A method and system for predicting key indicators of hydrometallurgical zinc smelting and cobalt removal process

Through the graph attention network, wet zinc smelting and cobalt removal process is constructed, which solves the problem of low online detection accuracy of cobalt ion concentration in the existing technology, and realizes accurate prediction of the wet zinc smelting and cobalt removal process, improving the Control effect.

CN114841059BActive Publication Date: 2024-11-26CENT SOUTH UNIV
View PDF 3 Cites 0 Cited by

Patent Information

Application Number
CN202210415449.9
Authority / Receiving Office
CN · China
Patent Type
Patents(China)
Current Assignee / Owner
Filing Date
2022-04-20
Publication Date
2024-11-26
Estimated Expiration
2042-04-20

AI Technical Summary

Technical Problem

The online detection accuracy of cobalt ion concentration during the existing wet zinc smelting process is low, resulting in poor control effect. The model fitting tracking trend is good but the accuracy is low, which cannot meet the requirements of industrial applications.

Method used

The graph attention network is used to construct a key index prediction model for the wet zinc smelting and cobalt removal process. By obtaining data from multiple process variables, including sensor acquisition and offline assay data, the process variable graph topology diagram is constructed, and data characteristics and relationship characteristics are extracted. Use a fully connected prediction layer to output the cobalt ion concentration prediction results to improve prediction accuracy.

Benefits of technology

An accurate prediction model for export cobalt ion concentration key indicators has been established, which improves prediction accuracy and can predict online in real time. It is suitable for complex nonlinear and interdependent industrial process variables, and optimizes control.

✦ Generated by Eureka AI based on patent content.

Smart Images

  • Figure CN114841059B_ABST
    Figure CN114841059B_ABST
Patent Text Reader

Abstract

The present application discloses a method and system for predicting key indicators of the cobalt removal process in hydrometallurgy, the method comprising: obtaining data of multiple process variables of a reactor in the hydrometallurgy process; inputting the data of the multiple process variables into a graph attention network to train and determine the parameters of the graph attention network to construct data features and relationship features of the multiple process variables; training and determining the parameters of the fully connected prediction layer of the graph attention network according to the data features and relationship features of the multiple process variables obtained after training the parameters of the graph attention network. The present application solves the problem of low prediction accuracy in the technical solutions for predicting key indicators of the cobalt removal process in hydrometallurgy in the related art, thereby establishing an accurate prediction model for the key indicators of the outlet cobalt ion concentration in view of the complex nonlinear and interdependent relationships between the variables of the cobalt removal process, thereby improving the prediction accuracy.
Need to check novelty before this filing date? Find Prior Art

Description

Technical Field

[0001] The present application relates to the field of cobalt removal during hydrometallurgy of zinc smelting, and in particular to a method and system for predicting key indicators of cobalt removal during hydrometallurgy of zinc smelting. Background Art

[0002] Zinc is an important non-ferrous metal raw material with a wide range of uses. At present, zinc production methods are mainly pyrometallurgical zinc smelting and hydrometallurgical zinc smelting, among which hydrometallurgical zinc smelting is the most important zinc smelting method in the world. The zinc output obtained by this technology accounts for 80% to 85% of the total output, and it is showing a growing trend with the development of technology. Among them, the cobalt removal process is an important step in hydrometallurgical zinc smelting. By adding zinc powder and catalysts, cobalt ions are gradually removed. The reasonable operation of the cobalt removal process requires online detection of the key indicator of cobalt ion concentration to reflect the current state of the cobalt removal process. However, the current cobalt ion concentration online detector is not only expensive to purchase, but also easy to cause rust and failure of instrument components due to the influence of complex and changeable environment, such as acid mist and high temperature, and the measured medium has strong corrosion, easy crystallization, dirt and other problems, which ultimately affects the reliability and stability of the instrument. Therefore, most companies often obtain the key indicator of export cobalt ion concentration through chemical analysis through offline experiments, which leads to untimely feedback of key indicator information and high time lag, which ultimately leads to a certain degree of blindness in on-site manual adjustment, so its control effect cannot be guaranteed. Therefore, it is very important to predict the key indicators of the cobalt removal process in hydrometallurgy so as to achieve the optimal guidance of the entire process.

[0003] In order to solve the above problems, two solutions are provided in the related art. The two solutions are respectively described below.

[0004] In solution one, a mechanism kinetic model of cobalt-hydrogen electrochemical competition for the cobalt removal process in hydrometallurgy is proposed. First, based on the characteristics of the cobalt removal process and combined with the basic kinetic model, the researchers established a mechanism kinetic model consisting of three models for the cobalt removal process, of which the three sub-models are divided into: electrochemical rate model, mixed potential model of zinc powder, cobalt ions and hydrogen ions, and hydrolysis rate model of zinc ions. These models are all constructed based on the principles of microbatteries and electrochemical kinetics. Secondly, the researchers designed a constrained estimation scheme for the parameters of the established combined model, transforming the estimation problem of a given parameter with a difficult-to-estimate value range into an estimation problem of an experimental metric whose value range can be estimated more easily from available experimental data. Then, the parameters of the model were estimated using actual data. Finally, the model simulated the changes in the outlet cobalt ion concentration, and the fitting index MSE was 7.142%.

[0005] In solution 2, an effective comprehensive prediction model for cobalt ion concentration is provided. The researchers first established a kinetic model based on mechanism research. Then, in order to overcome the problem that the mechanism kinetic model has good tracking trend but low accuracy, data-driven compensation is used to introduce the ORP (oxidation-reduction potential) that can be detected online into the kinetic model. Finally, the key indicator of the cobalt removal process in hydrometallurgy is predicted for the outlet cobalt ion, and the average relative error of the best prediction indicator is 6.34%.

[0006] There are certain problems in both solutions in the above-mentioned related technologies: the prediction method given in Solution 1 has certain limitations. Because the cobalt removal process is carried out in a large-capacity continuous reactor, the reaction time is long, the reaction mechanism is complex, and there are a large number of nonlinear parameters in the system. In addition to complex environmental factors and constantly changing boundary conditions, it is also affected by many factors. The simplified model cannot reflect the actual industrial process, and its low accuracy cannot meet the requirements of field applications. Although the method given in Solution 2 compensates for the low accuracy of the mechanism kinetic model based on data-driven, it only relies on the flat cascade input of a certain variable into the model, and does not consider multiple variables and the relationship between variables. The model is not capable of mining implicit information and is only suitable for small sample data. It cannot achieve high prediction accuracy for today's large sample industrial data. Summary of the invention

[0007] The embodiments of the present application provide a method and system for predicting key indicators of the cobalt removal process in hydrometallurgy, so as to at least solve the problem of low prediction accuracy in the technical solutions for predicting key indicators of the cobalt removal process in hydrometallurgy in related technologies.

[0008] According to one aspect of the present application, a method for predicting key indicators of a cobalt removal process in a hydrometallurgical zinc smelting process is provided, comprising: obtaining data of multiple process variables of a reactor in a hydrometallurgical zinc smelting process, wherein the multiple process variables include multiple first process variables collected by sensors and multiple second process variables obtained by offline testing, the multiple process variables are associated with the cobalt removal process, and the multiple second process variables include key indicators, which are outlet cobalt ion concentrations; inputting the data of the multiple process variables into a graph attention network to train and determine the parameters of the graph attention network to construct data features and relationship features of the multiple process variables, wherein the data features are the variable nodes in a process variable graph topology diagram constructed by the multiple process variables, and each variable node corresponds to a process variable; the relationship features are the association relationships between the variable nodes in the graph topology diagram; the graph attention network is used to extract the corresponding relationships between the key indicators and the multiple process variables; and training and determining the parameters of a fully connected prediction layer of the graph attention network according to the data features and relationship features of the multiple process variables obtained after training the parameters of the graph attention network, wherein the trained fully connected prediction layer is used to output a prediction result of cobalt ion concentration.

[0009] According to another aspect of the present application, a key indicator prediction system for the cobalt removal process of hydrometallurgy is also provided, comprising: an acquisition module, used to acquire data of multiple process variables of a reactor in the hydrometallurgy process, wherein the multiple process variables include multiple first process variables collected by sensors and multiple second process variables obtained by offline analysis, the multiple process variables are associated with the cobalt removal process, the multiple second process variables include key indicators, and the key indicator is the outlet cobalt ion concentration; a first training and determination module, used to input the data of the multiple process variables into a graph attention network to train and determine the parameters of the graph attention network to construct data of the multiple process variables Features and relationship features, wherein the data features are the various variable nodes in the process variable graph topology diagram constructed by the multiple process variables, and each variable node corresponds to a process variable; the relationship features are the association relationships between the various variable nodes in the graph topology diagram; the graph attention network is used to extract the corresponding relationship between the key indicators and the multiple process variables; a second training determination module is used to train and determine the parameters of the fully connected prediction layer of the graph attention network according to the data features and relationship features of the multiple process variables obtained after training the parameters of the graph attention network, wherein the trained fully connected prediction layer is used to output the cobalt ion concentration prediction result.

[0010] In an embodiment of the present application, a method is adopted to obtain data of multiple process variables of a reactor in a hydrometallurgical zinc smelting process, wherein the multiple process variables include multiple first process variables collected by sensors and multiple second process variables obtained by offline analysis, the multiple process variables are associated with a cobalt removal process, and the multiple second process variables include key indicators, wherein the key indicator is the outlet cobalt ion concentration; the data of the multiple process variables are input into a graph attention network to train and determine the parameters of the graph attention network to construct data features and relationship features of the multiple process variables, wherein the data features are the various variable nodes in a process variable graph topology diagram constructed by the multiple process variables, and each variable node corresponds to a process variable; the relationship features are the association relationships between the various variable nodes in the graph topology diagram; the graph attention network is used to extract the corresponding relationship between the key indicators and the multiple process variables; the parameters of the fully connected prediction layer of the graph attention network are trained and determined according to the data features and relationship features of the multiple process variables obtained after training the parameters of the graph attention network, wherein the trained fully connected prediction layer is used to output the cobalt ion concentration prediction result. The present application solves the problem of low prediction accuracy in the technical solutions for predicting key indicators of the cobalt removal process in hydrometallurgy of zinc in the related art, thereby establishing an accurate prediction model for the key indicators of the outlet cobalt ion concentration in view of the complex nonlinear and interdependent relationships between the variables in the cobalt removal process, thereby improving the prediction accuracy. BRIEF DESCRIPTION OF THE DRAWINGS

[0011] The drawings constituting a part of the present application are used to provide a further understanding of the present application. The illustrative embodiments and descriptions of the present application are used to explain the present application and do not constitute an improper limitation on the present application. In the drawings:

[0012] Figure 1 is a schematic diagram of a structure applied to a cobalt removal process according to an embodiment of the present application;

[0013] Figure 2 It is a flow chart of a method for constructing a key indicator prediction model for a hydrometallurgical zinc purification and cobalt removal process according to an embodiment of the present application;

[0014] Figure 3 It is a flowchart for constructing a graphical topology diagram according to an embodiment of the present application;

[0015] Figure 4 is a graph topology diagram according to an embodiment of the present application;

[0016] Figure 5 is a schematic diagram of a graph attention network calculation process according to an embodiment of the present application;

[0017] Figure 6This is a schematic diagram of the online prediction effect of key indicators based on the graph attention network according to an embodiment of the present application;

[0018] Figure 7 is a process variable association diagram according to an embodiment of the present application;

[0019] Figure 8 It is a flow chart of a method for predicting key indicators of a hydrometallurgical zinc smelting and cobalt removal process according to an embodiment of the present application. DETAILED DESCRIPTION

[0020] It should be noted that, in the absence of conflict, the embodiments and features in the embodiments of the present application can be combined with each other. The present application will be described in detail below with reference to the accompanying drawings and in combination with the embodiments.

[0021] It should be noted that the steps shown in the flowcharts of the accompanying drawings can be executed in a computer system such as a set of computer executable instructions, and that, although a logical order is shown in the flowcharts, in some cases, the steps shown or described can be executed in an order different from that shown here.

[0022] Figure 1 is a schematic diagram of a structure applied to a cobalt removal process according to an embodiment of the present application, Figure 1 The cobalt removal process flow of the smelter is shown. Figure 1 It can be seen that the cobalt removal process is mainly composed of four continuous stirred tank reactors and a thickener. The continuous stirred tank reactors are arranged in a stepped manner. The inlet of each continuous stirred tank reactor is located at the top, with a height difference of about 1 meter from the outlet. The inlet and outlet of the continuous stirred tank reactor are connected by a chute. The whole process is to add zinc powder and arsenic salt to the No. 1 cobalt removal reactor under high temperature and acidic conditions, and a complex redox reaction occurs in the reactor with the cobalt ions in the copper removal liquid and the residual copper ions. After cobalt removal in four continuous reactors, the zinc sulfate solution flows into the thickener, where liquid-solid separation is carried out; the overflow of the thickener is transported to the subsequent cadmium removal process, and the bottom flow containing the reaction product is circulated to the first reactor.

[0023] During the cobalt removal process, the temperature in the reactor is controlled at 80±2°C, and the total flow rate of the copper removal solution is usually 200-300m 3 / h, the underflow rate is usually fixed at 80m 3 / h, the concentrator diameter is 25m, and the volume of each reactor is 460m 3 The size of zinc powder added to each reactor is 80-200 mesh, and arsenic salt is only added to reactor 1#, and its dosage is calculated according to the mass balance formula ( where n Cu , n Co , and n NiThe total molar amount of copper ions, cobalt ions, and nickel ions entering the reactor per unit time, n As represents the molar amount of arsenic required per unit time, α Cu The ratio of copper ions participating in the reaction to all copper ions. The waste acid flow rate added to the 1# reactor is usually 0.4m 3 / h, the concentrations of copper ions and cobalt ions in the liquid after copper removal are usually 150-500 mg / L and 15-35 mg / L, the cobalt ion concentration at the outlet of reactor 1 is tested every 2 hours, and the copper ion concentration at the outlet is tested every 4 hours. The cobalt ion concentration of the thickener overflow is tested every hour; each reactor is equipped with an online redox potential detector, and the specific gravity of the underflow is tested online using a hydrometer.

[0024] The key technical indicator of the arsenic salt cobalt removal process is the cobalt ion concentration at the outlet of the cobalt removal process. The technical indicator of the cobalt ion concentration at the outlet of the cobalt removal process is about 0.5 mg / L, which reflects the purification effect of the arsenic salt cobalt removal process on the impurity cobalt. Considering that most of the cobalt ions in the liquid after copper removal are removed in the 1# reactor, a small amount is precipitated in the 2# reactor, and the 3# and 4# reactors play a fine-tuning role, and the reactions of the 2#, 3#, and 4# reactors are relatively stable, therefore, this embodiment mainly constructs a model for the 1# reactor to achieve the prediction of the key indicator of the cobalt ion concentration at the outlet of the 1# reactor. Similarly, the established method can be extended to other reactors.

[0025] In this embodiment, a method for predicting key indicators of the cobalt removal process of hydrometallurgy of zinc is provided. Figure 8 : is a flow chart of a method for predicting key indicators of a hydrometallurgical zinc smelting and cobalt removal process according to an embodiment of the present application, such as Figure 8 As shown, the process includes the following steps:

[0026] Step S802, obtaining data of multiple process variables of the reactor in the hydrometallurgical zinc smelting process, wherein the multiple process variables include multiple first process variables collected by sensors (for example, total flow of liquid after copper removal, waste acid, arsenic salt, zinc powder, redox potential) and multiple second process variables obtained by offline analysis (for example, inlet copper ion concentration, inlet cobalt ion concentration, outlet cobalt ion concentration data), the multiple process variables are associated with the cobalt removal process, the multiple second process variables include key indicators, and the key indicator is the outlet cobalt ion concentration. The multiple process variables measured in step S802 are used as training data to train the graph attention network.

[0027] Step S804, inputting the data of the multiple process variables into the graph attention network to train and determine the parameters of the graph attention network to construct the data features and relationship features of the multiple process variables, wherein the data features are the various variable nodes in the process variable graph topology diagram constructed by the multiple process variables, and each variable node corresponds to a process variable; the relationship features are the association relationships between the various variable nodes in the graph topology diagram; the graph attention network is used to extract the corresponding relationships between the key indicators and the multiple process variables.

[0028] In this step, a better processing method is to normalize the data of the multiple process variables to obtain normalized data; input the normalized data of the multiple process variables into the graph attention network to train and determine the parameters of the graph attention network.

[0029] Step S806, training and determining the parameters of the fully connected prediction layer of the graph attention network according to the data features and relationship features of the multiple process variables obtained after training the parameters of the graph attention network, wherein the trained fully connected prediction layer is used to output the cobalt ion concentration prediction result.

[0030] This embodiment aims at the problem that the prediction effect based on the mechanism kinetic model technology is poor and the relationship between industrial structure variables is ignored based on the data-driven technology. A method for predicting the key indicators of the hydrometallurgical zinc removal process based on the graph attention network is proposed. The industrial structure process variables can be deeply excavated to achieve accurate real-time online prediction of the key indicators of the outlet cobalt ion concentration, and establish a foundation for subsequent optimization control. Through the above steps, an available graph attention network can be trained. In the future, multiple process variables obtained in the actual production process can be input into the graph attention network, and the graph attention network can output the result of the cobalt ion concentration prediction. The graph attention network establishes the connection between the key indicator (export cobalt ion concentration) process variable and other process variables, so that other process variables in the production process (for example, the total flow rate of the liquid after copper removal, waste acid, arsenic salt, zinc powder, redox potential, inlet copper ion concentration, inlet cobalt ion concentration) are input into the graph attention network, and the output cobalt ion concentration prediction result can be obtained.

[0031] The present embodiment maps the relationship between each process variable feature and the variables of the 1# reactor into graph information based on the graph structure data. Subsequently, the time series of the corresponding sensor data of each variable node (graph node) is normalized. Then, the graph information and the normalized node data are input to the graph attention network for processing. The graph attention network can extract the complex nonlinearity and mutual dependence between the key indicators and the process variables, and a multi-head attention scheme is designed. The network of multi-head attention can extract the feature correlation of different dimensions. Finally, the feature information extracted by the graph attention network is input to the prediction output layer based on the full connection, and the prediction output layer outputs the cobalt ion concentration prediction result. The steps involved in this implementation are described below.

[0032] Step 1:

[0033] Figure 2 is a flowchart of a method for constructing a key indicator prediction model for a hydrometallurgical zinc purification and cobalt removal process according to an embodiment of the present application. Figure 2 In the process, we first determine the number of variables and their relationships, then construct the graph data, build the graph attention network based on the graph data, and input the feature information extracted by the graph attention network into the prediction output layer based on full connection, which outputs the prediction result of cobalt ion concentration. Figure 1 The given 1# process variable, considering the temperature, the underflow is a fixed value, which has little effect on it, so it is ignored. The total flow rate of the liquid after copper removal, the inlet copper ion concentration, the inlet cobalt ion concentration, the waste acid, the arsenic salt, the zinc powder, the redox potential and the outlet cobalt ion concentration key indicator variables are extracted to form 8 variables. Figure 3 It is a flowchart of constructing a graphical topology structure diagram according to an embodiment of the present application. Figure 3 The figure shows the process of constructing a process variable graph topology diagram based on graph structure data. In this process, the cobalt removal process variable data is normalized, and then the unit variable features are constructed, and the associated edges between the variables are constructed, and finally the graph topology structure is constructed. Figure 4 As shown, Figure 4 Each circle node in the figure represents the data feature of the process variable, and the line between two variable nodes represents the correlation between the two process variable features. Then, a multi-head attention mechanism graph attention network is constructed to extract data features and relationship features, and finally, it is input into the prediction layer constructed by the full connection, and the prediction layer obtains the final key indicator prediction results.

[0034] Step 2:

[0035] The total flow rate of the copper-removed liquid, waste acid, arsenic salt, zinc powder, and redox potential data are collected through real-time high-rate sensors, and the inlet copper ion concentration, inlet cobalt ion concentration, and outlet cobalt ion concentration data are collected through offline testing. Then, in order to ensure the consistency of dimensions, the data is normalized. Let x be the original data of the process variable data, then the normalized data is Then, the normalized data set is divided into training set and test set in a ratio of 8:2, where x min is the minimum value in the original data, x max is the maximum value in the original data:

[0036]

[0037] Step 3:

[0038] The dataset is used to train the key indicator prediction model of the cobalt removal process in hydrometallurgy of zinc smelting based on the graph attention network. This part has two purposes. The first is to train and determine the parameters of the graph attention network to extract the variable characteristics and relationship characteristics of the cobalt removal process. The second is to train and determine the parameters of the prediction layer so that the prediction layer can process the features from the graph attention network and obtain the final output results.

[0039] Figure 5 It is a schematic diagram of the graph attention network calculation process according to the embodiment of the present application. It should be noted that in the figure, calculations 1 and 3 are performed on the update of node 2. Although the accompanying drawings show the core part of the present embodiment for the node part calculation, it should be understood that the present invention can be implemented in various forms and should not be limited by the examples described here. On the contrary, these examples are provided to enable a more thorough understanding of the present invention and to fully convey the scope of the present invention to those skilled in the art.

[0040] like Figure 5 As shown, the attention network processes the data feature vectors h1 and h3 of the neighboring variable nodes 1 and 3 of the process variable node 2 data feature vector h2 through the attention mechanism. The attention mechanism assigns the corresponding attention coefficients to different neighboring nodes, and obtains the 1-to-2 attention coefficient α 12 , 3 to 2 attention coefficient α 32 Then, the attention coefficients of the neighboring variable nodes of the 2 nodes are weighted and summed to obtain the new feature vector h2 of the 2 nodes. This process fully considers the structural association relationship between variables, integrates the correlation of each node in the graph into the model, improves the feature extraction capability, and especially solves the defect that the traditional model cannot extract association relationships.

[0041] Then the calculation of the variable nodes in the graph is generalized, and the training of the parameters of the attention network is divided into two parts: calculating the attention coefficient and weighted summation.

[0042] The specific process of calculating the attention coefficient is as follows:

[0043] For vertex i, calculate the correlation coefficient e between itself and its neighboring nodes ij , which can be expressed as follows:

[0044] e ij =a([Wh i ‖Wh j ])

[0045] W represents the weight of feature transformation;

[0046] h i Represents the characteristic vector of variable node i, that is, the specific value of process variable i;

[0047] h j Represents the characteristic vector of variable node j, that is, the specific value of process variable j;

[0048] [·||·] represents the concatenation of the transformed features of variable nodes i and j;

[0049] a is the learnable attention coefficient weight vector.

[0050] In order to make the correlation between the variable vertex i node and the neighbor variable node j easier to calculate and compare, the attention coefficients calculated for all neighboring variable nodes are normalized by softmax to obtain the final attention coefficient a ij The calculation process is as follows:

[0051]

[0052] k∈N i Indicates that variable node k is any neighboring node of variable node i,

[0053] LeakyReLU is the activation function.

[0054] Through the above process, the attention coefficient has been calculated, and in the next step, a weighted sum is performed to obtain a new feature vector representation. The weighted summation method is used to calculate the linear combination of the output feature vectors corresponding to the attention coefficient, and the final result is used as the updated feature information h' of the vertex variable node i , the calculation formula is as follows:

[0055]

[0056] In order to enable the model to extract more information, a multi-head attention mechanism is used to calculate K groups of attention coefficients, and a splicing operation is used to integrate multiple attention heads. The calculation formula is as follows:

[0057]

[0058] || represents splicing

[0059] K represents the number of attention heads;

[0060] represents the attention coefficient of k groups;

[0061] W k Represents the weight of the feature transformation of the k-th group of attention mechanism;

[0062] σ is the activation function ELU.

[0063] Therefore, the feature representation of the process variable node i after the multi-head attention network is obtained as follows: That is, GAT outputs new features for each process variable node i by fusing domain information.

[0064] The second purpose is that the updated process variable features pass through the fully connected layer to obtain the final output result. The specific calculation process is as follows:

[0065] Y=wH`+b

[0066] Y represents the predicted value;

[0067] H` is the feature information matrix of the process variables after the graph attention network updates;

[0068] w is the learnable weight;

[0069] b is the bias.

[0070] The outlet cobalt ion concentration value predicted by the model is compared with the actual data collected in the training sample data, and the parameters in the graph attention network are adjusted by error back propagation. The loss function used in this embodiment is the mean square error MSE loss function, which is expressed as:

[0071]

[0072] Y represents the predicted value, y represents the actual value, and n represents the total number of samples.

[0073] The mean square error loss function is used to evaluate the degree of deviation between the model's predicted value and the true value. The network parameters are continuously adjusted through the error back propagation algorithm. The smaller the deviation, the smaller the loss value, the higher the model prediction accuracy, which proves that the model prediction effect is better.

[0074] Through multiple iterative training, when the deviation is small, the learnable attention coefficient weight vector a in the key indicator prediction model of hydrometallurgical zinc removal process based on graph attention network, the weight W of the feature transformation of k groups of attention mechanism k, k groups of attention coefficients The parameters of the fully connected layer w and bias b are continuously optimized and adjusted to reach the optimal value.

[0075] Step 4:

[0076] 20% of the data (excluding the training set) collected on-site from the hydrometallurgical zinc smelting cobalt removal process was used as the test set to test the trained graph attention network model. The specific process is as follows:

[0077] The characteristic process variables of each group in the test set and their relationships are converted into graph data and input into the trained model. The graph attention network is used to extract the relationships and features, and new feature information containing the process variable's own characteristics and domain characteristics is obtained. The new feature information is input into the prediction layer based on full connection, and the predicted value of the outlet cobalt ion concentration is finally output. The deviation degree between the predicted value and the actual value (including RMSE, MAPE and R 2 ), as the evaluation criterion for the test set.

[0078] Through this embodiment, an accurate prediction model for the key indicator of outlet cobalt ion concentration is established in view of the complex nonlinear and interdependent relationships between the variables in the cobalt removal process.

[0079] In this embodiment, the cobalt removal process is first analyzed, and eight important process variables are extracted, namely, the total flow rate of the liquid after copper removal, the inlet copper ion concentration, the inlet cobalt ion concentration, the waste acid, the arsenic salt, the zinc powder, the redox potential, and the outlet cobalt ion concentration key indicator variables. Subsequently, a process variable graph topology structure is constructed for the above eight variable nodes based on the graph structure data. Then, the multi-head attention mechanism graph attention network extracts data features and relationship features. Finally, it is input into the prediction layer constructed by the full connection, and the prediction layer obtains the final key indicator prediction results.

[0080] In order to illustrate the effect of this embodiment, the following experiment was conducted, and the conducted experiment is described below.

[0081] This method was applied to the actual data of the hydrometallurgical zinc smelting and cobalt removal process of a smelter, and the results were compared with three other types of industrial indicator prediction methods. The machine learning method ARIMA, the deep learning method LSTM, and the graph network model GCN are very representative classic models and are widely used in industrial processes. Therefore, this experiment compares the prediction effects of ARIMA, LSTM, GCN and the method proposed in this embodiment on the hydrometallurgical zinc smelting and cobalt removal process. The evaluation criteria include RMSE, MAPE and R 2 The three indicators are: RMSE reflects the degree of deviation between the predicted output and the labeled output; MAPE is a statistical indicator used to measure the accuracy of the prediction, which can reflect the accuracy of the prediction model; R2 The degree to which the total variance in the forecast output can be explained, R 2 The larger the predicted output, the better it can track the changing trend of the real value. In addition, this embodiment gives the contribution of process variables to the outlet cobalt ion concentration, which can better understand which process variables have a greater impact on key indicators in the current operating state, and can provide effective guidance information for later optimization and regulation.

[0082] The following is an explanation of the online prediction effect of key indicators based on the graph attention network.

[0083] Figure 6 is a schematic diagram of the online prediction effect of key indicators based on the graph attention network according to an embodiment of the present application, Figure 6 The online prediction effect is demonstrated. From the figure, it can be seen that the online prediction method of key indicators based on the graph attention network can well predict the actual export cobalt ion concentration. Although there is fluctuation, the fluctuation range is small. The overall prediction trend is consistent with the actual value, the tracking effect is good, and the deviation degree is low.

[0084] Figure 7 is a process variable association diagram according to an embodiment of the present application, Figure 7 Display the relationship diagram between each variable (see step 3 for the calculation process), that is, the correlation between each variable. It can be seen that the largest factor affecting the outlet cobalt ion concentration is the redox potential, followed by zinc powder, then arsenic salt and waste acid, and finally the total flow, inlet copper ions and inlet cobalt ions. This is consistent with the actual situation. Actual on-site workers usually add zinc powder through the redox potential for regulation. Arsenic salt acts as a catalyst to increase the reaction rate of the cobalt removal process, and waste acid provides an acidic environment for cobalt removal. The total flow is generally relatively stable, so the inlet cobalt ion and copper ion concentrations are also relatively stable. This proves that the graph attention network has learned the relationship between the actual production process variables.

[0085] In order to fully reflect the effect of the method in this embodiment, the RMSE, MAPE and R 2 The two indicators are compared. The results are shown in Table 1:

[0086] Table 1: Comparison of the method in this embodiment with related methods

[0087] Model RMSE MAPE <![CDATA[R 2 ]]> ARIMA 0.00091 0.34384 0.51146 LSTM 0.00074 0.32422 0.67432 GCN 0.00071 0.23839 0.77233 The method of this embodiment 0.00069 0.21329 0.80172

[0088] As can be seen from Table 1, the method of this embodiment is superior to other related methods in all indicators, which shows that the method of the present invention has a better fitting effect and can better follow the actual value. Because the method of this embodiment fully extracts the complex dependency of process variables, it is suitable for highly nonlinear and non-steady-state industrial processes such as the cobalt removal process.

[0089] In this embodiment, an electronic device is provided, including a memory and a processor. The memory stores a computer program, and the processor is configured to run the computer program to execute the method in the above embodiment.

[0090] The above program can be run in the processor, or it can be stored in the memory (or computer-readable medium), which includes permanent and non-permanent, removable and non-removable media. Information storage can be achieved by any method or technology. Information can be computer-readable instructions, data structures, program modules or other data. Examples of computer storage media include, but are not limited to, phase change memory (PRAM), static random access memory (SRAM), dynamic random access memory (DRAM), other types of random access memory (RAM), read-only memory (ROM), electrically erasable programmable read-only memory (EEPROM), flash memory or other memory technology, read-only compact disk read-only memory (CD-ROM), digital versatile disk (DVD) or other optical storage, magnetic cassettes, tape disk storage or other magnetic storage devices or any other non-transmission media that can be used to store information that can be accessed by a computing device. As defined in this article, computer-readable media does not include temporary computer-readable media (transitory media), such as modulated data signals and carrier waves.

[0091] These computer programs can also be loaded onto a computer or other programmable data processing device so that a series of operating steps are executed on the computer or other programmable device to produce a computer-implemented process, so that the instructions executed on the computer or other programmable device provide instructions for implementing the process. Figure 1 A process or multiple processes and / or boxes Figure 1 The steps of the functions specified in one or more blocks can be implemented by different modules corresponding to different steps.

[0092] Such a device or system is provided in this embodiment. The system is called a key indicator prediction system for the cobalt removal process in hydrometallurgy, and includes: an acquisition module for acquiring data of multiple process variables of the reactor in the hydrometallurgy process, wherein the multiple process variables include multiple first process variables collected by sensors and multiple second process variables obtained by offline testing, and the multiple process variables are associated with the cobalt removal process, and the multiple second process variables include key indicators, which are outlet cobalt ion concentrations; a first training and determination module for inputting the data of the multiple process variables into a graph attention network to train and determine the parameters of the graph attention network to construct data features and relationships of the multiple process variables. Features, wherein the data features are the various variable nodes in the process variable graph topology diagram constructed by the multiple process variables, and each variable node corresponds to a process variable; the relationship features are the association relationships between the various variable nodes in the graph topology diagram; the graph attention network is used to extract the corresponding relationship between the key indicators and the multiple process variables; a second training determination module is used to train and determine the parameters of the fully connected prediction layer of the graph attention network according to the data features and relationship features of the multiple process variables obtained after training the parameters of the graph attention network, wherein the trained fully connected prediction layer is used to output the cobalt ion concentration prediction result.

[0093] The system or device is used to implement the functions of the method in the above-mentioned embodiment. Each module in the system or device corresponds to each step in the method, which has been explained in the method and will not be repeated here.

[0094] For example, the first training determination module is used to: normalize the data of the multiple process variables to obtain normalized data; input the normalized data of the multiple process variables into the graph attention network to train and determine the parameters of the graph attention network.

[0095] Optionally, the first training module is used to train and determine the parameters of the graph attention network, including: calculating the attention coefficient and performing weighted summation.

[0096] The above embodiments solve the problem of low prediction accuracy in the technical solutions for predicting key indicators of the cobalt removal process in hydrometallurgy of zinc in the related art, thereby establishing an accurate prediction model for the key indicators of the outlet cobalt ion concentration in view of the complex nonlinear and interdependent relationships between the variables in the cobalt removal process, thereby improving the prediction accuracy.

[0097] The above are only embodiments of the present application and are not intended to limit the present application. For those skilled in the art, the present application may have various changes and variations. Any modification, equivalent replacement, etc. made within the spirit and principle of the present application should be included in the scope of the claims of the present application.

Claims

1. A method for predicting key indicators of a hydrometallurgical zinc smelting process for removing cobalt, characterized in that: include: Acquire data of multiple process variables of a reactor in a hydrometallurgical zinc smelting process, wherein the multiple process variables include multiple first process variables collected by sensors and multiple second process variables obtained by offline analysis, the multiple process variables are associated with a cobalt removal process, the multiple second process variables include a key indicator, and the key indicator is an outlet cobalt ion concentration; The data of the multiple process variables are input into a graph attention network to train and determine the parameters of the graph attention network, so as to construct data features and relationship features of the multiple process variables, wherein the data features are the variable nodes in the process variable graph topology map constructed by the multiple process variables, and each variable node corresponds to a process variable; the relationship features are the association relationships between the variable nodes in the graph topology map; the graph attention network is used to extract the corresponding relationship between the key indicators and the multiple process variables; The parameters of the fully connected prediction layer of the graph attention network are trained and determined according to the data features and relationship features of the multiple process variables obtained after training the parameters of the graph attention network, wherein the trained fully connected prediction layer is used to output the cobalt ion concentration prediction result.

2. The method according to claim 1, characterized in that Inputting the data of the plurality of process variables into a graph attention network to train and determine parameters of the graph attention network includes: Normalizing the data of the plurality of process variables to obtain normalized data; The normalized data of the multiple process variables are input into the graph attention network to train and determine the parameters of the graph attention network.

3. The method according to claim 1 or 2, characterized in that: Training and determining the parameters of the graph attention network includes: calculating the attention coefficient and performing weighted summation.

4. The method according to claim 3, characterized in that Calculating the attention coefficient includes: for variable node i, calculating the correlation coefficient e between itself and neighboring variable nodes ij , which is expressed as follows: e ij =a([Wh i ‖Wh j ]);in, W represents the weight of feature transformation; h i Represents the characteristic vector of variable node i, that is, the specific value of process variable i; h j Represents the characteristic vector of variable node j, that is, the specific value of process variable j; [·||·] represents the concatenation of the transformed features of variable nodes i and j; a is the learnable attention coefficient weight vector; The attention coefficients calculated for all adjacent variable nodes are normalized by softmax to obtain the final attention coefficient a ij , the calculation process is as follows: k∈N i Indicates that variable node k is any neighboring node of variable node i, LeakyReLU is the activation function.

5. The method according to claim 4, characterized in that The weighted summation includes: using the weighted summation to calculate the linear combination of the output feature vectors corresponding to the attention coefficient, and the final result is used as the updated feature information h′ of the variable node i , a multi-head attention mechanism is used to calculate K groups of attention coefficients, and a splicing operation is used to integrate multiple attention heads. The calculation formula is as follows: || represents splicing; K represents the number of attention heads; represents the attention coefficient of k groups; W k Represents the weight of the feature transformation of the k-th group of attention mechanism; σ is the activation function ELU.

6. A key indicator prediction system for the hydrometallurgical zinc removal process, characterized in that: include: An acquisition module is used to acquire data of multiple process variables of a reactor in a hydrometallurgical zinc smelting process, wherein the multiple process variables include multiple first process variables collected by sensors and multiple second process variables obtained by offline analysis, the multiple process variables are associated with a cobalt removal process, and the multiple second process variables include key indicators, which are outlet cobalt ion concentrations; A first training and determination module is used to input the data of the multiple process variables into the graph attention network to train and determine the parameters of the graph attention network to construct data features and relationship features of the multiple process variables, wherein the data features are the various variable nodes in the process variable graph topology map constructed by the multiple process variables, and each variable node corresponds to a process variable; the relationship features are the association relationships between the various variable nodes in the graph topology map; the graph attention network is used to extract the corresponding relationship between the key indicators and the multiple process variables; The second training determination module is used to train and determine the parameters of the fully connected prediction layer of the graph attention network according to the data characteristics and relationship characteristics of the multiple process variables obtained after training the parameters of the graph attention network, wherein the trained fully connected prediction layer is used to output the cobalt ion concentration prediction result.

7. The system according to claim 6, characterized in that The first training determination module is used for: Normalizing the data of the plurality of process variables to obtain normalized data; The normalized data of the multiple process variables are input into the graph attention network to train and determine the parameters of the graph attention network.

8. The system according to claim 6 or 7, characterized in that: The first training module is used to train and determine the parameters of the graph attention network, including: calculating the attention coefficient and performing weighted summation.

9. The system according to claim 8, characterized in that Calculating the attention coefficient includes: for variable node i, calculating the correlation coefficient e between itself and neighboring variable nodes ij , which is expressed as follows: e ij =a([Wh i ‖Wh j ]);in, W represents the weight of feature transformation; h i Represents the characteristic vector of variable node i, that is, the specific value of process variable i; h j Represents the characteristic vector of variable node j, that is, the specific value of process variable j; [·||·] represents the concatenation of the transformed features of variable nodes i and j; a is the learnable attention coefficient weight vector; The attention coefficients calculated for all adjacent variable nodes are normalized by softmax to obtain the final attention coefficient a ij , the calculation process is as follows: k∈N i Indicates that variable node k is any neighboring node of variable node i, LeakyReLU is the activation function.

10. The system according to claim 9, characterized in that The weighted summation includes: using the weighted summation to calculate the linear combination of the output feature vectors corresponding to the attention coefficient, and the final result is used as the updated feature information h′ of the variable node i , a multi-head attention mechanism is used to calculate K groups of attention coefficients, and a splicing operation is used to integrate multiple attention heads. The calculation formula is as follows: || represents splicing; K represents the number of attention heads; represents the attention coefficient of k groups; W k Represents the weight of the feature transformation of the k-th group of attention mechanism; σ is the activation function ELU.

Citation Information

Patent Citations

  • Hesitant fuzzy c-means clustering method for copper removal process based on state transition algorithm

    CN111368897A

  • Method and system for extracting information from graphs

    US20200104729A1

  • Prediction method and system based on heterogeneous graph neural network model

    WO2021179838A1