A copper smelting oxygen-enriched side-blown furnace matte grade prediction method based on multi-model fusion

Abstract

The application discloses a copper smelting oxygen-enriched side-blown furnace copper matte grade prediction method based on multi-model fusion, aiming at the problems of multi-scale variable coupling in the oxygen-enriched side-blown smelting process, insufficient time sequence modeling, small sample overfitting, poor existing prediction accuracy, obvious hysteresis and the like, relying on smelting multi-source heterogeneous production data, after cleaning, time sequence alignment and standardization pretreatment, combining with metallurgical mechanism to build composite features, and adopting mRMR criterion to complete key feature screening. The application builds a TCN-BiLSTM-XGBoost three-branch parallel coding architecture, cooperates with a sliding window and a channel attention mechanism to realize second-level and hour-level heterogeneous data adaptive fusion, the prediction accuracy and generalization performance of the application are better than those of existing single and integrated models, the influence of detection hysteresis can be effectively eliminated, the copper matte grade change can be accurately tracked, reliable technical support can be provided for copper smelting process online soft measurement and intelligent working condition regulation, and the application is suitable for various non-ferrous metal smelting key process index prediction scenes.

Description

Technical Field

[0001] This invention relates to the field of intelligent control technology for copper metallurgical production processes, specifically a method for predicting matte grade in oxygen-enriched side-blown furnaces for copper smelting based on multi-model fusion. Background Technology

[0002] Side-blown smelting is a multivariable, time-delayed, and strongly coupled system, with a 3-4 hour time lag between raw material input and matte output. Furthermore, the smelting process involves complex raw materials and is influenced by numerous associated factors. The matte grade (i.e., the copper mass fraction in matte) is a core process indicator characterizing the copper smelting state, directly determining the efficiency of subsequent blowing processes, material consumption, and final product quality. However, the significant 3-4 hour time lag between matte formation in the reaction zone and laboratory testing in industrial settings means that test results cannot promptly reflect the real-time smelting state, making it difficult to support real-time process adjustments.

[0003] Currently, methods for predicting the grade of copper matte have evolved from mechanism-driven to data-driven approaches, but they still face the following key technical challenges:

[0004] 1. Multi-scale variable coupling problem: The smelting process has multi-scale variables with second-level process parameters (such as air volume and temperature) and hour-level raw material data (such as concentrate composition). Traditional downsampling or interpolation methods are prone to causing serious information loss and it is difficult to achieve effective fusion of heterogeneous multi-frequency data.

[0005] 2. Insufficient modeling of temporal dependencies and bidirectional context: While existing temporal convolutional networks (TCNs) are good at capturing long-term temporal dependencies, they lack bidirectional context awareness; while recurrent neural networks (RNNs) suffer from gradient degradation, making it difficult to achieve joint modeling of temporal dependencies and cross-variable coupling.

[0006] 3. Small sample overfitting problem: The high cost of chemical analysis leads to a scarcity of labeled samples. Deep learning models are prone to overfitting under small sample conditions, resulting in decreased generalization performance.

[0007] Meanwhile, based on technical investigation, the main patents and existing research results related to this invention are as follows:

[0008] 1) Patent Title: A Method for Cooperative Continuous Smelting of Copper Concentrate in Three Furnaces. Patent Application Publication Number:

[0009] CN121653397A. This invention, through comprehensive perception of furnace conditions and improved basis for control parameter adjustment, achieves intelligent decision-making and precise control of process parameters based on smelting mechanisms and intelligent methods. When changes occur in the operating conditions of raw materials and upstream processes, it identifies and pre-processes these changes in advance, improving furnace condition and product stability, thereby increasing furnace heat utilization, reducing fuel consumption, and achieving energy conservation and environmental protection. However, the metallurgical calculation model used in this invention often results in significant deviations between the calculated and actual results due to modeling accuracy issues, and the generalization ability of the algorithm needs improvement.

[0010] 2) Patent Title: A Multivariate Time-Series Prediction Method for Copper Matte Quality Based on Digital Twin. Patent Publication Number: CN121210845A. This invention employs the multi-head self-attention mechanism of the Transformer model to capture the long-term dependencies of process parameters in time series, improving the modeling capability for complex time-series data; it introduces a Kolmogorov-Arnold Network (KAN) refinement module, using adaptive B-spline basis functions to perform nonlinear correction on the global features output by the Transformer, enhancing the model's adaptability to varying operating conditions; and it establishes a digital twin model of the copper smelting process to achieve virtual-real mapping of the copper smelting process and prediction of copper matte quality. However, practice in the "side + top" process shows that it is difficult to establish an accurate prediction model between process parameters and copper matte grade using only process parameters.

[0011] 3) Patent Title: A Data-Driven Method for Predicting Copper Matte Grade in Copper Smelting. Patent Number: CN118468148A. This invention uses a random forest and an XGBoost copper matte grade prediction model as base learners, and a linear regression model as a meta-learner to construct a stacking ensemble learning model for predicting copper matte grade. However, linear regression cannot capture the strong nonlinear relationship between copper matte grade and process parameters.

[0012] 4) Patent Title: A Method for Constructing and Applying a Copper Grade Prediction Model. Patent Number: CN116631525B. This invention first uses a mechanistic model-based copper grade prediction model to screen the slag type and operating conditions for directional slag refining to prepare 3N copper, and then performs directional slag refining on scrap copper raw materials. Its application scope is limited to the copper refining stage and cannot be adapted to the smelting stage.

[0013] 5) Patent Title: Soft Measurement Method, Apparatus, Electronic Equipment, and Storage Medium for Copper Matte Grade. Patent Number: CN116258087B. This invention constructs a modeling dataset based on corrected historical data, determines a similar sample set for the current operating point, and establishes a local soft measurement model through Gaussian process regression to predict the grade of the current operating point. Due to the difficulty in monitoring the changes in material composition and furnace conditions at the production site, the determination of the similar sample set often has biases, resulting in inaccurate predictions of copper matte grade.

[0014] Other invention patents are basically similar in content to the above patents, and none of them can use the entire production dataset for holistic grade prediction. To solve the above problems, a method for predicting the grade of matte in oxygen-enriched side-blown furnaces for copper smelting is now proposed based on multi-model fusion. Summary of the Invention

[0015] To address the aforementioned issues, a method for predicting matte grade in oxygen-enriched side-blown copper smelting furnaces based on multi-model fusion is presented.

[0016] The specific solution of this invention is: a method for predicting the grade of matte in an oxygen-enriched side-blown copper smelting furnace based on multi-model fusion, comprising the following steps:

[0017] S1. Data Acquisition and Preprocessing: Collect multi-source heterogeneous data of the entire production process of the oxygen-enriched side-blown furnace in copper smelting. The multi-source heterogeneous data includes DCS real-time operating data, offline detection data of matte composition, slag physicochemical detection data, and furnace feed material ratio data. The collected data is processed in sequence by outlier removal, missing value completion, time-series timestamp alignment, resampling, and normalization and standardization of dimensions to remove production noise data and obtain a complete, continuous, and effective modeling dataset.

[0018] S2. Multi-scale data registration and feature weighting: In response to the multi-scale characteristics of the modeling data, where high-frequency variables at the second level coexist with low-frequency variables at the hour level, a sliding window statistical registration strategy is adopted. Taking the sampling time of the hour-level variable as the anchor point, the mean, standard deviation, minimum value, and maximum value of the second-level variable within the corresponding time window are extracted and concatenated to form a unified dimension sample matrix. An overlapping registration strategy is introduced to adapt to cross-period related variables. Then, the registered feature matrix is ​​adaptively weighted through a channel attention mechanism to strengthen the weight of key features and suppress the interference of invalid features, resulting in a weighted optimized feature set.

[0019] S3. Feature Selection and Domain Knowledge Feature Construction: The maximum relevance and minimum redundancy mRMR criterion is combined with random forest feature importance scoring to select core features that are strongly correlated with the grade of copper matte and have low redundancy between features; based on the copper smelting metallurgical mechanism, three types of composite domain features are constructed: heat balance, oxygen balance, and slag characteristics. The original features and the mechanism composite features are integrated to form the final input feature set.

[0020] S4. Three-branch parallel feature encoding: Construct a TCN-BiLSTM-XGBoost three-branch parallel encoding architecture, and input the final input feature set into the three branch models for synchronous encoding: The TCN branch captures long-term dependency features of time series data through multi-layer dilated causal convolution combined with residual connections; the BiLSTM branch mines bi-directional contextual correlation features of time series data through forward and reverse bi-directional encoding; the XGBoost branch fits the complex nonlinear coupling relationship between process parameters and copper matte grade.

[0021] S5. Multi-branch feature adaptive fusion and prediction: Introducing a multi-head attention mechanism combined with learnable gating parameters, the encoding outputs of TCN branch, BiLSTM branch, and XGBoost branch are adaptively weighted and fused to obtain global fused features. The prediction model trained by the preset outputs the real-time prediction result of the grade of copper matte.

[0022] S6. Model Optimization Output: The model is iteratively trained using a preset optimizer and an early stop mechanism. The model performance is verified through multi-dimensional evaluation indicators, and a high-precision predicted value of matte grade is output, realizing real-time dynamic prediction of matte grade during the oxygen-enriched side-blown furnace smelting process.

[0023] Furthermore, in step S1, the outlier removal specifically includes: setting physical thresholds for each variable based on knowledge of the smelting field for hard truncation, identifying statistical outliers using the 3σ criterion, capturing local outlier samples using the isolated forest algorithm, and removing outlier data through multi-level filtering; the missing value completion specifically includes: using forward interpolation to complete continuous process variables, using mode filling for discrete material variables, and removing invalid variables with a missing value ratio exceeding 20%.

[0024] Furthermore, in step S2, the specific calculation method of the sliding window statistical strategy is as follows: The smelting cycle is set to ΔT, and the second-level variable within the i-th cycle... The statistical characteristics of the mean, standard deviation, minimum, and maximum values ​​are as follows:

[0025]

[0026]

[0027]

[0028] in, To determine the number of sampling points for second-level variables within the i-th melting cycle, an overlapping registration method with a sliding step size smaller than the window length is set to adapt to the correlation influence characteristics of variables in adjacent cycles.

[0029] Furthermore, in step S2, the weighted calculation method of the channel attention mechanism is as follows: Let the registered feature matrix be... The formula for calculating channel attention weights is:

[0030]

[0031] The weighted optimized feature matrix is Where σ is the sigmoid activation function and ReLU is the nonlinear activation function. , For learnable weight matrix, , For bias parameters, This is the global mean feature of the feature matrix.

[0032] Furthermore, in step S3, the optimization objective of the mRMR criterion is to simultaneously maximize the correlation between the features and the target variable of copper matte grade, and minimize the redundancy between features. The specific formula is as follows:

[0033]

[0034] Where D is the average correlation between the features and the target variable, R is the average redundancy between features, S is the candidate feature set, y is the target variable for copper matte grade, and I(·) is the mutual information function used to capture nonlinear relationships between variables.

[0035] Furthermore, in step S3, the specific formula for constructing the mechanistic composite feature is as follows:

[0036]

[0037]

[0038]

[0039] in, This refers to the concentrate feed rate. For the enthalpy change of the smelting reaction, This represents the actual oxygen flow rate. For theoretical stoichiometry of oxygen flow rate, Oxygen density, The specific heat capacity of oxygen, For oxygen supply temperature, The mass fraction of ferrous oxide in the slag. The mass fraction of silica in the slag. R is the activation energy for slag viscosity flow, R is the gas constant, and T is the molten pool temperature.

[0040] Furthermore, in step S4, the TCN branch adopts a dilated causal convolution and residual connection structure, and the output formula of the dilated convolution at time t is:

[0041]

[0042] Where d is the dilation coefficient, k is the kernel size, and multiple dilated convolutions are stacked with the dilation coefficient of the i-th layer set. Feel the wild size L is the number of convolutional layers; the residual block output is The gradient degradation problem in deep networks is eliminated by residual connections.

[0043] Furthermore, in step S4, the BiLSTM branch adopts a bidirectional temporal coding structure, and the hidden state calculation formulas for the forward LSTM and backward LSTM are as follows:

[0044]

[0045]

[0046] By concatenating the bidirectional hidden states along the feature dimension, a temporal comprehensive representation is obtained. This enables complete capture of bidirectional information about the temporal context.

[0047] Furthermore, in step S4, the prediction model and objective function of the XGBoost branch are:

[0048]

[0049]

[0050]

[0051] in, Here, M represents the predicted grade of copper matte, and M is the number of decision trees. For loss function, is the regularization term, γ and λ are the regularization coefficients, T is the number of leaf nodes in a single decision tree, and w is the weight of the leaf node, used to suppress model overfitting.

[0052] Furthermore, in step S5, the multi-head attention fusion mechanism includes multi-head attention calculation and gated weighted fusion, and the basic attention formula is:

[0053]

[0054]

[0055] In the formula, Q, K, and V are the query matrix, key matrix, and value matrix, respectively. Let h be the dimension of the key matrix and h be the number of attention heads; let the output features of the TCN, BiLSTM, and XGBoost branches be respectively... , , Adaptive fusion is achieved through learnable gating parameters α and β:

[0056] .

[0057] Furthermore, in step S6, the batch size, maximum number of training rounds, and early stopping threshold are set during the model training process. Training is terminated when the validation set metrics show no improvement for multiple consecutive rounds. The test set metrics of the fusion model after training meet the following requirements: R²≥0.94, RMSE≤0.42%, MAE≤0.33%, MAPE≤0.87%.

[0058] The present invention has the following beneficial effects:

[0059] 1. Adaptive fusion of heterogeneous multi-scale data: Compared with the existing technology of simple downsampling or interpolation processing of second-level and hour-level data, the present invention designs a data registration module based on sliding window and channel attention to realize adaptive fusion of heterogeneous multi-scale features, effectively retain information of each frequency band and avoid information loss.

[0060] 2. Three-branch parallel encoding architecture: This invention constructs a three-branch parallel encoding layer of TCN-BiLSTM-XGBoost. The TCN branch achieves exponential expansion of the receptive field through dilated causal convolution, effectively capturing long-term temporal dependencies; the BiLSTM branch encodes contextual information bidirectionally, making up for the shortcomings of TCN's unidirectional perception; and the XGBoost branch automatically handles nonlinear interactions between features. The three branches complement each other, breaking through the performance limits of a single model.

[0061] 3. Multi-head attention adaptive fusion: The multi-head attention mechanism is introduced to adaptively weight and fuse the three-branch outputs, capturing the complex relationships between features from different representation subspaces. It has a natural modeling advantage for the strong coupling characteristics of multiple variables in the copper smelting process.

[0062] Significantly improved prediction accuracy: Experimental results based on 8670 sets of industrial measured data show that the method of this invention achieves an R² of 0.6964, an RMSE as low as 0.7463, a MAE of 0.5325, and a MAPE of 0.87. It outperforms existing methods such as LSTM, GRU, TCN, XGBoost, LightGBM, and TabNet in all four evaluation indicators (R², RMSE, MAE, and MAPE), and the prediction accuracy and generalization performance are significantly improved. Attached Figure Description

[0063] Figure 1 This is a schematic diagram illustrating the working principle of the present invention; Detailed Implementation

[0064] Please refer to Figure 1A method for predicting matte grade in oxygen-enriched side-blown copper smelting furnaces based on multi-model fusion includes the following steps:

[0065] S1. Data Acquisition and Preprocessing: Collect multi-source heterogeneous data of the entire production process of the oxygen-enriched side-blown furnace in copper smelting. The multi-source heterogeneous data includes DCS real-time operating data, offline detection data of matte composition, slag physicochemical detection data, and furnace feed material ratio data. The collected data is processed in sequence by outlier removal, missing value completion, time-series timestamp alignment, resampling, and normalization and standardization of dimensions to remove production noise data and obtain a complete, continuous, and effective modeling dataset.

[0066] S2. Multi-scale data registration and feature weighting: In response to the multi-scale characteristics of the modeling data, where high-frequency variables at the second level coexist with low-frequency variables at the hour level, a sliding window statistical registration strategy is adopted. Taking the sampling time of the hour-level variable as the anchor point, the mean, standard deviation, minimum value, and maximum value of the second-level variable within the corresponding time window are extracted and concatenated to form a unified dimension sample matrix. An overlapping registration strategy is introduced to adapt to cross-period related variables. Then, the registered feature matrix is ​​adaptively weighted through a channel attention mechanism to strengthen the weight of key features and suppress the interference of invalid features, resulting in a weighted optimized feature set.

[0067] S3. Feature Selection and Domain Knowledge Feature Construction: The maximum relevance and minimum redundancy mRMR criterion is combined with random forest feature importance scoring to select core features that are strongly correlated with the grade of copper matte and have low redundancy between features; based on the copper smelting metallurgical mechanism, three types of composite domain features are constructed: heat balance, oxygen balance, and slag characteristics. The original features and the mechanism composite features are integrated to form the final input feature set.

[0068] S4. Three-branch parallel feature encoding: Construct a TCN-BiLSTM-XGBoost three-branch parallel encoding architecture, and input the final input feature set into the three branch models for synchronous encoding: The TCN branch captures long-term dependency features of time series data through multi-layer dilated causal convolution combined with residual connections; the BiLSTM branch mines bi-directional contextual correlation features of time series data through forward and reverse bi-directional encoding; the XGBoost branch fits the complex nonlinear coupling relationship between process parameters and copper matte grade.

[0069] S5. Multi-branch feature adaptive fusion and prediction: Introducing a multi-head attention mechanism combined with learnable gating parameters, the encoding outputs of TCN branch, BiLSTM branch, and XGBoost branch are adaptively weighted and fused to obtain global fused features. The prediction model trained by the preset outputs the real-time prediction result of the grade of copper matte.

[0070] S6. Model Optimization Output: The model is iteratively trained using a preset optimizer and an early stop mechanism. The model performance is verified through multi-dimensional evaluation indicators, and a high-precision predicted value of matte grade is output, realizing real-time dynamic prediction of matte grade during the oxygen-enriched side-blown furnace smelting process.

[0071] In this embodiment, step S1, the outlier removal specifically includes: setting physical thresholds for each variable based on knowledge of the smelting field for hard truncation, identifying statistical outliers using the 3σ criterion, capturing local outlier samples using the isolated forest algorithm, and removing abnormal data through multi-level filtering; the missing value completion specifically includes: using forward interpolation to complete continuous process variables, using mode filling for discrete material variables, and removing invalid variables with a missing ratio exceeding 20%.

[0072] In this embodiment, the specific calculation method of the sliding window statistical strategy in step S2 is as follows: The melting cycle is set to ΔT, and the statistical characteristics of the mean, standard deviation, minimum, and maximum values ​​of the second-level variable x_s within the i-th cycle are as follows:

[0073]

[0074]

[0075]

[0076] in, To determine the number of sampling points for second-level variables within the i-th melting cycle, an overlapping registration method with a sliding step size smaller than the window length is set to adapt to the correlation influence characteristics of variables in adjacent cycles.

[0077] In this embodiment, in step S2, the weighted calculation method of the channel attention mechanism is as follows: Let the registered feature matrix be... The formula for calculating channel attention weights is:

[0078]

[0079] The weighted optimized feature matrix is Where σ is the sigmoid activation function and ReLU is the nonlinear activation function. , For learnable weight matrix, , For bias parameters, This is the global mean feature of the feature matrix.

[0080] In this embodiment, in step S3, the optimization objective of the mRMR criterion is to simultaneously maximize the correlation between the features and the target variable of copper matte grade, and minimize the redundancy between features. The specific formula is as follows:

[0081]

[0082] Where D is the average correlation between the features and the target variable, R is the average redundancy between features, S is the candidate feature set, y is the target variable for copper matte grade, and I(·) is the mutual information function used to capture nonlinear relationships between variables.

[0083] In this embodiment, the specific formula for constructing the mechanistic composite feature in step S3 is as follows:

[0084]

[0085]

[0086]

[0087] in, This refers to the concentrate feed rate. For the enthalpy change of the smelting reaction, This represents the actual oxygen flow rate. For theoretical stoichiometry of oxygen flow rate, Oxygen density, The specific heat capacity of oxygen, For oxygen supply temperature, The mass fraction of ferrous oxide in the slag. The mass fraction of silica in the slag. R is the activation energy for slag viscosity flow, R is the gas constant, and T is the molten pool temperature.

[0088] In this embodiment, in step S4, the TCN branch adopts a dilated causal convolution and residual connection structure, and the output formula of the dilated convolution at time t is:

[0089]

[0090] Where d is the dilation coefficient, k is the kernel size, and multiple dilated convolutions are stacked with the dilation coefficient of the i-th layer set. Feel the wild size L is the number of convolutional layers; the residual block output is The gradient degradation problem in deep networks is eliminated by residual connections.

[0091] In this embodiment, in step S4, the BiLSTM branch adopts a bidirectional temporal coding structure, and the hidden state calculation formulas for the forward LSTM and backward LSTM are as follows:

[0092]

[0093]

[0094] By concatenating the bidirectional hidden states along the feature dimension, a temporal comprehensive representation is obtained. This enables complete capture of bidirectional information about the temporal context.

[0095] In this embodiment, in step S4, the prediction model and objective function of the XGBoost branch are:

[0096]

[0097]

[0098]

[0099] in, Here, M represents the predicted grade of copper matte, and M is the number of decision trees. For loss function, is the regularization term, γ and λ are the regularization coefficients, T is the number of leaf nodes in a single decision tree, and w is the weight of the leaf node, used to suppress model overfitting.

[0100] In this embodiment, step S5, the multi-head attention fusion mechanism includes multi-head attention calculation and gated weighted fusion, and the basic attention formula is:

[0101]

[0102]

[0103] In the formula, Q, K, and V are the query matrix, key matrix, and value matrix, respectively. Let h be the dimension of the key matrix and h be the number of attention heads; let the output features of the TCN, BiLSTM, and XGBoost branches be respectively... , , Adaptive fusion is achieved through learnable gating parameters α and β:

[0104] .

[0105] In this embodiment, in step S6, the batch size, maximum number of training rounds and early stopping threshold are set during the model training process. Training is terminated when the validation set metrics have no optimization for multiple consecutive rounds. The test set metrics of the fusion model after training meet the following conditions: R²≥0.94, RMSE≤0.42%, MAE≤0.33%, MAPE≤0.87%.

[0106] Taking the side-blown furnace smelting production line of a large domestic copper smelting enterprise as an example, the application steps of this invention in actual production are described.

[0107] (I) Dataset Construction

[0108] Production process data spanning one year was collected, including 40 process variables, and categorized into three types based on sampling frequency:

[0109]

[0110] After data cleaning and registration, a total of 8,670 valid labeled samples were obtained, which were divided into training set (6,937 sets, accounting for 80%), validation set (867 sets, accounting for 10%) and test set (867 sets, accounting for 10%) in chronological order.

[0111] (ii) Data preprocessing

[0112] 1. Outlier handling: Based on domain knowledge, physical thresholds are set for each variable to perform hard truncation. The 3σ criterion is used to identify statistical outliers, and the isolated forest algorithm is used to capture local outlier patterns.

[0113] 2. Handling missing values: For continuous variables, forward interpolation is used; for discrete variables, mode filling is used; variables with a missing value ratio exceeding 20% ​​are removed.

[0114] (III) Feature Selection Results

[0115] use The method ultimately retains 10 key variables, including target grade of matte, oxygen concentration, copper content in concentrate, molten pool temperature, iron-silicon ratio, and feed rate.

[0116] (iv) Model training and prediction

[0117] The preprocessed data is input into the TCN-BiLSTM-XGBoost fusion model for training.

[0118] (v) Prediction effect

[0119] The prediction performance of the fusion model on the test set is as follows:

[0120]

[0121] Based on the actual prediction curves, the method of this invention can accurately follow the overall evolution trend of copper matte grade. During the stable production phase, the absolute error of grade prediction is basically controlled within ±1. For periods of drastic grade fluctuations, the model can quickly capture the characteristics of grade mutations and simultaneously complete the response prediction, with no significant time lag in the full sample prediction process.

[0122] (vi) Comparative verification

[0123] The method of the present invention is compared with the prior art, and the results are as follows:

[0124]

[0125] The comparative results show that the method of the present invention is superior to the prior art in all evaluation indicators. The R² is improved by 32.8% compared with the second-best LSTM model, and the RMSE is reduced by 19.5%, which verifies the effectiveness and superiority of the method of the present invention.

Claims

1. A method for predicting matte grade in oxygen-enriched side-blown copper smelting furnaces based on multi-model fusion, characterized in that, Includes the following steps: S1. Data Acquisition and Preprocessing: Collect multi-source heterogeneous data of the entire production process of the oxygen-enriched side-blown furnace in copper smelting. The multi-source heterogeneous data includes DCS real-time operating data, offline detection data of matte composition, slag physicochemical detection data, and furnace feed material ratio data. The collected data is processed in sequence by outlier removal, missing value completion, time-series timestamp alignment, resampling, and normalization and standardization of dimensions to remove production noise data and obtain a complete, continuous, and effective modeling dataset. S2. Multi-scale data registration and feature weighting: In response to the multi-scale characteristics of the modeling data, where high-frequency variables at the second level coexist with low-frequency variables at the hour level, a sliding window statistical registration strategy is adopted. Taking the sampling time of the hour-level variable as the anchor point, the mean, standard deviation, minimum value, and maximum value of the second-level variable within the corresponding time window are extracted and concatenated to form a unified dimension sample matrix. An overlapping registration strategy is introduced to adapt to cross-period related variables. Then, the registered feature matrix is ​​adaptively weighted through a channel attention mechanism to strengthen the weight of key features and suppress the interference of invalid features, resulting in a weighted optimized feature set. S3. Feature Selection and Domain Knowledge Feature Construction: The maximum relevance and minimum redundancy mRMR criterion is combined with random forest feature importance scoring to select core features that are strongly correlated with the grade of copper matte and have low redundancy between features; based on the copper smelting metallurgical mechanism, three types of composite domain features are constructed: heat balance, oxygen balance, and slag characteristics. The original features and the mechanism composite features are integrated to form the final input feature set. S4. Three-branch parallel feature encoding: Construct a TCN-BiLSTM-XGBoost three-branch parallel encoding architecture, and input the final input feature set into the three branch models for synchronous encoding: The TCN branch captures long-term dependency features of time series data through multi-layer dilated causal convolution combined with residual connections; the BiLSTM branch mines bi-directional contextual correlation features of time series data through forward and reverse bi-directional encoding; the XGBoost branch fits the complex nonlinear coupling relationship between process parameters and copper matte grade. S5. Multi-branch feature adaptive fusion and prediction: Introducing a multi-head attention mechanism combined with learnable gating parameters, the encoding outputs of TCN branch, BiLSTM branch, and XGBoost branch are adaptively weighted and fused to obtain global fused features. The prediction model trained by the preset outputs the real-time prediction result of the grade of copper matte. S6. Model Optimization Output: The model is iteratively trained using a preset optimizer and an early stop mechanism. The model performance is verified through multi-dimensional evaluation indicators, and a high-precision predicted value of matte grade is output, realizing real-time dynamic prediction of matte grade during the oxygen-enriched side-blown furnace smelting process.

2. The method for predicting matte grade in oxygen-enriched side-blown copper smelting based on multi-model fusion as described in claim 1, characterized in that, In step S1, the outlier removal specifically includes: setting physical thresholds for each variable based on knowledge of the smelting field for hard truncation, identifying statistical outliers using the 3σ criterion, capturing local outlier samples using the isolated forest algorithm, and removing outlier data through multi-level filtering; the missing value completion specifically includes: using forward interpolation to complete continuous process variables, using mode filling for discrete material variables, and removing invalid variables with a missing value ratio exceeding 20%.

3. The method for predicting matte grade in oxygen-enriched side-blown copper smelting based on multi-model fusion as described in claim 1, characterized in that, In step S2, the specific calculation method of the sliding window statistical strategy is as follows: The smelting cycle is set to ΔT, and the statistical characteristics of the mean, standard deviation, minimum, and maximum values ​​of the second-level variable x_s within the i-th cycle are as follows: in, To determine the number of sampling points for second-level variables within the i-th melting cycle, an overlapping registration method with a sliding step size smaller than the window length is set to adapt to the correlation influence characteristics of variables in adjacent cycles.

4. The method for predicting matte grade in oxygen-enriched side-blown copper smelting based on multi-model fusion as described in claim 1, characterized in that, In step S2, the weighted calculation method of the channel attention mechanism is as follows: Let the registered feature matrix be... The formula for calculating channel attention weights is: The weighted optimized feature matrix is Where σ is the sigmoid activation function and ReLU is the nonlinear activation function. , For learnable weight matrix, , For bias parameters, This is the global mean feature of the feature matrix.

5. The method for predicting matte grade in oxygen-enriched side-blown copper smelting based on multi-model fusion as described in claim 1, characterized in that, In step S3, the optimization objective of the mRMR criterion is to simultaneously maximize the correlation between the features and the target variable of copper matte grade, and minimize the redundancy between features. The specific formula is as follows: Where D is the average correlation between the features and the target variable, R is the average redundancy between features, S is the candidate feature set, y is the target variable for copper matte grade, and I(·) is the mutual information function used to capture nonlinear relationships between variables.

6. The method for predicting matte grade in oxygen-enriched side-blown copper smelting based on multi-model fusion as described in claim 1, characterized in that, In step S3, the specific formula for constructing the mechanistic composite feature is as follows: in, This refers to the concentrate feed rate. For the enthalpy change of the smelting reaction, This represents the actual oxygen flow rate. For theoretical stoichiometry of oxygen flow rate, Oxygen density, The specific heat capacity of oxygen, For oxygen supply temperature, The mass fraction of ferrous oxide in the slag. The mass fraction of silica in the slag. R is the activation energy for slag viscosity flow, R is the gas constant, and T is the molten pool temperature.

7. The method for predicting matte grade in oxygen-enriched side-blown copper smelting based on multi-model fusion as described in claim 1, characterized in that, In step S4, the TCN branch adopts a dilated causal convolution and residual connection structure, and the output formula of the dilated convolution at time t is: Where d is the dilation coefficient, k is the kernel size, and multiple dilated convolutions are stacked with the dilation coefficient of the i-th layer set. Feel the wild size L is the number of convolutional layers; the residual block output is The gradient degradation problem in deep networks is eliminated by residual connections.

8. The method for predicting matte grade in oxygen-enriched side-blown copper smelting based on multi-model fusion as described in claim 1, characterized in that, In step S4, the BiLSTM branch adopts a bidirectional temporal coding structure, and the hidden state calculation formulas for the forward LSTM and backward LSTM are as follows: By concatenating the bidirectional hidden states along the feature dimension, a temporal comprehensive representation is obtained. This enables complete capture of bidirectional information about the temporal context.

9. The method for predicting matte grade in oxygen-enriched side-blown copper smelting based on multi-model fusion as described in claim 1, characterized in that, In step S4, the prediction model and objective function of the XGBoost branch are: in, Here, M represents the predicted grade of copper matte, and M is the number of decision trees. For loss function, is the regularization term, γ and λ are the regularization coefficients, T is the number of leaf nodes in a single decision tree, and w is the weight of the leaf node, used to suppress model overfitting.

10. The method for predicting matte grade in oxygen-enriched side-blown copper smelting based on multi-model fusion as described in claim 1, characterized in that, In step S5, the multi-head attention fusion mechanism includes multi-head attention calculation and gated weighted fusion, and the basic attention formula is: In the formula, Q, K, and V are the query matrix, key matrix, and value matrix, respectively. Let h be the dimension of the key matrix and h be the number of attention heads; let the output features of the TCN, BiLSTM, and XGBoost branches be respectively... , , Adaptive fusion is achieved through learnable gating parameters α and β: 。

Images