Method for constructing LF ladle refining prediction model based on knowledge graph

Abstract

The invention provides a method for constructing an LF ladle refining prediction model based on a knowledge graph, and the method comprises the steps: 1) obtaining LF refining process data, and creating process data nodes, node attributes and the relation between the nodes; 2) creating process data nodes, node attributes and relationships among the nodes into a Neo4j database; 3) acquiring a path association relationship with the reference sample by adopting a knowledge graph, and screening to obtain a process sample similar to the reference sample by combining a similar model with an actual accuracy requirement so as to form a similar sample set together with the reference sample; according to the method, the proportion of the molten steel temperature at the refining end point with the prediction error within + / -10 DEG C obtained through the prediction model is 95% or above, the proportion of the molten steel temperature at the refining end point with the prediction error within + / -5 DEG C is 90% or above, and the proportion of the molten steel sulfur content at the refining end point with the prediction error within + / -0.002% is 90% or above.

Description

Technical Field

[0001] The present invention belongs to the field of prediction model construction, and specifically relates to a method for constructing a LF ladle refining prediction model based on a knowledge graph. Background Art

[0002] The mechanism of the LF ladle refining process is extremely complex, the working conditions evolve dynamically, the data is multi-source and heterogeneous, the semantics of expert experience is fuzzy and non-standard, the working processes are complicated and coupled, the operating parameters are time-space dependent, the boundary conditions are complex and changeable, and there is a lack of real-time perception of key information. These results in the model characteristics not matching the actual working conditions. Traditional mechanism models and global machine learning prediction models are unable to achieve accurate modeling of the LF ladle refining process, which seriously restricts the high-performance intelligent operation of the refining process. Summary of the invention

[0003] In order to solve the problems existing in the prior art, the present invention provides a method for constructing a LF ladle refining prediction model based on a knowledge graph, aiming to solve the problem that the existing machine learning sample library is too large and some distorted samples affect the prediction accuracy, thereby improving the accuracy of the LF refining prediction model.

[0004] The present invention provides

[0005] A method for constructing a LF ladle refining prediction model based on a knowledge graph, characterized in that it comprises the following steps:

[0006] 1) Obtain LF refining process data, and create process data nodes, node attributes and the relationship between nodes based on the acquired data;

[0007] 2) The process data nodes, node attributes and relationships between nodes obtained in step 1) are created in the Neo4j database to obtain the LF refining production process knowledge graph;

[0008] 3) Determine the similar working condition benchmark sample, use the knowledge graph obtained in step 2) to obtain the path association relationship between the benchmark sample, and screen the process samples similar to the benchmark sample through the similarity model combined with the actual accuracy requirements to form a similar sample set together with the benchmark sample. The similarity model is S=W1*(1 / L)+W2*F+W3*A, where W1, W2, and W3 are the path length weight, the node association degree weight, and the node attribute similarity weight on the path, respectively. W1+W2+W3=1, L is the path length, F is the node association degree, and A is the node attribute similarity on the path;

[0009] 4) Using a machine learning algorithm to model the similar sample set obtained in step 3) to obtain a LF ladle refining prediction model.

[0010] As a preferred solution of the method for constructing an LF ladle refining prediction model based on a knowledge graph in the present invention, the process data node includes an equipment parameter node and a production data node, and the equipment parameter node includes a station node and a ladle number node; the production data node includes a furnace number node, a steel type node, an operation shift node, a shift node, an alloy node, an auxiliary material node, an inlet and outlet temperature measurement and sampling node, a specific process operation node, and other process parameter nodes.

[0011] As a preferred embodiment of the method for constructing an LF ladle refining prediction model based on a knowledge graph in the present invention, the attributes of the ladle number node include ladle age and turnover status; the attributes of the alloy node include alloy composition and addition amount; the attributes of the auxiliary material node include auxiliary material composition and addition amount; the attributes of the inlet and outlet temperature measurement sampling node include the temperature measurement result, composition and content of the molten steel; the attributes of the specific process operation node include the number of refining times, bottom blowing gas, electric furnace steel tapping time, ladle arrival time, initial power supply time, end power supply time, and ladle departure time; the attributes of the other process parameter nodes include steel tapping weight, power consumption, refining time, empty ladle time, static stirring time, returned molten steel, and liquidus temperature.

[0012] As a preferred solution of the method for constructing a LF ladle refining prediction model based on a knowledge graph of the present invention, step 1) specifically includes:

[0013] 101) The acquired process data is normalized, including filling missing values, removing outliers, unifying date formats, standardizing terms, and removing duplicate records, unnecessary spaces, and special characters. Then, Python code is used to create nodes and relationship definitions between nodes.

[0014] 102) Create a furnace number node and set the specific value of the furnace number as the furnace number node attribute;

[0015] 103) Create a steel type node and associate it with the furnace number node using the PROCESSES relationship (i.e., the relationship between the steel type node and the furnace number node is the PROCESSES relationship);

[0016] 104) Create a station node and associate it with the furnace number node using the LOCATED_AT relationship (i.e., the relationship between the station node and the furnace number node is the LOCATED_AT relationship);

[0017] 105) Create operation shift node and shift node, and associate them with furnace number node using HAS_SHIFT and HAS_TEAM relationships respectively (i.e. the relationships between operation shift node, shift node and furnace number node are HAS_SHIFT and HAS_TEAM relationships respectively);

[0018] 106) Create inlet and outlet temperature measurement sampling nodes and specific process operation nodes, and associate them with the furnace number node using the HAS_OPERATION relationship (that is, the relationship between the inlet and outlet temperature measurement sampling nodes, specific process operation nodes and the furnace number node is an OPERATION relationship);

[0019] 107) Create ladle number nodes, alloy nodes, and auxiliary material nodes, and associate them with the furnace number nodes using the USES relationship (i.e., the relationships between the ladle number node, alloy node, auxiliary material node, and furnace number node are all USES relationships);

[0020] 108) Create other process parameter nodes and associate them with the furnace number node using the ADOPT relationship (that is, the relationship between the other process parameter nodes and the furnace number node is the ADOPT relationship).

[0021] As a preferred solution of the method for constructing a LF ladle refining prediction model based on a knowledge graph of the present invention, the acquisition of the path length, the node association degree and the node attribute similarity on the path in step 3) includes:

[0022] 301) Obtain the number of connections L between the node relationships of different furnace numbers, that is, obtain the path length.

[0023] 302) Use Jaccard similarity method to obtain the node correlation degree F=|S of different furnace numbers A ∩S B | / |S A ∪S B |, where S A and S B are the node sets associated with two furnace number nodes A and B respectively; |S A ∩S B | represents the number of elements in the intersection of two associated node sets, that is, the number of nodes that A and B are associated with; |S A ∪S B | represents the number of elements in the union of two associated node sets, that is, the total number of non-duplicate nodes associated with A and B.

[0024] Each "heat number" node will be associated with multiple types of nodes, such as alloy nodes, auxiliary material nodes, inlet and outlet temperature measurement sampling nodes, specific process operation nodes, etc. First, it is necessary to accurately count the number of different types of nodes associated with each heat number node. For example, for the heat number node A, it is associated with 2 alloy nodes, 1 auxiliary material node, 2 inlet and outlet temperature measurement sampling nodes, and 4 specific process operation nodes; while the heat number node B is associated with 2 alloy nodes, 1 auxiliary material node, 3 inlet and outlet temperature measurement sampling nodes, and 3 specific process operation nodes. Through such statistics, we can clearly see the basic situation of each heat number node in terms of the type and number of associated nodes. The Jaccard similarity coefficient F is a common indicator for measuring the similarity of two sets. Here, the node set associated with each heat number node is regarded as a whole. The closer the Jaccard similarity is to 1, the more similar the two heat number nodes are in terms of the type and number of associated nodes. For example, if the node sets associated with two furnace number nodes are almost exactly the same, then the number of elements in their intersection and union will be very close, and the Jaccard similarity will be close to 1; conversely, if the number of intersection elements is small, the Jaccard similarity will be low, indicating that they have large differences in associated nodes.

[0025] 303) The formula A=1 / (1+D) is used to obtain the similarity of node attributes on the path, where D is the Euclidean distance of the numerical attribute vectors shared by the two nodes.

[0026] The similarity of node attributes on the path is used to measure the similarity between the attributes of adjacent nodes in the path. When evaluating the similarity of two furnace numbers, the steel grades they produce have similar composition requirements and similar process parameters (such as power consumption, refining time, etc.), which can reflect the similarity of the two "furnace number" nodes in terms of relevant business characteristics. Through a loop traversing the adjacent nodes in the path, find out their common numerical attributes, and calculate the Euclidean distance D of the corresponding attribute vectors of the two nodes. Since the larger the value of the Euclidean distance D, the greater the difference, in order to convert it to a similarity score between 0 and 1 (0 means completely dissimilar, 1 means completely similar), use A=1 / (1+D) for conversion. In this way, the smaller the distance, the closer the similarity score is to 1, indicating that the two furnace number nodes are more similar.

[0027] In the graph, the path length refers to the number of relationships in the path connecting two "heat number" nodes. If two heat numbers can be connected through very few intermediate links (such as stations, process operation nodes, etc.), they may have high similarities in process execution, production conditions, etc. For example, if two "heat number" nodes are connected by 3 relationships, then the length of this path is 3. In the calculation formula of the similarity score, the reciprocal of the path length 1 / L will be used for calculation. This is because the shorter the path, the greater its contribution to the similarity should be, so the reciprocal is taken to reflect this inverse relationship.

[0028] As a preferred solution of the method for constructing a LF ladle refining prediction model based on a knowledge graph of the present invention, the modeling in step 4) specifically includes:

[0029] The steel output weight, empty ladle time, LF power consumption, refining time, power-on time, station entry temperature, alloy addition amount and auxiliary material addition amount are used as model input items, and the station exit temperature is used as the output item to construct a local training model to obtain a prediction model for the refining endpoint temperature, thereby realizing the prediction of the molten steel temperature at the refining endpoint.

[0030] As a preferred solution of the method for constructing a LF ladle refining prediction model based on a knowledge graph of the present invention, the modeling in step 4) specifically includes:

[0031] The steel output weight, refining time, power-on time, soft blowing time, incoming sulfur content, lime addition, refining slag addition, and fluorite addition are used as model input items, and the endpoint sulfur content is used as the output item to construct a local training model to obtain a prediction model for the sulfur content at the refining endpoint, thereby realizing the prediction of the sulfur content in molten steel at the refining endpoint.

[0032] Compared with the prior art, the present invention has the following beneficial effects:

[0033] 1. During the LF refining process, a large amount of production data and process knowledge are accumulated in various forms. Traditional Excel spreadsheet data and scattered expert knowledge cannot meet the needs of refining modeling. Based on the above data and knowledge, the present invention uses relationship extraction and modeling technology to construct a LF refining process knowledge graph, thereby providing practical support for the working condition cognition, dynamic modeling and other needs of the steel metallurgical process.

[0034] 2. In the actual LF refining process, the working conditions are complex and volatile. As the production situation changes, the prediction accuracy and stability of the global prediction model will decrease significantly, making it difficult to adapt to the LF refining production process with variable production equipment and conditions. The local model based on the real-time learning algorithm proposed in the present invention does not need to be constructed in advance, but adopts the mode of "modeling and predicting at the same time". Whenever the sample to be tested needs to be predicted and output, the sample data most relevant to the working conditions of the sample to be tested is selected through the above-mentioned process knowledge map, so as to construct a local model for prediction. Therefore, the LF refining local model based on the real-time learning algorithm has stronger adaptability and prediction accuracy stability than the global model.

[0035] 3. The LF ladle refining prediction model constructed based on the knowledge graph in the present invention obtains a prediction error of the molten steel temperature at the refining end point within ±10°C, accounting for more than 95%, a prediction error of the molten steel temperature at the refining end point within ±5°C, accounting for more than 90%, and a prediction error of the sulfur content in the molten steel at the refining end point within the range of ±0.002%, accounting for more than 90%. BRIEF DESCRIPTION OF THE DRAWINGS

[0036] Figure 1 A partial schematic diagram of the knowledge graph obtained in Example 1 of the present invention. DETAILED DESCRIPTION

[0037] The following will be described clearly and completely in conjunction with the technical solutions in the embodiments. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without creative work are within the scope of protection of the present invention.

[0038] The embodiments and comparative examples of the present invention both use 17634 heats of molten steel produced in the LF refining workshop of a steel plant as samples for model establishment.

[0039] Example 1

[0040] A method for constructing a LF ladle refining prediction model based on a knowledge graph comprises the following steps:

[0041] 1) Obtain LF refining process data, and create process data nodes, node attributes and relationships between nodes based on the acquired data, including:

[0042] 101) Obtain the furnace number, time, furnace seat, team, steel type, ladle number, furnace age, ladle condition, empty ladle time, molten steel weight, inlet temperature, outlet temperature, refining cycle, heating time, static blowing time, cumulative argon volume, slag and alloy material addition, inlet molten steel composition content and / or outlet molten steel composition content of the LF refining production process, and standardize them;

[0043] 102) Create a furnace number node and set the specific value of the furnace number as the furnace number node attribute;

[0044] 103) Create a steel type node and associate it with the furnace number node using the PROCESSES relationship;

[0045] 104) Create a station node and associate it with the furnace number node using the LOCATED_AT relationship;

[0046] 105) Create operation shift and shift nodes, and associate them with the furnace number node using HAS_SHIFT and HAS_TEAM relationships respectively;

[0047] 106) Create inbound and outbound temperature measurement, sampling, and specific process operation nodes, and associate them with the furnace number node using the HAS_OPERATION relationship;

[0048] 107) Create ladle number, ladle status, and alloy auxiliary material nodes, and associate them with the furnace number node using the USES relationship;

[0049] 108) Create a process parameter node and associate it with the furnace number node using the ADOPT relationship.

[0050] 2) The process data nodes, node attributes and relationships between nodes obtained in step 1) are created in the Neo4j database to obtain the LF refining production process knowledge graph, where the knowledge graph formed by some nodes is as follows: Figure 1 As shown;

[0051] 3) Determine the first heat as a similar process benchmark sample, use the knowledge graph obtained in step 2) to obtain the path association relationship between the first heat, and establish a similarity model S=W1*(1 / L)+W2*F+W3*A, where W1, W2, and W3 are the path length weight, node association coefficient weight, and path node attribute similarity weight, respectively, and are assigned values ​​of 0.3, 0.3, and 0.4, L is the path length, F is the path co-occurrence frequency, and A is the path node attribute similarity. The acquisition of the path length, path co-occurrence frequency, and path node attribute similarity includes:

[0052] 301) Obtain the number of connections L of the node relationships between different furnace numbers, that is, obtain the path length. In this embodiment, after calculation and statistics, the distribution range of the path length L is 1-50;

[0053] 302) The Jaccard similarity method is used to obtain the degree of correlation between nodes of different furnace numbers F=|S A ∩S B | / |S A ∪S B|, in multiple production records, two heat numbers are always produced through the same or similar processes (such as passing through the same station, using the same ladle, adding the same alloy auxiliary materials, etc.), which shows that they have a high similarity in production mode. For example, the nodes associated with the heat number node A are: alloy node 1, alloy node 2, inlet and outlet temperature measurement sampling node 1, specific process operation node 1, that is, S A ={alloy node 1, alloy node 2, inlet and outlet temperature measurement sampling node 1, specific process operation node 1}; the nodes associated with the furnace number node B are: alloy node 2, inlet and outlet temperature measurement sampling node 1, specific process operation node 2, specific process operation node 3, that is, S B ={alloy node 2, inlet and outlet temperature measurement sampling node 1, specific process operation node 2, specific process operation node 3}, first calculate the intersection S A ∩S B ={alloy node 2, inlet and outlet temperature measurement sampling node 1}=2, then calculate the union S A ∪S B ={alloy node 1, alloy node 2, inlet and outlet temperature measurement sampling node 1, specific process operation node 1, specific process operation node 2, specific process operation node 3}=6, and finally calculate J(A,B)=|S according to the formula A ∩S B | / |S A ∪S B |=2 / 6=0.33. In this embodiment, after calculation and statistics, the distribution range of the node association coefficient F is 0-1;

[0054] 303) Formula A=1 / (1+D) is used to obtain the similarity of node attributes on the path. When evaluating the similarity of two furnace numbers, the steel grades they produce have similar composition requirements and similar process parameters (such as power consumption, refining time, etc.), which can reflect the similarity of the two "furnace number" nodes in terms of relevant business characteristics. Through a loop traversing the adjacent nodes in the path, their common numerical attributes are found, and the Euclidean distance of the corresponding attribute vectors of the two nodes is calculated. Since the larger the value of the Euclidean distance, the greater the difference, in order to convert it into a similarity score between 0 and 1 (0 means completely dissimilar, 1 means completely similar), A=1 / (1+D) is used for conversion. In this way, the smaller the distance, the closer the similarity score is to 1, indicating that the two furnace number nodes are more similar. In this embodiment, after calculation and statistics, the distribution range of the node attribute similarity A is 0-1;

[0055] 304) Based on the calculation results of the above three steps, the comprehensive similarity score S is used to judge the similarity between the two furnace number nodes, and the calculation formula is: S=W1*(1 / L)+W2*F+W3*A, where W1 and W2 are both given a weight of 0.3, and W3 is given a weight of 0.4. In this embodiment, after calculation and statistics, the distribution range of S is 0.005-1. Combined with the actual prediction accuracy requirements, the similarity standard is determined to be S≥0.9, and 29 groups are screened according to this standard to have similarity with the first furnace.

[0056] 4) Using the 29 groups of data obtained in step 304) combined with the first heat data for a total of 30 heats as a similar sample set, a machine learning algorithm is used, and the steel weight, empty bag time, LF power consumption, refining time, power-on time, station entry temperature, alloy (silicon manganese alloy, high carbon ferromanganese) addition amount and auxiliary material (lime) addition amount are used as model input items, and the station exit temperature is used as an output item to construct a local training model to obtain a prediction model for the refining endpoint temperature;

[0057] 5) According to the prediction model calculation in step 4), the prediction value of the molten steel temperature prediction model at the end point of LF refining under the process conditions shown in Table 1 is 1663°C.

[0058] It has been determined that the actual molten steel temperature at the end point of refining under the above process conditions is 1665°C. It can be seen that the deviation between the model prediction value and the actual value of this embodiment is 2°C, which is within ±5°C.

[0059] 99 different sets of process conditions were used to calculate using the prediction model obtained in step 4) and to carry out actual refining operations using the above process conditions. The actual molten steel temperature at the refining endpoint was measured. According to statistics, 92% of the 100 sets of prediction values ​​had a deviation of ±5°C from the actual values.

[0060] Table 1 Process conditions of Example 1

[0061] Steel weight / kg Empty packet time / min LF power consumption / MWH Refining time / min Power-on time / min Station entry temperature / ℃ Lime / kg Silicon manganese alloy / kg High carbon ferromanganese / kg 108973 34 12.747 47 35 1587 99 122 297

[0062] Example 2

[0063] The difference between this embodiment and embodiment 1 is that step 4) adopts a machine learning algorithm, takes the steel output weight, static stirring time, molten steel refining time, incoming molten steel temperature, incoming molten steel sulfur content, lime addition amount, refining slag addition amount and fluorite addition amount as model input items, and takes the outgoing molten steel sulfur content as an output item, so as to obtain a prediction model for the sulfur content of molten steel at the refining endpoint.

[0064] 5) According to the prediction model calculation in step 4), the sulfur content of the molten steel at the end point of LF refining is 0.0036% under the process conditions shown in Table 2.

[0065] It has been determined that the actual endpoint sulfur content of molten steel at LF refining under the above process conditions is 0.0032%. It can be seen that the deviation between the model prediction value and the actual value in this embodiment is 0.0004%.

[0066] 99 different sets of process conditions were used to calculate using the prediction model obtained in step 4) and to carry out actual refining operations using the above process conditions. The actual sulfur content of molten steel at the refining endpoint was measured. According to statistics, 94% of the 100 sets of prediction values ​​had a deviation from the actual value within the range of ±0.002%.

[0067] Table 2 Process conditions of Example 2

[0068] Steel weight / kg Static stirring time / min Refining time / min Station entry temperature / ℃ Sulfur content at the station / % Refining slag / kg Lime / kg Fluorite / kg 100257 24 72 1536 0.01569 150 121 80

[0069] Comparative Example 1

[0070] 1) The acquired data is normalized in the same manner as in Example 1;

[0071] 2) using the data obtained in step 1) as basic data, using the same machine learning algorithm as in Example 1, taking the steel tapping weight, empty ladle time, LF power consumption, molten steel refining time, power-on time, incoming molten steel temperature, alloy addition amount and auxiliary material addition amount as model input items, and taking the outgoing molten steel temperature as the output item to construct a local training model to obtain a prediction model for the molten steel temperature at the refining endpoint;

[0072] 3) According to the prediction model calculation in step 2), the prediction value of the molten steel temperature prediction model at the end point of LF refining under the process conditions shown in Table 1 is 1659°C.

[0073] It has been determined that the actual molten steel temperature at the end point of LF refining under the above process conditions is 1665°C. It can be seen that the deviation between the model prediction value and the actual value of this embodiment is 7°C, which is greater than ±5°C.

[0074] 99 different process conditions were continued to be used (the process conditions were the same as those in Example 1), and the prediction model obtained in step 2) was used to calculate and use the above process conditions to perform actual refining operations, and the actual molten steel temperature at the refining endpoint was measured. According to statistics, the deviation between the predicted values ​​and the actual values ​​of the 100 groups was within ±5°C, accounting for 77%.

[0075] It should be noted that, according to the above embodiments of the present invention, those skilled in the art can fully realize the full scope of the independent claims and dependent rights of the present invention, and the implementation process and method are the same as the above embodiments; and the part not described in detail in the present invention belongs to the well-known technology in the field. However, the protection scope of the present invention is not limited to this, and any changes or substitutions that can be easily thought of by any person familiar with the field within the technical scope disclosed by the present invention should be covered within the protection scope of the present invention.

Claims

1. A method for constructing a LF ladle refining prediction model based on a knowledge graph, characterized in that: The steps include: 1) Obtain LF refining process data, and create process data nodes, node attributes and the relationship between nodes based on the acquired data; 2) The process data nodes, node attributes and relationships between nodes obtained in step 1) are created in the Neo4j database to obtain the LF refining production process knowledge graph; 3) Determine the similar working condition benchmark sample, use the knowledge graph obtained in step 2) to obtain the path association relationship between the benchmark sample, and screen the process samples similar to the benchmark sample through the similarity model combined with the actual accuracy requirements to form a similar sample set together with the benchmark sample. The similarity model is S=W1*(1 / L)+W2*F+W3*A, where W1, W2, and W3 are the path length weight, the node association degree weight, and the node attribute similarity weight on the path, respectively. W1+W2+W3=1, L is the path length, F is the node association degree, and A is the node attribute similarity on the path; 4) Using a machine learning algorithm to model the similar sample set obtained in step 3) to obtain a LF ladle refining prediction model.

2. A method for constructing a LF ladle refining prediction model based on a knowledge graph according to claim 1, characterized in that: The process data nodes include equipment parameter nodes and production data nodes. The equipment parameter nodes include station nodes and ladle number nodes; the production data nodes include furnace number nodes, steel type nodes, operation shift nodes, shift nodes, alloy nodes, auxiliary material nodes, inlet and outlet temperature measurement and sampling nodes, specific process operation nodes, and other process parameter nodes.

3. A method for constructing a LF ladle refining prediction model based on a knowledge graph as claimed in claim 2, characterized in that: The attributes of the ladle number node include ladle age and turnover status; the attributes of the alloy node include alloy composition and addition amount; the attributes of the auxiliary material node include auxiliary material composition and addition amount; the attributes of the inlet and outlet temperature measurement and sampling node include the temperature measurement results, composition and content of the molten steel; the attributes of the specific process operation node include the number of refining times, bottom blowing gas, electric furnace steelmaking time, ladle arrival time, initial power supply time, end power supply time, and ladle departure time; the attributes of other process parameter nodes include steelmaking weight, power consumption, refining time, empty ladle time, static stirring time, returned molten steel, and liquidus temperature.

4. A method for constructing a LF ladle refining prediction model based on a knowledge graph according to any one of claims 1 to 3, characterized in that: Step 1) specifically includes: 101) The acquired process data is normalized, including filling missing values, removing outliers, unifying date formats, standardizing terms, and removing duplicate records, unnecessary spaces, and special characters. Then, Python code is used to create nodes and relationship definitions between nodes. 102) Create a furnace number node and set the specific value of the furnace number as the furnace number node attribute; 103) Create a steel type node and associate it with the furnace number node using the PROCESSES relationship; 104) Create a station node and associate it with the furnace number node using the LOCATED_AT relationship; 105) Create operation shift node and shift node, and associate them with furnace number node using HAS_SHIFT and HAS_TEAM relationships respectively; 106) Create inlet and outlet temperature measurement sampling nodes and specific process operation nodes, and associate them with the furnace number node using the HAS_OPERATION relationship; 107) Create ladle number node, alloy node, auxiliary material node, and associate them with the furnace number node using USES relationship; 108) Create other process parameter nodes and associate them with the furnace number node using ADOPT relationship.

5. A method for constructing a LF refinement prediction model based on a knowledge graph according to any one of claims 1 to 3, characterized in that: The acquisition of the path length, node association degree and node attribute similarity on the path in step 3) includes: 301) Obtain the number of connections L between the node relationships of different furnace numbers, that is, obtain the path length; 302) Use Jaccard similarity method to obtain the node correlation degree F=|S of different furnace numbers A ∩S B | / |S A ∪S B |, where S A and S B are the node sets associated with two furnace number nodes A and B respectively; |S A ∩S B | represents the number of elements in the intersection of two associated node sets, that is, the number of nodes that A and B are associated with; |S A ∪S B | represents the number of elements in the union of two associated node sets, that is, the total number of non-repeated nodes associated with A and B; 303) The formula A=1 / (1+D) is used to obtain the similarity of node attributes on the path, where D is the Euclidean distance of the numerical attribute vectors shared by the two nodes.

6. A method for constructing a LF ladle refining prediction model based on a knowledge graph as claimed in claim 5, characterized in that: The modeling in step 4) specifically includes: The steel output weight, empty ladle time, LF power consumption, refining time, power-on time, station entry temperature, alloy addition amount and auxiliary material addition amount are used as model input items, and the station exit temperature is used as the output item to construct a local training model to obtain a prediction model for the refining endpoint temperature, thereby realizing the prediction of the molten steel temperature at the refining endpoint.

7. The method for constructing a LF ladle refining prediction model based on a knowledge graph according to claim 5, characterized in that: The modeling in step 4) specifically includes: The steel output weight, refining time, power-on time, soft blowing time, incoming sulfur content, lime addition, refining slag addition, and fluorite addition are used as model input items, and the endpoint sulfur content is used as the output item to construct a local training model to obtain a prediction model for the sulfur content at the refining endpoint, thereby realizing the prediction of the sulfur content in molten steel at the refining endpoint.

Images