Molten steel quality real-time prediction method based on sample entropy and fractal feature fusion evolution

Abstract

The invention discloses a molten steel quality real-time prediction method based on sample entropy and fractal feature fusion evolution, and relates to the technical field of converter steelmaking molten steel quality prediction. A molten steel quality multi-index real-time prediction model is constructed and trained, fractal features and sample entropy features are innovatively introduced, converter steelmaking process data are deeply analyzed through a mathematical method, and key features are extracted and serve as model input. In the model construction stage, an LSTM network is selected, and the advantage of processing time sequence data of the LSTM network is utilized, so that the two key indexes of the carbon element content and the temperature of the molten steel are predicted at the same time; and the LSTM network is further optimized by adopting a multi-objective algorithm, the molten steel quality real-time prediction model based on sample entropy and fractal feature fusion evolution is obtained, the model can more quickly and accurately complete the real-time prediction task of the molten steel carbon element content and the temperature, and powerful support is provided for accurate control of the converter steelmaking process.

Description

Technical Field

[0001] The invention relates to the technical field of molten steel quality prediction for converter steelmaking, and in particular to a real-time prediction method for molten steel quality based on the fusion evolution of sample entropy and fractal features. Background Art

[0002] Converter steelmaking is a very complex process involving physical and chemical changes at high temperatures. The carbon content and temperature of molten steel have a serious impact on the quality of steel. Converter steelmaking quality control refers to controlling the composition and temperature of molten steel within a reasonable range. In the actual steelmaking process, there are many factors that affect the carbon content and temperature of the final molten steel, and the converter steelmaking process is a very complex multi-element multi-phase high-temperature reaction process, so it is impossible to detect the carbon content and temperature of molten steel in real time during the converter steelmaking process, which leads to the problem of low endpoint hit rate. Most small and medium-sized converter steel plants still rely on manual experience to maintain endpoint control, that is, the temperature of the flame is inferred by the brightness and shape of the converter furnace mouth flame. This method has low control accuracy and is unstable. The traditional single-output static control mode mainly calculates the endpoint carbon content and temperature of converter steelmaking by manually inputting parameters, and it cannot predict the carbon content and temperature of molten steel in real time. Although the dynamic prediction model based on machine learning performs well in the converter steelmaking process, it has not achieved simultaneous or real-time prediction of the carbon content and temperature of molten steel. Therefore, there is an urgent need for a real-time prediction technology for multiple indicators of molten steel quality in converter steelmaking that is suitable for the converter steelmaking process and has low cost and high hit rate. Summary of the invention

[0003] In view of the shortcomings of the prior art, the present invention provides a real-time prediction method for molten steel quality based on the fusion evolution of sample entropy and fractal features, aiming to achieve real-time prediction of multiple indicators of molten steel quality in converter steelmaking.

[0004] In order to solve the above technical problems, the technical solution adopted by the present invention is:

[0005] A real-time prediction method for molten steel quality based on sample entropy and fractal feature fusion evolution, the method comprising the following steps:

[0006] Step 1: Collect the actual production history data of the steel plant and construct the production data set of the converter steelmaking process;

[0007] Step 2: Perform sample entropy calculation and fractal feature extraction on the converter steelmaking process production data set to obtain the time characteristics and spatial characteristics of the converter steelmaking process production data set, and add the time characteristics and spatial characteristics to the converter steelmaking process production data set to obtain the data set Input;

[0008] Step 3: Use the data set Input to train the multi-input and output LSTM neural network. During the training, the carbon content and temperature of the molten steel in the Input are used as outputs, and other data in the Input are used as input information. The sliding window LSTM network multi-output method is used to establish a real-time prediction model for multiple indicators of converter steelmaking quality;

[0009] Step 4: Optimize the parameters of the real-time prediction model of multiple indicators of converter steelmaking quality through a multi-objective differential evolution algorithm, and establish a real-time prediction model of molten steel quality based on the fusion evolution of sample entropy and fractal features;

[0010] Step 5: Use the real-time prediction model of molten steel quality based on the fusion evolution of sample entropy and fractal features to predict the carbon content and temperature of molten steel in the actual production of converter steelmaking in real time;

[0011] Step 6: According to the actual production situation, the real-time prediction model of molten steel quality based on the fusion evolution of sample entropy and fractal features is updated and upgraded to maintain the prediction accuracy of the model.

[0012] The actual production history data of the steel plant includes:

[0013] Data of main raw materials of different batches before entering the furnace, including weight of scrap steel entering the furnace, weight, temperature and content of carbon, silicon, manganese, sulfur and phosphorus in molten iron entering the furnace;

[0014] The data generated by different heats during the entire blowing process include oxygen lance height, O 2 Flow, flue gas flow, CO flow, CO 2 Flow rate, N 2 - Ty flow, Ar-Ty flow, lump lime weight, light-burned dolomite weight, magnesia ball weight, sishen lump ore weight, lump dolomite weight, nickel plate weight, composite silicon carbide ball weight information;

[0015] The carbon content and temperature data of molten steel measured at the time of steel tapping for different furnaces.

[0016] The multi-index real-time prediction model for converter steelmaking quality is as follows:

[0017] {Y C ,Y T}=LSTM{W all ,O 2 ,T,C,Si,Mn,P,S} (3)

[0018] Among them, Y C , Y T They represent the carbon content and temperature of molten steel predicted by the real-time prediction model of multiple indicators for converter steelmaking quality; W allRepresents scrap weight, molten iron weight, oxygen lance height, flue gas flow, CO flow, CO 2 Flow rate, N 2 - Ty flow, Ar-Ty flow, lump lime weight, light-burned dolomite weight, magnesia ball weight, sishen lump ore weight, lump dolomite weight, nickel plate weight, composite silicon carbide ball weight information; 2 It represents the cumulative oxygen consumption in the steelmaking process; T represents the temperature of molten iron; C represents the carbon content in molten iron; Si represents the silicon content in molten iron; Mn represents the manganese content in molten iron; P represents the phosphorus content in molten iron; S represents the sulfur content in molten iron.

[0019] The step 2 comprises the following steps:

[0020] Step 2.1: Calculate the sample entropy of the converter steelmaking process production data set to obtain the sample entropy feature, and add the obtained sample entropy feature to the converter steelmaking process production data set to obtain a new data set Z u,p ;

[0021] Step 2.2: Apply fractal theory to Z u,p Analyze and obtain fractal features, and add the obtained fractal features to Z u,p Get the data set Input.

[0022] When calculating the sample entropy of the converter steelmaking process production data set as described in step 2.1, the column containing each type of data in the converter steelmaking process production data set is regarded as a time series of {x(n)}=x(1), x(2),…, x(N), where N represents the number of data points in each furnace; the initial value of the embedding dimension m is set to 2, and the tolerance r is selected as 0.15 times the standard deviation of the time series data {x(n)}.

[0023] The fractal theory is used to calculate Z u,p The method for analyzing and obtaining fractal characteristics is:

[0024] The data set Z u,p Each row in is considered as a time series. Each time series is divided three times with interval length L = 2, 4, and 8, and d intervals are obtained each time. Each interval is considered as a subsample. For each subsample, the average value E is calculated. p and standard deviation S p ;X jj,p The result of data normalization is obtained by subtracting the mean value of each subsample, where jj represents the data point of the current interval data, and the cumulative deviation sequence is calculated for each subsample. For each interval, calculate R P =Max(Yu,p )-Min(Y u,p ); calculate the average of d subsamples (R p / S p ):

[0025]

[0026] Take the logarithm, do a linear regression on log(R / S) and log(n), and calculate the regression equation:

[0027] log(R / S) n =log(a)+H log(n) (2)

[0028] Usually a is a constant, and the slope H in linear regression is the Hurst exponent, which is a fractal characteristic.

[0029] The step 4 described in which the parameters of the multi-index real-time prediction model of converter steelmaking quality are optimized by a multi-objective differential evolution algorithm, specifically, the number of network nodes and the number of layers of the multi-index real-time prediction model of converter steelmaking quality are selected as decision variables of the multi-objective differential evolution algorithm; and the root mean square error of the two outputs of the multi-index real-time prediction model of converter steelmaking quality is used as the two objective functions of the multi-objective differential evolution algorithm.

[0030] The step 6 comprises the following steps:

[0031] Step 6.1: Determine whether the carbon content and temperature of the molten steel obtained by sampling the auxiliary gun at the end of blowing in actual production meet the conditions for steel tapping. If yes, the molten steel is tapped and step 6.3 is executed; otherwise, step 6.2 is executed;

[0032] Step 6.2: Determine whether the prediction error of the carbon content and temperature of the molten steel output by the real-time prediction model of molten steel quality based on the fusion evolution of sample entropy and fractal features is within the preset deviation range; if so, continue to use the current real-time prediction model of molten steel quality based on the fusion evolution of sample entropy and fractal features in the next real-time prediction, and execute step 6.3; otherwise, return to step 4, adjust the decision variables and other parameters of the multi-objective optimization algorithm, and further optimize the multi-index real-time prediction model of converter steelmaking quality;

[0033] Step 6.3: Regularly update the converter steelmaking process production data set incrementally, and repeat steps 2 to 5 after the update.

[0034] Compared with the prior art, the present invention has the following beneficial effects:

[0035] (1) The present invention uses sample entropy theory to analyze the time series characteristics in the actual steelmaking production history data, and can effectively extract time characteristics. As an efficient time series analysis tool, sample entropy is particularly suitable for evaluating the complexity and uncertainty of data. By applying sample entropy to analyze the steelmaking production history data, it is possible to judge the dynamic characteristics and complexity of the system, and further explore the data characteristics at different time scales, thereby deepening the understanding of the steelmaking process.

[0036] (2) The present invention characterizes the fractal structural characteristics of the actual steelmaking production history data by applying fractal theory and calculating the fractal dimension, thereby revealing the spatial characteristics of the actual steelmaking production history data. These characteristics can effectively capture the changing laws and patterns of steelmaking data. The Hurst index, as a fractal feature, can detect the self-similarity of data, that is, the similar patterns exhibited by data on different time scales. Using the Hurst index, the changes in the steelmaking production history data can be clearly reflected and the nonlinear structure hidden therein can be revealed. These analysis results are of great significance for establishing a real-time prediction model for molten steel quality based on the fusion evolution of sample entropy and fractal features, and help to improve the prediction speed and accuracy of the model.

[0037] (3) The generalization ability of the multi-index real-time prediction model of converter steelmaking quality was optimized using a multi-objective evolutionary algorithm, and the optimal parameters and structure of the real-time prediction model of molten steel quality based on the fusion evolution of sample entropy and fractal features were found, which further improved the real-time prediction accuracy of molten steel quality.

[0038] (4) The method of the present invention can simultaneously predict the carbon content and temperature of molten steel, and the hit rate reaches the expected effect, which can assist the operation of the converter steelmaking process; it can quickly predict the carbon content and temperature of molten steel in real time based on the collected data, which is suitable for the actual background of converter steelmaking and can meet the needs of actual production. BRIEF DESCRIPTION OF THE DRAWINGS

[0039] Figure 1 This is a flow chart of the real-time prediction method of molten steel quality based on the fusion evolution of sample entropy and fractal features in this embodiment;

[0040] Figure 2 This is a flow chart of optimizing the real-time prediction model of multiple indicators of converter steelmaking quality using a multi-objective differential evolution algorithm in this embodiment;

[0041] Figure 3 Schematic diagram of a method for selecting an optimal solution for multiple sets of supportable solution points for a multi-objective optimization algorithm;

[0042] Figure 43 is a prediction error curve diagram of the present embodiment, wherein (a) is a real-time prediction error curve diagram of the carbon content of molten steel, and (b) is a real-time prediction error curve diagram of the temperature of molten steel. DETAILED DESCRIPTION

[0043] The specific implementation of the present invention is further described in detail below in conjunction with the accompanying drawings and examples. The following examples are used to illustrate the present invention, but are not intended to limit the scope of the present invention.

[0044] The core idea of ​​the present invention is: in order to realize the real-time prediction of the carbon content and temperature of molten steel in the converter steelmaking process, the present invention constructs and trains a real-time prediction model of molten steel quality with multiple indicators. The model innovatively introduces fractal features and sample entropy features, deeply analyzes the converter steelmaking process data through mathematical methods, extracts key features and uses them as model input. In the model construction stage, the LSTM (Long Short-Term Memory) network is selected to take advantage of its processing time series data to realize the simultaneous prediction of two key indicators, the carbon content and temperature of molten steel. In order to optimize the model structure to make it more suitable for specific prediction problems, the present invention further adopts a multi-objective algorithm to optimize the LSTM network to obtain a real-time prediction model of molten steel quality based on the fusion evolution of sample entropy and fractal features. This optimization process is designed to enable the network to learn more complex nonlinear relationships and better fit the diversified feature inputs including fractals and sample entropy. Through this series of innovations and optimizations, the real-time prediction model of molten steel quality based on the fusion evolution of sample entropy and fractal features can complete the real-time prediction task of molten steel carbon content and temperature more quickly and accurately, and provides strong support for the precise control of the converter steelmaking process.

[0045] like Figure 1 As shown, the real-time prediction method of molten steel quality based on the fusion evolution of sample entropy and fractal features in this embodiment includes the following steps:

[0046] Step 1: Collect the actual production history data of the steel plant and perform data classification preprocessing, and organize the preprocessed data into a production data set of the converter steelmaking process;

[0047] In this embodiment, 40 different heats are selected from the collected actual production history data of the steel plant as reference heats, and the abnormal data are deleted and then the data classification preprocessing is performed. The actual production history data of the steel plant collected includes:

[0048] Collect data on the main raw materials of different batches before entering the furnace, including the weight of scrap steel entering the furnace, the weight, temperature and content of carbon, silicon, manganese, sulfur and phosphorus in molten iron entering the furnace;

[0049] Collect data generated during the entire blowing process for different heats, including oxygen lance height, O 2 Flow, flue gas flow, CO flow, CO2 Flow rate, N 2- Ty flow, Ar-Ty flow, lump lime (MQG-N) weight, light burned dolomite (MQD-A) weight, magnesia ball (MMD-P) weight, sishen lump ore weight, lump dolomite weight, nickel plate weight, composite silicon carbide ball weight information;

[0050] The carbon content and temperature data of molten steel measured at the time of steel tapping for different furnaces are obtained, and the carbon content and temperature of molten steel at each sampling point in the converter steelmaking process are generated according to the mechanism formula.

[0051] Finally, the above actual production history data are used to construct the converter steelmaking process production data set.

[0052] Step 2: Perform sample entropy calculation and fractal feature extraction on the converter steelmaking process production data set to obtain the time characteristics and spatial characteristics of the converter steelmaking process production data set, and add the time characteristics and spatial characteristics to the converter steelmaking process production data set to obtain the data set Input;

[0053] Step 2.1: Calculate the sample entropy of the converter steelmaking process production data set to obtain the sample entropy feature, and add the obtained sample entropy feature to the converter steelmaking process production data set to obtain a new data set Z u,p ;

[0054] In this embodiment, the weight of the scrap steel entering the furnace, the weight of the molten iron entering the furnace, the temperature, the content of carbon, silicon, manganese, sulfur, phosphorus, the height of the oxygen lance, O 2 Flow, flue gas flow, CO flow, CO 2 Flow rate, N 2 The sample entropy feature is obtained by calculating the sample entropy feature of the converter steelmaking process production data set consisting of -Ty flow, Ar-Ty flow, mass lime (MQG-N) weight, light burned dolomite (MQD-A) weight, magnesia ball (MMD-P) weight, sishen block ore weight, mass dolomite weight, nickel plate weight, and composite silicon carbide ball weight, and the obtained sample entropy feature is added to the converter steelmaking process production data set to obtain a new data set Z u,p .

[0055] The calculation of sample entropy involves measuring the complexity of time series data, and the complexity and regularity of data are evaluated by calculating the similarity of data points in different dimensions. This embodiment adopts the calculation method of sample entropy to analyze the complexity of converter steelmaking process data. Specifically, a column of each type of data in the converter steelmaking process production data set is regarded as a time series of {x(n)}=x(1), x(2),…, x(N), where N represents the number of data points in each furnace. Two important parameters in the calculation of sample entropy are embedding dimension and tolerance. The selection of embedding dimension and tolerance directly affects the calculation result of sample entropy. In this embodiment, the embedding dimension m=2 is selected, and the tolerance r is selected as 0.15 times the standard deviation of the time series data {x(n)}.

[0056] Convert the time series data {x(n)} into a vector sequence {X}_m(i)={x(i),x(i+1),…,x(i+m-1)} with dimension m, where 1≤i≤N-m+1, and calculate the distance d[{X}_m(i),{X}_m(j)] between the vector sequences {X}_m(i) and {X}_m(j) at different times. Calculate the absolute value of the maximum difference between the corresponding elements of the two, where i and j represent two different data points in the current sequence data, and k represents the position index in the current sequence, ranging from 0 to m-1. For each vector sequence {X}_m(i), the vector sequence with a distance less than or equal to r is a similar sequence, and the number of similar sequences is recorded as B i (excluding {X}_m(i) itself). The number of sequences similar to sequence {X}_m(i) is B i The ratio of the total number of sequences N-m+1 is the vector matching ratio. Calculate the vector matching ratio of the i-th sequence: Calculate the average B of all vector matching ratios m (r):

[0057] Increase the dimension and calculate the new vector matching ratio: increase the embedding dimension of the vector to m = 3, repeat the construction of vector sequences and the calculation of the distance between vector sequences; then count the number of similar sequences and record them as A i , calculate the vector matching ratio of the i-th sequence: Calculate the average A of all new vector matching ratios m (r):

[0058] Finally, the sample entropy is obtained: The calculated sample entropy is added as the sample entropy feature to the production data set of the converter steelmaking process to obtain a new data set Z u,p, u=1,...,U,p=1,...,P. Each column is a feature. U represents the total number of sampling points in each furnace, u represents the uth sampling point; P represents the total number of features, and p represents the pth feature.

[0059] Step 2.2: Apply fractal theory to Z u,p Analyze and obtain fractal features, and add the obtained fractal features to Z u,p Get the data set Input;

[0060] In this embodiment, the data set Z obtained in step 2.1 is u,p Each row is considered as a time series. Each time series is divided three times with interval length L = 2, 4, and 8, and d intervals are obtained each time. Each interval is considered as a subsample. For each subsample, the average value E is calculated. p and standard deviation S p .X jj,p The result of data normalization is obtained by subtracting the mean value of each subsample, where jj represents the data point of the current interval data, and the cumulative deviation sequence is calculated for each subsample. For each interval, calculate R P =Max(Y u,p )-Min(Y u,p ). Calculate the average (R p / S p ):

[0061]

[0062] Take the logarithm, do a linear regression on log(R / S) and log(n), and calculate the regression equation:

[0063] log(R / S) n =log(a)+H log(n) (2)

[0064] Usually a is a constant, and the slope H in linear regression is the Hurst exponent, which is a fractal feature. By calculating the fractal dimension, we can analyze the fractal structure characteristics of steel production data. These characteristics can capture the changes in data and hidden nonlinear structures, so that the model can make predictions faster and more accurately. Adding fractal features to Z u,p The final data set Input is composed of the data set, which is divided into training data set and test data set in a ratio of 7:3.

[0065] Step 3: Use the sliding window long short-term memory network LSTM to establish a real-time prediction model for multiple indicators of converter steelmaking quality;

[0066] In this embodiment, a training data set is used to train a multi-input and output LSTM neural network. During training, the carbon content and temperature of the molten steel in the set are used as outputs, and other data in the set are used as input information. A sliding window LSTM network multi-output method is used to establish a real-time prediction model for the molten steel temperature and carbon content in the converter steelmaking process, such as the converter steelmaking quality multi-indicator real-time prediction model shown in formula (3), where the preset sliding window size is 5.

[0067] {Y C ,Y T}=LSTM{W all ,O 2 ,T,C,Si,Mn,P,S} (3)

[0068] Among them, Y C , Y T The carbon content and temperature of molten steel predicted by the real-time prediction model of multiple indicators for converter steelmaking quality; all Represents scrap weight, molten iron weight, oxygen lance height, flue gas flow, CO flow, CO 2 Flow rate, N 2 -Ty flow, Ar-Ty flow, lump lime (MQG-N) weight, light burned dolomite (MQD-A) weight, magnesia ball (MMD-P) weight, sishen lump ore weight, lump dolomite weight, nickel plate weight, composite silicon carbide ball weight information; 2 is the cumulative oxygen consumption in the steelmaking process, T is the molten iron temperature, C is the carbon content in the molten iron, Si is the silicon content in the molten iron, Mn is the manganese content in the molten iron, P is the phosphorus content in the molten iron, and S is the sulfur content in the molten iron.

[0069] Step 4: Optimize the parameters of the real-time prediction model of multiple indicators of converter steelmaking quality through a multi-objective differential evolution algorithm to obtain the current real-time prediction model of molten steel quality based on the fusion evolution of sample entropy and fractal features;

[0070] The number of network nodes and the number of layers of the real-time prediction model of the multi-index of converter steelmaking quality are selected as the decision variables of the multi-objective differential evolution algorithm; the root mean square error of the two outputs of the real-time prediction model of the multi-index of converter steelmaking quality is used as the two objective functions of the multi-objective differential evolution algorithm; the process of optimizing the parameters of the real-time prediction model of the multi-index of converter steelmaking quality by the multi-objective differential evolution algorithm is as follows: Figure 2 As shown below:

[0071] Initialize the various parameter values ​​of the converter steelmaking quality multi-index real-time prediction model to form an initial parameter combination; substitute the initial parameter combination into the converter steelmaking quality multi-index real-time prediction model, and then use the training data set to train the converter steelmaking quality multi-index real-time prediction model to obtain the prediction results of the carbon content and temperature of the molten steel, and calculate the root mean square error of the prediction results, so as to obtain the objective function value corresponding to the decision variable;

[0072] In this embodiment, the decision variables of the multi-objective differential evolution algorithm select the number of network nodes and layers of the real-time prediction model of converter steelmaking quality multi-index. Number of network nodes: the number of neurons in each layer, ranging from [10, 1000]. Number of layers: the number of layers of the neural network, ranging from [1, 4]. The objective function is the root mean square error (RMSE) of the two outputs of the prediction model, and these two objective functions need to be minimized. G = 100 represents the evolutionary generation,

[0073] Start to initialize and generate an initial population containing NN individuals. Each individual is a two-dimensional vector, representing a candidate solution. Each candidate solution corresponds to a set of network nodes and network layers of the LSTM network. The obtained candidate solution set is as follows:

[0074] P={(o 1 ,l 1 ),(o 2 ,l 2 ),…,(o NN ,l NN )} (4)

[0075] Among them, i is the number of network nodes, l i is the number of network layers, where i represents the i-th data point in the current sequence data.

[0076] Evaluate the population: For each individual in the population, build a prediction model using its decision variables and calculate its two objective function values.

[0077] Differential mutation: for each individual s in the population ri , select three random individuals s r1 ,s r2 ,s r3 , perform differential mutation and obtain the mutated individual v i .

[0078] v i =s r1 +F·(s r2 -s r3 ) (5)

[0079] Among them, F=0.7 is the scaling factor.

[0080] Crossover selection: for the mutated individual v i Perform crossover operation to generate test individuals c i .

[0081] For each individual s ri and the experimental individual c i Perform a selection operation, using crowding distance and dominance relationship to decide which individuals to keep to the next generation and update the population:

[0082] If c i Dominate ri , then use c i Replace s ri ; if s ri Domination i , then keep s ri ; If the two do not dominate each other, individuals are selected based on the crowding distance.

[0083] Termination condition judgment: The algorithm is terminated by presetting a maximum number of evolutionary iterations G. If the algorithm does not meet the termination condition, G=G+1, and the differential operation is re-performed to perform mutation and crossover selection operations to generate a descendant population; otherwise, the Pareto optimal solution set of the parameter optimization problem of the multi-indicator real-time prediction model for converter steelmaking quality is output and the operation is terminated.

[0084] This embodiment adopts the inverse point strategy to determine the optimal parameters from the Pareto optimal solution set. It is difficult to find which one is the most suitable solution in the Pareto optimal solution set. Therefore, this embodiment uses the inverse point strategy to select an ideal solution, namely the number of network nodes and the number of network layers. This strategy connects the extreme points at both ends of the non-dominated solution and selects the Pareto solution with the largest distance from the extreme line, which represents the optimal solution. It represents a good trade-off relationship between multiple conflicting objectives. The test data set is then used for verification, and the structure of the real-time prediction model of molten steel quality based on the fusion evolution of sample entropy and fractal features is finally determined.

[0085] Step 5: Use the current real-time prediction model of molten steel quality based on the fusion evolution of sample entropy and fractal features to make real-time predictions on the carbon content and temperature of molten steel in the actual production of converter steelmaking;

[0086] Step 6: According to the actual production situation, the real-time prediction model of molten steel quality based on the fusion evolution of sample entropy and fractal features is updated to maintain the prediction accuracy of the model;

[0087] Step 6.1: Determine whether the carbon content and temperature of the molten steel obtained by sampling the auxiliary gun at the end of blowing in actual production meet the conditions for steel tapping. If yes, the molten steel is tapped and step 6.3 is executed; otherwise, step 6.2 is executed;

[0088] In this embodiment, the composition of molten steel is tested. The carbon content and temperature of molten steel outputted by the real-time prediction model of molten steel quality based on the fusion evolution of sample entropy and fractal features according to actual production data are compared and analyzed with the carbon content and temperature of molten steel collected by the steel plant's auxiliary gun. If the accuracy and hit rate of the model reach the required values, the real-time prediction model of molten steel quality based on the fusion evolution of sample entropy and fractal features is established, and step 6.3 is executed; otherwise, step 6.2 is executed;

[0089] Step 6.2: Determine whether the prediction error of the carbon content and temperature of the molten steel output by the real-time prediction model of molten steel quality based on the fusion evolution of sample entropy and fractal features is within the preset deviation range; if so, continue to use the current real-time prediction model of molten steel quality based on the fusion evolution of sample entropy and fractal features in the next real-time prediction, and execute step 6.3; otherwise, return to step 4, adjust the decision variables and other parameters of the multi-objective optimization algorithm, further optimize the real-time prediction model of the multi-index of converter steelmaking quality, and then use it for the real-time prediction of molten steel quality based on the fusion evolution of sample entropy and fractal features;

[0090] Step 6.3: Regularly update the production data set of the converter steelmaking process, and repeat steps 2 to 5 after the update;

[0091] The converter steelmaking process production data set established in step 1 is incrementally updated regularly, and data or abnormal data that are irrelevant to the input of the real-time prediction model of molten steel quality based on the fusion evolution of sample entropy and fractal features are deleted to enhance the adaptability of the model.

[0092] Based on the actual production data of a steel plant, this embodiment repeatedly tests the real-time prediction model of molten steel quality based on the fusion evolution of sample entropy and fractal features, wherein Table 1 is a comparison between the prediction results of the real-time prediction model of molten steel quality based on the fusion evolution of sample entropy and fractal features calculated under different furnaces of converter steelmaking process production data sets and the actual detection values; Table 2 is a comparison between the prediction results of the real-time prediction model of molten steel quality based on the fusion evolution of sample entropy and fractal features calculated under different furnaces of converter steelmaking process production data sets and the actual detection values. RMSE is the root mean square error, which is the square root of the ratio of the square of the error between the predicted value and the actual value to the number of test samples n, and can accurately reflect the error size of the model. Its expression is where y i is the ith theoretical value, is the i-th predicted value. MAE stands for mean absolute error, which is the average of the absolute values ​​of the errors between the predicted value and the actual value. It is also an indicator of the magnitude of the model error. Its expression is: The goodness of fit is expressed as in is the average value of the theoretical value. In Table 2, the experimental data of different heats include: Heat 1 (2763), Heat 2 (1510), Heat 3 (380), Heat 4 (490) and Heat 5 (861), and the data volume in brackets is contained in the heat. Figure 3 Schematic diagram of a method for selecting optimal solutions for multiple sets of supportable solution points for a multi-objective optimization algorithm. Figure 4 The prediction curves of the experiment are shown, where (a) is a real-time prediction curve of the carbon content of molten steel, and (b) is a real-time prediction curve of the temperature of molten steel. It can be seen from the prediction curve that the real-time prediction model of molten steel quality based on the fusion evolution of sample entropy and fractal features can simultaneously complete the real-time prediction of the two indicators of molten steel carbon content and temperature, and the prediction accuracy is excellent. Experiments have proved that the method of the present invention can help steel companies improve the accuracy of molten steel carbon content and temperature prediction in the converter steelmaking process, improve production quality, and reduce energy consumption.

[0093] Table 1 Comparison of current prediction values ​​and theoretical values ​​of the real-time prediction model for molten steel quality based on the fusion evolution of sample entropy and fractal features

[0094]

[0095]

[0096] Table 2 Comparison results of the error index between the predicted value and the actual detection of the real-time prediction model of molten steel quality based on the fusion evolution of sample entropy and fractal features

[0097]

[0098] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention, rather than to limit it. Although the present invention has been described in detail with reference to the aforementioned embodiments, those skilled in the art should understand that they can still modify the technical solutions described in the aforementioned embodiments, or make equivalent replacements for some or all of the technical features therein. However, these modifications or replacements do not cause the essence of the corresponding technical solutions to deviate from the scope defined by the claims of the present invention.

Claims

1. A real-time prediction method for molten steel quality based on the fusion evolution of sample entropy and fractal features, characterized in that: The method comprises the following steps: Step 1: Collect the actual production history data of the steel plant and construct the production data set of the converter steelmaking process; Step 2: Perform sample entropy calculation and fractal feature extraction on the converter steelmaking process production data set to obtain the time characteristics and spatial characteristics of the converter steelmaking process production data set, and add the time characteristics and spatial characteristics to the converter steelmaking process production data set to obtain the data set Input; Step 3: Use the data set Input to train the multi-input and output LSTM neural network. During the training, the carbon content and temperature of the molten steel in the Input are used as outputs, and other data in the Input are used as input information. The sliding window LSTM network multi-output method is used to establish a real-time prediction model for multiple indicators of converter steelmaking quality; Step 4: Optimize the parameters of the real-time prediction model of multiple indicators of converter steelmaking quality through a multi-objective differential evolution algorithm, and establish a real-time prediction model of molten steel quality based on the fusion evolution of sample entropy and fractal features; Step 5: Use the real-time prediction model of molten steel quality based on the fusion evolution of sample entropy and fractal features to predict the carbon content and temperature of molten steel in the actual production of converter steelmaking in real time; Step 6: According to the actual production situation, the real-time prediction model of molten steel quality based on the fusion evolution of sample entropy and fractal features is updated and upgraded to maintain the prediction accuracy of the model.

2. The method according to claim 1, characterized in that The actual production history data of the steel plant includes: Data of main raw materials of different batches before entering the furnace, including weight of scrap steel entering the furnace, weight, temperature and content of carbon, silicon, manganese, sulfur and phosphorus in molten iron entering the furnace; Data generated by different furnaces during the entire blowing process, including oxygen lance height, O2 flow, flue gas flow, CO flow, CO2 flow, N2-Ty flow, Ar-Ty flow, lump lime weight, light-burned dolomite weight, magnesia ball weight, sishen lump ore weight, lump dolomite weight, nickel plate weight, composite silicon carbide ball weight information; The carbon content and temperature data of molten steel measured at the time of steel tapping for different furnaces.

3. The method according to claim 1, characterized in that The multi-index real-time prediction model for converter steelmaking quality is as follows: {Y C ,Y T} = LSTM{W all , O2, T, C, Si, Mn, P, S} (3) Among them, Y C , Y T They represent the carbon content and temperature of molten steel predicted by the real-time prediction model of multiple indicators for converter steelmaking quality; W all It represents the weight of scrap steel, the weight of molten iron entering the furnace, the height of the oxygen lance, the flue gas flow rate, the CO flow rate, the CO2 flow rate, the N2-Ty flow rate, the Ar-Ty flow rate, the weight of lump lime, the weight of light-burned dolomite, the weight of magnesia balls, the weight of sishen ore, the weight of lump dolomite, the weight of nickel plates, and the weight of composite silicon carbide balls; O2 represents the cumulative oxygen consumption in the steelmaking process; T represents the molten iron temperature; C represents the carbon content in the molten iron; Si represents the silicon content in the molten iron; Mn represents the manganese content in the molten iron; P represents the phosphorus content in the molten iron; S represents the sulfur content in the molten iron.

4. The method according to claim 1, characterized in that The step 2 comprises the following steps: Step 2.1: Calculate the sample entropy of the converter steelmaking process production data set to obtain the sample entropy feature, and add the obtained sample entropy feature to the converter steelmaking process production data set to obtain a new data set Z u,p ; Step 2.2: Apply fractal theory to Z u,p Analyze and obtain fractal features, and add the obtained fractal features to Z u,p Get the data set Input.

5. The method according to claim 4, characterized in that When calculating the sample entropy of the converter steelmaking process production data set as described in step 2.1, the column containing each type of data in the converter steelmaking process production data set is regarded as a time series of {x(n)}=x(1), x(2),…, x(N), where N represents the number of data points in each furnace; the initial value of the embedding dimension m is set to 2, and the tolerance r is selected as 0.15 times the standard deviation of the time series data {x(n)}.

6. The method according to claim 4, characterized in that The fractal theory is used to calculate Z u,p The method for analyzing and obtaining fractal characteristics is: The data set Z u,p Each row in is considered as a time series. Each time series is divided three times with interval length L = 2, 4, and 8, and d intervals are obtained each time. Each interval is considered as a subsample. For each subsample, the average value E is calculated. p and standard deviation S p ;X jj,p The result of data normalization is obtained by subtracting the mean value of each subsample, where jj represents the data point of the current interval data, and the cumulative deviation sequence is calculated for each subsample. For each interval, calculate R P =Max(Y u,p )-Min(Y u,p ); Calculate the average (R p / S p ): Take the logarithm, do a linear regression on log(R / S) and log(n), and calculate the regression equation: log(R / S) n =log(a)+H log(n) (2) Usually a is a constant, and the slope H in linear regression is the Hurst exponent, which is a fractal characteristic.

7. The method according to claim 1, characterized in that The step 4 described in which the parameters of the multi-index real-time prediction model of converter steelmaking quality are optimized by a multi-objective differential evolution algorithm, specifically, the number of network nodes and the number of layers of the multi-index real-time prediction model of converter steelmaking quality are selected as decision variables of the multi-objective differential evolution algorithm; and the root mean square error of the two outputs of the multi-index real-time prediction model of converter steelmaking quality is used as the two objective functions of the multi-objective differential evolution algorithm.

8. The method according to claim 1, characterized in that The step 6 comprises the following steps: Step 6.1: Determine whether the carbon content and temperature of the molten steel obtained by sampling the auxiliary gun at the end of blowing in actual production meet the conditions for steel tapping. If yes, the molten steel is tapped and step 6.3 is executed; Otherwise, proceed to step 6.2; Step 6.2: Determine whether the prediction error of the carbon content and temperature of the molten steel output by the real-time prediction model of molten steel quality based on the fusion evolution of sample entropy and fractal features is within the preset deviation range; if so, continue to use the current real-time prediction model of molten steel quality based on the fusion evolution of sample entropy and fractal features in the next real-time prediction, and execute step 6.3; otherwise, return to step 4, adjust the decision variables and other parameters of the multi-objective optimization algorithm, and further optimize the multi-index real-time prediction model of converter steelmaking quality; Step 6.3: Regularly update the converter steelmaking process production data set incrementally, and repeat steps 2 to 5 after the update.

Images