Blast furnace gas utilization rate time sequence prediction method

Abstract

The invention discloses a blast furnace gas utilization rate time sequence prediction method. The method comprises the following steps: S1, preprocessing gas utilization rate time sequence data; s2, a PSO algorithm is improved, and an LDIW-PSO algorithm is formed; s3, designing a KELM algorithm based on an ELM algorithm; s4, optimizing hyper-parameters of the KELM algorithm by adopting an LDIW-PSO algorithm, and determining an LDIW-PSO-KELM model and model parameters; and S5, realizing the prediction of the blast furnace gas utilization rate by adopting an LDIW-PSO-KELM model. The method can accurately predict the blast furnace gas utilization rate, and has important application values in the aspects of stabilizing blast furnace production, improving gas flow distribution, improving molten iron quality and yield, prolonging the service life of the blast furnace, saving energy, reducing emission and the like.

Description

Technical Field

[0001] This invention relates to intelligent technology for blast furnace smelting processes in the iron and steel industry, and more specifically, to a time series prediction method for blast furnace gas utilization rate. Background Technology

[0002] The blast furnace is a major vessel in modern metallurgical industry, a huge, vertical, and complex reactor for producing pig iron. Its process involves a continuous production flow where iron ore and other iron-containing compounds react with coke under high temperature and pressure to produce molten iron and blast furnace gas through a redox reaction. Blast furnace gas is generated by the redox reaction inside the blast furnace and discharged from the top. Blast furnace gas utilization rate refers to the ratio of carbon dioxide content to the total content of carbon monoxide and carbon dioxide. It reflects the degree of gas utilization within the blast furnace, gas flow distribution, and the overall efficiency of the ironmaking process. Gas utilization rate is a crucial indicator of the overall state of the blast furnace. Analyzing changes in blast furnace gas utilization rate during the blast furnace reaction process, and adjusting blast furnace operation accordingly, is of great significance for reducing energy consumption and increasing blast furnace output.

[0003] Domestic and international scholars have conducted extensive research and achieved certain results regarding the control and prediction of blast furnace gas utilization rate, as well as the establishment and optimization of prediction models. For example, patent application number 202310804301.9 provides a multi-timescale adjustment method and system for blast furnace gas utilization rate, including the following steps: S1: Acquire actual blast furnace data, preprocess the data to obtain blast furnace operating parameters and blast furnace gas utilization rate; S2: Acquire the time series A of blast furnace operating parameters and the time series B of blast furnace gas utilization rate, and calculate the dynamic time-warped distance matrix between time series A and time series B; S3: Obtain the optimal distance path in the dynamic time-warped distance matrix, and improve the prediction and control of blast furnace indicators by solving for the optimal distance path. This patented technology can effectively analyze the relationship between blast furnace operating parameters and blast furnace gas utilization rate through the dynamic time-warped distance matrix, and effectively guide the prediction and control of blast furnace indicators through the optimal distance path in the dynamic time-warped distance matrix.

[0004] Patent application number 202110007323.3 discloses a method, system, and computer equipment for predicting fluctuations in blast furnace gas utilization rate. The method includes: acquiring blast furnace operating parameters and raw material data; selecting gas utilization rate performance indicators based on the impact of the acquired blast furnace operating parameters and raw material data on gas utilization rate; processing and analyzing the selected gas utilization rate performance indicators using principal component analysis (PCA) to obtain a comprehensive characteristic index value for blast furnace gas utilization rate; and comprehensively evaluating the impact of sinter metallurgical properties on blast furnace gas utilization rate based on the comprehensive characteristic index value, predicting the fluctuation of blast furnace gas utilization rate during sinter stack replacement. This patented technology, through PCA, obtains a comprehensive characteristic index of four out of 13 performance indicators affecting blast furnace gas utilization rate, optimizes parameters, and improves blast furnace gas utilization rate.

[0005] Patent application number 201210208357.X discloses a method for predicting blast furnace gas utilization rate, including: inputting the physical properties of ore and coke and gas parameters; obtaining the diffusion coefficient in the layered structure; calculating the reduction reaction rate of iron ore and the dissolution reaction rate of coke based on the diffusion coefficient; obtaining the gas composition distribution based on the reduction reaction rate and the dissolution reaction rate of coke; determining whether the obtained gas composition converges; if not, recalculating the reduction reaction rate of iron ore and the dissolution reaction rate of coke; otherwise, predicting the gas utilization rate based on the gas composition. This patented technology is used to predict blast furnace gas utilization rate without coupling parameters such as the movement of the internal furnace charge and the temperature field. It can predict the gas utilization rate based on parameters such as the initial gas composition, temperature, and furnace charge structure.

[0006] Patent application number 202210349640.8 discloses a method for improving blast furnace gas utilization based on blast furnace burden surface monitoring, including the following steps: monitoring the shape of the blast furnace burden surface using radar and transmitting the blast furnace burden surface shape information to a host computer; the host computer identifies and clusters the blast furnace burden surface shape; the host computer matches the current burden surface shape with data stored in a database, where the database stores multiple categories formed by statistically calculating multiple burden surface shapes and the gas utilization rate corresponding to each category; based on the matching results of the current burden surface shape and the data in the database, the burden surface shape is adjusted through a burden distribution operation to improve the blast furnace gas utilization rate. This patented technology is beneficial for timely adjustment of blast furnace burden distribution, improving blast furnace gas utilization, and reducing blast furnace fuel consumption and costs.

[0007] Therefore, it is evident that traditional gas utilization prediction methods, primarily based on blast furnace mechanisms and data-driven approaches, cannot fully analyze the relationship between blast furnace gas flow and operational status indicators. Data-driven analysis methods mostly employ only a single prediction model, exhibiting limitations such as overfitting, underfitting, and data bias, resulting in poor model prediction accuracy and robustness. However, most studies are based on blast furnace smelting principles, assessing blast furnace gas utilization based on blast furnace reaction principles and their equilibrium equations. These methods often involve numerous assumptions, failing to consider the large time lag in the blast furnace smelting process and the presence of noisy data in blast furnace gas utilization. Furthermore, the selection of prediction models is relatively simplistic, neglecting the stability and accuracy of the prediction models themselves. Summary of the Invention

[0008] To address the shortcomings of existing technologies, the purpose of this invention is to provide a time series prediction method for blast furnace gas utilization rate, which can accurately predict blast furnace gas utilization rate and has important application value in stabilizing blast furnace production, improving gas flow distribution to increase molten iron quality and output, extending blast furnace life, and saving energy and reducing emissions.

[0009] To achieve the above objectives, the present invention adopts the following technical solution:

[0010] A method for predicting the time series of blast furnace gas utilization rate includes the following steps:

[0011] S1, preprocessing the time series data of gas utilization rate;

[0012] S2, improve the PSO algorithm to form the LDIW-PSO algorithm;

[0013] S3, Design a KELM algorithm based on the ELM algorithm;

[0014] S4. The hyperparameters of the KELM algorithm are optimized using the LDIW-PSO algorithm to determine the LDIW-PSO-KELM model and its parameters.

[0015] S5 uses the LDIW-PSO-KELM model to predict blast furnace gas utilization rate.

[0016] Preferably, in step S1, the gas utilization rate time series data is filtered and denoised using singular spectrum analysis, specifically including the following steps:

[0017] S11, embedded in phase space;

[0018] S12, Singular Value Decomposition;

[0019] S13, construct subspace;

[0020] S14, Refactoring.

[0021] Preferably, step S11 specifically includes:

[0022] For the time series of original gas utilization rate X T =(x1,x2,…,x T This is mapped to a vector sequence of length L, forming K = T - L + 1 vectors of length L:

[0023] Y i =(x1,x2,…,x i+L-1 ) · (1≤i≤K)

[0024] These vectors form the trajectory matrix:

[0025]

[0026] Where L is the window length, which is an integer multiple of the sampling period, K = T - L + 1.

[0027] Preferably, step S12 specifically includes:

[0028] Let S = YY · Then λ1, ..., λ L λ is an eigenvalue, and λ1≥…≥λ L ≥0, and U1,…,U L It is the orthogonal vector of matrix S corresponding to the eigenvalues;

[0029] Let d = rank(Y) = max{i,λ i If > 0, then the SVD of the trajectory matrix Y can be written as:

[0030] Y = Y1 + Y2 + ... + Y d

[0031]

[0032] Preferably, step S13 specifically includes:

[0033] Divide the index set {1,…,d} into m distinct subsets I1,…,I m Let I = {i1,…,i} p}, then the composite matrix corresponding to I. Then we have:

[0034]

[0035] Preferably, step S14 specifically includes:

[0036] expression Each matrix in Transform it into a new sequence of length T, that is, obtain the decomposed sequence;

[0037] Let P be an L×K matrix with elements p i,j ,1≤i≤L,1≤j≤K,IfL<K, otherwise

[0038] Let L * =min(L,K),K * =max(L,K), T=L+K-1, then use the following expression to calculate the average along the diagonal, and transform the matrix P into a sequence p1,…,p N ;

[0039]

[0040] Preferably, in step S2, the particle swarm optimization algorithm is improved using a linearly decreasing inertia weight method to optimize the key parameters of the kernel extreme learning machine. The position and velocity update equations for each particle are as follows:

[0041] x i (t+1)=x i (t)+v i (t+1)

[0042] v i (t+1)=ω·v i (t)+c1·r1·(pbest i -x i (t))+c2·r2·(gbest i -x i (t))

[0043] Where, x i (t) and v i (t) represents the position and velocity of particle i at the current moment, respectively; x i (t+1) and v i (t+1) represent the position and velocity of particle i at the next moment; ω represents the inertial weight; c1 and c2 represent the acceleration coefficients; r1 and r2 are random numbers in the range [0,1]; pbest i It is the individual optimal solution for particle i; gbest i It is the globally optimal solution for the entire group.

[0044] Preferably, the improved particle swarm optimization algorithm has the following inertia weight equation:

[0045]

[0046] Where d represents the number of iterations; K represents the total number of iterations; ω start ω represents the initial weights; end This indicates the end weight.

[0047] Preferably, in step S3, the learning objective function y(x) of the kernel extreme learning machine is represented by a matrix as follows:

[0048] y(x)=h(x)β=Hβ=Y

[0049]

[0050] Where x represents the input vector; H and h(x) represent the hidden node outputs; β represents the output weights; Y represents the desired output; g represents the activation function; and L represents the number of neurons in the hidden layer. Let H be the generalized inverse matrix.

[0051] Preferably, the kernel extreme learning machine introduces a regularization coefficient C and an identity matrix I, and the corresponding output weights are expressed as follows:

[0052]

[0053] Introducing kernel functions into ELM, the kernel matrix is:

[0054] Ω ELM(i,j) =HH · =k(x i ,x j )

[0055] The learning objective function of the kernel limit learning machine is obtained as follows:

[0056]

[0057] Where I represents the identity matrix; C represents the regularization coefficient; σ 2 Indicates kernel parameters.

[0058] This invention provides a time-series prediction method for blast furnace gas utilization rate, which utilizes a particle swarm optimization algorithm with linearly decreasing inertia weights to optimize the single-step prediction method of gas utilization rate using a kernel extreme learning machine. This invention proposes an integrated prediction model for time-series prediction of blast furnace gas utilization rate, providing a new approach and method for monitoring and predicting complex blast furnace production processes. Accurate prediction of blast furnace gas utilization rate using this method has significant application value in stabilizing blast furnace production, improving gas flow distribution to increase molten iron quality and yield, extending blast furnace life, and achieving energy conservation and emission reduction. Simultaneously, steel enterprises can better plan production schedules, optimize process parameters, reduce energy costs, and minimize unnecessary losses. Furthermore, it can reduce dependence on fossil fuels and greenhouse gas emissions, contributing to sustainable development. This invention also has the following beneficial effects:

[0059] (1) This invention applies singular spectrum analysis to the processing of blast furnace GUR time series data, effectively extracting the time series features in GUR data and separating noise.

[0060] (2) This invention uses real blast furnace production process data from a steel plant to design an LDIW-PSO-KELM prediction algorithm for the time series of blast furnace gas utilization rate. Attached Figure Description

[0061] Figure 1 This is a flowchart illustrating the time series prediction method for blast furnace gas utilization rate of the present invention.

[0062] Figure 2 This is a schematic diagram of the blast furnace production process;

[0063] Figure 3 This is a schematic diagram of the original time series of blast furnace gas utilization rate in an embodiment of the blast furnace gas utilization rate time series prediction method of the present invention;

[0064] Figure 4 This is a schematic diagram of the blast furnace gas utilization time series after SSA processing in an embodiment of the blast furnace gas utilization time series prediction method of the present invention;

[0065] Figure 5 This is a schematic diagram of the ELM prediction model framework in an embodiment of the blast furnace gas utilization time series prediction method of the present invention;

[0066] Figure 6 This is a schematic diagram of the single-step prediction model structure in an embodiment of the blast furnace gas utilization rate time series prediction method of the present invention;

[0067] Figure 7 This is a schematic diagram of the BP, ELM, and KELM prediction results in an embodiment of the blast furnace gas utilization time series prediction method of the present invention;

[0068] Figure 8 This is a schematic diagram of the KELM and PSO-KELM prediction results in an embodiment of the blast furnace gas utilization time series prediction method of the present invention;

[0069] Figure 9 This is a schematic diagram of the prediction results of PSO-KELM and LDIW-PSO-KELM in the embodiment of the time series prediction method for blast furnace gas utilization rate of the present invention. Detailed Implementation

[0070] To better understand the above-mentioned technical solutions of the present invention, the technical solutions of the present invention will be further described below in conjunction with the accompanying drawings and embodiments.

[0071] Combination Figure 1 As shown, the present invention provides a time series prediction method for blast furnace gas utilization rate, comprising the following steps:

[0072] S1 employs the Singular Spectrum Analysis (SSA) method to preprocess the time series data of gas utilization rate, extract the characteristic signals in the time series, and separate the noise components to make the data suitable for analysis, thus completing the industrial production data cleaning.

[0073] S2 improves the Particle Swarm Optimization (PSO) algorithm by using a linearly decreasing inertia weight method, balancing the exploration and utilization capabilities of the PSO algorithm, so that the PSO algorithm focuses more on global search in the early stage and more on local search in the later stage.

[0074] S3 combines the improved particle swarm optimization algorithm with kernel extreme learning machine (KELM), that is, it uses particle swarm optimization algorithm to optimize the hyperparameters of kernel extreme learning machine, so that it has good prediction ability and stability.

[0075] S4. Experimental verification of the prediction method for the time series of blast furnace gas utilization rate was carried out. Multiple prediction models were selected for comparative experiments to verify the superior accuracy and stability of the designed PSO-KELM model based on singular spectrum analysis for predicting blast furnace gas utilization rate.

[0076] S5 compares and demonstrates the improved PSO-KELM blast furnace gas utilization time series prediction model based on singular spectrum analysis.

[0077] The blast furnace ironmaking process involves the interaction of descending and rising gas flows within a high-temperature, high-pressure, airtight furnace. For example... Figure 2The schematic diagram of blast furnace ironmaking illustrates the principle: raw materials such as iron ore, coke, and limestone are fed into the blast furnace through chutes, while pulverized coal and high-temperature gases are blown in from the bottom. The coke is first heated to a high temperature and then undergoes a gasification reaction, producing a reducing gas called blast furnace gas, which contains carbon monoxide, hydrogen, and other components. The blast furnace gas reacts with the iron ore to produce pig iron and byproducts (such as slag and gas). The molten pig iron flows into the bottom of the furnace in liquid form, while the slag floats on the surface of the molten iron, forming a layer due to its lighter weight. The tail gas produced after the reducing gas reaction is discharged from the top of the blast furnace. The gases discharged from the top mainly include CO, CO2, and other gases. The ratio of CO to the sum of CO and CO2 is called GUR, defined as follows:

[0078]

[0079] Where, η CO This represents the gas utilization rate (RUR) of the blast furnace; φ CO and This indicates the volume percentage of CO2 and CO in the top gas.

[0080] The preprocessing of the gas utilization rate time series data using SSA in step S1 above specifically includes:

[0081] Due to the complex coupling and high-temperature, high-pressure environment present during the blast furnace reaction, CO utilization rate is affected, resulting in noise and outliers in production data. Singular Spectrum Analysis (SSA) is a time series analysis method. Its basic idea is to decompose a time series into several components with specific structures. These components reflect different patterns of change in the sequence and can effectively extract trends, periodicity, and noise components from the signal.

[0082] Step S1 of this invention uses SSA to filter and denoise blast furnace gas utilization data, mainly including the following steps:

[0083] S11, embedded in phase space. For the original gas utilization rate time series X T =(x1,x2,…,x T This is mapped to a vector sequence of length L, forming K = T - L + 1 vectors of length L:

[0084] Y i =(x1,x2,…,x i+L-1 ) · (1≤i≤K)

[0085] These vectors form the trajectory matrix:

[0086]

[0087] Where L is the window length, which is an integer multiple of the sampling period, K = T - L + 1.

[0088] S12, Singular Value Decomposition.

[0089] Let S = YY · Then λ1, ..., λ L λ is an eigenvalue, and λ1≥…≥λ L ≥0, and U1,…,U L It is the orthogonal vector of matrix S corresponding to the eigenvalues;

[0090] Let d = rank(Y) = max{i,λ i If > 0, then the SVD of the trajectory matrix Y can be written as:

[0091] Y = Y1 + Y2 + ... + Y d

[0092]

[0093] S13, construct the subspace.

[0094] Divide the index set {1,…,d} into m distinct subsets I1,…,I m Let I = {i1,…,i} p}, then the composite matrix corresponding to I. Then we have:

[0095]

[0096] S14, Refactoring. Refactor the expression... Each matrix in Transform it into a new sequence of length T, that is, obtain the decomposed sequence;

[0097] Let P be an L×K matrix with elements p i,j ,1≤i≤L,1≤j≤K,IfL<K, otherwise

[0098] Let L * =min(L,K),K * =max(L,K), T=L+K-1, then use the following expression to calculate the average along the diagonal, and transform the matrix P into a sequence p1,…,p N ;

[0099]

[0100] By processing raw data using SSA (Self-Solving Aspect), important information in the time series of gas utilization rates can be effectively separated and identified, leading to better analysis of relevant data. For example... Figure 3 and Figure 4 The original sequence of blast furnace GUR time data and the GUR time series after SSA processing are shown respectively.

[0101] Furthermore, to avoid excessive fluctuations in the data range that could negatively impact the stability and accuracy of the prediction model, the blast furnace gas rate data must be normalized to a range of [0,1]. The normalization formula is as follows:

[0102]

[0103] Where, maxη co minη represents the maximum value in the time series of gas utilization rate; co This represents the minimum value in the time series of coal gas utilization rate; N represents the length of the time series.

[0104] The improvements to the particle swarm optimization algorithm in step S2 above include:

[0105] Particle Swarm Optimization (PSO) is a heuristic optimization algorithm that finds the optimal solution to a problem by simulating the collective behavior of organisms such as flocks of birds or schools of fish. This invention utilizes an improved PSO algorithm to optimize key parameters of a kernel extreme learning machine. The position and velocity update equations for each particle are as follows:

[0106] x i (t+1)=x i (t)+v i (t+1)

[0107] v i (t+1)=ω·v i (t)+c1·r1·(pbest i -x i (t))+c2·r2·(gbest i -x i (t))

[0108] Where, x i (t) and v i (t) represents the position and velocity of particle i at the current moment, respectively; x i (t+1) and v i (t+1) represent the position and velocity of particle i at the next moment; ω represents the inertial weight; c1 and c2 represent the acceleration coefficients; r1 and r2 are random numbers in the range [0,1]; pbest iIt is the individual optimal solution for particle i; gbest i It is the globally optimal solution for the entire group.

[0109] This invention improves the particle swarm optimization (PSO) algorithm by using linearly decreasing inertia weights (LDIW). This LDIW enhances the global search capability during the PSO iteration process and reduces the possibility of the optimization result getting trapped in local minima. The improved PSO algorithm's inertia weight equation is as follows:

[0110]

[0111] Where d represents the number of iterations; K represents the total number of iterations; ω start ω represents the initial weights; end This indicates the end weight.

[0112] The design of the kernel extreme learning machine in step S3 above specifically includes:

[0113] This invention uses KELM as the basic prediction model, demonstrating outstanding performance in both speed and accuracy. ELM, introduced by Huang, is a high-performance feedforward network unique in that it contains only a single hidden layer. Compared to traditional neural networks, ELM possesses strong robustness in nonlinear fitting and rapid learning characteristics. Figure 5 The diagram shows the structure of the ELM (Extreme Learning Machine). The learning objective function y(x) is represented by a matrix as follows:

[0114] y(x)=h(x)β=Hβ=Y

[0115]

[0116] Where x represents the input vector; H and h(x) represent the hidden node outputs; β represents the output weights; Y represents the desired output; g represents the activation function; and L represents the number of neurons in the hidden layer. Let H be the generalized inverse matrix.

[0117] KELM is an improved algorithm based on ELM and combined with a kernel function. It introduces a regularization coefficient C and an identity matrix I, which enhances the model's stability while retaining the advantages of ELM. The corresponding output weights are represented as follows:

[0118]

[0119] Introducing kernel functions into ELM, the kernel matrix is:

[0120] Ω ELM(i,j) =HH · =k(x i ,x j )

[0121] The learning objective function of KELM is obtained as follows:

[0122]

[0123] Where I represents the identity matrix; C represents the regularization coefficient; σ 2 Indicates kernel parameters.

[0124] The design of the LDIW-PSO-KELM model in step S4 above specifically includes:

[0125] To address the issue that single ELM prediction models are highly sensitive to the selection of hyperparameters, this invention designs a KELM prediction method based on ELM and combined with a kernel function, which improves the prediction performance of the model while retaining the advantages of ELM. In addition, linearly decreasing inertial weights (LDIW) are used to improve the ability of the PSO algorithm to search for parameters of the KELM model globally and locally, where the mean absolute error (MAE) of the KELM model prediction is used as the fitness function of PSO.

[0126] Based on the above ideas, this invention establishes a blast furnace gas utilization time series prediction model based on LDIW-PSO-KELM. Specifically, this invention uses a single-step prediction method to predict the blast furnace gas utilization time series; that is, the input of this prediction model is the gas utilization rate of the previous 20 time points, and the output is the gas utilization rate of the next time point. A schematic diagram of this single-step prediction model is shown below. Figure 6 As shown.

[0127] The verification of the LDIW-PSO-KELM model in step S5 above specifically includes:

[0128] The gas utilization rate time series data in this embodiment of the invention is from a steel company, with 560 data samples and a sampling interval of 6 minutes. The first 70% of the sample data is selected as training data, and the last 30% is selected as test data.

[0129] A, Evaluation Indicators

[0130] To compare and differentiate the precision and accuracy of prediction results from various methods, root mean square error (RMSE), mean absolute percentage error (MAPE), and mean absolute error (MAE) were used as evaluation criteria.

[0131]

[0132] Among them, Y i and These represent the actual value and the predicted value, respectively; n represents the number of samples.

[0133] B, Prediction Results

[0134] The time series prediction of blast furnace gas utilization rate is performed using the LDIW-PSO-KELM model. To illustrate the feasibility and superiority of this method, the time series prediction results of four methods—BP neural network, ELM, KELM, and PSO-KELM—are compared.

[0135] like Figure 7 The chart shows a comparison of predictions from three methods: BP neural network, ELM, and KELM. The chart reveals that the BP neural network exhibits significant fluctuations in the sample data range of 80-120, indicating poor model stability. While ELM can track the development trend of the GUR time series relatively well, its prediction accuracy is not high. The KELM prediction model, however, improves the model's generalization ability and better adapts to data distribution. Figure 7 As shown by the orange curve, its prediction accuracy has been improved.

[0136] To more accurately find the optimal hyperparameters of KELM, this invention employs the PSO algorithm to optimize its parameters. For example... Figure 8 As shown in the figure, the prediction results of KELM and PSO-KELM are displayed. It can be found that the KELM prediction accuracy is higher after optimization by the PSO algorithm. Considering the problem that the PSO algorithm is prone to getting trapped in local optima, LDIW is introduced to improve its optimization performance in both global and local aspects.

[0137] Table 1 shows the optimal parameter combinations extracted from KELM. Figure 9 The prediction results of PSO-KELM and LDIW-PSO-KELM are shown in the figure. It can be seen that the improved LDIW-PSO-KELM model has better stability and accuracy in predicting GUR time series, and has a good prediction effect.

[0138] Table 1. Optimal KELM parameters for the improved PSO algorithm.

[0139] KELM parameters C <![CDATA[σ 2 ]]> 335.7714 11.7114

[0140] To more accurately illustrate the performance of each prediction model, a quantitative comparison of the prediction models for each GUR time series was conducted, and the results are shown in Table 2. The LDIW-PSO-KELM prediction model performed better. The mean absolute error (MAE), mean absolute percentage error (MAPE), and root mean square error (RMSE) of the prediction results using LDIW-PSO-KELM were all lower than those of other prediction models. The experimental results validated the effectiveness of the designed LDIW-PSO-KELM prediction model for GUR time series prediction, and demonstrated higher prediction accuracy.

[0141] Table 2 Comparison of Evaluation Indicators for Five Prediction Models

[0142] Evaluation criteria RMSE MAPE MAE BP 1.1260 0.0172 0.8187 ELM 0.5829 0.0097 0.4534 KELM 0.4965 0.0084 0.3903 PSO-KELM 0.3235 0.0048 0.2804 LDIW-PSO-KELM 0.2677 0.0037 0.2358

[0143] This invention combines ironmaking principles with the time series characteristics of blast furnace gas utilization (GUR). Using real blast furnace production process data from a steel plant, it obtains the time series characteristics of blast furnace gas utilization rate and separates noise through singular spectrum analysis. Based on this, a kernel extreme learning machine prediction model optimized with an improved particle swarm optimization algorithm is designed. Experiments and comparisons verify that the designed prediction model has better stability and accuracy in predicting blast furnace GUR time series.

[0144] Those skilled in the art should recognize that the above embodiments are merely illustrative of the present invention and are not intended to limit the present invention. Any variations or modifications to the above embodiments that are within the spirit and essence of the present invention will fall within the scope of the claims of the present invention.

Claims

1. A blast furnace gas utilization rate time series prediction method, characterized by, The method comprises the following steps: S1, pre-processing the time series data of the gas utilization rate; S2, improving the PSO algorithm to form the LDIW-PSO algorithm; S3, designing the KELM algorithm based on the ELM algorithm; S4, optimizing the hyperparameters of the KELM algorithm by using the LDIW-PSO algorithm, determining the LDIW-PSO-KELM model and the model parameters; S5, realizing the prediction of the blast furnace gas utilization rate by using the LDIW-PSO-KELM model.

2. The blast furnace gas utilization rate time series prediction method according to claim 1, characterized by, In the step S1, the singular spectrum analysis method is used to filter and denoise the time series data of the gas utilization rate, and the step specifically comprises the following steps: S11, embedding a phase space; S12, singular value decomposition; S13, constructing a subspace; S14, reconstruction.

3. The blast furnace gas utilization rate time series prediction method according to claim 2, characterized by, The step S11 specifically comprises: For the raw coal gas utilization rate time series X T = (x1, x2,..., x T ), it is mapped into a sequence of vectors of length L, forming K = T - L + 1 vectors of length L: Y i = (x1, x2,..., x i+L-1 ) · (1≤i≤K) These vectors constitute a trajectory matrix: Wherein, L is the window length, which is an integer multiple of the sampling period K=T-L+1.

4. The blast furnace gas utilization rate time series prediction method according to claim 3, characterized by, The step S12 specifically comprises: Let S = YY · Then λ1,..., λ L are eigenvalues of S, and λ1≥... ≥ λ L ≥ 0, and U1,..., U L are orthonormal vectors of the matrix S corresponding to the eigenvalues Let d = rank(Y) = max{i, λ i} and the SVD of the trajectory matrix Y is written as: Y = Y1+ Y2+... + Y d wherein 5. The blast furnace gas utilization rate time series prediction method according to claim 4, characterized by, The step S13 specifically comprises: The subscript set {1,..., d} is divided into m different subsets I1,..., Im m , let I = {i1,..., id p}, then the synthesis matrix has:

6. The blast furnace gas utilization rate time series prediction method according to claim 5, characterized by, The step S14 specifically comprises: Each matrix in the expression is transformed into a new sequence of length T, i.e. the decomposed sequence is obtained; is transformed into a new sequence of length T, i.e. the decomposed sequence is obtained; Let P be an L x K matrix with elements p i,j ,1≤i≤L,1≤j≤K, if L < K, then Let L * = min(L, K), K * = max(L, K), T = L + K - 1, then convert the matrix P into a sequence p1,..., p N ; 7. The blast furnace gas utilization rate time series prediction method according to claim 1, characterized by, In the step S2, the linearly decreasing inertia weight method is used to improve the particle swarm algorithm, and the key parameters of the kernel extreme learning machine are optimized, and the position and speed update equations of each particle are respectively: x i (t+1) = x i (t) + v i (t+1) v i (t+1) = ω · v i (t) + c1 · r1 · (pbest i -x i (t)) + c2 · r2 · (gbest i -x i (t)) where x i (t) and v i (t) represent the position and velocity of particle i at the current time, respectively; x i (t+1) and v i (t+1) represent the position and velocity of particle i at the next time, respectively; ω represents the inertia weight; c1 and c2 represent the acceleration coefficients; r1 and r2 are random numbers in the range [0, 1]; pbest i is the individual optimal solution of particle i; and gbest i is the global optimal solution of the entire population.

8. The blast furnace gas utilization rate time series prediction method according to claim 7, characterized by, The inertia weight equation of the improved particle swarm algorithm is: where d denotes the number of cycles; K denotes the total number of cycles; ω start denotes the initial weight; ω end denotes the terminal weight.

9. The blast furnace gas utilization rate time series prediction method according to claim 1, characterized by, In the step S3, the learning objective function y(x) of the kernel extreme learning machine is represented by a matrix as: y(x)=h(x)β=Hβ=Y where x represents an input vector; H and h(x) represent hidden node outputs; β represents output weights; Y represents a desired output; g represents an activation function; and L represents the number of neurons in the hidden layer. denotes the generalized inverse of H.

10. The blast furnace gas utilization rate time series prediction method according to claim 9, characterized by, The kernel extreme learning machine introduces a regularization coefficient C and a unit matrix I, and the corresponding output weight is expressed as: The kernel function is introduced into the ELM, and the kernel matrix is: Ω ELM(i,j) = HH · = k(x i , x j ) The learning objective function of the kernel extreme learning machine is obtained: where I denotes an identity matrix; C denotes a regularization coefficient; σ 2 denotes a kernel parameter.

Images