Ladle refining furnace molten steel temperature prediction method and system based on LSTM-CBR dynamic adjustment

Abstract

The invention relates to a ladle refining furnace molten steel temperature prediction method and system based on LSTM-CBR dynamic adjustment, and the method comprises the steps: obtaining LF refining process production data, carrying out the preprocessing, and building a molten steel production database; production process control parameters are determined by combining metallurgy mechanism analysis and big data mining, and the production process control parameters are divided according to different time sequences and smelting stages. Related parameters of the molten steel production database are transmitted and stored according to the production process control parameters, and a temperature narrow window database is formed. And extracting time sequence characteristics of each control parameter in the temperature narrow window database by using an LSTM model, and generating a dynamic weight vector so as to construct a dynamic parameter adjustment rule base. And according to collection input of operation data in the current smelting process, case retrieval is carried out on the temperature narrow window database through a CBR algorithm, similarity calculation weighting is carried out on historical cases by utilizing dynamic weights of a dynamic parameter adjustment rule base based on the current case, the optimal case furnace number is screened out, and prediction of the LF refining end point molten steel temperature is achieved.

Description

Technical Field

[0001] This invention belongs to the field of iron and steel metallurgy technology, specifically relating to a method and system for predicting the temperature of molten steel in a ladle refining furnace based on LSTM-CBR dynamic adjustment. Background Technology

[0002] In modern steelmaking processes, LF refining equipment has wide and effective applications, providing functions such as steel deoxidation, desulfurization, alloying, heating, and inclusion control. Temperature control is crucial in LF refining to ensure steel quality and smooth production flow. Currently, companies primarily use on-site temperature guns to obtain steel temperature data for production guidance. However, long-term use of thermocouples incurs significant costs and makes it difficult to provide real-time guidance for process temperature control. Furthermore, a large amount of valuable process temperature data remains unused, hindering further improvement in technological capabilities. Therefore, accurate prediction of steel temperature during the LF refining process is of paramount importance.

[0003] In existing technologies, traditional temperature prediction methods mainly rely on metallurgical mechanism models, static data-driven models, and hybrid data-mechanism models, but each has its shortcomings. Because some data in the smelting process cannot be quantitatively determined, mechanism models require numerous assumptions and simplifications, making it difficult to accurately describe complex physicochemical reactions and resulting in low prediction accuracy. Static data-driven models, such as traditional neural networks or case-based reasoning, cannot dynamically adapt to real-time parameter changes during smelting, leading to prediction bias and failing to provide a reasonable explanation for process changes at the mechanistic level. Furthermore, existing methods do not effectively combine temporal characteristics with case knowledge, making it difficult to balance long-term process patterns and short-term operational responses, and lacking a dynamic adjustment mechanism. Hybrid data-mechanism models can, to some extent, compensate for the errors in the assumptions and simplifications of mechanism models, but still only reflect the temperature of molten steel under static conditions, failing to reflect temperature changes during the actual dynamic smelting process. Therefore, it is necessary to develop a new method and system to overcome the existing problems. Summary of the Invention

[0004] To overcome the aforementioned problems in the existing technology, this invention provides a method and system for predicting the temperature of molten steel in a ladle refining furnace based on LSTM-CBR dynamic adjustment, which is used to solve the aforementioned problems in the existing technology.

[0005] A method for predicting the temperature of molten steel in a ladle refining furnace based on LSTM-CBR dynamic adjustment includes the following steps:

[0006] S1. Obtain LF refining production data and preprocess it to obtain a molten steel production database;

[0007] S2. Determine the production process control parameters that affect the temperature of molten steel, and divide the production process control parameters according to different time sequences and smelting stages;

[0008] S3. Based on the production process control parameters obtained after the division, the relevant parameters of the molten steel production database are transmitted and stored to form a temperature narrow window database;

[0009] S4. Use the LSTM model to extract the time-series features of each control parameter in the narrow temperature window database, generate dynamic weights that reflect the influence of each parameter on temperature, and construct a dynamic parameter adjustment rule base.

[0010] S5. Based on the collected and inputted smelting process operation data, the CBR algorithm is used to search for cases in the dynamic parameter adjustment rule base based on the current case, and the most similar furnace is determined to achieve dynamic temperature forecasting.

[0011] S6. Learn new case data online, fine-tune LSTM model parameters, and update the temperature narrow window database;

[0012] S7. Accurately predict the final steel temperature of LF refining using the trained LSTM-CBR model.

[0013] In addition to the aspects and any possible implementations described above, a further implementation is provided, wherein the preprocessing includes determining the data start and end points, data cleaning, data normalization, and time series data construction, and the specific steps include:

[0014] S1.1. Determination of data start and end points: The argon blowing flow rate in the production data is selected as the basis for judgment. When the flow rate is 0, it is taken as the start point, and when the flow rate changes to the maximum value, it is taken as the end point.

[0015] S2.2. Data cleaning: Remove outliers from the data at the determined start and end points, and fill in missing values ​​to obtain cleaned data;

[0016] S2.3. Data Normalization: Normalize the cleaned data to obtain normalized data;

[0017] S2.4. Time Series Data Construction: The normalized data is converted into a format suitable for the input of the LSTM model to obtain time series data.

[0018] In addition to the aspects and any possible implementations described above, a further implementation is provided in which metallurgical mechanism analysis and big data mining techniques are used in S2 to determine the production process control parameters.

[0019] In addition to the aspects described above and any possible implementation, a further implementation is provided in which the production process control parameters include ladle baking state, ladle lining state, slag layer thickness, molten steel quality, initial temperature, ambient temperature, heating power, alloy addition amount, slag addition amount, argon blowing flow rate, heating time, and / or cooling time.

[0020] In addition to the aspects described above and any possible implementation, a further implementation is provided in which the temperature narrow window database data items include ladle baking state parameters, ladle lining state parameters, slag layer thickness, molten steel quality, initial temperature, ambient temperature, heating power, alloy addition, slag addition, argon blowing flow rate, heating time, cooling time, electrode current, electrode voltage, and / or time nodes of each operation event.

[0021] In addition to the aspects described above and any possible implementations, a further implementation is provided in which the LSTM model includes an input layer, an LSTM layer, an attention layer, and an output layer:

[0022] The input layer is used to receive time series data;

[0023] The LSTM layer uses a single layer, and the number of cells is fixed;

[0024] The attention layer is used to compute the weights of the hidden states at each time step;

[0025] The output layer is connected after the LSTM layer, outputting a weight vector corresponding to the case attributes, and then normalizing it so that the sum of the weights is 1.

[0026] As described above, and in accordance with any possible implementation, a further implementation is provided, wherein the formula for the normalization process is:

[0027] Where, x norm This is the normalized data, where x is the original data. min and x max These are the minimum and maximum values ​​of the original data, respectively.

[0028] As described above, and in accordance with any possible implementation, a further implementation is provided, wherein the expression for the similarity calculation of the CBR algorithm is:

[0029]

[0030] Among them, W j x represents the attribute weight value of the j-th feature value output by the LSTM model. i,j Let x be the j-th feature value of historical case i in the temperature narrow window database. current,j Let d(c) be the j-th feature value of the current case. i,c current ) represents the distance between the current case and historical cases, c i Let c be the feature vector of the working conditions of historical case i. current This is the feature vector of the current case.

[0031] In addition to the aspects and any possible implementations described above, an implementation is further provided in which online learning in S6 includes: adding new cases to the case library after each smelting batch is completed;

[0032] Fine-tune the LSTM model parameters using new case data.

[0033] The present invention also provides a steel molten metal temperature prediction system for ladle refining furnace based on LSTM-CBR dynamic adjustment. The system is used to implement the method and includes the following modules: an acquisition module, used to acquire LF refining production data and preprocess it to obtain a steel molten metal production database;

[0034] The determination module is used to determine the production process control parameters that affect the temperature of molten steel, and to divide the production process control parameters according to different time sequences and smelting stages;

[0035] The transmission module is used to transmit and store relevant parameters of the molten steel production database based on the production process control parameters obtained after division, forming a temperature narrow window database;

[0036] The module is used to extract the temporal features of each control parameter in the narrow temperature window database using the LSTM model, generate dynamic weights that reflect the influence of each parameter on temperature, and build a dynamic parameter adjustment rule base.

[0037] The retrieval module is used to retrieve cases from the dynamic parameter adjustment rule base based on the collected input of current smelting process operation data and the current case using the CBR algorithm to determine the most similar furnace and realize dynamic temperature prediction.

[0038] The update module is used for online learning of new case data, fine-tuning of LSTM model parameters, and updating of the temperature narrow window database;

[0039] The prediction module is used to accurately predict the final temperature of molten steel at the end of LF refining using the trained LSTM-CBR model.

[0040] Beneficial effects of the present invention

[0041] The method of this invention includes: acquiring production data of the LF refining process and constructing a molten steel production database using clean production data after data preprocessing. Simultaneously, combining metallurgical mechanism analysis and big data mining techniques, the production process control parameters affecting the molten steel temperature are determined and divided according to different time sequences and smelting stages. Based on the determined production process control parameters, relevant parameters in the molten steel production database are transmitted and stored to form a temperature narrow window database. Further, an LSTM (Long Short-Term Memory) model is used to extract the temporal features of each control parameter in the temperature narrow window database, generating a dynamic weight vector reflecting the importance of each parameter, thereby constructing a dynamic parameter adjustment rule base. Based on the collected input of current smelting process operation data, a case retrieval is performed on the temperature narrow window database using the CBR (Case-Based Reasoning) algorithm. Based on the current case, the dynamic weights of the dynamic parameter adjustment rule base are used to calculate and weight historical cases based on similarity, selecting the optimal case heat to achieve prediction of the molten steel temperature at the end of the LF refining process. This invention generates case matching weights through an LSTM model, driving the dynamic optimization of CBR case retrieval in real time, achieving accurate prediction of the molten steel temperature during the LF refining process. Attached Figure Description

[0042] Figure 1 This is a flowchart of the method of the present invention;

[0043] Figure 2 This describes the automatic data acquisition process of the present invention. Detailed Implementation

[0044] To better understand the technical solution of this invention, the content of this invention includes, but is not limited to, the specific embodiments described below. Similar technologies and methods should be considered within the scope of protection of this invention. To make the technical problems to be solved, the technical solutions, and advantages of this invention clearer, a detailed description will be provided below in conjunction with the accompanying drawings and specific embodiments.

[0045] It should be understood that the embodiments described in this invention are merely some, not all, of the embodiments of this invention. All other embodiments obtained by those skilled in the art based on the embodiments of this invention without inventive effort are within the scope of protection of this invention.

[0046] The terminology used in the embodiments of this invention is for the purpose of describing particular embodiments only and is not intended to limit the invention. The singular forms “a,” “the,” and “the” as used in the embodiments of this invention and the appended claims are also intended to include the plural forms unless the context clearly indicates otherwise.

[0047] like Figure 1As shown, this invention provides a method for predicting the temperature of molten steel in a ladle refining furnace based on LSTM-CBR dynamic adjustment, comprising the following steps:

[0048] S1. Independently developed data source interface to automatically collect LF refining production data in real time, serving as the data foundation for subsequent model applications, including:

[0049] 1.1 Data Interface Development Methodology:

[0050] 1.1.1 Requirements Analysis: Determine the system's functions, required data, data format, etc.;

[0051] 1.1.2 Interface Design: Define the interface protocol, including request methods, data transmission formats, etc.;

[0052] 1.1.3 Interface Development: Based on the interface requirements and protocols, write the specific interface code;

[0053] 1.1.4 Function Implementation: Test the functionality of the interface to ensure its correctness and stability;

[0054] 1.1.5 Deployment and Release: Deploy the interface code to the server and release it to ensure that the interface can provide services normally.

[0055] 1.2 Data Transmission:

[0056] Based on a clear understanding of the system's data requirements and formats, the necessary data source addresses, target database addresses, and communication protocols are determined. Code is then written according to the source data addresses for primary, secondary, and tertiary data, the target data storage address (steel production database), and the requirements of the communication protocol to enable the transmission of the required data from the source data addresses to the steel production database address.

[0057] 1.3 Automatic data collection method:

[0058] The automatic data collection process is as follows Figure 2As shown, the primary PLC data is transmitted to the KepServerEX 6 software via the OPC protocol, and then standardized data access is provided for subsequent data acquisition services via the OPC UA protocol. The model database (secondary data source) relies on the Oracle database management system as a secondary data source. The MES database (tertiary data source) uses SQL structured query language for data conversion and transmission. A program developed using VB.NET serves as the core of data acquisition and integration, collecting data from KepServerEX 6, the model database, and the MES database via corresponding protocols, and writing the data to a MySQL database at a frequency of 30 seconds per iteration. The MySQL software serves as the final data storage location for subsequent data analysis and applications. As long as all programs are kept running, automatic, real-time transmission and storage can be achieved.

[0059] 1.4 The LF refining production data obtained includes:

[0060] The system collects and analyzes the following data: LF inlet composition, LF process composition, and LF endpoint composition; process temperature data (inlet temperature, process temperature, and endpoint temperature); LF charging data (lime, modifier, submerged arc slag, ferrosilicon, ferroticonium, ferrotitanium, medium-carbon ferromanganese, etc.); process operating parameters (bottom-blown argon, electrode effective power, current value, voltage value, power consumption, heating time, cooling time); molten steel volume; and process operation time nodes. By automatically collecting the above-mentioned LF refining production data for each heat, the system provides input and output values ​​for subsequent LSTM-CBR model training. The more heats collected, the higher the accuracy of the trained model. Simultaneously, the system automatically and in real-time collects the above production data during the current on-site smelting process, using it as model data.

[0061] S2. Obtain clean production data through data processing and establish a molten steel production database, specifically including:

[0062] 2.1.1 Determination of Data Start and End Points: To ensure the completeness and accuracy of data for each furnace, the cumulative argon blowing flow rate was used as the criterion for all collected data. When the flow rate is 0, it indicates that there is no ladle on the ladle car and production has not yet started. When the cumulative flow rate changes from 0 to a value, it indicates that the ladle has been set and production has begun. This signal is used as the start point for furnace data collection. When the flow rate reaches its maximum value, it is the end point, indicating that the ladle has been lifted and the smelting process has ended. By determining the start and end points for all data collection in this way, complete and accurate collection of production data for each furnace is achieved.

[0063] 2.1.2 Data cleaning: outlier detection and missing value handling.

[0064] To address the abnormal data collected, the first step involves setting thresholds for each module using metallurgical process rules to constrain and identify and remove abnormal data. The second step employs an improved 3σ criterion to process the production data generated during the LF refining process. Values ​​exceeding the range of (μ-3σ, μ+3σ) are considered outliers, and all data from the relevant furnaces are directly removed to improve data quality.

[0065]

[0066] Where n is the total number of samples for a certain feature in the dataset; x k This refers to the k-th sample data of a certain feature in the dataset; μ is the average value of a certain feature sample; σ is the standard deviation of a certain feature sample.

[0067] When the missing data is a single point of random missing data, cubic spline interpolation is used to supplement the data; when the missing data is a long period of missing data, it is either directly marked as invalid segment or supplemented by combining with the process data of the same period according to different modules.

[0068] In this embodiment, taking temperature data acquisition as an example, the following data anomalies exist: Through on-site tracking and data acquisition comparison, it was found that the acquired temperature was 7°C higher than the actual measured temperature, requiring compensation processing. Simultaneously, the temperature data acquired and transmitted in the database exhibited abrupt changes. After the temperature measurement operation was completed, the sensor temperature data fluctuated within a range near the actual temperature. Comparison with the actual on-site temperature measurement data revealed that the temperature transmission data was 1200°C when no temperature measurement was performed. This temperature was cleaned up, and the lowest temperature within the fluctuation range became the actual temperature of the measurement node. Therefore, the preprocessing of temperature data requires a multi-step process.

[0069] 2.1.3 Data Normalization: Min-Max scaling is used for normalization, placing different features within the same range, thus accelerating model training and improving stability. The transformation function is:

[0070]

[0071] Where, x norm Here, x is the normalized data, and x is the cleaned dataset. min and x max These are the minimum and maximum values ​​of x, respectively.

[0072] 2.1.4 Time series data construction: The normalized data is converted into a format suitable for the input of the LSTM model using the sliding window method.

[0073] 2.2 Steel Molten Production Database

[0074] The molten steel production database is established by defining smelting production nodes and ensuring the integrity of charging data, temperature measurement data, and process operation data for each furnace. Through data cleaning and normalization, it generates historical operation data (lf_info_history), equipment status data (tb_lf_info), and molten steel testing data (tb_lf_steel_metal). Based on time series, various events are correlated, and data sources such as real-time sensor data, laboratory test data, and operation records are integrated into the lf_result_datas (molten steel internal control and composition, temperature data) and lf_sub_datas (process operation parameters) data tables, achieving unified alignment of diverse data.

[0075] S3. Combining metallurgical mechanism analysis and big data mining techniques, the production process control parameters affecting the temperature of molten steel are determined, and these control parameters are divided according to different time sequences and smelting stages, specifically:

[0076] By identifying the heat input and heat output components of molten steel using metallurgical mechanisms, and analyzing the control parameters affecting their changes, we can further utilize big data mining techniques to perform Pearson correlation analysis on relevant steel production data and molten steel temperature to determine potential control parameters influencing molten steel temperature changes. Combining these two approaches, we can then determine the control parameters affecting molten steel temperature.

[0077] The parameters for controlling the temperature of molten steel include, but are not limited to: ladle baking status, ladle lining status, slag layer thickness, molten steel quality, initial temperature, ambient temperature, heating power, alloy addition, slag addition, argon blowing flow rate, heating time, and cooling time.

[0078] Based on the influence of various parameters on the temperature of molten steel at different time stages during the smelting process, the data are divided into long-time series and short-time series. Long-time series data refers to data that remains basically unchanged during the smelting process and has a relatively stable influence on the temperature; short-time series data refers to data that affects the temperature of molten steel within a short period of time after input.

[0079] The long-term sequence includes: ladle baking status, ladle lining status, slag layer thickness, molten steel quality, initial temperature, and ambient temperature;

[0080] Short time series include: heating power, alloy addition, slag addition, argon blowing flow rate, heating time, and cooling time.

[0081] S4. Based on the determined production process control parameters, relevant parameters are transmitted and stored in the molten steel production database to form a narrow temperature window database.

[0082] Based on the steel molten temperature control parameters determined above, the data table and field names in the temperature narrow window database are determined. SQL statements are used to automatically and in real-time transfer and store the control parameters from the steel molten production database to the temperature narrow window database. Data items in the temperature narrow window database include, but are not limited to:

[0083] The parameters for ladle baking, ladle lining, slag layer thickness, molten steel quality, initial temperature, ambient temperature, heating power, alloy addition, slag addition, argon blowing flow rate, heating time, cooling time, electrode current, electrode voltage, and time points of each operation event are recorded.

[0084] S5. Use the LSTM model to extract the time-series features of each control parameter in the narrow temperature window database, generate dynamic weights that reflect the influence of each parameter on temperature, and construct a dynamic parameter adjustment rule base.

[0085] The model structure of the LSTM model is as follows: Input layer → LSTM layer → Attention layer → Output layer (fully connected layer) → Softmax normalization. Each layer is described in detail below:

[0086] 4.1 Input Layer: Used to receive time series data. The number of neurons in the input layer is consistent with the dimension of the input features. The shape of the input data is [number of cases, time step, feature dimension].

[0087] The input vector x at each time step t =[x t1 ,x t2 ,x t3 ,x t4 ,x t4 ,x t5 ,x t6 ,x t7 ,x t8 ,x t9 ,x t10 ,x t11 ,x t12 ], corresponding to 12 respectively

[0088] parameter.

[0089] In this embodiment, the input layer feature dimension is 12, which are the ladle baking state parameters, ladle lining state parameters, slag layer thickness, molten steel quality, initial temperature, ambient temperature, heating power, alloy addition, slag addition, argon blowing flow rate, heating time, and cooling time.

[0090] 4.2 LSTM Layer: The core function of the LSTM layer is to extract long-term dependencies in the time series, extracting temporal features from the input data. Initially, a single-layer LSTM is chosen, with 128 units per layer. For each time step, the LSTM unit receives the current input x.t The hidden state h of the previous time step t-1 and the cell state c t-1 Calculate the hidden state h at the current time step. t and cell state c t The calculation process is as follows:

[0091] Input gate: i t =σ(W xi ·x t +W hi ·h t-1 +b i (4)

[0092] Forgotten Gate: f t =σ(W xf ·x t +W hf ·h t-1 +b f (5)

[0093] Output gate: o t =σ(W xo ·x t +W ho ·h t-1 +b o (6)

[0094] Cell state: c t =f t ·c{t-1}+i t ·tanh(W xc ·x t +W hc ·h t-1 +b c (7)

[0095] Hidden state: h t =o t ·tanh(c t (8)

[0096] Where, x t It is the input vector, where t represents the time step and h t-1 W represents the output of the LSTM at the previous moment. xi W xf W xo W xc It is the weight matrix of the input gate, forget gate, output gate, and candidate states, b i b f b o b c It is the bias vector of the input gate, forget gate, output gate, and candidate states.

[0097] W hi W hf W ho W hc It is in a hidden state h t-1 The weight matrix is ​​given by the input gate, forget gate, output gate, and candidate state. σ is the sigmoid function, which compresses the values ​​to the (0,1) interval, and tanh compresses the values ​​to the (-1,1) interval.

[0098] After all time steps, the hidden state output H = [h1, h2, ... ht] for all time steps is obtained.

[0099] 4.3 Attention Layer: Calculates the weight of the hidden state at each time step, highlighting the importance of different time steps and different operating parameters to the temperature of molten steel;

[0100] 4.3.1 Hidden state h at each time step t t Perform a linear transformation, then apply tanh activation, and finally pass it through a learnable vector u. a Perform a dot product to obtain the energy value (attention score) e at that time step. t The calculation formula is:

[0101] e t =u a ^T·tanh(W a ·h t +b a (9)

[0102] Among them, W a It is a weight matrix, b a It is the bias vector, u a It is a vector of dimension attention dim (attention dim is a hyperparameter with a pre-defined value), e t It is an energy value (attention score).

[0103] 4.3.2 Softmax normalization is applied to the energy values ​​at T time steps to obtain the attention weight α at each time step. t :

[0104]

[0105] Where, α t Let e ​​represent the attention weight at time step t. t Let t represent the energy value at time step t, T represent the total length of the time series, and t′ represent the time step traversal variable used to sum over all time steps (1 to T).

[0106] 4.3.3 Calculate the context vector:

[0107]

[0108] Where c is the hidden state h at each time step. t The weighted sum, α t h represents the attention weight at time step t. t This represents the hidden state of the LSTM at time t.

[0109] 4.4 Threshold Generation Layer:

[0110] The threshold will be used to dynamically adjust the similarity matching threshold for CBR case retrieval, thereby achieving adaptive adjustment of retrieval accuracy.

[0111] confidence = σ(W) conf ·c+b conf (12)

[0112] Where confidence is the matching confidence threshold, σ is the Sigmoid activation function, and b conf Confidence bias term, W conf The confidence weight vector and the context vector output by the attention layer.

[0113] Dynamically adjust search precision:

[0114] θ actual =θ base -k·(1-confidence) (13)

[0115] Where k is the adjustment coefficient, θ actual The dynamically adjusted threshold used in practice, θ base This is the baseline threshold.

[0116] 4.5 Output Layer: After the LSTM layer, a fully connected layer is added to transform the context vector c into a weight vector w′=[w1,w2,...,w n ], where the output weights are [w1, w2, ..., w n The weights of the influence of [ladle baking state parameters, ladle lining state parameters, slag layer thickness, molten steel quality, initial temperature, ambient temperature, heating power, alloy addition, slag addition, argon blowing flow rate, heating time, and cooling time] on temperature are as follows:

[0117] w′=W fc c+b fc (14)

[0118] Where w′ represents the unnormalized weight vector, W fcLet c represent the weight matrix of the fully connected layer, and b represent the context vector. fc This represents the bias vector of the fully connected layer.

[0119] 4.6 Softmax Normalization: Softmax normalization is applied to the output layer w′ to ensure that the sum of the weights is 1. The normalization formula is:

[0120]

[0121] Among them, w j w′ represents the normalized weight of the j-th input feature. j w′ k This represents the original weight score of the j-th and k-th input features.

[0122] Training objective: The LSTM-Attention module and the CBR retrieval module are trained together, with the contribution of each attribute in historical data to the accuracy of case matching as the supervision signal, and the loss function is the mean square error of the predicted temperature and the actual temperature.

[0123] By generating the weights of the influence of different temperature control parameters on temperature through the above process, a dynamic parameter adjustment rule base is constructed to store the influence weights of different temperature control parameters on temperature in each case at each stage, providing case retrieval guidance for the CBR model.

[0124] In this embodiment, the smelting process control parameters have different effects on the molten steel temperature at different smelting stages. Long-term parameters have a lower weight on the molten steel temperature in a single heat, but their weight changes periodically over a long smelting period. Short-term parameters have an immediate effect on the molten steel temperature in a single heat, and their weight on the molten steel temperature varies at different smelting stages.

[0125] The LF refining process can be divided into four stages: slag formation, alloying, heating, and static stirring. During the slag formation stage, heating power and slag addition significantly affect the molten steel temperature; during the alloying stage, the molten steel temperature is mainly affected by heating power and alloy addition; heating power has a significant impact on the molten steel temperature during the heating stage; and during the soft-blowing static stirring stage, the argon flow rate primarily affects the molten steel temperature. Furthermore, different stages are influenced by the same control parameter, but the weight of this influence varies at different stages. This study uses an LSTM model to predict long-term low-weight static attribute parameters and short-term high-weight sensitive attribute parameters, and combines this with an attention mechanism to quantify the weights of each control parameter and the weights of the same control parameter at different stages, revealing the mechanism by which process parameters function.

[0126] S6. Based on the real-time data of the current smelting process operation, the CBR algorithm is used to perform case retrieval on the temperature narrow window database to determine the most similar furnace and realize dynamic temperature forecasting.

[0127] The CBR algorithm is an artificial intelligence technique that uses past experience knowledge to reason about solving similar current problems. It finds the most relevant cases from the case library and modifies them as necessary to adapt to the specific situation of the current problem. During smelting production, based on the input of each feature value of the current furnace, the temperature narrow window database is searched. The weights of each parameter of the corresponding furnace number in the dynamic parameter adjustment rule library are obtained by matching the furnace number. By inputting the feature value of the current furnace, the feature value of historical cases, and the weights of each feature value of historical cases, the similarity of the CBR algorithm is calculated using Equation (14). The historical case with the highest similarity is selected to complete the dynamic prediction of the molten steel temperature of the current smelting furnace, and at the same time, the corresponding inlet parameters and process operation parameters of the furnace are matched.

[0128] The CBR algorithm retrieves matching logical expressions including:

[0129] Similarity calculation uses dynamically weighted Euclidean distance. The following formula is the retrieval logic formula of the CBR algorithm:

[0130]

[0131] Among them, W j x represents the attribute weight value of the j-th feature value output by the LSTM model. i,j Let x be the j-th feature value of historical case i in the temperature narrow window database. current,j Let d(c) be the j-th feature value of the current case. i ,c current ) represents the distance between the current case and historical cases. i Let c be the feature vector of the working conditions of historical case i. current This is the feature vector of the current case.

[0132] Dynamic adjustment of matching threshold: The LSTM model additionally outputs a matching confidence threshold, which represents the typicality of the current working condition, with a value range of [0,1]. The retrieval accuracy is dynamically adjusted based on the difference between the real-time working condition and the case database. When the input data differs significantly from historical typical working conditions, a low confidence level (close to 0) is output, and the similarity threshold for case retrieval is relaxed; conversely, the retrieval accuracy is increased (the similarity threshold is reduced).

[0133] S7. Learn new case data online, fine-tune LSTM model parameters, and update the temperature narrow window database.

[0134] The online learning mechanism involves collecting complete data from the current smelting furnace. current The data is then transferred and stored in a narrow temperature window database to supplement new case data. Simultaneously, the dynamic weights w of the control parameters for the current furnace batch are predicted using the LSTM model.j Updated and supplemented to the dynamic parameter adjustment rule library.

[0135] Online learning mechanisms include:

[0136] After each smelting batch is completed, the data of the current smelting furnace is collected to form a new case and added to the temperature narrow window database;

[0137] The LSTM model learns the degree of new information reception, the degree of historical information retention, the current state output strength, and the time step in new case data, thereby influencing the weights (W) of the input gate, forget gate, output gate, and attention in the LSTM model. xi W xf W xo W a Fine-tuning is performed to adapt to equipment aging and changes in raw materials. These parameter adjustments are made because various parameters such as equipment and raw materials will change to varying degrees over time, which will be reflected in the production data of new smelting furnaces. Furthermore, the model is used to learn these parameters and adjust for long-term drift.

[0138] S8. Using an LSTM-CBR model trained with a large amount of historical furnace data and a validation set temperature prediction loss (MSE) ≤ 10 (corresponding to a mean absolute error of temperature (MAE) ≤ 3℃), accurate prediction of the molten steel temperature at the end of LF refining is achieved.

[0139] The most similar case is selected by dynamically calculating the weights of the LSTM model and dynamically adjusting the CBR algorithm. The molten steel outlet temperature in the case is used to predict the final temperature of the molten steel in the current smelting furnace. The process operation parameters matched by the case are used to guide actual production.

[0140] In summary, this embodiment presents a method for predicting molten steel temperature in an LF refining furnace based on LSTM-CBR dynamic adjustment. It employs data preprocessing methods to reduce noise and normalize the production dataset, improving the quality of the input data. The LSTM model provides real-time weight guidance for the CBR algorithm, upgrading case matching from "static template matching" to "dynamic adaptive matching," effectively addressing the complex nonlinear and time-varying conditions during smelting and intuitively reflecting the relative importance of different operating parameters to molten steel temperature. The integration of the LSTM model's temporal modeling capabilities with the CBR algorithm's case knowledge enhances prediction accuracy and interpretability. An online learning mechanism is used to supplement and update the case library and fine-tune the LSTM model parameters. By combining these methods, the accuracy and stability of the combined model's predictions are achieved.

[0141] As an embodiment of the present invention, the present invention also provides a steel molten metal temperature prediction system for ladle refining furnace based on LSTM-CBR dynamic adjustment. The system is used to implement the method and includes the following modules: an acquisition module, used to acquire LF refining production data and preprocess it to obtain a steel molten metal production database;

[0142] The determination module is used to determine the production process control parameters that affect the temperature of molten steel, and to divide the production process control parameters according to different time sequences and smelting stages;

[0143] The transmission module is used to transmit and store relevant parameters of the molten steel production database based on the production process control parameters obtained after division, forming a temperature narrow window database;

[0144] The module is used to extract the temporal features of each control parameter in the narrow temperature window database using the LSTM model, generate dynamic weights that reflect the influence of each parameter on temperature, and build a dynamic parameter adjustment rule base.

[0145] The retrieval module is used to retrieve cases from the dynamic parameter adjustment rule base based on the collected input of current smelting process operation data and the current case using the CBR algorithm to determine the most similar furnace and realize dynamic temperature prediction.

[0146] The update module is used for online learning of new case data, fine-tuning of LSTM model parameters, and updating of the temperature narrow window database;

[0147] The prediction module is used to accurately predict the final temperature of molten steel at the end of LF refining using the trained LSTM-CBR model.

[0148] The foregoing description illustrates and describes several preferred embodiments of the present invention. However, as previously stated, it should be understood that the present invention is not limited to the forms disclosed herein and should not be construed as excluding other embodiments. It can be used in various other combinations, modifications, and environments, and can be altered within the scope of the inventive concept described herein through the foregoing teachings or techniques or knowledge in related fields. Any modifications and variations made by those skilled in the art that do not depart from the spirit and scope of the present invention should be within the protection scope of the appended claims.

Claims

1. A method for predicting the temperature of molten steel in a ladle refining furnace based on LSTM-CBR dynamic adjustment, characterized in that, Includes the following steps: S1. Obtain LF refining production data and preprocess it to obtain a molten steel production database; S2. Determine the production process control parameters that affect the temperature of molten steel, and divide the production process control parameters according to different time sequences and smelting stages; S3. Based on the production process control parameters obtained after the division, the relevant parameters of the molten steel production database are transmitted and stored to form a temperature narrow window database; S4. Use the LSTM model to extract the time-series features of each control parameter in the narrow temperature window database, generate dynamic weights that reflect the influence of each parameter on temperature, and construct a dynamic parameter adjustment rule base. S5. Based on the collected and inputted smelting process operation data, the CBR algorithm is used to search for cases in the dynamic parameter adjustment rule base based on the current case, and the most similar furnace is determined to achieve dynamic temperature forecasting. S6. Learn new case data online, fine-tune LSTM model parameters, and update the temperature narrow window database; S7. Accurately predict the final steel temperature of LF refining using the trained LSTM-CBR model.

2. The method according to claim 1, characterized in that, The preprocessing includes determining the data start and end points, data cleaning, data normalization, and time series data construction. Specific steps include: S1.

1. Determination of data start and end points: The argon blowing flow rate in the production data is selected as the basis for judgment. When the flow rate is 0, it is taken as the start point, and when the flow rate changes to the maximum value, it is taken as the end point. S2.

2. Data cleaning: Remove outliers from the data at the determined start and end points, and fill in missing values ​​to obtain cleaned data; S2.

3. Data Normalization: Normalize the cleaned data to obtain normalized data; S2.

4. Time Series Data Construction: The normalized data is converted into a format suitable for the input of the LSTM model to obtain time series data.

3. The method according to claim 1, characterized in that, In S2, metallurgical mechanism analysis and big data mining techniques are used to determine the production process control parameters.

4. The method according to claim 3, characterized in that, The production process control parameters include ladle baking status, ladle lining status, slag layer thickness, molten steel quality, initial temperature, ambient temperature, heating power, alloy addition, slag addition, argon blowing flow rate, heating time, and / or cooling time.

5. The method according to claim 1, characterized in that, The temperature narrow window database includes ladle baking status parameters, ladle lining status parameters, slag layer thickness, molten steel quality, initial temperature, ambient temperature, heating power, alloy addition, slag addition, argon blowing flow rate, heating time, cooling time, electrode current, electrode voltage, and / or time nodes of each operation event.

6. The method according to claim 1, characterized in that, The LSTM model includes an input layer, an LSTM layer, an attention layer, a threshold generation layer, and an output layer. The input layer is used to receive time series data; The LSTM layer uses a single layer, and the number of cells is fixed; The attention layer is used to compute the weights of the hidden states at each time step; The threshold generation layer is used to dynamically adjust the similarity matching threshold of CBR case retrieval, so as to achieve adaptive adjustment of retrieval accuracy; The output layer is connected after the LSTM layer, outputting a weight vector corresponding to the case attributes, and then normalizing it so that the sum of the weights is 1.

7. The method according to claim 2, characterized in that, The formula for the normalization process is: Where, x norm This is the normalized data, where x is the original data. min and x max These are the minimum and maximum values ​​of the original data, respectively.

8. The method according to claim 1, characterized in that, The expression for similarity calculation using the CBR algorithm is as follows: Among them, W j x represents the attribute weight value of the j-th feature value output by the LSTM model. i,j Let x be the j-th feature value of historical case i in the temperature narrow window database. current,j Let d(c) be the j-th feature value of the current case. i ,c current ) represents the distance between the current case and historical cases, c i Let c be the feature vector of the working conditions of historical case i. current This is the feature vector of the current case.

9. The method according to claim 1, characterized in that, Online learning in S6 includes adding new cases to the case library after each smelting batch is completed; Fine-tune the LSTM model parameters using new case data.

10. A steel ladle refining furnace molten steel temperature prediction system based on LSTM-CBR dynamic adjustment, characterized in that, The system is used to implement the method according to any one of claims 1-9, and includes the following modules: an acquisition module, used to acquire LF refining production data and preprocess it to obtain a molten steel production database; The determination module is used to determine the production process control parameters that affect the temperature of molten steel, and to divide the production process control parameters according to different time sequences and smelting stages; The transmission module is used to transmit and store relevant parameters of the molten steel production database based on the production process control parameters obtained after division, forming a temperature narrow window database; The module is used to extract the temporal features of each control parameter in the narrow temperature window database using the LSTM model, generate dynamic weights that reflect the influence of each parameter on temperature, and build a dynamic parameter adjustment rule base. The retrieval module is used to retrieve cases from the dynamic parameter adjustment rule base based on the collected input of current smelting process operation data and the current case using the CBR algorithm to determine the most similar furnace and realize dynamic temperature prediction. The update module is used for online learning of new case data, fine-tuning of LSTM model parameters, and updating of the temperature narrow window database; The prediction module is used to accurately predict the final temperature of molten steel at the end of LF refining using the trained LSTM-CBR model.

Images