Construction method of blast furnace gas utilization rate prediction model considering time delay effect
By constructing a blast furnace gas utilization rate prediction model that considers the time delay effect, using singular spectrum analysis and partial autocorrelation function to extract time delay features, and combining it with the Transformer model, the problems of prediction accuracy and stability of blast furnace gas utilization rate are solved, and efficient blast furnace operation optimization and energy efficiency improvement are achieved.
Patent Information
- Application Number
- CN202510954691.7
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2025-07-11
- Publication Date
- 2025-12-16
AI Technical Summary
Existing methods for predicting blast furnace gas utilization rates are inaccurate and have unstable prediction results, making it difficult to effectively reflect the complex thermo-mechanical-chemical coupling mechanism within the blast furnace and to handle high-dimensional, nonlinear, and multi-disturbance systems.
A prediction model for blast furnace gas utilization considering time delay effects is constructed. Key time delay features are extracted through singular spectrum analysis and partial autocorrelation function, and prediction is performed by combining an attention-based Transformer model, thereby enhancing the model's adaptability to the complex operating dynamics of blast furnaces.
It improves the accuracy and stability of gas utilization rate prediction, enhances the intelligent technical support for blast furnace optimization operation and energy efficiency improvement, and is suitable for intelligent operation of blast furnaces.
Smart Images
Figure CN121145162A_ABST
Abstract
Description
Technical Field
[0001] This application belongs to the field of blast furnace gas utilization rate prediction technology, and more specifically, relates to a method for constructing a blast furnace gas utilization rate prediction model that considers the time lag effect. Background Technology
[0002] The blast furnace is one of the most crucial thermal equipment in modern metallurgical industry, a large vertical reactor used for the continuous production of pig iron. During the smelting process, raw materials such as iron ore, coke, and flux are sequentially added from the top of the furnace, undergoing a series of complex oxidation-reduction reactions under the high temperature and pressure environment inside, ultimately producing molten iron and blast furnace gas. Blast furnace gas is a combustible gas produced during the reaction process inside the furnace, mainly discharged through the top of the furnace, and widely used as a secondary energy source for heating, power generation, and other processes.
[0003] Blast furnace gas utilization rate is an important indicator for measuring the effective utilization of reactant gases in the furnace, usually expressed as the ratio of carbon dioxide content to the total content of carbon monoxide and carbon dioxide. Its changes not only reflect gas flow distribution and reaction efficiency but are also closely related to fuel consumption and production levels. Accurate prediction and control of blast furnace gas utilization rate helps optimize blast furnace operating parameters, reduce energy consumption, and improve production efficiency, making it a key aspect of intelligent blast furnace operation. However, in actual production, gas utilization rate is affected by multiple factors, including burden ratio, combustion intensity, gas flow distribution, and temperature changes. The relationships between these factors are complex and highly nonlinear. Furthermore, the blast furnace operation involves numerous disturbances, and the collected raw data often exhibits high noise and significant non-stationarity, posing a significant challenge to the accurate modeling and prediction of gas utilization rate. Existing prediction methods mainly fall into two categories: Mechanism-based methods rely on mathematical modeling of the gas flow and reaction processes within the blast furnace. While these methods have a strong theoretical foundation, they require numerous simplifications and assumptions during modeling, making it difficult to accurately reflect the complex thermo-mechanical-chemical coupling mechanisms within the blast furnace. Furthermore, they are highly sensitive to parameters and prone to introducing biases. Traditional data-driven methods, such as linear regression and backpropagation neural networks, can utilize actual collected data for modeling, but their modeling capabilities are limited and their prediction results are unstable when dealing with high-dimensional, nonlinear, and multi-perturbation systems.
[0004] Therefore, how to solve the problems of inaccurate prediction accuracy and unstable prediction effect of traditional blast furnace gas utilization prediction methods is a pressing issue that needs to be addressed. Summary of the Invention
[0005] In view of the shortcomings of the existing technology, the purpose of this application is to provide a method for constructing a blast furnace gas utilization rate prediction model that considers the time delay effect, which can effectively improve the accuracy and stability of gas utilization rate prediction.
[0006] To achieve the above objectives, in a first aspect, this application provides a method for constructing a blast furnace gas utilization rate prediction model that considers time delay effects, comprising the following steps: S10: Obtain and filter the historical time series data of blast furnace gas utilization rate, and extract the trend and periodic feature signals in the data. S20, perform detrending operation on the data processed in step S10, extract long-range correlation in the time series, and then combine the partial autocorrelation function to quantify the lag effect of historical time points on the current variable, thereby determining the time lag step of gas utilization rate after the adjustment of blast furnace operating parameters. S30: Using the data processed in step S10 as the dataset, the blast furnace operating parameters and the time delay step determined in step S20 as input features, and the gas utilization rate as the output feature, a pre-established Transformer model based on the attention mechanism is trained to obtain a blast furnace gas utilization rate prediction model.
[0007] The intelligent prediction method for blast furnace gas utilization rate considering time lag effect provided in this application fully explores the key information in the historical data of gas utilization rate that has a delayed impact on the current state. Combined with advanced time series modeling technology, it can achieve accurate prediction of the trend of gas utilization rate change. In addition, by introducing a time lag feature modeling mechanism, the model's adaptability to the complex operating dynamics of the blast furnace is effectively improved, and the accuracy and stability of the prediction are enhanced, providing intelligent technical support for blast furnace operation optimization and energy efficiency improvement.
[0008] As a further preferred embodiment, the blast furnace operating parameters include air pressure, heat load, fuel ratio, coke ratio, ore-coke ratio, blast kinetic energy, coal ratio, furnace top temperature, and furnace top pressure.
[0009] As a further preferred option, in step S10, the singular spectral analysis method is used to perform data filtering and noise reduction by embedding and singular value decomposition of historical time series data.
[0010] As a further preferred option, step S10 specifically includes: S11, historical time series data of blast furnace gas utilization rate Mapped to a length of The vector sequence, taking an integer multiple of the sampling period. A length of vector And these vectors are used to form a trajectory matrix. ; S12, let ,but for eigenvalues, and ,and It is a matrix The orthogonal vectors corresponding to these eigenvalues; And order Then the trajectory matrix SVD is ;in, , ; S13, set the index Divided into Different subsets ,make Then the corresponding Composite matrix Then there is ; S14, each matrix in the above formula Transform into a length of The new sequence is obtained, that is, the decomposed sequence, and then let For one A matrix with elements of .if , ,otherwise ; And order , Then, using the following formula to calculate the average of the diagonals, we can transform the matrix... Convert to sequence ; .
[0011] As a further preferred option, in step S20, the step of performing detrending operation on the data processed in step S10 and extracting long-range correlation in the time series is as follows: First, the logarithmic relationship between the fluctuation function F(s) and the window size s is calculated using autotrend analysis to obtain the slope α; then, based on α, it is determined whether there is long-range correlation and the long-range correlation in the time series is extracted.
[0012] As a further preferred embodiment, step S20, which involves performing a detrending operation on the data processed in step S10 to extract long-range correlations in the time series, specifically includes: S21, Cumulative Gas Utilization Rate Time Series: Assuming the gas utilization rate time series is... First, calculate the cumulative time series. , The mean of the sequence; S22, Split the time series: Divide the cumulative time series Divided into lengths of There are a total of non-overlapping windows. One window; S23, for each window Using the least squares method to fit a local trend Then, the trend is removed from the accumulated signal of each window to obtain the first... v Standard deviation of detrended volatility for each window :
[0013] S24, Wave function calculation: for all windows The average of the squared fluctuations is: .
[0014] As a further preferred embodiment, in step S20, the formula for calculating the partial autocorrelation function is:
[0015] In the formula, Indicates the first k The extent to which historical data has an impact on the current moment; Based on of The least squares prediction; the same Based on of The least squares prediction value; () represents the correlation coefficient.
[0016] As a further preferred embodiment, after training the pre-established attention-based Transformer model in step S30, the following method is also included: The Adam optimization algorithm is used to adjust the model parameters, and cross-validation and hyperparameter optimization techniques are combined to ensure model convergence and achieve optimal performance. The adjusted model parameters include the Query, Key, and Value projection matrix parameters and position encoding parameters in the attention layer, as well as the weights and biases of the feedforward neural network layer and the weights of the prediction output layer.
[0017] As a further preferred embodiment, step S30, before obtaining the blast furnace gas utilization rate prediction model, further includes: The experimental results were analyzed in depth and compared with the prediction results without considering the time delay effect. The experimental results were analyzed in depth in combination with the prediction indicators, and the model structure and hyperparameters were adjusted according to the experimental results.
[0018] Secondly, this application provides a system for constructing a blast furnace gas utilization rate prediction model that considers the time delay effect, including: The data processing module is used to acquire and filter and denoise the time series data of blast furnace gas utilization rate, and extract trend and periodic feature signals from the data. The time lag step extraction module is used to perform detrending operations on the data processed by the data processing module, extract long-range correlations in the time series, and then combine the partial autocorrelation function to quantify the lag effect of historical time points on the current variable, thereby determining the time lag step of the gas utilization rate after the adjustment of the basic parameters of blast furnace operation. The training module uses the data processed by the data processing module as the dataset, the basic parameters of blast furnace operation and the time delay step determined by the time delay step extraction module as input features, and the gas utilization rate as the output feature to train a pre-established Transformer model based on the attention mechanism, thereby obtaining a blast furnace gas utilization rate prediction model.
[0019] It is understandable that the beneficial effects of the second aspect mentioned above can be found in the relevant descriptions in the first aspect mentioned above, and will not be repeated here. Attached Figure Description
[0020] Figure 1 This is a flowchart of the method for constructing the blast furnace gas utilization prediction model considering the time delay effect provided in this application; Figure 2 This is a schematic diagram of the blast furnace production process provided in the embodiments of this application; Figure 3 These are the original time series and SSA reconstructed time series of blast furnace gas utilization rate provided in the embodiments of this application; Figure 4 This is a diagram showing the experimental results of DFA provided in the embodiments of this application; Figure 5 This is a diagram of PACF experimental results provided in an embodiment of this application; Figure 6 This is a structural diagram of the Transformer provided in the embodiments of this application; Figure 7 These are comparative experimental figures of Transformer models considering and not considering time delay provided in the embodiments of this application; Figure 8 This is a comparative experimental diagram of the Transformer model and the SVM model considering time delay provided in the embodiments of this application. Detailed Implementation
[0021] To make the objectives, technical solutions, and advantages of this application clearer, the following detailed description is provided in conjunction with the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are merely illustrative and not intended to limit the scope of this application.
[0022] It should be noted that the time delay effect in this application refers to the fact that the current output of the system (such as blast furnace gas utilization rate (GUR)) not only depends on the current input or operating conditions, but is also significantly affected by its past historical state or historical input. This historical dependence has a certain time delay characteristic.
[0023] This application, through research, has found the following reasons for the inaccuracy and unstable prediction results of traditional blast furnace gas utilization rate prediction methods: First, most studies have failed to adequately consider the time lag effect prevalent in blast furnace processes, namely, the significant correlation between the current state of gas utilization and production parameters at multiple historical moments. Ignoring this characteristic can easily lead to models failing to capture the true dynamic changes, resulting in increased prediction bias, delayed response, and reduced practical value of the predictions.
[0024] Secondly, traditional machine learning methods such as SVM and XGBoost exhibit insufficient feature representation capabilities when dealing with blast furnace systems characterized by strong nonlinearity and complex operating conditions, making it difficult to effectively model the long-term dependence and sequential variation trends of gas utilization. Meanwhile, while common deep learning models such as RNN and GRU possess some temporal modeling capabilities, they suffer from low training efficiency, severe gradient vanishing problems, and poor performance in modeling long-distance dependencies.
[0025] Furthermore, existing technologies still rely mainly on empirical methods or simple lag terms in constructing input features, lacking a systematic approach to extract key time-delay features. They also have weak noise handling capabilities and are susceptible to abnormal fluctuations, which can affect the overall model performance.
[0026] The purpose of this application is to address the aforementioned problems by proposing a Transformer prediction model based on significant time delay characteristics. This model automatically extracts key time delay terms using partial autocorrelation function (PACF) analysis, effectively characterizing the dependence of blast furnace gas utilization on historical conditions. Simultaneously, a self-attention mechanism is introduced to model long-term time-series dependencies, enhancing the model's ability to model complex fluctuations and its resistance to interference. Ultimately, this achieves a more accurate and stable prediction of blast furnace gas utilization, demonstrating good engineering applicability and promotional value.
[0027] like Figure 1 As shown, this application provides a method for constructing a blast furnace gas utilization rate prediction model that considers the time lag effect. This method can also be applied to the prediction of time series of other indicators (such as furnace temperature and permeability index), and mainly includes the following steps: Step S10, Time series preprocessing: Obtain and filter the historical time series data of blast furnace gas utilization rate and remove noise, extract the trend and periodic feature signals in the time series of blast furnace gas utilization rate, and remove noise components.
[0028] In one embodiment, step S10 can employ Singular Spectrum Analysis (SSA) to perform data filtering and denoising by embedding and singular value decomposition of historical time series data, i.e., feature enhancement, effectively extracting the main signal (trend and periodic characteristic signals) components, removing noise and non-stationary trends, improving the time series quality of the data, and laying a solid foundation for subsequent modeling.
[0029] Step S20, Determining the time lag step: Perform detrending operation on the data processed in step S10, extract the long-range correlation in the time series, and then combine the partial autocorrelation function (PACF) to quantify the lag effect of historical time points on the current variable, thereby determining the time lag step of the gas utilization rate after the adjustment of the blast furnace operating parameters.
[0030] In step S20, self-trending analysis (DFA) can be used to detrend the data after feature enhancement processing to extract long-range correlations in the time series.
[0031] In this application, the purpose of detrending the data after feature enhancement is to demonstrate that historical timeframes are necessary in the prediction problem of GUR (qualitative analysis); the role of PACF is to determine which historical timeframes have the greatest impact on the present (quantitative analysis). Based on the feature enhancement in step S10, further analysis of the processed time series using detrending and partial autocorrelation function methods can effectively identify significant lag relationships in the series.
[0032] In one embodiment, step S20, determining the time delay step size, can be as follows: First, using the DFA method, calculate the logarithmic relationship between the fluctuation function F(s) and the window size s to obtain the slope α. Then, based on α, determine whether long-range correlation exists, i.e., whether a persistent, autocorrelated pattern exists; next, combine this with PACF to reveal the relationship between the current moment and the past... k The strength of the direct linear relationship between time points helps identify which historical time points have the greatest impact on the present, thereby determining the optimal input sequence length (i.e., delay step length) in the subsequently built Transformer model.
[0033] In this embodiment, the DFA method is introduced to analyze the long-range correlation of the gas utilization rate sequence, while PACF is used to determine the influence range of historical moments on the current moment, thereby determining the optimal time delay step size based on the Transformer time series prediction model. This analysis process not only considers the self-similarity of the time series but also effectively captures the memory characteristics of the system, which helps to improve the scientificity and rationality of the prediction time series window.
[0034] Step S30, Transformer model establishment and training: Using the data processed in step S10 as the dataset, the blast furnace operating parameters and the time delay step determined in step S20 as input features, and the gas utilization rate as the output feature, a pre-established Transformer model based on the attention mechanism is trained to obtain the blast furnace gas utilization rate prediction model.
[0035] In step S30, a Transformer network architecture based on a multi-head self-attention mechanism is first designed and established. This structure assigns different weights to key moments in the input sequence through the attention mechanism, thereby effectively modeling the nonlinear and long-term dependencies in the blast furnace gas utilization rate sequence. A position encoding module is added to enhance the model's ability to perceive time-series information. The next step is to use the feature data processed in steps S10 and S20 (using basic blast furnace operation parameters and a determined time delay step as input features, and GUR data as output features to construct multi-input single-output sample pairs) to train the Transformer model, thus obtaining the blast furnace gas utilization rate prediction model.
[0036] Furthermore, to improve model performance, after training, the Adam optimization algorithm can be used to adjust model parameters (primarily the Query, Key, and Value projection matrix parameters in the attention layer, positional encoding parameters, weights and biases of the feedforward neural network layer, and weights of the prediction output layer). This is combined with cross-validation and hyperparameter optimization techniques to ensure model convergence and optimal performance. In addition, the model can be evaluated and compared, involving in-depth analysis of experimental results and comparison with prediction results that do not consider time lag effects. The experimental results are analyzed in conjunction with prediction metrics, and the model structure and hyperparameters are adjusted based on the results to further optimize the Transformer network's performance and improve its prediction accuracy and stability on complex time-series data.
[0037] The intelligent prediction method for blast furnace gas utilization rate considering time lag effect provided in this application fully explores the key information in the historical data of gas utilization rate that has a delayed impact on the current state. Combined with advanced time series modeling technology, it can achieve accurate prediction of the trend of gas utilization rate change. In addition, by introducing a time lag feature modeling mechanism, the model's adaptability to the complex operating dynamics of the blast furnace is effectively improved, and the accuracy and stability of the prediction are enhanced, providing intelligent technical support for blast furnace operation optimization and energy efficiency improvement.
[0038] The present application will now be described in detail with reference to specific embodiments.
[0039] The blast furnace production process flow is as follows: Figure 2As shown. The blast furnace ironmaking process involves the interaction of descending and ascending gas flows within a high-temperature, high-pressure, airtight furnace. The principle is that 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 pig iron flows into the furnace bottom 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 the gas utilization rate (GUR), defined as follows:
[0040] here It is the gas utilization rate of the blast furnace. and Volume percentages of CO2 and CO in the top gas.
[0041] By predicting blast furnace gas utilization rates, steel companies can better plan production, optimize process parameters, reduce energy costs, and minimize unnecessary losses. Simultaneously, it can reduce dependence on fossil fuels and lower greenhouse gas emissions, contributing to sustainable production development.
[0042] The complex couplings and high-temperature, high-pressure environment within the blast furnace reaction process can significantly impact CO utilization, potentially generating noise and outliers. Singular spectral analysis (SSA), a time series analysis method, effectively extracts trend and periodic components from signals while filtering out noise. Therefore, it is necessary to utilize SSA to filter and denoise blast furnace gas utilization data.
[0043] The basic idea of Singular Spectrum Analysis (SSA) is to decompose a time series into several components with specific structures, which reflect different patterns of change in the series. SSA mainly consists of four steps: Step 1: Embedding in phase space: For the original GUR time series Map it to a length of The vector sequence forms A length of vector:
[0044] These vectors form the trajectory matrix:
[0045] here The window length is an integer multiple of the sampling period. . Step 2: Singular Value Decomposition: make ,but for eigenvalues, and ,and It is a matrix The orthogonal vectors corresponding to these eigenvalues.
[0046] make Then the trajectory matrix The SVD can be written as:
[0047] here, , .
[0048] Step 3: Construct subspace: Set the index Divided into Different subsets
[0049] make Then the corresponding Composite matrix Then there is
[0050] Step 4: Reconstruct: Reconstruct each matrix in the above formula. Transform into a length of The new sequence is obtained, that is, the decomposed sequence. Let For one A matrix with elements of .if , ,otherwise
[0051] make , Then, using the following formula to calculate the average of the diagonals, we can transform the matrix... Convert to sequence .
[0052] SSA (Separate Analytical Gas Analysis) can effectively separate and identify important information in the time series of gas utilization rates, enabling better analysis of relevant data. For example... Figure 3The images show the original GUR time series of blast furnace data and the GUR time series after SSA processing, respectively. Only the data preprocessing results of 2000 sample points are shown here; the actual number of samples is 8466.
[0053] 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:
[0054] here This represents the maximum value in the time series of gas utilization rate. This represents the maximum value in the time series of coal gas utilization rate. This represents the length of the time series.
[0055] This embodiment provides an analysis of the time lag effect on blast furnace gas utilization: Detrending volatility analysis (DFA) is an effective tool for analyzing long-range correlations in time series, particularly suitable for non-stationary time series data. Its core idea is to remove the global trend from the data through segmented detrending, thereby highlighting the residual volatility and long-range dependencies in the time series. In the analysis of the time lag effect of blast furnace gas utilization rate, DFA combined with the partial autocorrelation function (PACF) can effectively reveal the degree of influence of historical moments on the current gas utilization rate, and further analyze its long-range correlation structure at different time scales. DFA mainly extracts the long-range correlation in the time series through detrending operations, while PACF is used to quantify the lag effect of historical time points on the current variable. The combination of the two can accurately characterize the lag response characteristics of gas utilization rate after adjustment of operating parameters, providing a solid theoretical basis for constructing predictive models or control strategies that consider time lag effects. The mathematical framework of DFA is as follows: STEP 1: Cumulative Gas Utilization Rate Time Series. Assume the gas utilization rate time series is as follows: First, calculate the cumulative time series. :
[0056] in, The mean of the sequence STEP 2: Segment the time series. Divide the cumulative time series... Divided into lengths of Non-overlapping windows. A total of A window.
[0057] STEP3: For each window Using the least squares method to fit a local trend Then remove the trend from the accumulated signal of each window:
[0058] STEP 4: Calculate the fluctuation function. (For all windows) The average of the squared fluctuations is:
[0059] The fluctuation function is used to quantify how much residual fluctuation of a sequence is within a certain window length (time scale), thereby determining whether the entire sequence has long-range dependence.
[0060] The partial autocorrelation function (PACF) is used to measure the current value in a time series. and lag value of period The correlation between them is purely linear, eliminating interference from all intermediate lags. The formula is as follows:
[0061] in, Indicates PACF at the th k The value of the order; Based on of The least squares prediction; the same Based on of The least squares prediction value; () represents the correlation coefficient.
[0062] This embodiment performs in-depth analysis of the time series using the DFA and PACF methods. The DFA results are as follows: Figure 4 As shown, the fluctuation function of the sequence exhibits a power-law relationship with the window length, and the fitting slope α = 1.04, indicating that the sequence has strong long-range correlation. This suggests that the model needs to consider historical data during input to avoid omitting remote information that has a significant impact on the current state. Figure 5 The results of the PACF experiment are shown. Analysis reveals that the partial autocorrelation coefficients for lags 1 and 2 are significant, indicating that the first and second historical lag steps are most representative for predicting the current time. Combining the two methods, this embodiment selects the first two lag points as input features to better capture the dynamic characteristics of the sequence.
[0063] This embodiment provides a Transformer-based blast furnace gas utilization prediction model: The Transformer is a deep learning model based on an attention mechanism. Initially used for natural language processing tasks, it has since been widely applied to time series prediction. Compared to traditional recurrent neural networks (RNNs) and their variants such as LSTM, the Transformer model completely eliminates the recurrent structure, enabling parallel data processing and significantly improving training efficiency and the ability to model long-range dependencies. The Transformer model is particularly suitable for processing time series data with complex dynamic characteristics, such as blast furnace data. Its key advantage lies in its self-attention mechanism, which can dynamically allocate weights between different time steps, allowing for focused attention on important information. The basic structure of the Transformer is as follows: Figure 6 As shown, it mainly includes input-output embedding, encoder-decoder components, and logistic regression. When processing blast furnace data, the Transformer can flexibly model multi-scale, non-linear sequence patterns, thereby improving the accuracy and robustness of predictions. By stacking multiple encoder modules, the Transformer model can extract higher-order features layer by layer, effectively capturing the deep dynamic patterns in the blast furnace operation process.
[0064] Regarding the input / output embedding part: Blast furnace gas utilization data exhibits significant time-series characteristics, its values being influenced not only by current moment factors but also by the evolutionary process of historical states. Traditional recurrent neural networks (RNNs) implicitly incorporate time step position information into the network structure by iteratively processing the input sequence, thus possessing certain advantages in capturing sequence dependencies. However, the serial computational nature of RNN structures limits their efficiency in large-scale data processing, making it difficult to meet the real-time prediction needs of high-frequency, high-dimensional industrial data. To address this issue, Transformer abandons the serial structure of RNNs, employing a self-attention mechanism to achieve global modeling and parallel computation of information at each time step in the sequence, significantly improving computational efficiency and modeling capabilities. However, at the same time, the Transformer model itself lacks time sequence awareness and cannot directly understand the temporal relationship between data points in the sequence. To solve this challenge, Transformer introduces word embedding (WE) and positional encoding (PE) mechanisms. Specifically, the model first maps the original one-dimensional value of blast furnace gas utilization rate into a high-dimensional feature vector (e.g., 512-dimensional) through word embedding to enhance its expressive power. Then, using a set of positional encoding formulas based on sine and cosine functions, a unique and identifiable set of absolute positional information is generated for each time step in the sequence. Finally, the Transformer adds the positional encoding to the embedding vector bit by bit, enabling the model to explicitly perceive the temporal order features of the sequence without recursive structure, thereby more accurately capturing the dynamic evolution of blast furnace gas utilization rate. The positional encoding formula is as follows:
[0065]
[0066] In the formula, This is the position index of the data at a given time step within the input sequence; The dimension of the input sequence word embedding, Let be one dimension of the vector.
[0067] The core of the encoder-decoder section provided in this embodiment is the self-attention mechanism. The self-attention mechanism maps the input sequence to query-key-value pairs and calculates the dot product of each question and all keys to obtain weights, thereby learning the relative importance of each data point relative to all other data points in the sequence. Its scaled dot product attention calculation process is as follows:
[0068] In the formula: Represents a query; Representative key; Representative value, and in terms of word embedding dimension As a scaling factor, it can make the gradient more stable during training. The Transformer neural network, through its self-attention mechanism, makes the model pay more attention to the intrinsic relationships between the input sequence data, reducing the probability of the model making anomalous predictions. This provides a chance filtering capability for power load forecasting tasks, making the network model more stable and robust.
[0069] Unlike other models that use only a single attention module, the Transformer uses multi-head attention modules to operate in parallel. The original query, keys, and values are respectively used with the learned... H Projecting different vectors onto , and In a 3D space, attention weights are calculated for each projected query, key, and value according to the scaling dot product attention formula, and then output as attention weights. The model then concatenates all outputs and projects them back into the 3D space. Wei said:
[0070]
[0071] The structure of the Transformer neural network encoder is composed of Figure 6 The left side consists of the following four steps, assuming the input sequence is the blast furnace operating parameter, ore-coke ratio: (1) Perform word embedding and position encoding on the input sequence:
[0072] (2) Self-attention mechanism first calculates the query Key matrix Value matrix :
[0073] In the formula: , , The weight matrix is continuously learned and updated as the blast furnace gas utilization prediction model is trained. It is obtained by multiplying the embedded input sequence with the weight matrix and performing a linear transformation. , , .
[0074] Then, the self-attention score is calculated using the scaling dot product attention formula:
[0075] (3) Residual connectivity and layer normalization:
[0076] (4) A fully connected linear mapping with ReLU as the activation function is used to obtain the input of the encoder to the hidden layer:
[0077] Occlusion-based self-attention: The overall computational flow of the decoder is roughly the same as that of the encoder. Decoders in traditional sequence-to-sequence models often use RNN models, which are time-driven; during training, the model can only see the current... t The input at any given time has no future value in sight. However, during training, the entire sequence is exposed to the decoder, potentially leading to premature leakage of the ground truth. Therefore, a masking operation is needed to mask the input sequence to the decoder. The masked self-attention score is obtained by element-wise addition of the self-attention matrix and the masking matrix.
[0078] The logistic regression part provided in this embodiment consists of a linear transformation and a Softmax mapping. Its function is to regress the decoder's output vector space and remap it to the predicted probability of blast furnace gas utilization rate at the next moment.
[0079] It should be noted that the blast furnace gas utilization time series data used in this embodiment comes from the actual production process of a large domestic steel enterprise, and the data has strong representativeness and engineering application value. This dataset contains 8466 consecutive samples with a sampling interval of 1 minute, comprehensively recording the dynamic changes of the blast furnace during long-term operation. To ensure the effectiveness of model training and testing, the first 90% of the samples (7619 data points) were used as the training set, and the remaining 10% (847 data points) were used as the test set to examine the model's generalization ability on unseen data.
[0080] In constructing the DFA-PACF-Transformer blast furnace gas utilization rate prediction model, in addition to using the historical sequence of blast furnace gas utilization rate as the basic input variable, this embodiment further introduces various key blast furnace operating parameters to improve the model's input information dimensionality and prediction accuracy. Specifically, through Pearson correlation coefficient analysis, the linear correlation between multiple observable operating variables and the target variable was quantitatively screened, and nine core operating parameters with significant impact on blast furnace gas utilization rate were selected: blast pressure, heat load, fuel ratio, coke ratio, ore-coke ratio, blast kinetic energy, coal ratio, furnace top temperature, and furnace top pressure. These parameters comprehensively reflect the blast furnace operating characteristics from multiple perspectives, including blast furnace thermal state, raw material and fuel ratio, and furnace top operating conditions.
[0081] Based on the above processing, the final model's input features include: gas utilization rates for the first two time points (with the optimal lag step determined by DFA and PACF analysis) and nine blast furnace operating variables, forming a prediction network architecture with a total of 11 input dimensions and 1 output dimension. The output is the predicted gas utilization rate for a future time point. This fusion structure, while retaining the Transformer model's temporal modeling capabilities, further enhances the model's responsiveness and prediction accuracy to complex blast furnace operating conditions through feature enhancement.
[0082] To accurately distinguish between the precision and accuracy of the prediction results of the proposed model, the root mean square error (RMSE), mean absolute error (MAE), and coefficient of determination R0 were used respectively. 2 (R-Square) is used as the evaluation standard.
[0083]
[0084]
[0085]
[0086] in and These are the actual value and the predicted value, respectively. This represents the average of the actual values of blast furnace gas utilization rate. This represents the number of samples.
[0087] This embodiment presents the time series prediction results of blast furnace gas utilization rate based on the DFA-PACF-Transformer model considering time delay characteristics. To verify the importance and effectiveness of time delay characteristics in the modeling process, this embodiment constructs two sets of comparative experiments: one is a Transformer model constructed by introducing the optimal lag step size determined by DFA fractal analysis and PACF partial autocorrelation function; the other is a standard Transformer model that does not consider time delay characteristics and only uses the current input features for prediction. Specific experimental results are as follows: Figure 7As shown, the comparison of prediction results between the Transformer model incorporating time lag (red solid line) and the Transformer model without time lag (green dashed line) is illustrated. It is evident that the Transformer model incorporating time lag better matches the actual values at multiple fluctuation points (black solid line), especially demonstrating greater accuracy in the large fluctuations within the 600-700 range. The model considering time lag is more sensitive to capturing the changing trends and local fluctuations in gas utilization rates, effectively reflecting the dynamic characteristics of the system. In contrast, the Transformer model without time lag exhibits offset or amplitude errors at its peaks, indicating that ignoring historical lag factors affects the model's prediction accuracy. In summary, incorporating PACF (partial autocorrelation function) to extract time lag characteristics significantly enhances the time series modeling capability of the Transformer model, improving the prediction's fit and robustness.
[0088] Furthermore, considering the time lag effect, to further evaluate the performance difference between deep learning methods and traditional machine learning methods in predicting blast furnace gas utilization, this embodiment also introduces the classic Support Vector Machine (SVM) model for comparison. All models were tested on the same dataset, with a unified preprocessing procedure and the same training and testing partitioning. Cross-validation and grid search were used to ensure optimal parameter configuration for each model. Specific experimental results are shown in Figure 8. The figure shows that the Transformer model's overall prediction curve fits the true curve better, especially in the fitting accuracy of peak and trough values. While the SVM can capture the overall trend, it exhibits significant deviations in several places, showing some underfitting issues, particularly in the 200-750 range where the fluctuation amplitude control is not accurate enough. The Transformer, relying on multi-head attention and the introduction of time lag features, possesses stronger temporal modeling and nonlinear fitting capabilities, significantly outperforming the SVM model based on a fixed kernel function. Therefore, the Transformer model that incorporates time delay not only has better performance under the deep learning framework, but also significantly outperforms traditional machine learning methods, making it more suitable for industrial data prediction tasks with complex dynamic characteristics, such as blast furnace gas utilization rate.
[0089] To further quantitatively verify the effect of introducing time lag characteristics on improving model performance, this embodiment compares three key evaluation metrics of different models on the test set: RMSE, MAE, and R². The results are shown in Table 1-1 below. As can be seen from Table 1-1, after introducing time lag, the Transformer model shows significant optimization in all three metrics. RMSE and MAE decreased by approximately 54.9% and 52.4%, respectively, and R² increased from 0.9694 to 0.9938, indicating that the model's fitting ability and generalization ability for gas utilization rate are greatly enhanced. Under the same lag characteristic conditions, the Transformer model has a significant advantage over the traditional SVM model. RMSE and MAE decreased by 72.3% and 68.4%, respectively, and R² increased by approximately 7.4%, indicating that deep time series modeling methods can better characterize the nonlinear and dynamic features in the data. Although the Transformer without considering time lag is superior to SVM, it still has significant disadvantages in many metrics, further verifying that the time lag information extracted by introducing PACF is a key factor in improving prediction performance.
[0090] Table 1-1 Key Evaluation Indicators for Different Models
[0091] In summary, the quantitative experimental results are consistent with the image fitting results, which fully demonstrates that in the task of predicting blast furnace gas utilization, the introduction of time delay characteristics can significantly improve the accuracy and stability of the model, and the Transformer model has more advantages than traditional methods.
[0092] This embodiment combines the ironmaking process mechanism with the time series characteristics of blast furnace gas utilization rate (GUR). Based on real blast furnace operation data from a steel plant, it extracts time-lag features using the partial autocorrelation function (PACF) and incorporates them into a Transformer deep prediction model, achieving high-precision modeling and prediction of blast furnace GUR. Experimental comparisons verify that the Transformer model incorporating time lags outperforms traditional models in both accuracy and stability. The main contributions of this embodiment are as follows: (1) Based on the time dependence characteristics of blast furnace gas utilization data, this embodiment introduces detrended fluctuation analysis (DFA) and partial autocorrelation function (PACF) to extract lag features, which effectively improves the model's ability to dynamically model time series.
[0093] (2) This embodiment proposes a Transformer prediction model based on time delay feature enhancement, which has better nonlinear fitting ability and generalization ability than traditional methods, faster training speed, and significantly improves the accuracy and stability of GUR prediction.
[0094] (3) This embodiment uses real blast furnace operation data from a steel company. The comparative experimental results show that the Transformer model with time delay is significantly better than the model without time delay and the traditional SVM model, which verifies the feasibility and effectiveness of this embodiment in predicting actual blast furnace operation.
[0095] The main improvements in this embodiment are as follows: 1) A method for predicting blast furnace gas utilization rate by combining Partial Autocorrelation Analysis (PACF) with time-lag feature enhancement and Transformer structure is proposed. PACF is used to extract significant lag features from the GUR time series, clarify the time dependence of variables, and make the input features more time-series oriented, providing a more representative modeling foundation for the Transformer model. 2) This embodiment innovatively introduces time-lag features into the input of the Transformer model, improving the model's ability to model the nonlinear changes and long- and short-term dependencies of blast furnace gas utilization rate. Through the self-attention mechanism, the Transformer can accurately capture the global dependencies between features, achieving more efficient time series modeling. 3) This embodiment is experimentally verified based on real blast furnace operation data. The Transformer model with time-lag features introduced outperforms the model without time-lag features and the traditional SVM method in evaluation indicators such as RMSE, MAE, and R², demonstrating the feasibility, superiority, and engineering promotion value of this method in actual industrial data. 4) This embodiment constructs a Transformer-based intelligent prediction framework for blast furnace gas utilization, which has strong real-time prediction capabilities and engineering deployability. It can provide accurate data support and optimization suggestions for blast furnace operation, and help achieve efficient utilization of gas resources and intelligent control of the smelting process.
[0096] The advantages of this embodiment are as follows: 1) In existing technologies, most recurrent neural networks such as LSTM and GRU are used for time series modeling. Although they can model certain time dependencies, they are difficult to effectively capture long-distance global features and have low training efficiency. In contrast, this embodiment introduces a Transformer model with a self-attention mechanism, which can process sequence data in parallel and efficiently model long-term time dependencies. Combined with significant time lag features, it further enhances the ability to model the trend of blast furnace gas utilization, thereby significantly improving prediction accuracy and generalization ability. 2) Unlike existing methods that do not consider time lag features or only manually set lag terms, this embodiment systematically extracts key lag features related to gas utilization through partial autocorrelation analysis (PACF) and combines it with end-to-end modeling of the Transformer model, effectively improving the expressiveness and time-series relevance of the model input. This approach not only improves the stability and noise resistance of the model but also avoids information loss or error accumulation caused by subjectively setting lag terms, making it more adaptable and practical. Those skilled in the art will readily understand that the above description is merely a preferred embodiment of this application and is not intended to limit this application. Any modifications, equivalent substitutions, and improvements made within the spirit and principles of this application should be included within the protection scope of this application.
Claims
1. A method for constructing a blast furnace gas utilization rate prediction model considering time delay effects, characterized in that, Includes the following steps: S10: Obtain and filter the historical time series data of blast furnace gas utilization rate, and extract the trend and periodic feature signals in the data. S20, perform detrending operation on the data processed in step S10, extract long-range correlation in the time series, and then combine the partial autocorrelation function to quantify the lag effect of historical time points on the current variable, thereby determining the time lag step of gas utilization rate after the adjustment of blast furnace operating parameters. S30: Using the data processed in step S10 as the dataset, the blast furnace operating parameters and the time delay step determined in step S20 as input features, and the gas utilization rate as the output feature, a pre-established Transformer model based on the attention mechanism is trained to obtain a blast furnace gas utilization rate prediction model.
2. The method for constructing a blast furnace gas utilization rate prediction model considering time delay effects as described in claim 1, characterized in that, The blast furnace operating parameters include air pressure, heat load, fuel ratio, coke ratio, ore-coke ratio, blast kinetic energy, coal ratio, furnace top temperature, and furnace top pressure.
3. The method for constructing a blast furnace gas utilization rate prediction model considering time delay effects as described in claim 1, characterized in that, In step S10, the singular spectral analysis method is used to filter and denoise the historical time series data by embedding and singular value decomposition.
4. The method for constructing a blast furnace gas utilization rate prediction model considering time delay effects as described in claim 1, characterized in that, Step S10 is as follows: S11, historical time series data of blast furnace gas utilization rate Mapped to a length of The vector sequence, taking an integer multiple of the sampling period. A length of vector And these vectors are used to form a trajectory matrix. ; S12, let ,but for eigenvalues, and ,and It is a matrix The orthogonal vectors corresponding to these eigenvalues; And order Then the trajectory matrix SVD is ;in, , ; S13, set the index Divided into Different subsets ,make Then the corresponding Composite matrix Then there is ; S14, each matrix in the above formula Transform into a length of The new sequence is obtained, that is, the decomposed sequence, and then let For one A matrix with elements of .if , ,otherwise ; And order , Then, using the following formula to calculate the average of the diagonals, we can transform the matrix... Convert to sequence ; 。 5. The method for constructing a blast furnace gas utilization rate prediction model considering time delay effects as described in claim 1, characterized in that, In step S20, the steps for detrending the data processed in step S10 and extracting long-range correlations from the time series are as follows: First, autotrend analysis is used to calculate the logarithmic relationship between the fluctuation function F(s) and the window size s to obtain the slope α; then, based on α, it is determined whether there is a long-range correlation and the long-range correlations from the time series are extracted.
6. The method for constructing a blast furnace gas utilization rate prediction model considering time delay effects as described in claim 1, characterized in that, In step S20, the detrending operation is performed on the data processed in step S10 to extract the long-range correlation in the time series. Specifically, this step involves: S21, Cumulative Gas Utilization Rate Time Series: Assuming the gas utilization rate time series is... First, calculate the cumulative time series. , The mean of the sequence; S22, Split the time series: Divide the cumulative time series Divided into lengths of There are a total of non-overlapping windows. One window; S23, for each window Using the least squares method to fit a local trend Then, the trend is removed from the accumulated signal of each window to obtain the first... v Standard deviation of detrended volatility for each window : S24, Wave function calculation: for all windows The average of the squared fluctuations is: 。 7. The method for constructing a blast furnace gas utilization rate prediction model considering time delay effects as described in claim 1, characterized in that, In step S20, the formula for calculating the partial autocorrelation function is: In the formula, This indicates the degree to which the k-th historical data point has an impact on the current moment; Based on of The least squares prediction; the same Based on of The least squares prediction value; () represents the correlation coefficient.
8. The method for constructing a blast furnace gas utilization rate prediction model considering time delay effects as described in claim 7, characterized in that, Step S30, following the step of training the pre-built attention-based Transformer model, also includes: The Adam optimization algorithm is used to adjust the model parameters, and cross-validation and hyperparameter optimization techniques are combined to ensure model convergence and achieve optimal performance. The adjusted model parameters include the Query, Key, and Value projection matrix parameters and position encoding parameters in the attention layer, as well as the weights and biases of the feedforward neural network layer and the weights of the prediction output layer.
9. The method for constructing a blast furnace gas utilization rate prediction model considering time delay effects as described in claim 7, characterized in that, Before obtaining the blast furnace gas utilization rate prediction model in step S30, the following steps are also included: The experimental results were analyzed in depth and compared with the prediction results without considering the time delay effect. The experimental results were analyzed in depth in combination with the prediction indicators, and the model structure and hyperparameters were adjusted according to the experimental results.
10. A system for constructing a blast furnace gas utilization rate prediction model considering time delay effects, characterized in that, include: The data processing module is used to acquire and filter and denoise the time series data of blast furnace gas utilization rate, and extract trend and periodic feature signals from the data. The time lag step extraction module is used to perform detrending operations on the data processed by the data processing module, extract long-range correlations in the time series, and then combine the partial autocorrelation function to quantify the lag effect of historical time points on the current variable, thereby determining the time lag step of the gas utilization rate after the adjustment of the basic parameters of blast furnace operation. The training module uses the data processed by the data processing module as the dataset, the basic parameters of blast furnace operation and the time delay step determined by the time delay step extraction module as input features, and the gas utilization rate as the output feature to train a pre-established Transformer model based on the attention mechanism, thereby obtaining a blast furnace gas utilization rate prediction model.