightGBM

Abstract

The present invention discloses a method for predicting silicon content in molten iron based on TD-LightGBM, and belongs to the technical field of industrial process monitoring. The method comprises: first, using the decision tree model structure of a lightweight gradient booster to judge the importance of the feature according to the number of times the feature is used to segment training data; then, combining a time difference algorithm to establish a dynamic feature extraction model; finally, using a local weighted algorithm to establish a local model based on the Euclidean distance between samples. Predicting the silicon content in molten iron during the ironmaking process based on this model can effectively remove redundant items and drift phases in the blast furnace environmental parameters. Compared with the existing prediction scheme for silicon content in molten iron, the present invention further improves the accuracy of the prediction.

Description

Technical Field

[0001] The invention relates to a method for predicting silicon content in molten iron based on TD-LightGBM, and belongs to the technical field of industrial process monitoring. Background Art

[0002] Blast furnace ironmaking is one of the most energy-intensive industrial processes in the world. Its main purpose is to melt iron ore into molten iron. One of the key parameters in the blast furnace ironmaking process is the silicon content in the molten iron. The silicon content in the molten iron is an important indicator for evaluating the stability of the blast furnace condition and the quality of pig iron. Therefore, accurately predicting the silicon content in the molten iron helps control the thermal state of the blast furnace and thus ensure the stable operation of the blast furnace. Due to the high temperature and high pressure environment in the furnace, the measurement of silicon content becomes very difficult.

[0003] In recent years, with the explosive growth of data information in industrial processes, data-based soft measurement models are increasingly used to predict quality variables that are difficult to measure online due to poor measurement environments and large time lags. At present, data-driven modeling can be mainly divided into artificial intelligence methods and multivariate statistical analysis methods. Among them, artificial intelligence soft measurement methods are currently a hot research topic, and modeling methods mainly include artificial neural networks, support vector machines, etc. This type of model has strong nonlinear fitting capabilities, but due to the complexity of calculations and lack of reliability basis, it is still rarely used in actual industrial soft measurement. Multivariate statistical analysis is a relatively mature method, including principal component regression, partial least squares regression, etc. It has been widely used in actual industrial processes.

[0004] Constructing a soft sensor model to predict industrial data requires not only the screening of input variables and samples, but also the training of model parameters, which makes the construction of high-performance soft sensor models very time-consuming. When a high-performance soft sensor model is put into use, its prediction performance will slowly decline with factors such as changes in process state characteristics, catalyst deactivation, raw material changes, and instrument drift. Therefore, the maintenance of the soft sensor model is very important for maintaining its performance. Scholars have proposed many model update strategies, which can be divided into two types: time and space adaptation. Time adaptive models mainly include recursive models and sliding window models, which use the latest sample data for model update. The instant learning model uses the historical samples with the greatest spatial correlation with the query sample to update the model, so it belongs to the space adaptive model.

[0005] Liu Yi et al. proposed a preprocessing method combining two-layer variable space analysis to solve the problem that noise and outliers in blast furnace ironmaking data will reduce the accuracy of soft sensor models. They also used support vector regression to establish a soft sensor model (Liu Yi, Li Ping, Gao Zengliang. Improved support vector regression for blast furnace molten iron quality prediction [J]. Journal of Zhejiang University (Engineering Edition), 2012, 46(05): 830–836.). Guo Dongwei et al. proposed a molten iron dynamic soft sensor modeling method based on sparse robust least squares support vector machine and multi-objective genetic parameter optimization (Guo Dongwei, Zhou Ping. Soft sensor modeling of molten iron silicon content based on sparse robust LS-SVR and multi-objective optimization [J]. Journal of Engineering Science, 2016, 38(09): 1233-1241.).

[0006] Although the above methods can improve the accuracy of molten iron silicon content prediction to a certain extent, they do not take into account the problems of variable redundancy and process variable drift that often exist in the blast furnace ironmaking process, and they cannot update the established model in real time, resulting in the accuracy of molten iron silicon content prediction in practical applications is still greatly affected. Summary of the invention

[0007] In order to solve the problem that the current method for predicting silicon content in molten iron ignores the variable drift during long-term use of the instrument and cannot be updated in real time, resulting in low prediction accuracy, the present invention provides a method for predicting silicon content in molten iron, the method comprising:

[0008] Step 1: Collect the historical actual operation data of the iron-making blast furnace and construct the historical sample set and query sample set;

[0009] Step 2: Obtain the input data x(1), x(2), ..., x(t+1) of the historical sample set, and the output data y(1), y(2), ..., y(t+1), use LightGBM to build a decision tree model, and obtain a feature vector, where t is the sampling time of the historical sample set;

[0010] Step 3: Sort the feature vectors extracted by the LightGBM decision tree model according to the gain of each feature, and select the first f feature vectors;

[0011] Step 4: Calculate the differential terms Δx(1), Δx(2), …, Δx(t) and Δy(1), Δy(2), …, Δy(t) of the feature vectors obtained in step 3, and the differential term Δx(q-1) = x(q) - x(q-1) of the query sample input, where q is the sampling time of the query sample set;

[0012] Step 5: Calculate the Euclidean distance between the query sample difference item in step 4 and each historical input sample difference item and the corresponding weight value, and select the N historical difference items most relevant to the query sample as the training set ΔXN and ΔY N , and obtain the weight matrix Ω;

[0013] Step 6: Construct a local weighted ridge regression LWRR model and use the training set ΔX in step 5 N , ΔY N and weight matrix Ω to train the local weighted ridge regression LWRR model;

[0014] Step 7: Obtain real-time data of blast furnace operation, and use the LWRR model trained in step 6 to obtain the differential prediction value of silicon content in molten iron at the current moment and the previous moment

[0015] Step 8: Get the final predicted value of silicon content in molten iron Where y(q-1) is the actual value of the silicon content in the molten iron at the previous moment.

[0016] Optionally, the LightGBM decision tree model judges the importance of each feature according to the number of times it is used to segment the training data, sorts the data according to the gradient using the unilateral gradient sampling algorithm, retains the first a% of the data with larger gradients to form a subset A; for the residual set A formed by the (1-a)% of the data with smaller gradients c , randomly formed with size b·|A c |, and calculate the information gain:

[0017]

[0018] Among them, A l ={x(t)∈A:x j (t)≤d},A r ={x(t)∈A:x j (t)>d},B l ={x(t)∈B:x j (t)≤d},B r ={x(t)∈B:x j (t)>d}, g(t) is the gradient, x(t) represents the sampled value of the auxiliary variable at time t, d is the point where the optimal gain is constant found in the split calculation, Used to normalize the gradient of B to A c , It is a set A l The number of samples, x j (t) is the auxiliary variable of the j-th feature at time t.

[0019] Optionally, the local weighted ridge regression LWRR model includes: using a local weighted algorithm to assign different weights to each sample according to the Euclidean distance between the query sample and the historical sample; using a ridge regression algorithm and introducing a Tikhonov–Phillips regularization term to solve the collinearity problem in linear regression.

[0020] Optionally, the weight assigned to each sample by the local weighted ridge regression LWRR model is:

[0021]

[0022] in, is an adjustable parameter in the weight function, is the standard deviation of the distance, d i is the distance between the query sample and the i-th historical sample; the weight of each sample is expressed as a weight vector w = [w1, w2, ..., w n ], and the weight matrix is ​​recorded as Ω = diag(w).

[0023] Optionally, the predicted output value of the query sample in the local weighted ridge regression LWRR model The estimation process includes:

[0024] Step 71: De-averaging, calculating the weighted average of sample input and output, and performing de-averaging processing on the input and output of the training sample and the input of the query sample;

[0025] Step 72: Set the initial estimate of the query sample output to

[0026] Step 73: Define the cost function using the Tikhonov–Phillips regularization term;

[0027]

[0028] Step 74: The cost function J is Find the partial derivatives and get the estimated values ​​of the parameters;

[0029]

[0030] Step 75: Calculate the predicted output value

[0031]

[0032] Where X∈R n×m For a given set of auxiliary variables, Y∈R n×1 is the dominant variable set corresponding to the auxiliary variable set, m and n are the dimension and number of samples of the auxiliary variable, respectively, x(q)∈R 1×mis the input of the query sample, λ is the regular term coefficient of ridge regression, is the average silicon content of molten iron in historical samples, are the ridge regression parameters, is a set of predicted values ​​for silicon content in molten iron.

[0033] Optionally, the ridge regression regularization term coefficient λ is calculated using the ridge trace method.

[0034] Optionally, the adjustable parameters in the weight function are Obtained using cross validation method.

[0035] Optionally, the number of local training samples N is obtained using a cross-validation method.

[0036] Optionally, the actual operating data of the blast furnace include: coal powder amount, blast air volume, air permeability, blast temperature, feed speed, blast pressure and oxygen enrichment.

[0037] Optionally, the historical actual operation data of the iron-making blast furnace is collected, and a total of 500 samples are collected, wherein the first 400 samples are used to construct a historical sample set, and the remaining 100 samples are used to construct a query sample set.

[0038] The beneficial effects of the present invention are:

[0039] The method for predicting silicon content in molten iron of the present invention solves the problems of variable redundancy and variable drift through the TD-LightGBM method, thereby achieving the purpose of removing redundant variables and eliminating variable drift noise. At the same time, LightGBM, as a supervised feature extraction algorithm, can more effectively extract auxiliary variables related to the silicon content in molten iron; through the local weighted ridge regression LWRR method, the nonlinear problem of the blast furnace ironmaking process is solved, and the model can be updated in real time, thereby achieving the purpose of improving the model accuracy over time. Compared with existing prediction schemes, the present invention further improves the prediction accuracy in actual industrial scenarios. BRIEF DESCRIPTION OF THE DRAWINGS

[0040] In order to more clearly illustrate the technical solutions in the embodiments of the present invention, the following briefly introduces the drawings required for use in the description of the embodiments. Obviously, the drawings described below are only some embodiments of the present invention. For ordinary technicians in this field, other drawings can be obtained based on these drawings without creative work.

[0041] Figure 1 It is a schematic diagram of the LightGBM histogram algorithm of the present invention.

[0042] Figure 2 It is a schematic diagram of the leaf-based growth strategy of the LightGBM feature extraction model of the present invention.

[0043] Figure 3 It is a model training diagram of the time difference method of the present invention.

[0044] Figure 4 It is a curve diagram showing the relationship between the predicted value and the actual value of the silicon content in the molten iron according to the second embodiment of the present invention.

[0045] Figure 5 It is a comparison diagram of prediction error curves of four methods in Example 2 of the present invention. DETAILED DESCRIPTION

[0046] In order to make the objectives, technical solutions and advantages of the present invention more clear, the embodiments of the present invention will be further described in detail below with reference to the accompanying drawings.

[0047] Embodiment 1:

[0048] This embodiment provides a method for predicting silicon content in molten iron, comprising:

[0049] Step 1: Collect the historical actual operation data of the iron-making blast furnace and construct the historical sample set and query sample set;

[0050] Step 2: Obtain the input data x(1), x(2), ..., x(t+1) of the historical sample set, and the output data y(1), y(2), ..., y(t+1), use LightGBM to build a decision tree model, and obtain a feature vector, where t is the sampling time of the historical sample set;

[0051] Step 3: Sort the feature vectors extracted by the LightGBM decision tree model according to the gain of each feature, and select the first f feature vectors;

[0052] Step 4: Calculate the differential terms Δx(1), Δx(2), …, Δx(t) and Δy(1), Δy(2), …, Δy(t) of the feature vectors obtained in step 3, and the differential term Δx(q-1) = x(q) - x(q-1) of the query sample input, where q is the sampling time of the query sample set;

[0053] Step 5: Calculate the Euclidean distance between the query sample difference item in step 4 and each historical input sample difference item and the corresponding weight value, and select the N historical difference items most relevant to the query sample as the training set ΔX N and ΔY N , and obtain the weight matrix Ω;

[0054] Step 6: Construct a local weighted ridge regression LWRR model and use the training set ΔX in step 5 N , ΔY N and weight matrix Ω to train the local weighted ridge regression LWRR model;

[0055] Step 7: Obtain real-time data of blast furnace operation, and use the LWRR model trained in step 6 to obtain the differential prediction value of silicon content in molten iron at the current moment and the previous moment

[0056] Step 8: Get the final predicted value of silicon content in molten iron Where y(q-1) is the actual value of the silicon content in the molten iron at the previous moment.

[0057] Embodiment 2:

[0058] This embodiment provides a method for predicting the silicon content in molten iron. First, a prediction model is constructed. The construction scheme is as follows:

[0059] (I) Feature extraction model based on LightGBM

[0060] LightGBM is an algorithm for extracting feature variables. It is based on the optimization framework of the gradient boosting decision tree model. It trains multiple tree models by adding them together. Each new tree model is used to predict the error of the previous model to obtain a prediction model with better generalization. It is often used in multi-classification tasks, currency price prediction, and network intrusion detection. In addition, LightGBM can judge the importance of each feature based on the number of times it is used to split the training data, that is, it uses the unilateral gradient sampling algorithm to sort the data according to the gradient, and retains the first a% of the data with larger gradients to form a subset A. For the residual set A formed by (1-a)% of the data with smaller gradients c , randomly formed with size b·|A c |'s subset B and calculate its information gain, as shown in formula (1).

[0061]

[0062] In the formula, A l ={x(t)∈A:x j (t)≤d},A r ={x(t)∈A:x j (t)>d},B l ={x(t)∈B:x j (t)≤d},B r ={x(t)∈B:x j (t)>d}, g(t) is the gradient, x(t) represents the sampled value of the auxiliary variable at time t, d is the point where the optimal gain is constant found in the split calculation, Used to normalize the gradient of B to A c .

[0063] First, the input and output are normalized and the initial gradient value is set to 0. Then, the continuous floating-point feature values ​​obtained in feature engineering are discretized into k integers and packaged to create a histogram with a width of k. The process is as follows: Figure 1 shown.

[0064] Then, a leaf-by-leaf growth strategy with depth restriction is adopted to select the node with the largest information gain for growth. Figure 2 As shown in Figure 1, compared with the traditional layer-by-layer growth strategy, the leaf-by-leaf growth strategy avoids unnecessary calculations and improves computational efficiency. Finally, the gradient value is updated and repeated until all trees are built.

[0065] 2. Time Difference Model

[0066] The time difference model TD model is often used to solve the drift problem of process variables. Its core idea is to establish a mathematical relationship between Δx(t-1) and Δy(t-1), fully considering the time difference dynamics of process variables. The modeling process is as follows: Figure 3 As shown. Δx(t-1) and Δy(t-1) represent the first-order difference terms of the auxiliary variable and the quality variable respectively, and y(t) represents the sampled value of the dominant variable at time t. After the model training is completed, when a new sample x(t+1) arrives, the first-order difference Δx(t) is calculated to obtain the predicted output value of the dominant variable.

[0067] (III) Local Weighted Ridge Regression LWRR Algorithm

[0068] The local weighted algorithm assigns different weights to each sample according to the Euclidean distance between the query sample and the historical sample; the ridge regression algorithm introduces the Tikhonov–Phillips regularization term to solve the collinearity problem in linear regression. Therefore, LWRR overcomes the nonlinearity and time-varying characteristics of the variables while solving the process collinearity problem. Given an auxiliary variable set X∈R n×m and the corresponding dominant variable set Y∈R n×1 , m and n are the dimensions and number of samples of auxiliary variables respectively, and the input of query sample is x(q)∈R 1×m In this embodiment, the auxiliary variable set is the data set of coal powder quantity, blast air volume, air permeability, blast temperature, feed speed, blast pressure and oxygen enrichment in blast furnace operation, and the dominant variable set is the data set of silicon content in molten iron. The distance d between the query sample and the training sample is obtained by using the Euclidean distance i The LWRR algorithm needs to assign different weights to each training sample based on the sample similarity metric. The weight value is shown in formula (2):

[0069]

[0070] in, is an adjustable parameter in the weight function, is the standard deviation of the distance. The weight of each sample is represented as a weight vector w = [w1, w2, ..., w n ], and record the weight matrix as Ω = diag (w). Query the predicted output value of the sample This can be estimated by following the steps below:

[0071] (1) De-averaging. Calculate the weighted average of the input and output, and perform de-averaging on the input and output of the training samples and the input of the query samples;

[0072] (2) Set the initial estimate of the query sample output to

[0073] (3) Define the cost function using the Tikhonov–Phillips regularization term;

[0074]

[0075] (4) The cost function J is Find the partial derivatives and get the estimated values ​​of the parameters;

[0076]

[0077] (5) Calculate the predicted output value

[0078]

[0079] In the LWRR algorithm, there are three main parameters that need to be optimized: the number of local training samples N, the ridge regression regularization term coefficient λ, and the weight adjustable parameter The regularization term coefficient λ is calculated using the ridge trace method, and the number of local training samples N and the weights are adjustable parameters. Obtained using cross validation method.

[0080] After the model is built, the silicon content in molten iron is predicted through the following steps, including:

[0081] Step 1: Collect the historical actual operation data of the iron-making blast furnace and construct the historical sample set and query sample set;

[0082] This embodiment collects actual operation data of Baotou Steel's No. 6 blast furnace, including: coal powder amount, blast air volume, air permeability, blast temperature, feed speed, blast pressure and oxygen enrichment. A total of 500 samples are collected, the first 400 samples are used as historical sample sets, and the remaining 100 samples are used as query sets.

[0083] Step 2: Obtain the input data x(1), x(2), ..., x(t+1) (coal powder amount, blast air volume, air permeability, blast temperature, feed speed, blast pressure and oxygen enrichment data) of the historical sample set, and output data y(1), y(2), ..., y(t+1) (silicon content in molten iron data), use LightGBM to establish a decision tree model, and obtain a feature vector, where t is the sampling time of the historical sample set;

[0084] Step 3: Sort the feature vectors extracted by the LightGBM decision tree model according to the gain of each feature, and select the first f feature vectors as input;

[0085] Step 4: Calculate the time difference terms Δx(1), Δx(2), ..., Δx(t) and Δy(1), Δy(2), ..., Δy(t) of the input and output data obtained in step 3, and the difference term Δx(q-1) = x(q) - x(q-1) of the query sample input, where q is the sampling time of the query sample set;

[0086] Step 5: Calculate the Euclidean distance between the query sample difference item in step 4 and each historical input sample difference item and the corresponding weight value, and select the N historical difference items most relevant to the query sample as the training set ΔX N and ΔY N , and obtain the weight matrix Ω;

[0087] Step 6: Construct a local weighted ridge regression LWRR model and use the training set ΔX in step 5 N , ΔY N and weight matrix Ω to train the local weighted ridge regression LWRR model;

[0088] Step 7: Obtain real-time data of blast furnace operation, and use the LWRR model trained in step 6 to obtain the differential prediction value of silicon content in molten iron at the current moment and the previous moment

[0089] Step 8: Get the final predicted value of silicon content in molten iron Where y(q-1) is the actual value of the silicon content in the molten iron at the previous moment.

[0090] This embodiment is tested in blast furnace ironmaking data, and its main purpose is to predict the silicon content in molten iron. Due to the high temperature and high pressure environment in the furnace, it is very difficult to measure the silicon content.

[0091] In order to further illustrate the beneficial effects that can be achieved by the present invention, a series of experiments were conducted to compare the prediction method based on LightGBM-LWRR of the present invention with the prediction methods based on TD-RR, TD-LWPLS, and TD-Light-LWRR.

[0092] Figure 4 The prediction effects of the silicon content in molten iron by the LightGBM-LWRR prediction method of the present invention and four methods including TD-RR, TD-LWPLS and TD-Light-LWRR are given. Due to the large variable drift problem in the blast furnace ironmaking process, the fitting accuracy of the LightGBM-LWRR model is the worst. TD-RR only uses the TD model, which cannot solve the problems of time-varying process variables and variable redundancy, and the prediction accuracy is also low. Although TD-LWPLS can solve these problems, its prediction accuracy is lower than that of the TD-LightGBM-LWRR regression model.

[0093] The prediction error curves of the four methods, TD-LWRR, LightGBM-LWRR, TD-LWPLS and TD-LightGBM-LWRR, are shown in Figure 2. Figure 5 As shown in the figure, it can be found that after using the TD method to remove the drift in the data, the model prediction results are significantly improved; after using LightGBM to remove variable redundancy, the prediction performance is also significantly improved.

[0094] Table 1 shows the prediction results of silicon content in molten iron in blast furnace ironmaking process using the experimental method.

[0095] Table 1 Prediction results of silicon content in molten iron using nine methods

[0096]

[0097] It can be seen from Table 1 that the LWRR model under the TD-LightGBM data preprocessing algorithm of the present invention can obtain the best prediction effect of molten iron silicon content in a blast furnace ironmaking environment, and it is also proved that LightGBM-LWRR has better regression accuracy than LWPLS. Therefore, the molten iron silicon content prediction method based on TD-LightGBM of the present invention can effectively improve the prediction accuracy.

[0098] Some steps in the embodiments of the present invention may be implemented using software, and the corresponding software program may be stored in a readable storage medium, such as a CD or a hard disk.

[0099] The above description is only a preferred embodiment of the present invention and is not intended to limit the present invention. Any modifications, equivalent substitutions, improvements, etc. made within the spirit and principle of the present invention should be included in the protection scope of the present invention.

Claims

1. A method for predicting silicon content in molten iron, characterized in that: The method comprises: Step 1: Collect the historical actual operation data of the iron-making blast furnace and construct the historical sample set and query sample set; Step 2: Obtain the input data x(1), x(2), ..., x(t+1) of the historical sample set, and the output data y(1), y(2), ..., y(t+1), use LightGBM to build a decision tree model, and obtain a feature vector, where t is the sampling time of the historical sample set; Step 3: Sort the feature vectors extracted by the LightGBM decision tree model according to the gain of each feature, and select the first f feature vectors; Step 4: Calculate the differential terms Δx(1), Δx(2), …, Δx(t) and Δy(1), Δy(2), …, Δy(t) of the feature vectors obtained in step 3, and the differential term Δx(q-1) = x(q) - x(q-1) of the query sample input, where q is the sampling time of the query sample set; Step 5: Calculate the Euclidean distance between the query sample difference item in step 4 and each historical input sample difference item and the corresponding weight value, and select the N historical difference items most relevant to the query sample as the training set ΔX N and ΔY N , and obtain the weight matrix Ω; Step 6: Construct a local weighted ridge regression LWRR model and use the training set ΔX in step 5 N , ΔY N and weight matrix Ω to train the local weighted ridge regression LWRR model; Step 7: Obtain real-time data of blast furnace operation, and use the LWRR model trained in step 6 to obtain the differential prediction value of silicon content in molten iron at the current moment and the previous moment Step 8: Get the final predicted value of silicon content in molten iron Where y(q-1) is the actual value of silicon content in molten iron at the previous moment; It also includes: the LightGBM decision tree model judges the importance of each feature according to the number of times each feature is used to segment the training data, uses a unilateral gradient sampling algorithm to sort the data according to the gradient, and retains the first a% of the data with larger gradients to form a subset For the residual set formed by the data with smaller (1-a)% gradient Randomly form a size Subset of And calculate the information gain: in, g(t) is the gradient, x(t) represents the sampled value of the auxiliary variable at time t, d is the point with the best gain unchanged found in the split calculation, For normalization The gradient and Is a collection The number of samples, x j (t) is the auxiliary variable of the jth feature at time t; It also includes using a local weighted algorithm to assign different weights to each sample based on the Euclidean distance between the query sample and the historical sample; using the ridge regression algorithm and introducing the Tikhonov–Phillips regularization term to solve the collinearity problem in linear regression; The weight matrix is ​​recorded as Ω=diag(w), w=[w1,w2,…,w n ], the weight assigned to each sample by the local weighted ridge regression LWRR model is: in, is an adjustable parameter in the weight function, is the standard deviation of the distance, d i is the distance between the query sample and the i-th historical sample.

2. The method according to claim 1, characterized in that The predicted output value of the query sample in the local weighted ridge regression LWRR model The estimation process includes: Step 71: De-averaging, calculating the weighted average of sample input and output, and performing de-averaging processing on the input and output of the training sample and the input of the query sample; Step 72: Set the initial estimate of the query sample output to Step 73: Define the cost function using the Tikhonov–Phillips regularization term; Step 74: The cost function J is Find the partial derivatives and get the estimated values ​​of the parameters; Step 75: Calculate the predicted output value in, For a given set of auxiliary variables, is the dominant variable set corresponding to the auxiliary variable set, m, n are the dimensions and number of samples of auxiliary variables, is the input of the query sample, λ is the regularization term coefficient of ridge regression, is the average silicon content of molten iron in historical samples, is the ridge regression parameter, is a set of predicted values ​​for silicon content in molten iron.

3. The method according to claim 2, characterized in that The ridge regression regularization term coefficient λ is calculated using the ridge trace method.

4. The method according to claim 2, characterized in that: Adjustable parameters in the weight function Obtained using cross validation method.

5. The method according to claim 1, characterized in that The parameters that need to be optimized in the local weighted ridge regression LWRR model include the number of local training samples, and the number of local training samples is obtained by using a cross-validation method.

6. The method according to claim 1, characterized in that The actual operation data of the blast furnace include: coal powder amount, blast air volume, air permeability, blast temperature, feed speed, blast pressure and oxygen enrichment.

7. The method according to claim 1, characterized in that The historical actual operation data of the iron-making blast furnace is collected, and a total of 500 samples are collected, among which the first 400 samples are used to construct a historical sample set, and the remaining 100 samples are used to construct a query sample set.

Images