Copper blowing process oxygen-enriched air control optimization method based on improved Informer and Kalman filtering

Abstract

The invention relates to the technical field of copper blowing process oxygen-enriched air control optimization, and discloses a copper blowing process oxygen-enriched air control optimization method based on improved Informer and Kalman filtering, and the method comprises the steps: obtaining the historical data of an oxygen-enriched bottom blowing furnace, inputting the historical data into an improved Informer model for prediction, and obtaining the future oxygen-enriched air input amount; the real-time observation data of the oxygen-enriched bottom blowing furnace are obtained, Kalman filtering dynamic correction is conducted in combination with the future oxygen-enriched air input quantity, state estimation based on Kalman filtering and the future oxygen-enriched air input quantity are obtained, model prediction control is executed, and the optimal control quantity of the oxygen-enriched air valve opening degree is dynamically adjusted. According to the method, a prediction result of the improved Informer and state estimation of Kalman filtering are combined, a multi-step rolling optimization framework is designed, the valve opening range and the physical constraint of change rate limitation are solved and applied through quadratic programming, and smooth adjustment of the valve opening is achieved so as to ensure the stability of the blowing process.

Description

TECHNICAL FIELD

[0001] The application relates to the technical field of oxygen-enriched air control optimization in a copper blowing process, in particular to a copper blowing process oxygen-enriched air control optimization method based on an improved Informer and Kalman filtering. BACKGROUND

[0002] The copper smelting industry is an important pillar of China's non-ferrous metal industry, and its technological progress is of great significance to improving resource utilization efficiency, reducing environmental pollution, and enhancing international market competitiveness. The oxygen-enriched bottom-blown furnace copper blowing technology has become a trend in the industry due to its high efficiency and energy saving. However, its complex working mechanism and high dynamic characteristics pose a great challenge to traditional control methods, resulting in high energy consumption and product quality fluctuations in the production process. This research introduces deep learning technology to predict and control the key parameters of the oxygen-enriched bottom-blown furnace copper blowing process, aiming to break through the limitations of traditional methods and improve the intelligent level of the process. The oxygen-enriched bottom-blown furnace copper blowing technology has become one of the mainstream technologies in modern copper smelting processes. The oxygen-enriched bottom-blown furnace copper blowing process involves multivariable coupling, nonlinear dynamics, and uncertainty factors, making it a typical complex industrial system. Applying deep learning technology to parameter prediction and control optimization in this process can enrich the theoretical system of industrial process modeling and control. However, the process involves complex physical and chemical reactions, and the input of oxygen-enriched air as a key control parameter exhibits high-frequency fluctuation characteristics, directly affecting the blowing efficiency and furnace stability. Traditional control methods such as PID are simple to implement, but they are slow to respond to high-frequency disturbances and process mutations, making it difficult to meet real-time and precision requirements. SUMMARY

[0003] To overcome the shortcomings of the prior art, the present application provides a copper blowing process oxygen-enriched air control optimization method based on an improved Informer and Kalman filtering, which has the advantages of ensuring the stability of the blowing process and solving the above technical problems.

[0004] To achieve the above purpose, the present application provides the following technical scheme: a copper blowing process oxygen-enriched air control optimization method based on an improved Informer and Kalman filtering, including an Informer model, specifically including the following steps: S1: Obtain the historical data of the oxygen-enriched bottom-blown furnace and input it into the improved Informer model for prediction to obtain the future oxygen-enriched air input; S2: Obtain the real-time observation data of the oxygen-enriched bottom-blown furnace and combine the future oxygen-enriched air input predicted in S1 to perform Kalman filtering dynamic correction to obtain state estimation based on Kalman filtering; S3: Based on the state estimation from the Kalman filter and the future oxygen-enriched air input in S1, execute model predictive control to dynamically adjust the optimal control quantity for the opening of the oxygen-enriched air valve.

[0005] As a preferred technical solution of the present invention, the S1 improved Informer model specifically includes a feature embedding layer, a multi-head attention mechanism, and a feedforward neural network connected in sequence, and a layer normalization is set after the multi-head attention mechanism and the feedforward neural network respectively.

[0006] As a preferred embodiment of the present invention, the feature embedding layer performs high- and low-frequency feature decomposition and embedding on the input historical data of the oxygen-enriched bottom-blown furnace, and the specific steps are as follows: S1.1: Each variable sequence Applying the Fast Fourier Transform, we obtain the frequency domain representation; S1.2: Separate low-frequency components and high frequency components These constitute low-frequency and high-frequency sequences, respectively; in, , They represent respectively; S1.3: Input the low-frequency sequence and the high-frequency sequence into the multilayer perceptron for embedding, and generate the corresponding embedding vectors. The specific expressions are as follows: in, and These represent low-frequency time series subsequences and high-frequency time series subsequences, respectively. This represents the embedding vector of the low-frequency sequence. Represents the embedding vector of a high-frequency sequence. This indicates the low-frequency components separated during the copper blowing process. High-frequency time series subsequences Convert to embedding vector Operation; S1.4: Based on and The total embedding vector of variables is generated through weighted fusion. ; The multi-head attention mechanism specifically involves: during the Informer model training process, the query vector... Weighted adjustment of scores, and retention of only the scores before weighting in subsequent calculations. The key participates in the calculation: in, Indicates the query vector Weighted adjusted score, of which, Represents the key vector. denotes a variable weight factor, is a key vector dimension; The feedforward neural network is specifically: the GLUE activation function is used to replace the ReLU activation function, and the specific expression of the GLUE activation function is as follows: Wherein, denotes the numerical value input to the activation function, denotes the ratio of a circle, denotes the hyperbolic tangent function; The layer normalization has the specific expression as follows: Wherein, denotes a variable token set, denotes a feature vector of a single variable token, denotes the mean of , and denotes the variance of .

[0007] As a preferred technical scheme of the present application, the specific steps of the Kalman filter dynamic correction in S2 are as follows: S2.1: Construct a state space model, and the specific expression is as follows: Wherein, denotes the remaining state vector inside the blowing furnace, denotes the remaining state vector inside the blowing furnace at the next moment (t+1 moment), is the state dimension, denotes the input amount of oxygen-enriched air, denotes the actually measured input amount of oxygen-enriched air, containing sensor noise, is a state transition matrix, which can be obtained by fitting historical data, is a control input matrix, which represents the influence of the input amount of oxygen-enriched air on the state, is an observation matrix, is process noise, which is a Gaussian white noise with a mean of zero and a covariance of , is observation noise, which is a Gaussian white noise with a mean of zero and a covariance of ; S2.2: Based on the state estimation and the control input of the last moment, the prior state at the current moment is predicted through the state transition equation, and the prior state estimation is calculated as: Wherein; is the state estimate of the last moment (t-1 moment), is the control input of the last moment (t-1 moment), A is a state transition matrix, and B is a control input matrix. is the state estimate of the last moment (t-1 moment), is the control input of the last moment (t-1 moment), A is a state transition matrix, and B is a control input matrix.

[0008] The prior covariance matrix is updated as follows: Wherein: is the covariance matrix of the state estimate of the last moment, A is a state transition matrix, and Q is a process noise covariance.

[0009] S2.3: Based on the actual observation value collected and the future oxygen-enriched air input amount predicted by S1 to form a comprehensive observation value , and the specific expression is as follows: Wherein, indicates a weight coefficient; S2.4: Based on the comprehensive observation value , the state estimate is updated as follows: Wherein, is a Kalman gain, and C is an observation matrix.

[0010] As a preferred technical scheme of the application, S3 includes converting the optimization problem of dynamically adjusting the optimal control amount of the oxygen-enriched air valve opening degree into a quadratic programming, and the specific expression is as follows: Wherein, is a constraint matrix, H is a quadratic term weight matrix, is a first-order coefficient vector, U is an auxiliary vector of control input, indicates a minimum value function, and the superscript indicates transposition; S3 further includes solving the quadratic programming by using a QP solver to obtain a final optimal valve control sequence ; Wherein, is an optimal valve opening degree sequence obtained after optimization solving, indicates the optimal control opening of the oxygen-enriched air valve at t moment.

[0011] Compared with the prior art, the application provides a copper blowing process oxygen-enriched air control optimization method based on improved Informer and Kalman filtering, which has the following beneficial effects: The application improves the Informer model, and designs high and low frequency feature decomposition, improved probability sparse self-attention mechanism, GELU activation function, and variable independent layer normalization and other optimization strategies for the high frequency and multi-scale characteristics of the copper blowing process data, enhances the model's ability to capture the timing dynamics of the oxygen-rich air input, reduces the computational complexity, improves the industrial real-time adaptability, and combines Kalman filtering to build a state space model and design a weighted fusion strategy to dynamically integrate the improved Informer prediction results and actual observation data, effectively correcting sensor noise and model bias, further improving the accuracy of oxygen-rich air input estimation, and finally introducing model predictive control (MPC) based on the improved Informer prediction value and Kalman filter state estimation, designing a multi-step rolling optimization framework, solving by quadratic programming and imposing physical constraints such as valve opening range and change rate, realizing smooth adjustment of valve opening, and ensuring the stability of the blowing process. BRIEF DESCRIPTION OF DRAWINGS

[0012] Figure 1 is a schematic diagram of the overall architecture of the application; Figure 2 is a schematic diagram of the improved Informer oxygen-rich air input prediction architecture for the copper blowing process of the application; Figure 3 is a comparison diagram of the oxygen-rich air input control value and the actual value of the oxygen-rich bottom blowing furnace copper blowing process under the control of each model of the application; Figure 4 is a comparison diagram of the change of the oxygen concentration in the furnace under the control of each optimization model of the application; Figure 5 is a diagram showing the change in cosine similarity under the control of the oxygen-rich air control method of the copper blowing process of the application at different control steps. DETAILED DESCRIPTION

[0013] The technical solutions in the embodiments of the application will be described clearly and completely below with reference to the drawings in the embodiments of the application. Obviously, the described embodiments are only part of the embodiments of the application, not all. Based on the embodiments in the application, all other embodiments obtained by those skilled in the art without creative labor are within the scope of protection of the application.

[0014] Please refer to Figures 1-5 , the copper blowing process oxygen-rich air control optimization method based on improved Informer and Kalman filter, including Informer model, specifically including the following steps: S1: Obtain the historical data of the oxygen-rich bottom blowing furnace and input it into the improved Informer model to predict the future oxygen-rich air input; Input historical data The historical data includes oxygen-enriched air input, valve opening degree, etc. is the length of the historical window, N is the number of variables, including oxygen-enriched air input, valve opening degree, etc., which are directly mapped to the predicted values of the future time steps after being extracted by the encoder ; The S1 improved Informer model specifically includes a feature embedding layer, a multi-head attention mechanism, and a feedforward neural network connected in sequence, and layer normalization is set after the multi-head attention mechanism and the feedforward neural network, respectively. In Informer, multivariate at the same time is embedded as a time token, and the time dependence is captured through the attention mechanism. However, in the copper blowing process, the oxygen-enriched air input, valve opening degree, and oxygen concentration set value have different physical meanings and dynamic characteristics. This embedding method may flatten the independence between variables and fail to accurately reflect the time sequence characteristics of each variable, especially in high-frequency sampling scenarios, where trend and periodic characteristics are mixed. Therefore, this paper proposes an optimization method based on high and low frequency feature decomposition and embedding. The trend and periodic characteristics of the variable sequence are decomposed through Fourier transform and then embedded. The feature embedding layer decomposes and embeds the high and low frequency characteristics of the input oxygen bottom blowing furnace historical data. The specific steps are as follows: S1.1: Decompose each variable sequence Apply fast Fourier transform to get the frequency domain representation; S1.2: Separate the low frequency component and the high frequency component , respectively, to form the low frequency sequence and the high frequency sequence; wherein , represent, respectively; S1.3: Input the low frequency sequence and the high frequency sequence into the multi-layer perception respectively for embedding to generate the corresponding embedding vectors. The specific expression is as follows: wherein, and represent the low frequency time sequence subsequence and the high frequency time sequence subsequence, represents the embedding vector of the low frequency sequence, represents the embedding vector of the high frequency sequence, represents the operation of converting the low frequency and high frequency time sequence subsequence separated from the copper blowing process into embedding vectors ; S1.4: Generate the total embedding vector of the variable based on and by splicing or weighted fusion ; Traditional Informer models employ probabilistic sparse self-attention, which, compared to Transformer models, reduces time complexity from... reduce By filtering the dominant query and focusing only on a subset of keys, long-term series prediction tasks can be effectively handled. However, in high-frequency data scenarios related to copper smelting processes, historical windows... The length can reach thousands of steps, even The complexity still easily leads to memory overflow, and the original mechanism lacks specificity in the sparse selection among variables, failing to fully capture the dynamic correlation between oxygen-enriched air input and other process parameters. Therefore, this paper further improves the mechanism design based on Informer's probabilistic sparse attention, enhancing computational efficiency and task adaptability, such as... Figure 1 As shown.

[0015] The improvement in this paper lies in the introduction of trainable variable weight factors. For each variable, the query selection strategy in sparse attention is dynamically adjusted during model training. Specifically, the multi-head attention mechanism in attention computation involves adjusting the query vector during Informer model training. Weighted adjustment of scores, and retention of only the scores before weighting in subsequent calculations. The key participates in the calculation:

[0016] in, Indicates the query vector Weighted adjusted score, of which, Represents the key vector. Indicates the variable weight factor. It is a key vector The dimension; Informer uses a feedforward network (FFN) in both the encoder and decoder to process the representation of time tokens. However, its main function is limited to extracting local features between variables, failing to utilize global information on the time series of a single variable. Furthermore, it typically employs the ReLU activation function, which is prone to gradient vanishing in complex nonlinear modeling. To adapt to the long-term dependence and nonlinear characteristics of oxygen-enriched air input in copper blowing, this paper enhances the functionality of the feedforward network, enabling it to focus on sequence representation learning and apply it to the time dimension of the variable sequence. The feedforward network in this paper not only extracts periodic and trend global order representation, but also implicitly stores time sequence information to replace the explicit position encoding in Informer. In addition, to further improve the non-linear modeling ability, this paper introduces GELU (Gaussian Error Linear Unit) activation function in FFN to replace the traditional ReLU. GELU combines the sparsity of ReLU and the smoothness of Sigmoid, and the specific expression is as follows: wherein, represents the value input to the activation function, represents the ratio of the circumference of a circle, represents the hyperbolic tangent function, which can better capture the non-stationary fluctuations of the oxygen-rich air input in the copper blowing high-frequency data, and improve the fitting ability of the model to complex dynamics; Layer normalization was originally proposed to improve the convergence speed and training stability of deep networks. In traditional Transformer prediction models such as Informer, layer normalization is applied to the multi-variable representation at the same time point, gradually integrating the features of each variable. However, in the copper blowing process, the time points of variables such as oxygen-rich air input and valve opening may not represent the same physical event. This operation will introduce non-causal or delayed interaction noise between processes, leading to prediction bias. Therefore, this paper optimizes the application of layer normalization, which is applied to the independent representation of each variable sequence.

[0017] Specifically, for a variable token set wherein, is the embedding dimension, and the layer normalization in this paper is defined as: wherein, represents the variable token set, represents the feature vector of a single variable token, represents the mean of , and represents the variance of ; Unlike the time step normalization in Informer, this method independently normalizes the representation of each variable sequence to approximate a Gaussian distribution, which has been proven to be more effective in dealing with non-stationary problems. In the copper blowing scene, the measurement units and distributions of oxygen-rich air input and other process parameters are quite different, and this improvement can significantly reduce the impact of these inconsistencies and avoid the problem of excessive smoothing between time steps. S2: Obtain real-time observation data of the oxygen-rich bottom blowing furnace, and combine the predicted future oxygen-rich air input in S1 to perform Kalman filter dynamic correction to obtain state estimation based on Kalman filter; The core of Kalman filter is to describe the dynamic evolution of the system through a state-space model. In the copper smelting process, the change of oxygen-enriched air input depends not only on historical data but also on the hidden state of the furnace. Therefore, this paper abstracts the smelting process as a linear time-invariant system and establishes a state-space model as follows: where, represents the remaining state vector inside the smelting furnace, is the state dimension; represents the oxygen-enriched air input as a controllable variable of the system; represents the actual measured oxygen-enriched air input, but contains sensor noise. In addition, is the state transition matrix, which describes the evolution law of the hidden state over time, and can be obtained by fitting historical data; is the control input matrix, which represents the influence of oxygen-enriched air input on the state, is the observation matrix, which maps the state to the observation space; is the process noise, which is assumed to be a Gaussian white noise with mean zero and covariance , reflecting the random disturbance in the system; is the observation noise, which is assumed to be a Gaussian white noise with mean zero and covariance , reflecting the sensor measurement error; Kalman filter realizes the optimal estimation of the state by fusing prior estimation and observation data through prediction and update steps. In this study, Kalman filter not only uses actual observation values, but also combines the prediction results of the improved Informer model to further improve the estimation accuracy. The design and implementation process is as follows: The implementation of Kalman filter starts with the prediction step, based on the state estimation and control input at the last time, the prior state at the current time is predicted through the state transition equation, and the prior state estimation is calculated as: At the same time, in order to quantify the uncertainty of prediction, the prior covariance matrix is updated as: In the update stage, Kalman filter further modifies the state estimation by fusing actual observation data and prediction information. The actual observation value collected and the future oxygen-enriched air input predicted by the improved Informer model are considered to generate more reliable estimates. For this purpose, this paper designs a weighted fusion strategy to combine the current observation and prediction value to form a comprehensive observation value: where, is an adjustable weight coefficient, used to balance the contribution of actual observation and predicted value. If the sensor noise is large, can be reduced to enhance the dependence on prediction; conversely, if the prediction accuracy of the improved Informer model is higher, then is increased, and then the Kalman gain is calculated: The gain determines the weight of the observation data on the state correction, and its value reflects the trust degree allocation between the prior estimate and the observation. Then, based on the comprehensive observation value , the state estimation is updated as: By comparing the deviation between the predicted output and the comprehensive observation , the prior state is dynamically corrected, thereby reducing the influence of noise and prediction deviation. At the same time, the posterior covariance is updated as: Thereafter, based on the updated state , the corrected enriched air input estimate is backstepped through the observation equation: In this way, the traditional Kalman filter utilizes the prediction ability of the improved Informer model, not only making up for the errors that may be introduced by single observation, but also capturing the internal dynamic characteristics of the copper blowing process through the state space model; S3: Based on the state estimation of Kalman filter and the future enriched air input in S1, the model predictive control is executed to dynamically adjust the optimal control amount of the enriched air valve opening degree; The control objective of this study is to optimize the enriched air valve opening degree (unit: %), to indirectly control the enriched air input, and to ensure the efficiency and stability of the copper blowing process, where the efficiency is reflected in maintaining the oxygen concentration in the furnace through appropriate valve opening degree, promoting sulfide oxidation and impurity removal; the stability requires smooth changes in valve opening degree to avoid excessive fluctuations in input leading to instability of the furnace condition. To achieve this goal, the physical constraint conditions of process parameters need to be considered comprehensively, which are summarized as follows: (1) The valve opening degree range is limited to 0-100%, which is derived from the actual production records of the enriched air bottom blowing furnace in Guixi Smelter of Jiangxi Copper, reflecting the physical limitations of the actual control equipment.

[0018] (2) The enriched air oxygen concentration setpoint range is 60-80%, which is a reasonable interval for oxygen supply in the copper blowing process.

[0019] (3) Considering the continuity of the blowing process, the rate of change of the opening degree of the oxygen-enriched air valve. It also needs to be limited to avoid excessive adjustments impacting the system; In this study, MPC utilized an improved Informer model to predict future oxygen-enriched air input. The state estimated by Kalman filtering Build a prediction model and design optimization objectives.

[0020] The prediction model is the core of MPC, built upon a state-space model based on Kalman filtering, for current state estimation. The state has already been obtained through Kalman filtering; future state prediction is then performed through iterative calculation. in, For the future valve opening control sequence, the initial values ​​can be indirectly calculated from the predicted input. The optimization objective aims to balance control deviation and energy consumption, defined as the future... Cost function for each time step: in To achieve the desired oxygen-enriched air valve opening, This indicates the rate of change of valve opening, used to penalize excessive adjustment actions. The regularization coefficient balances control accuracy and stability. This multi-step optimization method comprehensively considers future trends, avoids local optima, and suppresses input fluctuations by defining the control oxygen-enriched valve opening vector. , target input vector The objective function can then be expressed as: in, It is a difference matrix used to calculate Its form is: After expansion, the objective function can be further simplified to a standard quadratic form: in, for The identity matrix, It is a symmetric positive definite matrix, called a Hessian matrix. The vector of linear coefficients, and the constant term. This can be ignored during optimization; The constraints stem from the physical limitations of the process parameters in the oxygen-enriched bottom-blown copper smelting process. Here, the valve opening range corresponds to the upper and lower boundaries, set as follows: and Meanwhile, the range of oxygen concentration set value indirectly restricts the input, which is considered to be linearly related to the oxygen concentration, and can be converted into Additional constraints. In addition, the limit of input rate Avoids excessive adjustment, set to These constraints can be summarized as linear inequalities: Integrated into matrix form: Where, , Is a full 1 vector.

[0021] In summary, the optimization problem of MPC can be formalized as a quadratic programming: This problem has convexity, and this paper chooses to use QP solver to solve, and gets the final optimal valve control sequence According to the rolling optimization principle of MPC, only the current control input Is executed, And repeat the process at the next time step Using the updated state And the predicted value And re-optimization.

[0022] The QP-based solution method proposed in this study has the following advantages, among which the quadratic form of the objective function and the linear constraints guarantee the existence and uniqueness of the global optimal solution, and the positive definiteness of the Hessian matrix Ensures the numerical stability of the optimization problem. In addition, by adjusting Can flexibly balance the control accuracy and input smoothness, compared with directly using the predicted value of the neural network, by combining the state space model and the constraint conditions, the process restrictions of copper blowing can be considered in multi-step prediction, avoiding the control error caused by a single prediction point; The pseudo code of the improved Informer and Kalman filter copper blowing process oxygen-enriched air control optimization method is shown in Table 1.

[0023] Table 1 Pseudo code of this chapter algorithm ; This algorithm describes the closed-loop control optimization method of oxygen-enriched air input in the copper blowing process, which realizes dynamic adjustment through the cooperative work of the prediction module (Informer), state estimation module (KALMAN) and optimization control module (MPC). The whole process runs in a loop, as follows: In the prediction module (Informer), the system predicts future oxygen-enriched air input demand based on historical data. This module uses fast Fourier transform to decompose the data into low and high frequency components, extracts features through multi-layer perceptron and fuses them, and then generates a prediction sequence of input quantities with the help of an encoder. This step can capture the dynamic trend of input quantities and provide forward-looking guidance for subsequent control.

[0024] The state estimation module (Kalman) corrects the deviation between the predicted value and the actual observation value through Kalman filtering technology. Specifically, the module combines the state prediction equation with a weighted fusion mechanism to dynamically integrate real-time sensor data and predicted values in proportion, continuously updating the estimated state of the current input quantity. This design significantly improves the system's robustness to production environment fluctuations and measurement noise, ensuring that the state estimation is closer to the real working conditions.

[0025] Subsequently, the optimization control module (MPC) generates optimal valve control instructions by solving a constrained quadratic programming problem based on the corrected state estimation. The optimization process considers tracking target input quantities, minimizes control fluctuations, and satisfies valve opening threshold and change rate constraints. The final output control quantity is applied in real time to adjust the oxygen-enriched air valve, dynamically balancing the input quantity demand and process constraints.

[0026] The operation of the entire system is centered on the main loop process, with each time step sequentially executing prediction, state correction, and optimization control. The prediction module provides trend reference for future input demand, the state estimation module fuses real-time observations to correct the current state, and the optimization control module generates precise instructions. The closed-loop cooperation of the three ensures that the oxygen-enriched air input quantity remains stable in the dynamically changing copper smelting process, ultimately achieving the goals of energy efficiency optimization and product quality improvement; The data set used in this study is the actual production record of the oxygen-enriched bottom blowing furnace of Guixi Smelter of Jiangxi Copper. To address the noise and outliers that may exist in actual industrial data, this study uses a robust normalization method to standardize each variable data in the time series dimension. The specific formula is: where, is the median, is the interquartile range. This formula converts each variable into a distribution centered on the median and scaled by the interquartile range, which is more resistant to outliers than traditional normalization methods.

[0027] The preprocessed data was used in three stages of the model: improved Informer prediction, Kalman filter state estimation correction, and MPC optimization control. In the prediction stage, the input consisted of six variables from a historical window: oxygen-enriched air input, oxygen concentration setpoint, oxygen-enriched air valve opening, auxiliary air volume, furnace start / stop, and raw material feeding start / stop. The length of the historical time series, the output is the future. The oxygen-enriched air input at each time step is calculated; the estimation and correction stage integrates the actual observed oxygen-enriched air input and the predicted value, and outputs the corrected input estimate; the control optimization stage outputs the optimal oxygen-enriched air valve opening control sequence based on the state estimation.

[0028] The dataset is divided into a training set (70%, the first 166,315 time steps), a validation set (15%, the middle 35,632 time steps), and a test set (15%, the last 35,646 time steps) in chronological order to avoid time series data leakage and ensure that the experimental results can truly reflect the model's performance in real industrial scenarios. In the improved Informer model, high- and low-frequency eigenvalue decomposition is achieved by applying a Fast Fourier Transform (FFT) to each variable sequence, with the cutoff frequency... The value is set to 0.1. The low-frequency and high-frequency components are embedded through two layers of MLP, with a hidden layer dimension of 64 and an output embedding dimension of [missing value]. The encoder structure retains 3 layers, each containing an improved probabilistic sparse self-attention module and a feedforward network module. The number of attention heads is set to 8, with each head having a dimension of 16. The total dimension and the embedding dimension are calculated as follows. Consistent; the improved probabilistic sparse self-attention mechanism incorporates variable weight factors. Initialize to uniform distribution Through dynamic adjustment during training, the first log(T_1) (approximately 10) keys are selected for computation to ensure a sparsity rate exceeding 90%. Each layer of the feedforward network (FFN) contains two linear transformations: the first layer expands the dimension from 128 to 512, and the second layer compresses it back to 128. The dropout rate is set to 0.1 to prevent overfitting. Layer normalization is applied independently to each variable sequence to smooth the parameters. Set to 1×10⁻ 5 Training was performed using the AdamW optimizer with an initial learning rate of 1×10⁻ 4 With the cosine annealing scheduling strategy, the minimum learning rate is 1×10⁻ 6 The batch size is 32, the total number of training rounds is 500, and the early stopping patience value is 20 rounds.

[0029] For the Kalman filter module, the state dimension of the state-space model Set to 6, state transition matrix and control input matrix Process noise covariance is initialized by historical data least square method and fine-tuned during training is initialized as Observation noise covariance is initialized as is set as 0.05, which is adjusted to the optimal value by experiments, and is the comprehensive observation weight coefficient is set as 0.7, and the prediction and update at each step are based on the current state and the prediction value of improved Informer Real-time calculation of Kalman gain In the MPC optimization control module, the rolling optimization step is 1, and the regularization coefficient of the objective function is set as 0.5 to balance the control bias and the rate of change of valve opening, and the rate constraint The oxygen concentration range of 25%-40% is converted to by linear mapping Constraint, the solver uses OSQP to handle the quadratic programming problem, and the maximum number of iterations is set to 1000 and the tolerance is set to 1×10⁻ 6 During the training process, the improved Informer is first independently pre-trained, and then jointly optimized with Kalman filtering and MPC, the validation set is used for hyperparameter tuning, and the test set is used for final performance evaluation; To verify the effectiveness of the oxygen-enriched air control optimization method based on improved Informer and Kalman filtering proposed in this study, this section experiment takes the mean square error (MSE) of the oxygen-enriched air input control result and the actual demand and the cosine similarity (Cosine Similarity) between them as evaluation indexes, reflecting the accuracy and trend consistency of the control strategy.

[0030] The selected comparison methods are as follows, including: (1) classical proportional-integral-derivative (PID) control, the target is to adjust the valve opening based on the deviation of the real-time measured oxygen-enriched air input and the set value, and the parameters The Ziegler-Nichols method was used to set the PID parameters to 1.2, 0.5, and 0.3, respectively, and the control frequency was consistent with the data sampling frequency. (2) Extended Kalman Filter (EKF) control, which improved the traditional KF for the nonlinear characteristics of the copper blowing process, linearized the nonlinear system equation through Taylor expansion, set the state dimension to 6, and set the initial covariance and KF-Control to be the same, but the Jacobian matrix needed to be calculated every iteration. (3) LSTM + Kalman Filter Control (LSTM + KF), which first used the LSTM model to predict the system state, and then optimized the valve opening based on the estimated value, with noise covariance Q and R set to 0.01 and 0.05, respectively. (4) Deep Deterministic Policy Gradient (DDPG), which is a continuous action space control method based on reinforcement learning, learns the optimal control policy through the Actor-Critic architecture, and the reward function is defined as the negative sum of the square of the deviation between the oxygen-rich air input and the target value. The experimental results of each model are shown in Table 4-2, which records the mean and standard deviation of the MSE and cosine similarity of each method based on the test set 5 times.

[0031] Table 2 Performance comparison of different control methods Table 2 shows that the MSE of PID control is 0.1152, the cosine similarity is 0.572, and the standard deviation is the smallest, but the precision is very limited. The main reason is that it relies on real-time feedback and cannot effectively predict future trends, which can easily cause oscillation under high-frequency disturbance, resulting in direct distortion of the control in the later period; the MSE of EKF is 0.0389, and the cosine similarity is 0.891, which is improved compared to PID, thanks to the noise filtering of state estimation, but it lacks prediction correction of observation values, limiting the modeling ability of future valve control amount nonlinear dynamics. The LSTM + KF method reduces the MSE to 0.0327 by adding a neural network prediction method to build a higher quality state update, but the LSTM model itself has limited ability to extract high-frequency time series features in the oxygen-rich bottom blowing furnace, and the prediction of the KF method is sensitive to the initial parameters; The MSE of DDPG is 0.0224, showing strong adaptability, and the performance is close to the models in this chapter, indicating that reinforcement learning can learn some potential rules through long-term optimization, but its training time is long and it is highly dependent on the reward function design, which is limited in the actual oxygen-rich bottom blowing furnace copper blowing scene with high real-time requirements. The MSE of the method in this chapter is 0.0215, and the cosine similarity reaches 0.937, both of which are better than all the above methods. Figure 3The control curve of the oxygen-enriched air input between time steps 5000 and 5256 in the test set is compared with the true demand. PID control has obvious lag in trend following, resulting in direct distortion of control in the later period and complete deviation from the true demand. KF-Control and EKF-Control can smooth the noise well, but the former has slow response speed in mutation, and the latter, although improved, still cannot completely eliminate the deviation, especially when the oxygen concentration set value mutates, the adjustment process overshoots and oscillates. The control curve of DDPG is close to the true demand, but there is overfitting phenomenon in local fluctuations, while the control curve of the method in this chapter is closest to the true demand, maintaining high accuracy in the smooth section and quickly responding and stabilizing in the mutation section, which reflects the advantages of prediction and state estimation combination; As the core environmental state of the blowing process, the oxygen-enriched air content directly reflects the control ability of the control strategy on the furnace condition. To further evaluate the actual effect of each model optimization control strategy in the oxygen-enriched bottom-blown furnace copper converting process, this study compared the matching degree of the oxygen-enriched air content (i.e., oxygen concentration) in the furnace under different control methods with the true demand, Figure 4 The comparison curve of the oxygen-enriched air content in the furnace under each model optimization control strategy with the true value between time steps 0 and 140 in the test set is shown in Figure 6; As Figure 4 shown in Figure 6, the oxygen-enriched air content curve of PID control shows a large deviation in following the true demand, especially when the oxygen concentration set value mutates, the response lag is obvious, the curve deviates far from the true value, and even oscillates in the later period, reflecting its insufficient adaptability to high-frequency disturbances and nonlinear dynamics. The EKF and LSTM+KF methods improve the tracking effect of oxygen concentration through state estimation and prediction ability, but EKF still has overshoot phenomenon near the mutation point, while LSTM+KF can capture the trend well, but has a large deviation in local high-frequency fluctuations, indicating its limited modeling ability for complex time series characteristics. The DDPG method makes the oxygen concentration curve close to the true value due to the long-term optimization ability of reinforcement learning, but still has slight fluctuations at some time steps, indicating that its robustness to instantaneous disturbances needs to be improved. In contrast, the control method based on the improved Informer and Kalman filter proposed in this chapter has the oxygen-enriched air content curve closest to the true demand, not only maintaining high accuracy in the smooth section, but also quickly responding and stabilizing in the mutation section, which reflects the synergistic advantages of prediction, state estimation, and multi-step optimization. This result further verifies the superiority of the method in maintaining the stability of the oxygen concentration in the furnace, promoting the oxidation of sulfides, and removing impurities.

[0032] In addition, in order to explore the control accuracy of the model in this chapter under different prediction steps, this study takes the cosine similarity as the index to analyze the performance under control steps of 10, 50, 100, 200, 500 and 1000, and each step is repeated 5 times to evaluate the accuracy and stability. The results are shown in Figure 5 As shown in the table, the control accuracy is the highest under short steps, and with the increase of steps, the average cosine similarity gradually decreases, and the standard deviation slightly increases. Especially after the step is more than 200, the decline is slightly more obvious due to the limitation of the modeling ability of the model on long-term dependence and the influence of high-frequency disturbance in the data, but the cosine similarity is still 0.8531 under the long step of 1000, which verifies the superiority of the model under long step control.

[0033] Although embodiments of the present application have been shown and described, it is to be understood that various modifications, substitutions, replacements and changes can be made to these embodiments without departing from the principles and spirit of the present application, and the scope of the present application is defined by the appended claims and their equivalents.

Claims

1. A method for control optimization of oxygen-enriched air in copper converting process based on improved Informer and Kalman filter, comprising an Informer model, characterized in that: Specifically comprising the following steps: S1: obtaining the oxygen-rich bottom blowing furnace historical data and inputting into the improved Informer model for prediction to obtain the future oxygen-rich air input amount; S2: obtaining the real-time observation data of the oxygen-rich bottom blowing furnace, and combining the future oxygen-rich air input amount predicted in S1 to perform Kalman filter dynamic correction to obtain state estimation based on Kalman filter; S3: based on the state estimation based on Kalman filter and the future oxygen-rich air input amount in S1, performing model predictive control to dynamically adjust the optimal control amount of the oxygen-rich air valve opening degree.

2. The improved Informer and Kalman filter based control optimization method for oxygen-enriched air in copper converting process as claimed in claim 1 wherein: The S1 improved Informer model specifically comprises a feature embedding layer, a multi-head attention mechanism and a feedforward neural network connected in sequence, and layer normalization is arranged after the multi-head attention mechanism and the feedforward neural network respectively.

3. The improved Informer and Kalman filter based control optimization method for oxygen enriched air in copper converting process as claimed in claim 2 wherein: The feature embedding layer performs high and low frequency feature decomposition and embedding on the input oxygen-rich bottom blowing furnace historical data, and the specific steps are as follows: S1.1: each variable sequence apply a fast Fourier transform to obtain a frequency domain representation; S1.2: separating low frequency components and high frequency components , respectively, to form low frequency sequences and high frequency sequences wherein, represents an input frequency characteristic component, represents a high-low frequency characteristic component threshold value; S1.3: input the low frequency sequence and the high frequency sequence into the multilayer perception machine respectively for embedding to generate corresponding embedding vectors, and the specific expression is as follows: wherein, and respectively represent a low-frequency timing sub-sequence and a high-frequency timing sub-sequence, represents an embedding vector of the low-frequency sequence, represents an embedding vector of the high-frequency sequence, represents a function to low frequency , high frequency timing sub-sequences converted into embedded vectors operations; S1.4: based on and generating a total embedding vector for the variable by concatenation or weighted fusion ; The multi-head attention mechanism specifically comprises: in the training process of the Informer model, the query vector The weighted adjustment score is obtained, and only the weighted adjustment score is retained in subsequent calculation The key before the weighted adjustment score is retained is involved in calculation: wherein, represents a query vector is a weighted adjusted score, represents a key vector, represents a variable weight factor, is a key vector of dimensionality The feedforward neural network specifically replaces the ReLU activation function with the GLUE activation function, and the specific expression of the GLUE activation function is as follows: wherein, denotes a value input to an activation function, denotes pi, denotes a hyperbolic tangent function; The layer normalization, the specific expression is as follows: wherein, denotes a set of variable tokens, denotes a feature vector of a single variable token, denotes a mean of denotes a variance of 4. The improved Informer and Kalman filter based control optimization method for oxygen enriched air in copper converting process as claimed in claim 1 wherein: The specific steps of the Kalman filter dynamic correction in S2 are as follows: S2.1: constructing a state space model, the specific expression is as follows: the numerical value input to the activation function in, Let represent the remaining state vector inside the furnace at time t. This represents the remaining state vector inside the furnace at the next moment. For the state dimension, This indicates the amount of oxygen-enriched air input. This represents the actual measured amount of oxygen-enriched air input, including sensor noise. The state transition matrix can be obtained by fitting historical data. The control input matrix represents the effect of the oxygen-enriched air input on the state. For the observation matrix, Let be the process noise, with a mean of zero and a covariance of . Gaussian white noise, For observation noise, let have a mean of zero and a covariance of . Gaussian white noise; S2.2: State estimation based on the previous time step and control input The prior state estimate is computed by predicting the current time step state from the previous time step state using the state transition equation: wherein is the state estimate of the previous time instant, is the control input of the previous time instant, A is the state transition matrix and B is the control input matrix. The prior covariance matrix is updated as: wherein is the covariance matrix of the state estimation at the previous time, A is the state transition matrix, and Q is the process noise covariance. S2.3: Based on the actual observation value collected and the S1 predicted future oxygen-rich air input amount to form a comprehensive observation value The specific expression is as follows: wherein, represents a weight coefficient, represents the oxygen-enriched air input amount at the next time point; S2.4: Based on the integrated observations Update the state estimate: wherein, is the Kalman gain, C is the observation matrix, denotes a predicted output value based on the state estimation of the previous time.

5. The improved Informer and Kalman filter based control optimization method for oxygen enriched air in copper converting process as claimed in claim 1 wherein: The S3 includes converting the optimization problem of the optimal control amount of the dynamically adjusted oxygen-rich air valve opening degree into a quadratic programming, and the specific expression is as follows: wherein the constraint matrix, H is a quadratic weight matrix, is a vector of linear coefficients, U is an auxiliary vector of control inputs, denotes a minimum function, the superscript denotes the transpose; The S3 further includes solving the quadratic programming using a QP solver to obtain a final optimal valve control sequence ; wherein, To optimize the optimal valve opening sequence obtained by solving, represents the optimal control opening of the oxygen-enriched air valve at time t.

Images