Prediction and control method for precise tapping of blast furnace

Abstract

The invention belongs to the technical field of intelligent prediction of blast furnace ironmaking, and discloses a blast furnace accurate tapping prediction and control method and system for optimizing a long short-term memory network (LSTM) based on principal component analysis (PCA) and a sparrow search algorithm (SSA). According to the method, multi-source data such as raw material conditions and operation parameters are acquired through a data acquisition module, abnormal values and missing values are processed and normalized through a preprocessing module, feature dimension reduction is performed by using a PCA module to eliminate variable collinearity, global optimization is performed on LSTM hyper-parameters by means of an SSA module to overcome the defect of manual parameter adjustment, and the accuracy of parameter adjustment is improved. An SSA-LSTM combined model is constructed to achieve accurate prediction of the tapping time and the tapping amount; and meanwhile, the opening degree of the tapping hole is dynamically adjusted through the hydraulic driving device, and automatic control over the tapping process is achieved in combination with integrated data acquisition, intelligent prediction, dynamic regulation and control and safety protection functions, so that the problem that traditional tapping depends on experience is solved.

Description

Technical Field

[0001] The present invention relates to the technical field of intelligent prediction for blast furnace ironmaking, and specifically to a blast furnace iron tapping prediction method and control system based on principal component analysis and sparrow search algorithm optimization of long short-term memory network. Background Art

[0002] Predicting blast furnace iron output is a key component of the intelligent upgrade of the steel industry. Accurate predictions optimize production scheduling, reduce energy costs, and ensure safe equipment operation. However, the blast furnace smelting process exhibits strong nonlinearity, multivariable coupling, and hysteresis effects. Iron output is influenced by the interactions of dozens of factors, including raw material composition, blast parameters, and furnace temperature fluctuations. Traditional mechanistic models struggle to accurately model this process, and single data-driven models also suffer from issues such as feature redundancy and parameter sensitivity.

[0003] In the existing technology, on the one hand, the use of models such as BP neural network, random forest and support vector regression to predict the amount of iron tapping is susceptible to feature redundancy, parameter sensitivity and insufficient ability to capture long-term dependencies when processing high-dimensional time series data. Although the long short-term memory network (LSTM) has advantages in time series modeling, its hyperparameters rely on experience-based tuning, resulting in prediction accuracy and robustness that need to be improved. On the other hand, the means of controlling the iron tapping process are limited. The iron tapping speed is difficult to accurately control. Too fast an iron tapping speed will cause pressure fluctuations in the furnace, affect the life of the furnace lining, and may also cause the molten iron to carry too much slag; too slow an iron tapping speed will extend the iron tapping time and reduce production efficiency. Therefore, there is an urgent need for a blast furnace iron tapping prediction method and control system that can effectively process high-dimensional data and optimize model parameters. Summary of the Invention

[0004] The purpose of the present invention is to provide an effective method and control system for predicting blast furnace iron tapping, to solve the problems of feature redundancy, parameter sensitivity and insufficient ability to capture long-term dependence in the prior art, to improve the prediction accuracy and robustness of blast furnace iron tapping, and to effectively control the iron tapping process, to improve the stability and production efficiency of blast furnace iron tapping, and to reduce production costs.

[0005] To achieve the above object, the present invention provides the following technical solutions:

[0006] 1. The blast furnace iron tapping prediction system includes: data acquisition module, data preprocessing module, principal component analysis (PCA) module, sparrow search algorithm (SSA) optimization module, long short-term memory network (LSTM) module and prediction result output module.

[0007] The data acquisition module is used to collect multi-source data during the blast furnace smelting process, including process parameters such as raw material conditions, operating parameters, status indicators, and iron tapping records.

[0008] The data preprocessing module is used to perform outlier processing, missing value filling and normalization processing on the collected data.

[0009] The principal component analysis module is used to perform feature dimensionality reduction on the preprocessed data, determine whether the data is suitable for PCA through KMO suitability test and Bartlett sphericity test, standardize the suitable data, calculate the covariance matrix and perform eigenvalue decomposition, and select the principal component with a cumulative contribution rate greater than 85% as the key feature.

[0010] The sparrow search algorithm module is used to optimize the hyperparameters of LSTM, including the number of hidden layer nodes, learning rate, number of iterations, etc., and searches for the optimal hyperparameter combination in the solution space by simulating the foraging behavior of a sparrow group.

[0011] The long short-term memory network module is used to build an LSTM prediction model, use the hyperparameters optimized by SSA to train the model, and output the predicted value of the blast furnace iron output by inputting the key features extracted by PCA.

[0012] The prediction result output module is used to perform denormalization processing on the prediction value, restore it to the actual dimension, and output the prediction result.

[0013] A method for predicting blast furnace iron tapping, comprising the following steps:

[0014] S1: Data acquisition and preprocessing

[0015] Multi-source data from the blast furnace smelting process was collected and processed for outlier handling, missing value filling, and normalization. Outlier handling used the 3σ criterion, replacing outliers with the mean. Missing value handling used linear interpolation, multiple imputation, and mode interpolation, depending on the data type. Normalization used the Min-Max method, mapping the data to the [0, 1] interval.

[0016] S2: Principal Component Analysis (PCA) feature dimensionality reduction

[0017] The preprocessed data were subjected to KMO suitability test and Bartlett sphericity test. If the KMO value was greater than 0.7 and the Bartlett test was significant, the data were standardized, the covariance matrix was calculated and the eigenvalue decomposition was performed. The data were sorted from large to small by eigenvalue, and the principal components with a cumulative contribution rate greater than 85% were selected as key features.

[0018] S3: Sparrow Search Algorithm (SSA) Optimize LSTM Hyperparameters

[0019] Initialize the sparrow population, and each sparrow position corresponds to a set of LSTM hyperparameter combinations. Use the prediction error (mean square error MSE) as the fitness function, and perform global optimization through the position update strategy of the discoverer, follower, and sentinel to obtain the optimal hyperparameter combination.

[0020] S4: LSTM model training and prediction

[0021] An LSTM network was constructed, and the key features extracted by PCA were input. The model was trained using the hyperparameters optimized by SSA. The input sequence was generated through a sliding window, and the dataset was divided into a training set and a test set. After training, the test set was predicted and the predicted values ​​were denormalized to obtain the actual predicted value of the blast furnace iron output.

[0022] 2. Tapping process control system: This consists of a taphole control device, a molten iron flow control device, and an automated control unit. The taphole control device precisely controls the opening and closing of the taphole based on the predicted tapping time, using either hydraulic or electric drive to ensure accuracy and stability. The molten iron flow control device, installed in the tapping trough, monitors the molten iron flow in real time and precisely controls it by adjusting the taphole size and using flow control valves to ensure a stable tapping speed. The automated control unit receives signals from the prediction system and feedback from various devices to automatically control the entire tapping process, achieving precise and intelligent tapping operations.

[0023] 3. Supporting Monitoring and Feedback System: Equipped with multiple sensors, the system monitors key parameters of the tapping process (such as molten iron temperature, slag content, and taphole conditions) in real time, feeding this data back to the data analysis module and automated control unit. The automated control unit then adjusts the tapping process control strategy based on this feedback, achieving closed-loop control and ensuring the tapping process is always in optimal condition. BRIEF DESCRIPTION OF THE DRAWINGS

[0024] In order to more clearly illustrate the embodiments of the present invention or the technical solutions in the prior art, the following briefly introduces the drawings required for use in the embodiments. Obviously, the drawings described below are only some embodiments of the present invention. Those skilled in the art can also derive other drawings based on these drawings without inventive work.

[0025] Figure 1 : Flowchart of the embodiment of the patent of this invention;

[0026] Figure 2 : PCA-SSA-LSTM model data processing flow chart;

[0027] Figure 3 :Schematic diagram of LSTM network structure;

[0028] Figure 4 : Comparison chart of multi-model prediction results; DETAILED DESCRIPTION

[0029] Example 1: System composition, process and parameter settings

[0030] The data flow chart of the blast furnace iron tapping prediction model is composed of Figure 1 As shown in the figure, the data acquisition module collects key data such as furnace temperature, pressure, and charge composition in real time by deploying temperature sensors (measuring range 0-2000℃, accuracy ±1℃), pressure sensors (measuring range 0-5MPa, accuracy ±0.01MPa), and composition analyzers (detecting the content of elements such as iron, silicon, and calcium, with an accuracy of ±0.1%) at the blast furnace body, furnace top, and iron outlet. At the same time, it integrates multiple process parameters such as blast furnace raw material conditions (coke load, ore grade, etc.), operating parameters (wind temperature, wind pressure, etc.), status indicators (furnace top temperature, permeability index, etc.), and iron tapping records (iron tapping time, iron tapping amount, iron tapping speed, etc.), and stores real-time monitoring data and historical iron tapping data in a database, providing multi-dimensional data support for subsequent analysis and processing.

[0031] The data preprocessing module processes the collected data. First, the 3σ criterion is used to detect and process outliers, and the outliers are replaced by the mean. For missing values, linear interpolation is used for short time series data, multiple filling method is used for long time series data, and mode interpolation method is used for no time series data. Finally, the Min-Max normalization method is used to map the data to the [0,1] interval.

[0032] The principal component analysis module performed KMO and Bartlett's sphericity tests on the preprocessed data. The KMO value was 0.78, and the Bartlett's P value was 0, indicating that the data were suitable for PCA. The data was standardized, the covariance matrix was calculated, and eigenvalue decomposition was performed. The first 15 principal components with a cumulative contribution of 86.72% were selected as key features.

[0033] The sparrow search algorithm module initializes 50 sparrows, sets the maximum number of iterations to 100, uses MSE as the fitness function, and optimizes the LSTM hyperparameters such as the number of hidden layer nodes, learning rate, and number of iterations through the update strategies of the discoverer, follower, and guard.

[0034] The LSTM module constructs an LSTM network including input gate, forget gate, output gate and cell state. Its core flow chart is as follows: Figure 1 As shown. The forget gate processes the hidden state h of the previous moment through the σ activation function t-1 and the current input x t , the output result is the same as the cell state C at the previous moment t-1Multiply and decide to keep C t-1 What information is updated? Then, the input gate determines the update content through the σ activation function, and uses tanh to generate candidate values. The two are multiplied to obtain the information to be updated. Then, the forget gate output is combined with C t-1 The product of , plus the updated information generated by the input gate, is used to obtain the current cell state C t ; Finally, the output gate processes h through the σ activation function t-1 and x t , the result is similar to tanh(C t ) to generate the current hidden state h t , and output h t and C t For subsequent calculations, 15 principal component features were input and trained using the optimized hyperparameters. The ratio of the training set to the test set was 8:2. After training, predictions were made on the test set, and the prediction results were output after denormalization.

[0035] Tapping process control system: A hydraulically driven taphole control device is installed. Its cylinder has a maximum thrust of 500kN and a stroke of 500mm, enabling precise control of the taphole opening angle (with an accuracy of ±1°). An electromagnetic flowmeter is installed in the tapping trough as a molten iron flow control device, with a measurement range of 0-100t / h and an accuracy of ±0.5%. The automated control unit uses a programmable logic controller (PLC), which implements automated control of the taphole control device and the molten iron flow control device through a programmed control program.

[0036] Supporting monitoring and feedback systems: Infrared temperature sensors are installed in the taphole to measure the molten iron temperature (measuring range: 1300-1500°C, accuracy: ±5°C). Image recognition technology is used to monitor the slag content in the molten iron (detection accuracy: ±2%). A displacement sensor is installed near the taphole to monitor taphole wear (accuracy: ±0.1mm). These monitoring data are fed back to the PLC and data analysis module in real time via a data transmission network.

[0037] The overall system process of the invention begins with the data acquisition module acquiring raw data, which is then cleaned and converted by the data preprocessing module to improve quality, and then simplified by the principal component analysis (PCA) module for dimensionality reduction. Subsequently, the model parameters are optimized using the sparrow search algorithm (SSA) optimization module, and input into the LSTM prediction module for prediction. The dynamic control system implements control based on the prediction results, and finally the supporting detection and feedback system detects the effect and provides feedback. The feedback information prompts the front-end modules (such as data preprocessing, model optimization, etc.) to adjust, forming a cyclic iteration, realizing a complete closed loop from data processing, analysis, prediction to control, and ensuring continuous optimization and dynamic adaptation of the system.

[0038] Example 2: Prediction method steps

[0039] Data collection and preprocessing: Data were collected from 00:00 on January 5, 2025 to 24:00 on April 8, 2025, and a total of 6557 groups of valid samples were obtained. The data were processed for outlier processing, missing value filling and normalization.

[0040] Principal component analysis: KMO values ​​and Bartlett's test statistics were calculated to confirm that the data were suitable for PCA. Standardization, covariance matrix calculation, and eigenvalue decomposition were performed to select the first 15 principal components.

[0041] SSA optimizes LSTM hyperparameters: Through population initialization, discoverer update, follower update, guard update and iterative optimization, the optimal hyperparameter combination is obtained: the number of hidden layer nodes is 128, the learning rate is 0.001, and the number of iterations is 200.

[0042] LSTM model training and prediction: Build an LSTM model, input the principal component features, and use the optimized hyperparameters to train the model. The mean absolute percentage error (MAPE) on the test set is 7.93%, and the root mean square error (RMSE) is 52.7 tons. The prediction accuracy is significantly better than the traditional model. Figure 3 shown.

[0043] Tap preparation stage: Operators prepare for tapping in advance based on the predicted tapping time, such as preparing the molten iron ladle and cleaning the tapping yard. 30 minutes before tapping, the automated control unit initiates the preheating process for the taphole control device to ensure smooth taphole opening.

[0044] Tapping process control stage: When the predicted tapping time is reached, the automated control unit controls the tapping hole control device to slowly open the tapping hole, while the molten iron flow control device monitors the molten iron flow in real time. According to the preset tapping speed (such as 5-8t / min), the automated control unit accurately controls the molten iron flow by adjusting the tapping hole opening angle and the opening of the flow control valve. During the tapping process, the supporting monitoring and feedback system monitors the molten iron temperature, slag content, and tapping hole condition in real time. If the molten iron temperature is too low or the slag content is too high, the automated control unit adjusts the tapping speed or takes appropriate slag removal measures; if the tapping hole shows abnormal wear, an alarm is promptly issued and the tapping strategy is adjusted.

[0045] At the end of tapping, when the predicted tapping volume is nearing completion or a significant decrease in molten iron flow is detected, the automated control unit controls the taphole control device to gradually close the taphole until it is completely closed. After tapping is complete, the tapping data is recorded and analyzed to optimize the prediction model and tapping process control strategy.

[0046] The above description is merely a preferred embodiment of the present invention and is not intended to limit the present invention. Although the present invention has been described in detail with reference to the aforementioned embodiments, those skilled in the art will be able to modify the technical solutions described in the aforementioned embodiments or replace some of the technical features therein with equivalents. Any modifications, equivalent replacements, improvements, etc. made within the spirit and principles of the present invention shall be included within the scope of protection of the present invention.

Claims

1. A blast furnace iron tapping prediction method and control system, characterized in that: The method comprises the following steps and devices: Data acquisition module: collects multi-source data during the blast furnace smelting process, including raw material conditions, operating parameters, status indicators, iron tapping records and other process parameters; Data preprocessing module: performs outlier processing, missing value filling and normalization on the collected data; Principal component analysis module: performs feature dimensionality reduction on preprocessed data and extracts key features; Sparrow search algorithm optimization module: Global optimization of the hyperparameters of the long short-term memory network; LSTM prediction module: Builds an LSTM network, uses optimized hyperparameters to train the model, and outputs prediction results. Dynamic control system: adjusts the taphole opening, blast parameters and slag-iron separation equipment based on the prediction results to maintain the target tapping speed and implement safety protection.

2. The blast furnace iron tapping prediction method and system according to claim 1, characterized in that: The data preprocessing module includes: Outlier processing unit: uses the 3σ criterion to identify outliers and replaces them with the mean; Missing value filling unit: linear interpolation, multiple filling and mode interpolation methods are used according to the data type; Normalization unit: Min-Max normalization method is used to map the data to the [0,1] interval.

3. The blast furnace iron tapping prediction method and system according to claim 1, characterized in that: The PCA module includes: Suitability test unit: Use KMO test and Bartlett's sphericity test to determine whether the data is suitable for dimensionality reduction; Feature dimensionality reduction unit: After standardizing the data, the covariance matrix is ​​calculated, eigenvalue decomposition is performed, and the principal components with a cumulative contribution rate of >85% are selected as input features.

4. The blast furnace iron tapping prediction method and system according to claim 1, characterized in that: The SSA optimization module includes: Population initialization unit: randomly generates a sparrow population, each individual corresponds to an LSTM hyperparameter combination; Fitness function unit: uses mean square error as fitness value to evaluate prediction error; Dynamic optimization unit: updates the positions of the sparrows through the "discoverer-follower-watcher" mechanism and searches for the optimal hyperparameter combination.

5. The blast furnace iron tapping prediction method and system according to claim 1, characterized in that: The LSTM prediction module includes: Network construction unit: LSTM network structure including input gate, forget gate, output gate and cell state; Time series modeling unit: Generates input sequences through a sliding window and divides them into training and test sets (ratio 8:2); Prediction output unit: Denormalize the prediction results and restore them to the actual dimension of the predicted iron output.

6. The method according to claim 1, wherein The dynamic control system dynamically adjusts the taphole opening through a hydraulic servo mechanism with an adjustment accuracy of ±1°, and is equipped with a pressure over-limit alarm and a taphole wear detection mechanism, and optimizes the reduction reaction in the furnace in combination with the blast oxygen enrichment rate.

7. The method according to claim 1 further includes a closed-loop feedback step, which compares the real-time monitoring data with the predicted value, and triggers the online update of the model and optimization of the control strategy when the error is large.

Images