A rapid prediction method for the performance of submerged arc furnaces for hot smelting of nickel iron ore in the steel process industry

Abstract

The present invention relates to the field of steel process industry, and specifically to a method for predicting the performance of a submerged arc furnace for hot smelting of nickel iron ore. In view of the difficulty in obtaining the performance of submerged arc furnaces due to complex and dangerous operating environments, and the long time-consuming multi-field simulation of electricity, heat and quality in the furnace, the present invention proposes a method for quickly predicting the performance of a submerged arc furnace for hot smelting of nickel iron ore in the steel process industry. The method couples a multi-head attention mechanism to mine the deep mapping relationship between input (electrode voltage, diameter-to-height ratio, electrode depth) and output parameters (smelting time parameters). The performance prediction accuracy of the submerged arc furnace for hot smelting of nickel iron ore is achieved to reach 96%, and the operating performance evaluation cycle is shortened to seconds to minutes. The relationship between input parameters and output parameters is mined, and compared with electrode depth and diameter-to-height ratio, electrode voltage has the greatest impact on the operating performance of the submerged arc furnace. The present invention can effectively reduce the operating performance evaluation cost of the submerged arc furnace for hot smelting of nickel iron ore.

Description

Technical Field

[0001] The invention belongs to the technical field of artificial intelligence and steel process optimization, relates to performance prediction of a submerged arc furnace, and in particular to a method for quickly predicting the performance of a submerged arc furnace for hot smelting of nickel-iron ore in the steel process industry. Background Art

[0002] Submerged arc furnaces are generally considered to be one of the most promising ore-fired furnaces in high-temperature smelting processes due to their high heating power and the wide variety of charge types they can process. However, the operating environment inside a submerged arc furnace is complex, and high heat and strong electromagnetics act together on the operation of the furnace, making it difficult to obtain performance inside the furnace. At the same time, the coupling mechanism of the electric-heat-mass multi-physical fields inside the submerged arc furnace is complex. The analysis of the operating performance of the submerged arc furnace under different working conditions requires a large amount of computing resources and takes a long time to calculate, which creates certain difficulties for real-time prediction of the operating conditions inside the furnace. In summary, the complex and dangerous operating environment of the submerged arc furnace and the current situation where a lot of resources are required to realize multi-field coupling simulation calculations make it difficult to predict the performance of the submerged arc furnace in real time, and it is difficult to efficiently guide the operation of the submerged arc furnace. Therefore, how to effectively realize the rapid prediction of the performance of the submerged arc furnace for hot smelting of nickel iron ore under different working conditions to guide the operation of the submerged arc furnace is a problem that technicians in this field urgently need to solve. Summary of the invention

[0003] In order to overcome the shortcomings of the above-mentioned prior art, the purpose of the present invention is to provide a method for quickly predicting the performance of a nickel-iron ore hot smelting submerged arc furnace in the steel process industry. Based on a deep table learning algorithm, the method solves the problems of repeated calling of calculation models, low calculation efficiency and large consumption of calculation resources in the full operating range of the nickel-iron ore hot smelting submerged arc furnace. The method realizes efficient evaluation of the operating performance of the nickel-iron ore hot smelting submerged arc furnace and real-time monitoring of the operating performance of the nickel-iron ore hot smelting submerged arc furnace, and shortens the estimation period of the operating performance of the nickel-iron ore hot smelting submerged arc furnace.

[0004] In order to achieve the above object, the technical solution adopted by the present invention is:

[0005] A method for quickly predicting the performance of a submerged arc furnace for hot smelting of nickel iron ore in the steel process industry comprises the following steps:

[0006] Step 1, determining the input parameters of the nickel iron ore hot smelting submerged arc furnace;

[0007] Step 2: Establish an electric field model to obtain the current density in the submerged arc furnace for hot smelting of nickel iron ore expression;

[0008] Step 3, constructing a heat-mass transfer model of a nickel-iron ore hot smelting submerged arc furnace based on the conservation of energy, momentum and mass, so as to obtain the expression of the temperature distribution T of the charge in the furnace and the smelting time parameter τ;

[0009] Step 4: Establish the electric-thermal-mass boundary conditions and solve the current density in the furnace Substitute into step 3 to obtain the temperature distribution T and smelting time parameter τ;

[0010] Step 5, defining a deep table learning model based on the input parameters, and preprocessing the data obtained in step 4, representing the input parameters as a one-dimensional vector, each numerical value in the vector corresponds to a different input label, and converting each numerical feature into a corresponding embedding representation;

[0011] Step 6, using the data obtained in step 4 to train the deep table learning model, and using the trained model to quickly predict the performance of the nickel iron ore hot smelting submerged arc furnace.

[0012] Compared with the prior art, the present invention has the following beneficial effects:

[0013] (1) The present invention combines the electric field model, heat and mass transfer model and deep learning algorithm of the nickel-iron ore hot smelting submerged arc furnace, which can realize the rapid prediction of the smelting time parameters of the nickel-iron ore hot smelting submerged arc furnace within a wide parameter operating range, and the performance prediction accuracy reaches 96%. The method of the present invention has strong versatility.

[0014] (2) The method of the present invention provides a dynamically adaptable way to mine the combination relationship between different features of each group of input data, and can adaptively discover the input feature combination that is most critical to the output result, and assign different weight coefficients to different features. According to the size of the weight coefficient, the correlation relationship between different features is dynamically mined. Compared with the electrode depth and the diameter-to-height ratio, the electrode voltage has the greatest impact on the operating performance of the submerged arc furnace.

[0015] (3) The method of the present invention has high calculation efficiency, which enables designers to quickly and efficiently complete the evaluation and calculation of the performance of the nickel iron ore hot smelting submerged arc furnace, avoid repeated calls of the calculation model, and realize the efficient evaluation of the operating performance of the nickel iron ore hot smelting submerged arc furnace. The estimated cycle of the operating performance of the nickel iron ore hot smelting submerged arc furnace is shortened to seconds to minutes. This method has important guiding significance and engineering application value for the design, optimization and performance evaluation of the nickel iron ore hot smelting submerged arc furnace. BRIEF DESCRIPTION OF THE DRAWINGS

[0016] Figure 1 This is a schematic diagram of the process flow of a method for rapid prediction of submerged arc furnace performance for hot smelting of nickel iron ore.

[0017] Figure 2 This is a schematic diagram of the structure of a submerged arc furnace for hot smelting of nickel-iron ore.

[0018] Figure 3 This is a block diagram of the neural network structure combined with the multi-head attention mechanism.

[0019] Figure 4 The performance prediction results of submerged arc furnace for hot smelting of nickel iron ore. DETAILED DESCRIPTION

[0020] The embodiments of the present invention are described in detail below with reference to the accompanying drawings and examples.

[0021] In view of the difficulty in obtaining the performance of submerged arc furnaces due to complex and dangerous operating environments, and the long time-consuming multi-field simulation of electricity, heat and mass in the furnace, the present invention proposes a method for quickly predicting the performance of nickel-iron ore hot smelting submerged arc furnaces in the steel process industry. A deep learning model is constructed using historical nickel-iron ore hot smelting submerged arc furnace operating performance data to explore deeper mapping relationships between input and output data, thereby achieving efficient calculation and prediction of the operating performance of nickel-iron ore hot smelting submerged arc furnaces. At the same time, the input data of the nickel-iron ore hot smelting submerged arc furnace operating tasks naturally has the characteristics of structured data, that is, the input data has clearly defined data labels and is easy for software and humans to efficiently access and understand. Deep table learning is further combined with the nickel-iron ore hot smelting submerged arc furnace operating performance analysis method to explore direct or indirect association relationships between different data labels, thereby achieving rapid, effective and more accurate result prediction. Finally, the present invention combines the electric-heat-mass multi-physical field coupling model of the nickel-iron ore hot smelting submerged arc furnace with a new algorithm of table learning and multi-head attention mechanism to evaluate the operating performance of the smelting submerged arc furnace, improve the development efficiency of the nickel-iron ore hot smelting submerged arc furnace, shorten the prediction and evaluation cycle of the operating performance of the nickel-iron ore hot smelting submerged arc furnace, and reduce the development cost of the nickel-iron ore hot smelting submerged arc furnace operating performance prediction.

[0022] Specifically, Figure 1 As shown, the main steps of the rapid prediction method of the present invention are as follows:

[0023] Step 1), determining the input parameters of the submerged arc furnace for hot smelting of nickel iron ore.

[0024] In the present invention, the input parameters of the nickel iron ore hot smelting submerged arc furnace are non-time series, mainly including electrode voltage U, furnace material diameter-to-height ratio ψ and electrode depth l. At the same time, the input parameters are also the input parameters of the deep table learning model established subsequently.

[0025] The input parameters of the present invention are set by the operator according to the operating specifications and experience. Figure 2 The structure schematic diagram of the submerged arc furnace for hot smelting of nickel iron ore is shown, and the key input parameters are explained. The submerged arc furnace for hot smelting of nickel iron ore involved in the present invention has a furnace lining 5, and the furnace is composed of a furnace gas layer 2, an electroslag layer 3 and a melt layer 4 from top to bottom, and an electrode 1 is inserted therein. The electrode voltage and electrode depth have an influence on the distribution of the electric field and the thermal field in the subsequent steps. The furnace charge diameter-to-height ratio is the ratio of the furnace charge area diameter d to the furnace charge height h, and is directly related to the shape of the furnace charge area.

[0026] The input parameters of the present invention are selected, and the correlation between different input parameters is dynamically mined through the multi-head attention mechanism in the subsequent step, and the weight relationship between the input parameter combinations is analyzed, which can further improve the accuracy of the prediction results.

[0027] Step 2), based on Maxwell's equations, establish an electric field model of the nickel iron ore hot smelting submerged arc furnace to obtain the current density of the nickel iron ore hot smelting submerged arc furnace expression.

[0028] During high-temperature smelting, the heat required for smelting comes from the arc heat at the electrode end and the Joule heat generated by the resistance when the current flows through the charge. That is, the electric field is the source of heat for smelting. Therefore, the electric field model of the nickel-iron ore hot smelting submerged arc furnace is first established to obtain the current distribution state in the furnace. The establishment of the electric field model is based on Maxwell's equations, which specifically include Ampere's circuit law, Faraday's law of electromagnetic induction, Gauss's law and the law of flux continuity. Equations (1)-(4) are shown as follows:

[0029]

[0030]

[0031]

[0032]

[0033] In the formula, H e Indicates the magnetic field strength, A·m -1 ; J represents current density, A·m -2 ; t represents the melting time, s; represents the electric displacement vector, C·m -2 ; e represents the charge, C; and Represents the electric field strength / V·m -1 and electromagnetic induction intensity / T.

[0034] Based on Maxwell's equations, the charge continuity equation can be obtained, as shown in equation (5);

[0035]

[0036] The charge in the furnace is electrically neutral, so the total charge in the furnace is considered to be 0, and the continuity equation of the current in the furnace can be expressed as formula (6);

[0037]

[0038] The potential distribution in the furnace is as shown in equations (7) and (8);

[0039]

[0040]

[0041] In the formula, σ e is the conductivity, S·m -1 ; is the electric potential, V;

[0042] The control equation of the electric potential in the furnace is equation (9);

[0043]

[0044] The control equation of the electric potential in the three-phase ore-fired furnace is (10);

[0045]

[0046] The current density is calculated as (11)-(13);

[0047]

[0048]

[0049]

[0050] In the formula, J x , J y , J z are the current densities of the x-axis, y-axis, and z-axis respectively.

[0051] Step 3), based on the energy conservation equation, momentum conservation equation, and mass conservation equation, a heat-mass transfer model of a nickel-iron ore hot smelting submerged arc furnace is constructed to obtain the temperature distribution T of the charge in the furnace and the expression of the smelting time parameter τ.

[0052] To obtain the temperature change and melting process of the charge in the furnace, we first need to establish the mass conservation equation as shown in equation (14);

[0053]

[0054] Where ρ is the density, kg·m -3 ; is the speed, m·s -1 ; t represents the smelting time, s, based on which the required smelting time parameter τ can be obtained as formula (15);

[0055]

[0056] In the formula, the melting time parameter τ is defined as the melting time t and the maximum melting time t max (In this embodiment, t max =12600s);

[0057] The momentum equation of the charge is equation (16);

[0058]

[0059] Where, p is pressure, Pa; μ is year, Pa·s; is the buoyancy, N; is the momentum balance term, N;

[0060] The energy conservation equation of the charge is equation (17), from which the temperature T required for step 3 can be obtained;

[0061]

[0062] Where λ is thermal conductivity, W·m -1 ·K -1 ;q A and q J are the arc heat source term and the Joule heat source term, which are obtained according to the current density and Gauss body heat source model respectively; H(T) represents the enthalpy equation;

[0063] The arc heat source term is expressed as formula (18);

[0064]

[0065] In the formula, r a is the arc radiation radius, m; r A is the maximum radiation radius of the arc on the charge surface, m; l is the depth of the arc immersed in the charge, that is, the required input parameter electrode depth, m; l A is the maximum depth of arc immersion into the charge, m; U is the electrode voltage, that is, the required input parameter electrode voltage, V;

[0066] The Joule heat source term is expressed as the current density and conductivity σ e The relationship is shown in formula (19);

[0067]

[0068] Step 4), establish the electrical-thermal-mass boundary conditions to solve the current density J in the furnace in step 2, and substitute the current density into step 3 to obtain the temperature T and the melting time parameter τ.

[0069] The electrical-thermal-mass boundary conditions for this step are:

[0070] In order to obtain the electric field distribution and temperature field distribution, the electromagnetic boundary conditions are first established. The current or voltage at the top of the three-phase electrode changes according to the trigonometric function curve. The current of a phase in the three-phase electrode is expressed as formula (20);

[0071] Ii =I max sin(2πωt+θ i ) (20)

[0072] Where, ω is the frequency of three-phase alternating current, which is 50Hz; θ i is the phase angle of the i-th phase electrode, and the phase angle difference between two adjacent points of the three-phase alternating current is 120°; I max is the maximum current.

[0073] The furnace shell and charge surface adopt the boundary condition of parallel electric line, then the surface potential is expressed as formula (21);

[0074]

[0075] In the formula, represents the surface unit normal vector;

[0076] The thermal-mass boundary conditions are further established, where the heat flow boundary condition is expressed as formula (22);

[0077] q h =hΔT=h(TT g ) (twenty two)

[0078] In the formula, q h Represents the heat flux density of convective heat transfer, W·m -2 ; h is the convective heat transfer coefficient, W·m -2 ·K -1 ; T g is the furnace gas temperature, K;

[0079] There is radiation heat transfer between the furnace gases above the furnace, and the heat flux density of radiation heat transfer is expressed as formula (23);

[0080]

[0081] In the formula, q r is the heat flux density of radiation heat transfer; ε r is the emissivity; σ r is the Stefan-Boltzmann constant;

[0082] The velocity boundary condition is to impose a no-slip boundary condition on the contact surface between the melt, the electrode and the furnace body. The velocity expression is formula (24);

[0083]

[0084] The melt surface velocity expression is (25);

[0085]

[0086] The interface between the electrode and the charge is the interface continuity boundary condition, as shown in equation (26);

[0087]

[0088] In the formula, λ elc is the electrode thermal conductivity, T elc is the electrode temperature, λ b is the thermal conductivity of the charge, T b is the charge temperature.

[0089] Step 5), based on step 1, define a deep table learning model, and preprocess the data obtained in step 4, and represent the input parameters of electrode voltage, charge diameter-to-height ratio, and electrode depth as one-dimensional vectors. Each value in the vector corresponds to a different input label and can be called a feature. Each feature is converted into a corresponding embedding representation.

[0090] In this step, the basic method of defining the deep table learning model and preprocessing the data is as follows:

[0091] Define and initialize the parameter parser for the submerged arc furnace for hot smelting of ferronickel ore;

[0092] Define the path for the sorted data, divide and preprocess it into a format that the model can handle;

[0093] The data obtained in steps 1)-4) are preprocessed to avoid the influence of singular values ​​that are too large or too small in the data on data modeling;

[0094] The original input parameters and output parameters are stored in a file in .excel format. In order to make the data easier to read, the xlrd tool in python is used to convert the data in the original file into Python list format data.

[0095] 80% of all data is used as training set, 4% of all data is used as validation set, and the rest is used as test set;

[0096] Set the dataset parameters: set the dataset input size to 5 and the output size to 4; to ensure both the efficiency and cost of the operation, set the number of threads that load data simultaneously to 4;

[0097] Define a fully connected neural network model and set the model's hyperparameters: the number of hidden neurons is 128, the number of iterations is 100, the number of epochs for early stopping training is 1, the random sampling batch size is 64, the learning rate is 0.0003, the random seed is 2000, and the number of repetitions of the random seed is 1 for parameter initialization and randomization;

[0098] Map the embedded representation of input data features into the logarithmic space, assign different weight coefficients to different features, and dynamically mine the association between different features according to the size of the weight coefficients; use the powerful adaptability, fault tolerance and self-organization ability of the fully connected neural network to construct the nonlinear relationship between input and output;

[0099] The above inputs are: electrode voltage, electrode depth, and charge diameter-to-height ratio;

[0100] The above output is: melting time parameter (ratio of melting time to maximum melting time 12600s);

[0101] In order to facilitate real-time monitoring of the model's operating status, a storage log is compiled into the program and its path is defined and set;

[0102] The input of this link can be expressed as a one-dimensional vector of length m. Each value in the vector corresponds to a different input label, and the input label has been clearly defined in advance. Before the one-dimensional vector is input into the deep table learning model, each feature needs to be converted into a corresponding embedding representation with a dimension of d, that is, Considering that each dimension of the input data in this link is numerical data, the above embedding representation process can be expressed as e i =x i e i Among them, e i is the embedding representation of the i-th dimension feature;

[0103] The correlation between the above inputs can be mined as follows: the predicted output results are not only obtained based on a single feature, but also calculated based on higher-order combinations of different features, such as the pairwise combination of different features. However, in each set of input and output data, it is not possible to predetermine the scale of each feature combination. For example, it is impossible to infer in advance whether the prediction result based on a single feature is more accurate than the prediction result based on the pairwise combination of features. Therefore, a dynamic and adaptable method is needed to mine the combination relationship between different features of each set of input data, and adaptively discover the input feature combination that is most critical to the output result. In order to achieve this goal, the following method is proposed: Figure 3 The neural network structure combined with the multi-head attention mechanism is shown. The present invention maps the embedded representation of the input data features into the logarithmic space, assigns different weight coefficients to different features, and dynamically mines the correlation between different features according to the size of the weight coefficient. The specific calculation process can be expressed as formula (27);

[0104]

[0105] Where ⊙ represents the Hadamard product, w ijIt is used to measure the correlation between the embedding representation of the j-th dimension feature and the embedding representation of the i-th dimension feature, y i The embedding of the i-th dimension feature represents the corresponding output result when mining the association relationship;

[0106] For each weight coefficient w in the above formula ij , the present invention uses a multi-head attention mechanism for automated mining. Specifically, for each feature of the input data, the present invention associates a weight vector with a dimension of m v i It is shared by different input data and can be learned in the actual model training process. At the same time, the present invention associates a query vector with a dimension of d for each feature. And the attention calibration weight is generated through the bidirectional attention flow mechanism. The specific calculation process can be expressed as formula (28) and (29);

[0107]

[0108]

[0109] Where W att is the bidirectional attention function The weight matrix of , dimension is d × d, and the weight matrix is ​​shared by different input features; Δ n represents a probability simplex of dimension d, which is calculated as When the parameter α≠1, the function It can be calculated as Otherwise, the function It can be calculated as

[0110] On this basis, the present invention can calculate the weight coefficient between each input feature and other input features to measure the correlation between different input features. The specific calculation process can be expressed as formula (30):

[0111]

[0112] Based on the calculation process of the above attention mechanism, the present invention constructs K groups of different attention weight parameters to mine the correlation between different input features from different perspectives. The attention weight matrix can be obtained as follows: At the same time, the query vector can be calculated as The output results corresponding to all input features are

[0113] Through the multi-head attention mechanism, we can obtain the intermediate analysis results of multiple perspectives, totaling K×m groups, and the dimension of each group is d. On this basis, the present invention concatenates all the prediction results in the first place, and further obtains the analysis result y, with the dimension of K×m×d;

[0114] Furthermore, the present invention uses a two-layer fully connected layer structure to define a multi-layer perceptron model, and constructs a nonlinear mapping relationship between the intermediate analysis results and the performance parameters. The model can be expressed as formula (31) and formula (32):

[0115] H=σ(yW (1) +b (1) ) (31)

[0116] O=HW (2) +b (2) (32)

[0117] In the formula, is the weight of the first fully connected layer structure, is the bias of the first fully connected layer structure, σ(·) is the activation function, is the weight of the second fully connected layer structure, is the bias of the second fully connected layer structure, are the performance parameters corresponding to c groups of different operating parameters, each group of performance parameters is a vector of length q;

[0118] Use the ReLU function as the activation function to perform piecewise nonlinear transformation of the first fully connected layer structure of the model. The ReLU function provides good prediction results for various deep learning models by simply retaining positive elements and discarding negative elements;

[0119] In order to improve the gradient descent speed of the model during training, the key operating parameters of the input model are normalized. At the same time, in order to improve the model's prediction accuracy for performance parameters, the performance parameters in the training data are normalized;

[0120] Step 6), load and train the deep table learning model.

[0121] To facilitate data loading, the Dateloader object in Pytorch is used to encapsulate the training set, validation set, and test set and convert them into tensor format;

[0122] Since the order of magnitude of the data varies greatly, in order to reduce the impact of overly large or undersized singular values ​​in the data on the accuracy of the model, each piece of data is first standardized, and the mean, standard deviation, random batch size, and number of threads of the data in each data set are loaded;

[0123] Use 80% of all the data obtained in step 4 as training data for model training and parameter optimization, 4% of the data for model verification, and the remaining data for model testing.

[0124] The hyperparameters of the trained model are further optimized as follows:

[0125] Grid Search is used to obtain the hyperparameters with good prediction results among all candidate hyperparameters. The number of neurons in the hidden layer is 128, the learning rate is 0.003, the number of samples input to the model each time is 64, and the parameter of the multi-head attention mechanism is K=6.

[0126] Step 7), deploy and run the deep table learning model.

[0127] Since it is necessary to calculate the gradient of the model loss function, the loss function used in this model is the L2 loss function, that is, the mean square error. This loss function represents the square of the difference between the predicted value and the true value. The specific expression is formula (33);

[0128]

[0129] A deep table learning model was constructed, and the mapping relationship between operating parameters and performance parameters was obtained through multiple rounds of iterative optimization. The activation function used in the model was the RELU function.

[0130] In terms of selecting the optimization algorithm, this model uses the Adam optimization algorithm to calculate the adaptive learning rate of each parameter. This algorithm combines the Momentum and RMSprop gradient descent methods. The Adam optimization algorithm uses the exponentially weighted moving average in the process of estimating the momentum and second moment of the gradient, so that the gradient and learning rate can be dynamically adjusted at the same time, which can achieve faster parameter convergence speed and more effective training effect, and can effectively solve the problems of learning rate disappearance and slow convergence existing in existing optimizers;

[0131] Train and run the model on the data set, and verify and test the updated model after training. In the process, the optimal error is continuously updated. If the patience value is reached in advance, the model will stop running in advance to save computing costs.

[0132] Record the running time and loss, output the calculated predicted value and true value, select the batch, calculate each batch to get the loss function, do a back propagation based on the loss function, and use the optimizer to optimize the whole process;

[0133] All batched data are concatenated, all predicted data are restored to the order of magnitude of the original data, and the mean absolute percentage error MAPE of the entire model is obtained, which is specifically expressed as formula (34), and the model is evaluated in turn;

[0134]

[0135] In the formula, n is the data sample number, N is the number of data samples, True value, is the predicted value;

[0136] After inputting any parameters of the nickel iron ore hot smelting submerged arc furnace into the trained model, the performance parameters such as the smelting time parameters of the nickel iron ore hot smelting submerged arc furnace can be predicted. Figure 4 The figure shows an implementation case based on the algorithm proposed in the present invention, from which it can be seen that the prediction accuracy between the predicted value and the actual value of the smelting time of the submerged arc furnace for hot smelting of nickel iron ore reaches 96%.

[0137] In summary, the present invention couples the multi-head attention mechanism to explore the deep mapping relationship between input (electrode voltage, diameter-to-height ratio, electrode depth) and output parameters (smelting time parameters). The performance prediction accuracy of the submerged arc furnace for hot smelting of nickel iron ore is achieved to 96%, and the operating performance evaluation cycle is shortened to seconds to minutes. The relationship between input parameters and output parameters is explored. Compared with electrode depth and diameter-to-height ratio, electrode voltage has the greatest impact on the operating performance of the submerged arc furnace. The present invention can effectively reduce the operating performance evaluation cost of the submerged arc furnace for hot smelting of nickel iron ore.

Claims

1. A method for quickly predicting the performance of a submerged arc furnace for hot smelting of nickel iron ore in the steel process industry, characterized in that: The steps include: Step 1, determining the input parameters of the nickel iron ore hot smelting submerged arc furnace, wherein the input parameters are non-time series data, including electrode voltage U, charge diameter-to-height ratio ψ and electrode depth l; Step 2: Establish an electric field model to obtain the current density in the submerged arc furnace for hot smelting of nickel iron ore expression; Step 3, constructing a heat-mass transfer model of a nickel-iron ore hot smelting submerged arc furnace based on the conservation of energy, momentum and mass, so as to obtain the expression of the temperature distribution T of the charge in the furnace and the smelting time parameter τ; Step 4: Establish the electric-thermal-mass boundary conditions and solve the current density in the furnace Substitute into step 3 to obtain the temperature distribution T and smelting time parameter τ; Step 5, defining a deep table learning model based on the input parameters, and preprocessing the data obtained in steps 1 to 4, representing the input parameters as a one-dimensional vector, each numerical value in the vector corresponds to a different input label, and converting each numerical feature into a corresponding embedding representation; Step 6, using the data obtained in step 4 to train the deep table learning model, and using the trained model to quickly predict the performance of the submerged arc furnace for hot smelting of nickel iron ore; the output of the deep table learning model is the smelting time parameter τ; The step 5 defines a deep table learning model based on the input parameters, and preprocesses the data obtained in steps 1 to 4, as follows: Define and initialize the parameter parser for the submerged arc furnace for hot smelting of ferronickel ore; Set the path for the sorted data, divide and preprocess it to convert it into a format that the model can handle; store the original input parameters and output parameters in a file in .excel format, and use the xlrd tool in python to convert the data in the original file into Python list format data; Set the dataset parameters; Define a fully connected neural network model and set the model's hyperparameters; The deep table learning model adopts a neural network structure combined with a multi-head attention mechanism to map the embedded representation of the input parameter value into the logarithmic space, and assigns different weight coefficients to different numerical features. The correlation between different numerical features is dynamically mined according to the size of the weight coefficient. The calculation process is expressed as follows: Where ⊙ represents the Hadamard product, w ij It is used to measure the correlation between the embedding representation of the j-th dimension feature and the embedding representation of the i-th dimension feature, y i The embedding of the i-th dimension feature represents the corresponding output result when mining the association relationship; For each weight coefficient w in the formula ij , using the multi-head attention mechanism for automated mining, for each feature of the input data, it is associated with a weight vector of dimension m v i Shared by different input data, and a query vector of dimension d is associated with each feature And generate attention calibration weights through a bidirectional attention flow mechanism.

2. The rapid prediction method for the performance of a submerged arc furnace for hot smelting of nickel iron ore in the steel process industry according to claim 1, characterized in that: The step 2 is to establish an electric field model based on Maxwell's equations and obtain the current density in the furnace of the submerged arc furnace for hot smelting of nickel iron ore. The expression is as follows: Among them, σ e is the conductivity, is the electric potential, is the electric field strength, is the electromagnetic induction intensity, t represents the melting time; The nickel iron ore hot smelting submerged arc furnace is a three-phase submerged arc furnace, so the current density in the furnace is It is expressed as: Among them, J x , J y , J z are the current densities of the x-axis, y-axis, and z-axis respectively.

3. The rapid prediction method for the performance of a submerged arc furnace for hot smelting of nickel iron ore in the steel process industry according to claim 2, characterized in that: In step 3, a heat-mass transfer model of a nickel iron ore hot smelting submerged arc furnace is constructed to obtain the temperature distribution T of the charge in the furnace and the smelting time parameter τ as follows: The melting time parameter τ is defined as the melting time t and the maximum melting time t max The ratio of , H(T) represents the enthalpy equation, is the velocity, λ is the thermal conductivity; q A is the arc heat source term, expressed as: q J is the Joule heat source term, expressed as: Among them, r a is the arc radiation radius, r A is the maximum radiation radius of the arc on the charge surface, l is the depth of the arc immersed in the charge, that is, the electrode depth, l A is the maximum depth of the arc immersed in the charge, and U is the electrode voltage.

4. The rapid prediction method for the performance of a submerged arc furnace for hot smelting of nickel iron ore in the steel process industry according to claim 3, characterized in that: In step 4, the electrical-thermal-mass boundary conditions are established as follows: First, the electromagnetic boundary conditions are established. The expression of a phase current in the three-phase electrode is as follows: I i =I max sin(2πωt+θ i ) In the formula, ω is the frequency of three-phase alternating current change, θ i is the phase angle of the i-th phase electrode, and the phase angle difference between two adjacent points of the three-phase alternating current is 120°; I max is the maximum current; The furnace shell and charge surface adopt the boundary condition of parallel electric line, and the surface potential expression is as follows: In the formula, represents the surface unit normal vector; Secondly, establish the thermal-mass boundary conditions, where the heat flow boundary condition expression is as follows: q h =hΔT=h(T-T g ) In the formula, q h represents the heat flux density of convective heat transfer, h is the convective heat transfer coefficient, T g is the furnace gas temperature; There is radiation heat exchange between the furnace gas in the air above the furnace. The heat flux density expression of radiation heat exchange is as follows: In the formula, q r is the heat flux density of radiation heat transfer, ε r is the emissivity, σ r is the Stefan-Boltzmann constant; The velocity boundary condition is to impose a no-slip boundary condition on the contact surface between the melt, the electrode and the furnace body. The velocity expression is as follows: The melt surface velocity expression is as follows: Finally, the interface between the electrode and the charge is taken as the interface continuity boundary condition, and the expression is as follows: Among them, λ elc is the electrode thermal conductivity, T elc is the electrode temperature, λ b is the thermal conductivity of the charge, T b is the charge temperature.

5. The rapid prediction method for the performance of a submerged arc furnace for hot smelting of nickel iron ore in the steel process industry according to claim 1, characterized in that: The data set parameters are: set the data set input size to 5, the output size to 4; set the number of threads for loading data simultaneously to 4; The hyperparameters of the model are set as follows: number of hidden neurons is 128, number of iterations is 100, number of epochs for early stopping training is 1, random sampling batch size is 64, learning rate is 0.0003, random seed is 2000, and number of repetitions of random seed is 1 for parameter initialization and randomization; Program storage logs into the program and define their paths to monitor the running status of the model in real time.

6. The rapid prediction method for the performance of a submerged arc furnace for hot smelting of nickel iron ore in the steel process industry according to claim 1, characterized in that: In step 6, the Dateloader object in Pytorch is used to encapsulate the data set and convert it into a tensor format; each piece of data is first standardized, and the mean value, standard deviation, random batch size, and number of threads of the data in each data set are loaded.

7. The rapid prediction method for the performance of a submerged arc furnace for hot smelting of nickel iron ore in the steel process industry according to claim 1, characterized in that: Step 6, optimizing the hyperparameters of the trained model: Use grid search to obtain good hyperparameters for prediction results among all candidate hyperparameters. The number of hidden layer neurons is 128, the learning rate is 0.003, the number of samples input to the model each time is 64, and the parameter of the multi-head attention mechanism is 6.

8. The rapid prediction method for the performance of a submerged arc furnace for hot smelting of nickel iron ore in the steel process industry according to claim 1, 6 or 7, characterized in that: In step 6, the trained model is deployed and the prediction results are evaluated as follows: All batched data are concatenated, all predicted data are restored to the order of magnitude of the original data, and the mean absolute percentage error (MAPE) of the entire model is obtained to evaluate the model; the expression of MAPE is as follows: In the formula, n is the data sample number, N is the number of data samples, True value, is the predicted value; Substituting any nickel-iron ore hot smelting submerged arc furnace operating parameters into the trained model, real-time prediction of the furnace operating conditions is performed.

Images