Blast furnace hot metal silicon content prediction method based on time series graph convolution network

Abstract

The application discloses a blast furnace molten iron silicon content prediction method based on a time sequence diagram convolution network. First, the blast furnace ironmaking data is sequenced through a time window, and the sequenced time segment is input into a double-channel parallel network architecture to extract the coupling relationship between variables and the dynamics within variables. For the spatial channel of the coupling relationship between variables, a graph structure learning module first converts the time segment into an irregular graph structure, and carries out graph convolution operation on the basis of the learned graph structure. For the time sequence channel of the dynamics capturing, a long short-term memory network is used to capture the time sequence information in the data. This architecture realizes the spatiotemporal feature collaborative extraction of the blast furnace ironmaking data, is beneficial to the more accurate modeling of the complex conditions of the blast furnace by the model, improves the representation ability of the model, and thus improves the prediction accuracy of the model.

Description

TECHNICAL FIELD

[0001] The present application belongs to the field of soft measurement modeling of process industry, and particularly relates to a method for extracting space-time features in blast furnace ironmaking data by using a time series graph convolution network, and realizing prediction of silicon content in molten iron. BACKGROUND

[0002] Blast furnace ironmaking is a process for providing important raw materials for steel production, and occupies an important position in the development of the entire national economy. The goal of ironmaking is to achieve smooth and high-quality operation of the blast furnace, and in this process, real-time and accurate measurement of quality variables is crucial, which can reflect the operation state of the blast furnace and the product quality. The silicon content in molten iron has become an important quality index that operators often pay attention to in actual production. However, due to the lack of suitable sensors, the silicon content is difficult to be directly measured. In actual industrial processes, measurement is often completed by laboratory technicians, which not only requires huge manpower and material resources, but also faces the problem of sampling lag. Therefore, a data-driven method can be used to model and predict the silicon content by using the data generated in the history process.

[0003] However, the data of blast furnace ironmaking is complex, there is dynamicity within the variables, and there is a coupling relationship between the variables. The existence of dynamicity makes the values of the same variable dependent on the time scale, and the value of the variable at the current time will be affected by the variable values at the historical time. The coupling relationship between variables means that variables are not independent of each other, but interact with each other, and finally the furnace condition of the blast furnace needs to be judged by comprehensively considering each variable. The existing silicon content prediction method generally focuses on the dynamics of ironmaking data, but ignores the exploration of the coupling relationship between variables. However, these two characteristics widely exist in the blast furnace ironmaking process, and modeling only one of them is not conducive to accurately representing the internal state of the blast furnace. SUMMARY

[0004] In order to overcome the shortcomings of the prior art, the purpose of the present application is to provide a blast furnace molten iron silicon content prediction method based on a time series graph convolution network. Ten variables such as the amount of bosh gas and top pressure are selected as input variables to predict the value of silicon content. This method uses a dual-channel parallel network architecture to comprehensively extract the space-time features in the data, converts the time series into a graph structure through the spatial channel, and carries out graph convolution operation under the constraint of the graph structure to capture the coupling relationship between variables; the time dependence in the data is modeled through the time series channel. On the basis of space-time features, the final silicon content prediction value is obtained through the fully connected network.

[0005] The technical solutions of the present application to achieve its technical purposes are as follows:

[0006] A blast furnace molten iron silicon content prediction method based on a time series graph convolution network, the steps are as follows:

[0007] Step 1: Correlation analysis of process parameters and silicon content in blast furnace ironmaking, and selection of process variables for prediction;

[0008] Step 2: Serializing the process variables of blast furnace ironmaking through time window;

[0009] Step 3: Inputting the time series segment into parallel time and space channels; in the space channel, first use the graph structure learning module to construct a graph structure representing the coupling relationship between variables, and then perform graph convolution operation based on the graph structure and time series to obtain the spatial features of blast furnace ironmaking data; in the time series channel, use the Long Short-Term Memory Network (LSTM) to capture the time series information in the data and obtain the time series features;

[0010] Step 4: Feature fusion of time series features and spatial features through horizontal connection to complete the collaborative extraction of spatio-temporal features;

[0011] Step 5: Based on the spatio-temporal features, complete the prediction of silicon content through a fully connected network.

[0012] Step 1: Pearson correlation analysis is used to select process parameters and silicon variables: furnace gas volume, top pressure, oxygen enrichment pressure, full pressure difference, cold wind temperature, hot wind temperature, top temperature, blast humidity, coal injection amount and last time silicon content.

[0013] Step 2: Serializing the ironmaking data through a fixed size T time window, predicting the silicon content at T time through the input at {1, 2, ……, T}.

[0014] Step 3 contains two parallel channels: time series channel and space channel, and the serialized time segments are input into these two channels; the space channel contains a graph structure learning module and a graph convolution network, and the time channel is composed of a Long Short-Term Memory Network; the graph structure learning module converts the time series at {1, 2, ……, T} into a non-Euclidean graph structure, where nodes represent variables, node features are time series segments corresponding to variables at {1, 2, ……, T}, and node edges represent the coupling relationship between variables; the graph convolution neural network performs message aggregation operation on the variable pairs with coupling relationship under the constraint of graph structure, capturing the spatial features in the ironmaking data; in the time series channel, the Long Short-Term Memory Network extracts important time series features within T .

[0015] In step 4, the fusion of different semantic features is achieved through the horizontal connection module, which facilitates the fusion of features between different channels and the collaborative training of networks within two channels.

[0016] In step 5, the temporal and spatial features are connected in parallel and then input into a fully connected network to complete the prediction output of silicon content.

[0017] The beneficial effects of this invention are as follows: A graph structure learning module is used to transform ironmaking time series segments into a graph structure. The existence of the graph structure can explicitly represent the coupling relationships between variables. This graph structure acts as a constraint, limiting the way the graph convolutional network aggregates information between different variables, thus enabling the extraction of coupling relationships between variables under the action of the graph convolutional network. A long short-term memory network is used to capture the temporal dependencies within variables. This network architecture achieves comprehensive extraction of spatiotemporal features, more accurately representing the internal state of the blast furnace, and ultimately improving the accuracy of the model's prediction of silicon content. Attached Figure Description

[0018] Figure 1 The diagram shown illustrates the blast furnace data serialization process.

[0019] Figure 2 The diagram shows the architecture of a temporal graph convolutional network.

[0020] Figure 3 The diagram shown is a flowchart of a time-series convolutional network.

[0021] Figure 4 The silicon content prediction results are shown. Detailed Implementation

[0022] The present invention will be further described below with reference to the accompanying drawings and embodiments.

[0023] The specific steps of the method for predicting silicon content in blast furnace hot metal based on temporal graph convolutional networks are as follows:

[0024] Step 1: Pearson correlation analysis was used to analyze the correlation between process parameters and silicon content. Variables with a large impact on silicon content were selected as inputs. The final confirmed input variables were: furnace gas volume, top pressure, oxygen-enriched pressure, total differential pressure, cold air temperature, hot air temperature, top temperature, blast humidity, pulverized coal injection volume, and silicon content at the previous moment.

[0025] Step 2: Represent the original blast furnace ironmaking data as follows , L This is the length of the dataset. (Using a fixed size) T Time window (X, Y) Perform serialization, the serialization method is as follows Figure 1 As shown. By Figure 1 It can be seen that time windows of a fixed size are shifted sequentially.T Silicon content at time y t The corresponding input is x 1, x 2, ……, x T} where , N is the number of input variables. The model will predict the silicon content at time T through the input at time T {1, 2, ……, }.

[0026] Step 3: contains two parallel channels: time channel and space channel, are input into the two channels respectively, in order to make the graph structure be able to represent the coupling relationship between variables, the input of the space channel needs to be transposed to . The network architecture contained in this step is shown in Figure 2 , and the corresponding flow chart is shown in Figure 3 . The space channel contains a graph structure learning module and a graph convolution network, and the time channel is composed of a long short-term memory network.

[0027] The graph structure learning module converts the time series at time T {1, 2, ……, T} into a non-Euclidean graph structure, in which the nodes represent variables, the features of the nodes are the time series of the variable at time A {1, 2, ……, p}, and the edges between the nodes represent the coupling relationship between variables. The edges between variable pairs are represented by an adjacency matrix q , and when there is an edge between variable A pq = A qp =1, otherwise 0. In order to ensure that the learned graph structure is related to the downstream prediction task, a matrix M is introduced in this module. This matrix adjusts the input and is learned together with other weight matrices in the network to form an end-to-end network. The way this matrix is introduced is as follows:

[0028]

[0029] where o is the Hadamard product. After adjustment, the radial basis distance function is used to calculate the distance between variables, and the calculation method is as follows:

[0030]

[0031] whereK (.) is a variable p and a variable q the radial basis distance between them, γ is a hyper-parameter. After obtaining the distance matrix, set the threshold ε, and convert the distance matrix into an adjacency matrix representing the connection between nodes A . The conversion method is defined as:

[0032]

[0033] After that, the graph convolutional neural network performs message aggregation operations on the variable pairs with coupled relationships under the constraints of the graph structure, capturing the spatial features in the ironmaking data. The extraction operation of spatial features is as follows:

[0034]

[0035] In the time sequence channel, important time sequence features in T are extracted by the long short-term memory network. The extraction method of time sequence features can be represented as:

[0036]

[0037] wherein, and x t represent the forget gate, input gate, output gate, hidden state and input of t time. Cell state C t is updated by the candidate state . b and W represent the weights and biases of the network.

[0038] Step 4: The fusion of different semantic features is realized by the transverse connection module, which is convenient for the fusion of features between different channels and the collaborative training of the network in the two channels. The transverse connection can be defined as: wherein, Z 1 is the time sequence feature extracted by the first layer long short-term memory network, S 1 is the spatial feature extracted by the first layer graph convolutional network, H is the fused feature. The fusion function Ω is element-wise summation, and the fused feature continues to propagate forward in the original channel.

[0039] In step 5, the time sequence feature and the spatial feature are connected in parallel and input into the fully connected network, and the prediction output of the silicon content is completed through the fully connected network.

[0040] The method is used to predict the silicon content in the actual blast furnace production process, and the experimental results of the prediction are as followsFigure 4 As shown, the solid line is the true silicon content value, and the dotted line is the model prediction value. It can be seen that the model prediction achieves good results. The root mean square error (RMSE) is selected to evaluate the prediction effect. The RMSE on the data set is 0.0338, which can meet the accuracy requirements of the actual industrial site.

[0041] The above-described embodiments only express several embodiments of the present application, which are described in detail and specifically, but should not be understood as a limitation on the scope of the application. It should be noted that for ordinary skilled persons in the art, without departing from the concept of the present application, a number of modifications and improvements can be made, which are within the scope of the present application. Therefore, the protection scope of the present application should be subject to the appended claims.

Claims

1. A blast furnace hot metal silicon content prediction method based on a time series graph convolution network, characterized by, The steps are as follows: Step 1: Correlation analysis of process parameters and silicon content of blast furnace ironmaking, selection of process variables for prediction; Step 2: The process variables of blast furnace ironmaking are serialized by time window; Step 3: The time series segment is input into the parallel time channel and space channel; In the space channel, first, the graph structure learning module is used to construct the graph structure representing the coupling relationship between variables, and on the basis of the graph structure and time series, graph convolution operation is performed to obtain the spatial features of blast furnace ironmaking data; In the time sequence channel, the long short-term memory network (LSTM) is used to capture the time sequence information in the data to obtain the time sequence features; Step 4: The time sequence features and spatial features are fused through horizontal connection to complete the collaborative extraction of spatio-temporal features; Step 5: On the basis of spatio-temporal features, the prediction of silicon content is completed through the fully connected network.

2. The method of claim 1, wherein: Step 1 uses Pearson correlation analysis to analyze the correlation between process parameters and silicon content, and selects the following process parameter variables: furnace gas volume, top pressure, oxygen enrichment pressure, total pressure difference, cold wind temperature, hot wind temperature, top temperature, blast humidity, coal injection amount and last time silicon content.

3. The method of claim 1, wherein: Step 2 by fixing the size T The ironmaking data is serialized using time windows, through {1, 2, ..., ... T Input prediction at time} T Silicon content at any given time.

4. The method of claim 3, wherein: Step 3 includes two parallel channels: a time channel and a space channel, and the serialized time slices are respectively input into the two channels; the space channel includes a graph structure learning module and a graph convolution network, and the time channel is composed of a long short-term memory network; the graph structure learning module converts the time series at {1, 2, ……, T} moments into a non-Euclidean graph structure, in which the nodes represent variables, the features of the nodes are the time series of the variables at {1, 2, ……, T} moments, and the edges of the nodes represent the coupling relationship between the variables; the graph convolution neural network performs a message aggregation operation on the variable pairs with the coupling relationship under the constraint of the graph structure, and captures the spatial features in the smelting data; In the temporal channel, the important temporal features in a time instant are extracted by a long short-term memory network. T In the spatial channel, the important spatial features in a time instant are extracted by a spatial transformer network.

5. The method of claim 1, wherein: In step 4, the fusion of different semantic features is realized through the horizontal connection module, which facilitates the fusion of features between different channels and the collaborative training of networks within the two channels.

6. The method of claim 1, wherein: In step 5, the time sequence features and spatial features are connected in parallel and input into the fully connected network, and the prediction of silicon content is completed through the fully connected network.

Images