Converter steelmaking endpoint prediction method and device and storage medium

Abstract

The application discloses a converter steelmaking endpoint prediction method and device and a storage medium. The method comprises the following steps: performing multi-scale importance modeling and graph attention mechanism modeling on the multi-dimensional process characteristics of the current heat, generating importance weights and structure perception weights of each dimension process characteristic; multiplying the importance weights and the structure perception weights after fusion with the corresponding dimension process characteristics to generate a feature regulation vector of the current heat; performing a multi-scale convolution operation on the feature regulation vector to extract local mode features under different receptive fields, and outputting a spatial feature vector after enhancement by a self-attention mechanism; using multiple groups of Transformer encoders to process heat feature sequences with different historical window lengths respectively to extract cross-heat time sequence feature vectors; performing dimension alignment on the spatial feature vector and the time sequence feature vector, performing weighted fusion of each dimension by a learnable fusion weight, and outputting a steelmaking endpoint state prediction value of the current heat by a regression prediction network.

Description

Technical Field

[0001] This application relates to the field of data processing technology, and in particular to a method, apparatus and storage medium for predicting the end point of converter steelmaking. Background Technology

[0002] Converter steelmaking is a typical integrated smelting and refining steelmaking process, such as... Figure 1 As shown, the process can generally be divided into two main stages: First, in the blowing stage, scrap steel and molten iron are added to the converter. Oxygen blowing triggers a vigorous oxidation reaction to remove impurities and raise the temperature. Simultaneously, auxiliary materials are added to absorb some heat, forming a slag system conducive to dephosphorization and desulfurization reactions. As the reaction progresses, the steelmaking process gradually approaches the blowing endpoint. At this point, the temperature of the molten steel must meet the blowing requirements to ensure smooth transitions to subsequent operations. Second, in the alloying and tapping stage, alloys are added based on the endpoint state to ensure the steel composition meets standard requirements. In this process, the endpoint state is both a direct result of the upstream oxygen blowing and auxiliary material control and a key basis for formulating the downstream alloy addition strategy, forming a pivotal link throughout the entire process. If the endpoint state deviates from expectations, it will not only lead to waste of upstream raw materials but may also cause drastic adjustments to the alloy addition amount, resulting in a chain reaction of problems such as composition fluctuations and increased costs. Therefore, constructing a high-precision endpoint state prediction model not only helps to anticipate the smelting trend and improve endpoint accuracy but is also a fundamental prerequisite and key bridge for achieving synergistic optimization of the "oxygen blowing—auxiliary materials—alloy" three-stage process.

[0003] From a process control perspective, the core task of predicting the endpoint state in converter steelmaking is to accurately forecast the temperature and carbon concentration of molten steel, while also considering the composition levels of key elements such as manganese, phosphorus, and sulfur. Temperature and carbon concentration together determine the heat balance and reactivity of the molten steel, which are core indicators for ensuring endpoint control accuracy and smelting stability. Phosphorus and sulfur are typical harmful impurities that need to be effectively removed before the endpoint to reduce the burden on subsequent refining and improve the quality of the finished product. Manganese, as a typical alloying element, has its endpoint content affecting alloy feeding schemes and costs.

[0004] To achieve the above objectives, a large number of studies have focused on the construction of endpoint prediction methods. For temperature and carbon concentration prediction, early methods mostly use neural network structure modeling. For example, some scholars have proposed a BP neural network framework combined with an improved particle swarm algorithm to enhance the fitting ability of complex variable coupling. Some scholars have further introduced a topological optimization mechanism to improve the depth and generalization ability of feature expression. With the development of ensemble learning technology, some scholars have used feature correlation analysis to select key variables and constructed a prediction model based on the extreme gradient boosting algorithm, which has improved the explainability while ensuring the nonlinear expression ability. For the endpoint prediction problem of elements such as manganese, phosphorus, and sulfur, various methods have been proposed. For example, some scholars have used particle swarm optimization to improve the training efficiency and generalization ability of BP networks; some scholars have used firefly algorithm and extreme learning machine to realize feature sensitivity modeling; some scholars have constructed a fusion structure by integrating random forests, adaptive boosting algorithms, and stacked ensembles, effectively improving the stability and precision of phosphorus content prediction. In summary, although existing research is constantly improving in algorithm selection and feature processing, the current endpoint prediction method still has two key bottlenecks: on the one hand, the variables in the converter steelmaking process have strong coupling, and the dependence of different targets on variables has significant differences. Most existing methods use a static approach, only giving variables fixed weights, making it difficult to dynamically model the importance distribution of features, leading to insufficient feature expression quality and difficulty in suppressing redundant interference; on the other hand, there are time series correlations such as thermal residual and process rhythm continuation between heats, but most models still treat heats as independent samples and only model based on the current heat, failing to effectively use historical evolution information, thus limiting the stability and generalization ability of the model.

[0005] In view of the technical problems in the prior art that the traditional converter steelmaking endpoint state prediction method is difficult to effectively cope with high-dimensional variable coupling relationship, different prediction target dependence structure heterogeneity, and cross-heat evolution information being forgotten, an effective solution has not yet been proposed. SUMMARY

[0006] Embodiments of the present disclosure provide a converter steelmaking endpoint prediction method, device and storage medium to at least solve the technical problems in the prior art that the traditional converter steelmaking endpoint state prediction method is difficult to effectively cope with high-dimensional variable coupling relationship, different prediction target dependence structure heterogeneity, and cross-heat evolution information being forgotten.

[0007] According to an aspect of an embodiment of the present disclosure, a converter steelmaking endpoint prediction method is provided, including: performing multi-scale importance modeling on multi-dimensional process features of a current heat, to generate importance weights of each dimension of the process features; modeling structural coupling relationships between the multi-dimensional process features through a graph attention mechanism, to generate structure perception weights of each dimension of the process features; fusing the importance weights and the structure perception weights, to obtain feature weights of each dimension of the process features; multiplying each dimension of the process features with the corresponding feature weight, to generate a feature regulation vector of the current heat; performing multi-scale convolution operation on the feature regulation vector of the current heat, to extract local pattern features under different receptive fields, and output a spatial feature vector of the current heat after enhancement through a self-attention mechanism; using multiple groups of Transformer encoders to respectively process heat feature sequences of different historical window lengths, to extract time sequence feature vectors across heats; wherein the heat feature sequence is composed of the feature regulation vectors corresponding to the current heat and the previous k heats; and performing dimension alignment on the spatial feature vector and the time sequence feature vector, performing weighted fusion dimension by dimension through a learnable fusion weight, and outputting a steelmaking endpoint state prediction value of the current heat through a regression prediction network.

[0008] According to another aspect of an embodiment of the present disclosure, a storage medium is also provided, which includes a stored program, wherein the program is executed by a processor when running.

[0009] According to another aspect of an embodiment of the present disclosure, a converter steelmaking endpoint prediction device is also provided, including: a feature soft selection module for target difference, configured to perform multi-scale importance modeling on multi-dimensional process features of a current heat, to generate importance weights of each dimension of the process features; model structural coupling relationships between the multi-dimensional process features through a graph attention mechanism, to generate structure perception weights of each dimension of the process features; fuse the importance weights and the structure perception weights, to obtain feature weights of each dimension of the process features; multiply each dimension of the process features with the corresponding feature weight, to generate a feature regulation vector of the current heat; a multi-scale spatial structure perception module, configured to perform multi-scale convolution operation on the feature regulation vector of the current heat, to extract local pattern features under different receptive fields, and output a spatial feature vector of the current heat after enhancement through a self-attention mechanism; a multi-granularity time sequence dependence modeling module, configured to use multiple groups of Transformer encoders to respectively process heat feature sequences of different historical window lengths, to extract time sequence feature vectors across heats; wherein the heat feature sequence is composed of the feature regulation vectors corresponding to the current heat and the previous k heats; and a space-time feature adaptive fusion prediction module, configured to perform dimension alignment on the spatial feature vector and the time sequence feature vector, perform weighted fusion dimension by dimension through a learnable fusion weight, and output a steelmaking endpoint state prediction value of the current heat through a regression prediction network.

[0010] According to another aspect of the embodiments of the present disclosure, a converter steelmaking endpoint prediction device is also provided, comprising: a processor; and a memory connected with the processor, for providing the processor with instructions to process the following processing steps: performing multi-scale importance modeling on multi-dimensional process characteristics of a current heat, to generate importance weights of each dimension of process characteristics; modeling structural coupling relationships between the multi-dimensional process characteristics through a graph attention mechanism, to generate structure-aware weights of each dimension of process characteristics; fusing the importance weights and the structure-aware weights to obtain feature weights of each dimension of process characteristics; multiplying each dimension of process characteristics with the corresponding feature weight to generate a feature regulation vector of the current heat; performing multi-scale convolution operation on the feature regulation vector of the current heat to extract local pattern features under different receptive fields, and outputting a spatial feature vector of the current heat after enhancement by a self-attention mechanism; using multiple groups of Transformer encoders to process heat feature sequences of different historical window lengths respectively, to extract time sequence feature vectors across heats; wherein the heat feature sequence is composed of the feature regulation vectors corresponding to the current heat and the previous k heats; and performing dimension alignment on the spatial feature vector and the time sequence feature vector, performing dimension-by-dimension weighted fusion through learnable fusion weights, and outputting a steelmaking endpoint state prediction value of the current heat through a regression prediction network.

[0011] The application first performs multi-scale gating weight generation and graph structure attention coupling analysis on the multi-dimensional steelmaking process characteristics of the current heat, generates a feature regulation vector to realize high-dimensional variable decoupling and target difference perception; then extracts local process combination patterns through multi-core parallel convolution, and constructs a spatial feature vector after enhancement by channel-independent self-attention; at the same time, a multi-granularity Transformer encoder is used to process the historical heat sequence composed of the feature regulation vectors of the historical heats, and extract a time sequence feature vector for representing process evolution across heats. Finally, the spatio-temporal features are fused by dynamic gating weighting, and the endpoint temperature and multi-element content prediction values are output simultaneously by a regression network. Among them, the multi-scale gating network distinguishes the key variable dependence of different prediction targets, and combines the graph attention mechanism to explicitly model the physical correlation between process variables, realizes the mining of key feature response structures for different output tasks, dynamically adjusts the importance distribution of each variable, improves the key variable identification ability under multi-objective, and solves the high-dimensional variable coupling and target-dependent heterogeneity problem; the convolutional neural network is used to extract the structural pattern of the process variables of the current heat, representing the static dependence of the "spatial dimension", at the same time, the Transformer is introduced to model the historical evolution process across heats, capturing the dynamic association of the "time dimension", solving the problem of forgetting the cross-heat evolution information; through the adaptive fusion mechanism, the spatio-temporal feature contribution weight is adjusted according to the dynamic working condition, realizing the deep fusion of multi-source information and the unified modeling of the endpoint state, supporting multi-objective collaborative optimization. Thus, the dynamic decoupling of strongly coupled process variables, the deep mining of cross-heat evolution information, and the collaborative and accurate prediction of endpoint temperature and multi-element content are realized. Further, the technical problems in the prior art that the traditional converter steelmaking endpoint state prediction method cannot effectively cope with the high-dimensional variable coupling relationship, the heterogeneous dependence structure of different prediction targets, and the forgetting of cross-heat evolution information are solved. BRIEF DESCRIPTION OF DRAWINGS

[0012] The drawings described herein are used to provide further understanding of the present disclosure, and form a part of the present application. The illustrative embodiments of the present disclosure and their descriptions serve to explain the present disclosure, and do not constitute an improper limitation on the present disclosure. In the drawings:

[0013] Figure 1 is a full process flow diagram of the existing converter steelmaking process;

[0014] Figure 2 is a hardware structure block diagram of a computing device for implementing the method according to Embodiment 1 of the present disclosure;

[0015] Figure 3 is a flowchart of the converter steelmaking endpoint prediction method according to Embodiment 1 of the present application;

[0016] Figure 4 is a framework schematic diagram of the converter steelmaking endpoint prediction method according to Embodiment 1 of the present application;

[0017] Figure 5 is a framework schematic diagram of a target difference oriented feature soft selection module according to the embodiment 1 of the present application;

[0018] Figure 6 is a framework schematic diagram of a multi-scale spatial structure perception module according to the embodiment 1 of the present application;

[0019] Figure 7 is a framework schematic diagram of a multi-granularity timing dependency modeling module according to the embodiment 1 of the present application;

[0020] Figure 8 is a method error distribution curve diagram of an endpoint temperature prediction task according to the embodiment 1 of the present application;

[0021] Figure 9 is a method error distribution curve diagram of an endpoint carbon element content prediction task according to the embodiment 1 of the present application;

[0022] Figure 10 is a hit type (carbon temperature) proportion comparison diagram of different methods according to the embodiment 1 of the present application;

[0023] Figure 11 is a method error distribution curve diagram of an endpoint phosphorus element content prediction task according to the embodiment 1 of the present application;

[0024] Figure 12 is a method error distribution curve diagram of an endpoint sulfur element content prediction task according to the embodiment 1 of the present application;

[0025] Figure 13 is a hit type (phosphorus sulfur) proportion comparison diagram of different methods according to the embodiment 1 of the present application;

[0026] Figure 14 is a method error distribution curve diagram of an endpoint manganese element content prediction task according to the embodiment 1 of the present application;

[0027] Figure 15 is a schematic diagram of a converter steelmaking endpoint prediction device according to the embodiment 2 of the present application;

[0028] Figure 16 is a schematic diagram of a converter steelmaking endpoint prediction device according to the embodiment 3 of the present application. DETAILED DESCRIPTION

[0029] In order to make the technical solution of the present disclosure better understood by the person skilled in the art, the technical solutions in the embodiments of the present disclosure will be clearly and completely described below in combination with the drawings in the embodiments of the present disclosure. Obviously, the described embodiments are only a part of the embodiments of the present disclosure, not all the embodiments. Based on the embodiments in the present disclosure, all other embodiments obtained by the person skilled in the art without creative labor should fall within the scope of protection of the present disclosure.

[0030] It should be noted that the terms "first", "second" and the like in the description and claims of the present disclosure and the above-described drawings are used to distinguish similar objects, and do not necessarily indicate a specific order or a chronological sequence. It should be understood that the data thus used can be interchanged under appropriate circumstances, so that the embodiments of the present disclosure described herein can be implemented in an order other than those illustrated or described herein. In addition, the terms "include" and "have" and any variations thereof are intended to cover non-exclusive inclusion, for example, a process, method, system, product or device including a series of steps or units does not have to be limited to those steps or units clearly listed, but can include other steps or units not clearly listed or inherent to these processes, methods, products or devices.

[0031] Embodiment 1

[0032] According to the present embodiment, a method embodiment of a converter steelmaking endpoint prediction method is provided. It should be noted that the steps shown in the flowchart of the drawings can be executed in a computer system such as a set of computer executable instructions, and although a logical sequence is shown in the flowchart, in some cases, the steps shown or described herein can be executed in an order different from that herein.

[0033] The method embodiment provided by the present embodiment can be executed in a server or similar computing device. Figure 2 A hardware structure block diagram of a computing device for implementing a converter steelmaking endpoint prediction method is shown. As Figure 2 shown, the computing device can include one or more processors (the processor can include but not limited to a microprocessor MCU or a programmable logic device FPGA processing device), a memory for storing data, a transmission device for communication function, and an input / output interface. The memory, transmission device and input / output interface are connected with the processor through a bus. In addition, it can also include a display, a keyboard and a cursor control device connected with the input / output interface. Those skilled in the art can understand that Figure 2 The structure shown is only schematic, which does not limit the structure of the above-mentioned electronic device. For example, the computing device can include more or less components than Figure 2 shown in the figure, or have a structure different from Figure 2different configurations are shown.

[0034] It should be noted that the one or more processors and / or other data processing circuitry described above can be referred to herein generally as "data processing circuitry." The data processing circuitry can be embodied in whole or in part as software, hardware, firmware, or any combination thereof. Furthermore, the data processing circuitry can be a single standalone processing module, or it can be incorporated in whole or in part within any one of other elements of a computing device. As referred to in the embodiments of the present disclosure, the data processing circuitry functions as a processor to control, for example, selection of a variable resistance terminal path connected to an interface.

[0035] The memory can be used to store software programs and modules of application software, such as program instructions / data storage means corresponding to the converter steelmaking endpoint prediction method in the embodiments of the present disclosure. The processor executes various functional applications and data processing by running the software programs and modules stored in the memory, i.e., implements the converter steelmaking endpoint prediction method of the application program described above. The memory can include a high-speed random access memory, and can also include a non-volatile memory, such as one or more magnetic storage devices, flash memories, or other non-volatile solid-state memories. In some examples, the memory can further include a memory disposed remotely with respect to the processor, which can be connected to the computing device through a network. Examples of the network include, but are not limited to, the Internet, an intranet, a local area network, a mobile communication network, and combinations thereof.

[0036] The transmission device is used to receive or send data via a network. Specific examples of the network can include a wireless network provided by a communication provider of the computing device. In one example, the transmission device includes a network interface controller (NIC), which can be connected to other network devices through a base station so as to be able to communicate with the Internet. In one example, the transmission device can be a radio frequency (RF) module, which is used to communicate with the Internet in a wireless manner.

[0037] The display can be, for example, a touch screen type liquid crystal display (LCD), which can enable a user to interact with the user interface of the computing device.

[0038] It should be noted that in some optional embodiments, the above Figure 2 The computing device shown can include hardware elements (including circuitry), software elements (including computer code stored on a computer readable medium), or a combination of both hardware and software elements. It should be noted that in some embodiments, the functions described herein can be implemented as part of an operating system or a specific application, component, method, or process Figure 2 is merely one example of a particular implementation and is intended to provide an example of the types of components that can be present in the computing device described above.

[0039] In the above operating environment, according to a first aspect of the present embodiment, a converter steelmaking endpoint prediction method is provided. Figure 3 A flowchart of the method is shown, Figure 4 An architectural diagram of the method is shown, referring to Figure 3 and Figure 4 As shown, the method comprises:

[0040] S302: Multi-scale importance modeling is performed on the multi-dimensional process characteristics of the current heat, to generate importance weights of each dimension of process characteristics; a graph attention mechanism is used to model the structural coupling relationship between the multi-dimensional process characteristics, to generate structure-aware weights of each dimension of process characteristics; the importance weights and the structure-aware weights are fused to obtain feature weights of each dimension of process characteristics; each dimension of process characteristics is multiplied by the corresponding feature weight to generate a feature regulation vector of the current heat;

[0041] Optionally, the operation of performing multi-scale importance modeling on the multi-dimensional process characteristics of the current heat to generate importance weights of each dimension of process characteristics comprises: dividing the multi-dimensional process characteristics of the current heat into local-scale characteristics, global-scale characteristics and statistical-scale characteristics; wherein the local-scale characteristics are variables dynamically regulated in the blowing process, the global-scale characteristics are static variables reflecting the burden structure, and the statistical-scale characteristics are statistical variables representing the process evolution trend; and generating the importance weights of the local-scale characteristics through a first type of gated subnetwork; generating the importance weights of the global-scale characteristics through a second type of gated subnetwork; and generating the importance weights of the statistical-scale characteristics through a third type of gated subnetwork; wherein the first type of gated subnetwork adopts a single-layer nonlinear transformation structure, the second type of gated subnetwork adopts a multi-layer perception structure, and the third type of gated subnetwork adopts a Softplus activation function.

[0042] Optionally, the operation of modeling the structural coupling relationship between the multi-dimensional process characteristics through a graph attention mechanism to generate structure-aware weights of each dimension of process characteristics comprises: taking each dimension of characteristics in the multi-dimensional process characteristics of the current heat as a graph node, establishing a connection edge between any two nodes to form a fully connected graph structure; and inputting the fully connected graph structure into a graph attention network to generate the structure-aware weights of each dimension of process characteristics.

[0043] Optionally, the operation of fusing the importance weights and the structure-aware weights to obtain the feature weights of each dimension of process characteristics comprises: for each dimension of process characteristics, dynamically fusing the importance weights and the structure-aware weights through an exclusive learnable fusion weight to obtain the corresponding feature weight.

[0044] Specifically, the present application is found through research and analysis that the data characteristics in the industrial steelmaking process have multiple scales of representation, including both local process quantities that change with the heat, and global variables that reflect process settings, as well as statistical characteristics extracted from historical processes. Directly using the same processing method for all variables can easily ignore the differences in their action granularity and modeling needs, leading to unstable modeling results and poor interpretability. Based on this, the present application divides the multi-dimensional process characteristics of the current heat into three types of subspaces, namely: local scale characteristics, global scale characteristics, and statistical scale characteristics.

[0045] Among them, the local scale characteristics include process variables such as oxygen consumption and charging weight of the current heat, which are mainly derived from the dynamic adjustment behavior in the middle of steelmaking, and are usually adjusted frequently during the process execution, reflecting the operation response characteristics in a short time. Such variables require stronger local perception and dynamic adaptation capabilities for the model.

[0046] The global scale characteristics include features such as hot metal weight, scrap ratio, and initial element content in hot metal, which remain stable over a longer time range, reflecting background information such as furnace charge structure and original working conditions, and constitute a static description of the global state.

[0047] The statistical scale characteristics include statistical variables such as the mean of temperature, carbon / manganese / phosphorus content, etc. in the current heat, which are usually calculated from data within multiple time periods of the heat, and reflect the potential trends and control rhythm in the process evolution. Compared with variables directly involved in operation control, such characteristics are more gentle and global.

[0048] Then, as shown in Figure 5 , for each type of subspace, an independent gating subnetwork is designed. Specifically, to adapt to the semantic differences of the three types of characteristic subspaces, the present application designs gating subnetworks with different structures to capture fast response, global stability influence, and time series trend characteristics respectively. The calculation methods of the three gating weights are shown in formulas (1), (2), and (3).

[0049] (1)

[0050] (2)

[0051] (3)

[0052] Among them, , , represent the local scale, global scale, and statistical scale, , , respectively, represent the gating weights of the input features of the local, global, and statistical subspaces,

[0053]

[0054] ​​​​​​​​​​​​​​​​​​​​​​​​​In addition, to avoid the training bias caused by the uneven intensity of the gate, the three types of sub-networks are respectively introduced with independent normalization and channel-level amplitude constraint mechanism to ensure their relatively balanced and stable regulation ability in the overall feature space. During the training process, all the weights of the gate sub-networks are automatically optimized by the loss function back propagation, without external label or prior rule intervention, which has strong adaptability and end-to-end training characteristics.

[0055] Further, the present application also finds that there is often a structural coupling relationship between process variables in the steelmaking process. For example, there may be a nonlinear linkage relationship between the hot metal temperature and the carbon content, and a small change in the structure of the raw materials may cause a phased fluctuation in the concentration of elements such as phosphorus and sulfur. The relationship between these variables is not always explicitly coded by position or order, but is implicit in the industrial control logic and physical mechanism. To model the implicit coupling relationship between the above-mentioned characteristics, the present application introduces a graph attention mechanism, which regards the input variables as nodes in the graph, and explicitly constructs the information channel between variables through a learnable attention mechanism.

[0056] Specifically, a fully connected graph containing nodes is constructed, and the current input vector (i.e., multi-dimensional process characteristics) is Each dimension of the process characteristic is regarded as a node in the graph , and the edge is between any two nodes. Since the correlation between variables is usually not fully understood before modeling, the present application constructs a complete graph (fully connected graph), that is, there is a connection edge between any two variable nodes. This graph construction method ensures the maximum coverage of the structure modeling, providing an expression basis for automatically learning the importance relationship between variables through the attention mechanism. To evaluate the semantic dependence strength between variables, the present application adopts a learnable attention mechanism, which is the edge weight between any two nodes and in the graph, as shown in equations (4) and (5).

[0057] (4)

[0058] (5)

[0059] wherein and represent the original features of nodes and , is an input linear transformation matrix, is an attention weight vector, and the symbol represents a splicing operation. This mechanism essentially calculates the node The relevance of its adjacent nodes , is a nonlinear activation function, that is, the edge weight (attention score) between the two, represents the exponential function, represents the number of nodes, in order to make the sum of the adjacent edge weights of each node equal to 1, so as to facilitate the interpretable weighted aggregation, the softmax operation is used for normalization. After the edge weight coefficient is calculated, the structure perception weight of the node can be obtained by weighted aggregation of its neighbor node information, as shown in equation (6).

[0060] (6)

[0061] wherein, is the structure perception weight after introducing the structure information, is a Sigmoid activation function, which implicitly encodes the importance of the variable in the global structure and the relevant dependent features. Compared with the method of generating gate weight based on single variable response value, the method allows the model to automatically focus on other variables with high semantic coupling degree with the current variable, and generates a gate vector with better global perception ability. At the same time, the model is allowed to learn high-order interaction patterns between variables, breaking through the limitations of local perception and linear assumption, so that the variable selection process has structure recognition ability.

[0062] And, in order to further improve the expression efficiency and robustness of the multi-channel gating mechanism, the application introduces a dynamic weighted fusion strategy, which adaptively integrates the gate weight of one of the local scale gating, global scale gating, statistical scale gating and graph structure gating, so as to form a final gating result with better discriminability and context perception ability. Specifically, for any feature dimension , the feature weight comes from two sources: the importance weight output by the scale channel to which it belongs, and the structure perception weight output by the graph structure channel, a learnable fusion weight vector is introduced for each feature, and after softmax normalization, the weighted contribution for controlling the two gates is obtained, as shown in equation (7).

[0063] (7)

[0064] wherein, represents the feature weight obtained by fusing the importance weight and the structure perception weight, that is, the feature weight of the i-th process feature;

[0065] The fusion process is stable and controllable by Softmax regularization. The dynamic fusion mechanism not only retains the perception advantages of each gating network for specific semantic features, but also automatically allocates attention focus according to the feature dependency structure of different working conditions and target tasks, realizing the deep integration of four types of semantics: global, local, trend and structure. The fused feature weight is multiplied by the process feature of the corresponding dimension to obtain the final regulated feature vector, i.e., the feature regulation vector of the current heat.

[0066] In the high-dimensional input space, some features are retained in the model, but their real contribution is very low, even redundant or pseudo-correlated. If not controlled, it is easy to cause model overfitting, training divergence or poor result interpretability. Therefore, the L1 sparse regularization term is introduced into the loss function to constrain the final fused gating weight , as shown in formula (8).

[0067] (8)

[0068] wherein, is the final total loss, is the regression loss of the main task, is a hyperparameter that controls the influence of the regularization term on the total loss. represents L1 regularization. Through the regularization term, the model is subjected to "sparse pressure" during training, tending to compress the gating weight of some invalid variables to close to 0, thereby realizing "soft selection" of feature dimensions. Unlike hard filtering, this mechanism allows weak features to be retained at the beginning and gradually suppressed at the end of training, which helps to stabilize the training convergence and information exploration. In addition, this mechanism can also improve the feature interpretability and engineering readability of variable selection, providing important decision basis and variable reduction support for subsequent multi-objective optimization.

[0069] S304: Perform a multi-scale convolution operation on the feature regulation vector of the current heat to extract local pattern features under different receptive fields, and output a spatial feature vector of the current heat after enhancement by a self-attention mechanism;

[0070] Optionally, the operation of performing a multi-scale convolution operation on the feature regulation vector of the current heat to extract local pattern features under different receptive fields, and outputting a spatial feature vector of the current heat after enhancement by a self-attention mechanism, includes: expanding the feature regulation vector of the current heat into a one-dimensional feature map; performing convolution operations on the one-dimensional feature map in parallel to obtain n groups of feature maps under different receptive fields; concatenating the n groups of feature maps along the channel dimension to obtain a fused feature map; and inputting the fused feature map into a channel-independent self-attention network, and flattening the attention-enhanced fused feature map into a one-dimensional vector as the spatial feature vector of the current heat.​

[0071] In the embodiment of the present application, it is found through research that in the converter steelmaking process, although the static process variables of each heat are represented in one dimension, there are often local combination relationships and characteristic substructures inside, such as the proportional relationship between the total amount of charging and the amount of oxygen supply, the coupled influence of scrap content and initial temperature, etc. These local patterns have the characteristics of multi-scale and multi-receptive field expression requirements, and the traditional one-layer convolutional neural network often has difficulty in capturing key patterns at different scales at the same time due to the fixed receptive field, which limits the full expression of process information by the model. To improve the spatial perception ability of the current heat input features, the present application designs a multi-scale convolutional structure, as shown in Figure 6 The structure forms a feature pyramid by introducing parallel convolution paths with different kernel widths, and finally obtains a feature map, enhancing the model's ability to capture spatial relationships between features.

[0072] Specifically, let the feature regulation vector of the current heat be Firstly, it is expanded into a one-dimensional feature map, and then n groups of convolution operations with different receptive fields are performed in parallel, as shown in formula (9).

[0073] (9)

[0074] Wherein, represents the feature map obtained by the th convolution, represents the convolution operation of the th path, and after each convolution, m layers of convolution with the same kernel width are connected for deep feature extraction, and the same number of output channels is used, finally obtaining n groups of feature maps , respectively representing the local pattern expression under different receptive fields. Subsequently, the n feature maps are spliced in the channel dimension. In order to further enhance the global perception ability of the spliced features between different positions, a lightweight self-attention mechanism is introduced to globally model the spliced feature map. This module takes the heat feature map after channel fusion as input, constructs a correlation matrix between positions, generates dynamic attention weights through Query-Key dot product, and then aggregates the Value feature map through weighted aggregation, realizing information interaction and fusion between different positions. This attention module maintains the independence between channels and does not introduce additional channel coupling. Finally, the multi-scale feature map enhanced by attention is flattened in the channel and position dimensions to construct a fixed-length heat feature vector as the final representation output of the spatial main branch, as shown in formula (10).

[0075] (10)

[0076] Wherein, is the spliced feature map after convolution, is the flattening operation, For attention calculation, a multi-scale spatial feature vector is finally obtained The vector serves as a spatial representation of the current heat characteristics and will be used for subsequent feature fusion and prediction tasks. This branch focuses on extracting key static structures in the current heat process information, which is an important channel for spatial perception in the entire endpoint prediction model.

[0077] S306: A plurality of Transformer encoders are used to process the heat feature sequences of different historical window lengths respectively to extract cross-heat time sequence feature vectors; wherein the heat feature sequence is composed of the feature regulation vectors corresponding to the current heat and the previous k heats;

[0078] Optionally, the operation of using a plurality of Transformer encoders to process the heat feature sequences of different historical window lengths respectively to extract cross-heat time sequence feature vectors includes: setting a plurality of different window lengths, and for each window length, constructing a corresponding heat feature sequence; wherein the heat feature sequence includes the feature regulation vectors corresponding to the current heat and the previous k heats, and k is different in different heat feature sequences; inputting each heat feature sequence into an independent Transformer encoder for processing to output a feature representation corresponding to the window length; and concatenating the feature representations corresponding to the plurality of different window lengths in the feature dimension to obtain a cross-heat time sequence feature vector.

[0079] In the embodiments of the present application, it is found through research that there is often significant temporal correlation between different heats in the converter steelmaking process. For example, factors such as the state of the residual heat of the furnace lining, the residual slag of the previous heat, and the thermal load of the equipment often have an impact on the oxygen utilization efficiency, the smelting reaction rate, and the temperature loss of the current heat. If the heat is regarded as an independent sample, it is difficult to capture such cross-heat information implied in the time evolution process, affecting the model's ability to perceive changes in working conditions and the stability of the prediction. In actual production, the temporal influence of the smelting process cannot be completely covered by a fixed window length. Short-term information has a quick impact on the direct heat balance and reaction rate of the current heat, while medium- and long-term factors (such as equipment wear, system operation period, etc.) may gradually accumulate their influence over a longer range of heats. Therefore, this application introduces a variety of granular sequence lengths to model short-term, medium-term, and long-term temporal patterns, thereby enhancing the model's sensitivity and expressiveness to multi-scale temporal changes.

[0080] To model the multi-granularity historical heat sequence, a Transformer encoder is introduced in the temporal branch to capture time evolution rules of different scales with its strong context perception ability and long-term dependency modeling ability. For example Figure 7As shown, it mainly consists of feature embedding, position encoding, multi-head attention mechanism and feedforward network, wherein the embedding layer is used for high-dimensional representation of the input features of each heat; the position encoding introduces sequence order information, so that the model has time sequence perception ability; the multi-head attention mechanism can model the dependency relationship between different positions in parallel, effectively improving the modeling ability of the model for cross-step time sequence features.

[0081] Specifically, let the current the feature regulation vector of the heat be , set multiple different window lengths, and for each window length, construct the corresponding heat feature sequence. For example, but not limited to, the window lengths are 5, 7, 9, when the window length is 5, the constructed heat feature sequence consists of the feature regulation vector of the heat , the feature regulation vectors corresponding to the previous 5 heats , and , that is . Then, first embed each sequence into a unified dimensional space through linear mapping, and then respectively send it into multiple independent Transformer encoders for modeling, as shown in formula (11).

[0082] (11)

[0083] wherein, represents the historical sequence input of the heat at the granularity, represents the Transformer encoder for the granularity, is the feature output at the corresponding granularity, and finally the output vectors of multiple different granularities are spliced to construct a multi-granularity time sequence feature representation, as shown in formula (12).

[0084] (12)

[0085] wherein, represents the time sequence feature vector of the multi-granularity time sequence dependency modeling, represents the splicing operation, and the multi-granularity design further enhances the perception ability of the model for short-term, medium-term and long-term heat evolution trends, avoiding the limitation of fixed window on modeling ability.

[0086] S308: Align the spatial feature vector and the time sequence feature vector in dimension, and perform dimension-by-dimension weighted fusion through learnable fusion weights, and output the prediction value of the current heat end state through the regression prediction network.

[0087] Optionally, the spatial feature vector and the time sequence feature vector are dimensionally aligned, weighted and fused by a learnable fusion weight, and the operation of outputting a prediction value of the steelmaking endpoint state of the current heat by the regression prediction network comprises: compressing or expanding the spatial feature vector and the time sequence feature vector to the same dimension by linear projection; splicing the aligned spatial feature vector and the time sequence feature vector, and then inputting the same into a gating network to generate a learnable fusion weight vector; weighted and fused by the learnable fusion weight vector to obtain a fusion feature vector; and inputting the fusion feature vector into the regression prediction network to output the prediction value of the steelmaking endpoint state of the current heat.

[0088] In the embodiment of the present application, it is found through research that the process state of the current heat is not only determined by the structural relationship between its static variables, but also deeply affected by the evolution process of the previous heat, and the spatial structure and the time dependence have natural complementarity in the prediction task. In order to fully integrate the static process characteristics of the current heat and the dynamic evolution information in the historical sequence, the present application designs a spatio-temporal feature adaptive fusion prediction module based on the modeling of the spatial branch and the time branch, realizing the deep collaborative modeling of multi-source features. The unified integration of the features extracted in the previous two modules is an important step to realize the final prediction. Among them, the multi-scale spatial structure perception module focuses on modeling the combination relationship between the process variables in the current heat, and has strong expression ability and local pattern sensitivity; while the multi-granularity time sequence dependence modeling module focuses on the cross-time sequence dependence structure between the heats, and has good context modeling ability and historical information memory. The two have obvious complementarity in feature source and modeling perspective, the former emphasizes local structure perception, and the latter strengthens global time sequence modeling, and after fusion, it is expected to build a heat-level feature representation with better discriminability and generalization ability.

[0089] Unlike the traditional direct splicing method, the present application introduces an adaptive weighted fusion mechanism. After unifying the feature dimensions, the spatial features and the time features are weighted and integrated by learnable fusion weights, improving the discriminability and adaptability of feature fusion. Specifically, let the multi-scale spatial structure perception module output a spatial feature vector , the multi-granularity time sequence dependence modeling module output a time sequence feature vector , and after linear projection or compression operation mapping to the same dimension, the fusion feature vector is defined as follows:

[0090] (13)

[0091] wherein, represents element-wise multiplication, and the fusion weight vector , to control the contribution ratio of two types of features in each dimension. The weight is generated by a set of fusion gating networks, whose input is the spliced vector of two types of features, and the structure is as follows:

[0092] (14)

[0093] wherein, represents feature splicing, and represent weight matrix and bias term. This gating mechanism allows the model to dynamically adjust the fusion strength of spatial and temporal features according to different samples, achieving fine-grained feature regulation. The fused representation is input to the regression prediction module to further obtain the prediction result of the endpoint state through nonlinear mapping. The regression prediction module is composed of n layers of full connection, and its structure is represented as follows:

[0094] (15)

[0095] wherein, is the final prediction result, represents the weight, represents the bias term.

[0096] To verify the effectiveness and generalization of the method proposed in the present application, the present application experiment uses real data of actual production of a converter steelmaking in a domestic steel plant, and the steel grade is HRB400E. The production data includes molten iron information, scrap information, smelting process information, oxygen supply, smelting endpoint information, and other production variables. The time is from January 1, 2023 to May 1, 2024, and 10991 pieces of valid data are selected as the original data set, and part of the data is shown in Table 1. Select 20 variables meaningful to the endpoint state as the feature regulation vector of the current heat, as shown in Table 2.

[0097] Table 1 Production variable table

[0098]

[0099] Table 2 Input variable table

[0100]

[0101] To evaluate the prediction performance of the model, the mean absolute error , the root mean square error , and the accuracy rate within the error range are used as evaluation criteria, and are defined as follows:

[0102] (16)

[0103] (17)

[0104] (18)

[0105] wherein, and denote the true and predicted values of the target variable, represents the number of test set samples, and is an indicator function, the value of the indicator function is 1 if the expression in the parentheses is true, otherwise it is 0, is the absolute value of the given error range. describes the average absolute distance between the predicted and actual values; describes the average error between the predicted and true values, and is more sensitive to outliers; a direct indicator of prediction performance, the larger the indicator, the better the prediction performance.

[0106] In this experiment, the original data set is first preprocessed to construct the input structure required by the model and remove mechanism outliers. For multi-granularity time series, four kinds of historical window lengths (3 / 5 / 7 / 9 heats) are set to capture short, medium and long-range inter-heater dependencies. In the multi-scale spatial structure perception module, a parallel four-path convolution structure is used, with convolution kernel sizes of 1 / 3 / 5 / 7. Each path contains 4 convolution layers, with the number of channels increasing layer by layer. In the multi-granularity time series dependency modeling module, an 8-layer encoder layer is used, and the number of multi-head attention heads is set to 8. During training, the optimizer is selected as Adam (with a fixed learning rate of 0.001), the batch size is 16, the maximum training round is 200, the early stopping strategy is used, and the basic loss function is L1 loss.

[0107] To comprehensively evaluate the performance of the proposed endpoint state prediction model under different prediction tasks, this application divides the modeling task into three sub-task groups, focusing on temperature and carbon content, phosphorus and sulfur content, and manganese content endpoint prediction problems. Each group of experiments selects a representative classical model or mainstream deep learning method in the field as a comparison baseline, denoted as SGCT-Net, and uniformly uses the mean absolute error , root mean square error and accuracy rate within the error range as evaluation indicators. In the accuracy calculation, the tolerance range is set according to the process control requirements of each variable: temperature is set to ±15℃, carbon and manganese are set to ±0.05%, phosphorus is set to ±0.01%, and sulfur is set to ±0.005%. Within this tolerance range, the absolute value of the error between the predicted value and the true value is not more than the threshold, and the sample is considered to be accurately predicted.

[0108] To further measure the applicability and stability of the model in the multivariate collaborative control scene, the joint hit rate index is introduced in each group of experiments: the proportion of samples with simultaneous hits of temperature and carbon , phosphorus and sulfur variables is calculated to reflect the comprehensive performance of the model in local multi-objective consistency control.

[0109] To verify the effectiveness of the proposed model in the traditional key control variable-end point temperature and carbon content prediction task, the first group of comparative experiments is designed. This group of tasks is the focus of current steelmaking field research, and the prediction accuracy is directly related to the end control hit rate and the stability of the smelting quality. Three representative comparative models in this field are selected for performance comparison, which are:

[0110] FOA-GRNN: This method uses fruit fly optimization algorithm to optimize the kernel width parameter of generalized regression neural network to improve the prediction accuracy of end point temperature and carbon mass fraction in steelmaking;

[0111] LWOA-TSVR: This method combines Logistic whale optimization algorithm and ε-insensitive double support vector regression, optimizes the model hyperparameters through LWOA to improve the generalization performance, and ε-TSVR has good nonlinear modeling ability and noise resistance, which is suitable for prediction scenarios with fluctuations and abnormalities in steelmaking data;

[0112] JITL: This method dynamically selects similar data from historical samples based on the input features of the current heat at each prediction, and builds a personalized regression model based on these local samples, thereby improving the adaptability of the model in nonlinear and time-varying working conditions.

[0113] The experimental results are shown in Table 3.

[0114] Table 3 Comparison of end point temperature and carbon content prediction results

[0115]

[0116] According to the experimental results in Table 3, the SGCT-Net model proposed in this application is superior to the comparative models in the prediction task of end point temperature and carbon content. Taking the end point temperature as an example, the and of SGCT-Net are reduced by 13.3% and 14.3% respectively compared with the second-best FOA-GRNN; within the error tolerance range of ±15℃, the accuracy is 85.30%, which is improved by 9.58 percentage points compared with FOA-GRNN. For the end point carbon content, SGCT-Net also has and The best performance, respectively, 13.2% and 7.5% lower than FOA-GRNN, the accuracy rate is as high as 94.21%. In addition, in the joint hit rate index In the above, SGCT-Net reaches 80.85%, which is significantly higher than FOA-GRNN, LWOA-TSVR and JITL, indicating that the model not only can predict key variables, but also shows superior ability in multi-objective consistency control. This performance advantage is mainly due to the use of SGCT-Net to extract the local combination pattern between process variables in the current heat by using the multi-scale spatial structure perception module, and to mine the cross-heat evolution law by combining the multi-granularity time series dependence modeling module; finally, through the spatiotemporal feature adaptive fusion module, the deep fusion and differential weighting modeling of spatial and temporal information are realized, thereby improving the prediction stability and generalization ability of the model under complex and volatile working conditions.

[0117] The prediction error distribution curve is shown in Figure 8 and Figure 9 The error density distribution of SGCT-Net and three kinds of comparative models in the end point temperature and end point carbon content prediction task is shown. The horizontal coordinate in the figure is the prediction error (predicted value minus true value), and the vertical coordinate is the sample density in the error interval. Each curve represents the error distribution characteristics of the model in this task, and the black dotted line represents the position of the error of 0, and in the ideal case, more samples should be densely distributed near this line, and the more concentrated the curve, the more accurate and less volatile the model prediction.

[0118] From Figure 8 It can be seen that the prediction error distribution of SGCT-Net is most dense near the error of 0, and the curve is highly contracted in the central region, with more than 90% of the samples having an error falling within the interval [–10℃, +10℃], and few samples far from the center. This indicates that SGCT-Net has strong accuracy and error control ability in end point temperature prediction. In contrast, the error distribution of FOA-GRNN and LWOA-TSVR is wider, with the density on both sides of the tail significantly higher than that of SGCT-Net, and there are many large error samples, and JITL has a lower density in the central region, indicating that its prediction results are more volatile. Figure 9The prediction error distribution of the endpoint carbon element is shown in FIG. 6. The SGCT-Net performs best around the error of 0, and the error density is quickly concentrated in the interval of [-0.025%, +0.025%], and there are basically no abnormal samples with large deviations from the center. This distribution pattern shows that the SGCT-Net has very strong robustness and stability in modeling of the carbon content, which is a low-amplitude and high-sensitivity variable. In the comparative models, the distributions of FOA-GRNN and LWOA-TSVR are biased towards the positive error area, and there is a systematic overestimation tendency. The distribution of JITL is relatively dispersed near the center, and lacks consistency. The advantage of the SGCT-Net mainly comes from the introduction of the multi-scale spatial structure perception module, which accurately captures the local combination patterns between variables, and effectively reduces the sensitivity to process fluctuations in prediction.

[0119] From Figure 10 It can be seen from FIG. 7 that the SGCT-Net proposed in the present application is obviously leading in the "double hit" index, reaching about 80%, which is about 10-20 percentage points higher than other models, indicating that it has stronger prediction ability on the two core variables of temperature and carbon. At the same time, in the samples of "only hit one of them" or "all miss", the proportion of SGCT-Net is the smallest, indicating that the prediction results are not only accurate, but also more consistent. This performance is mainly due to the introduction of the target difference modeling and spatio-temporal feature fusion mechanism of the SGCT-Net, which constructs different feature channels for temperature and carbon, so that the model realizes more coordinated feature matching and error control between different targets, thereby effectively improving the multi-target consistent hit rate.

[0120] To further verify the adaptability and robustness of the method in modeling of trace impurity elements, the present application sets up a second group of comparative experiments, focusing on the joint prediction task of endpoint phosphorus element content and sulfur element content. Phosphorus and sulfur, as typical harmful impurities, have low content and strong volatility, are affected by multiple factors such as slag composition and temperature fluctuations, and have high modeling difficulty and low tolerance to prediction error, which is a challenging task in intelligent modeling. Therefore, the following three models are selected as comparative methods in this experiment:

[0121] PCA-BP: This method first performs principal component analysis (PCA) on high-dimensional input variables in the smelting process to reduce noise interference by extracting main information components, and then uses a back propagation neural network (BP) to build a nonlinear prediction model;

[0122] Stacking: Based on the thermodynamic and kinetic analysis of converter dephosphorization, this method selects key process parameters as input features from production data, combines random forest, adaptive boosting algorithm and multiple boosting algorithm to build a Stacking ensemble regression model to predict the converter endpoint phosphorus content;

[0123] MLI-GA-BPNN: This method integrates metallurgical mechanisms and industrial data. First, it uses local linear embedding for feature selection and dimensionality reduction based on the dephosphorization mechanism. Then, it employs a backpropagation (BP) neural network to construct a prediction model and optimizes the model structure and parameters using a genetic algorithm. Mechanistic information is incorporated into the fitness function and loss function, improving the model's prediction accuracy and physical consistency.

[0124] The experimental results are shown in Table 4.

[0125] Table 4. Comparison of Experimental Results of Predicted Phosphorus and Sulfur Content at the Endpoint

[0126]

[0127] As shown in Table 4, SGCT-Net achieved significant performance advantages in both the endpoint phosphorus and sulfur content prediction tasks. Specifically, in the endpoint phosphorus content prediction, SGCT-Net... and Compared to MLI-GA-BPNN, the accuracy was reduced by approximately 22.2% and 19.7%, respectively; within an error range of ±0.01%, the accuracy reached 89.98%, significantly higher than Stacking and PCA-BP. For the endpoint sulfur content, it improved by more than 4 percentage points compared to the second-best performing Stacking. Notably, the combined hit rate for phosphorus and sulfur was significantly higher. In the above tests, SGCT-Net achieved a score of 81.07%, significantly outperforming MLI-GA-BPNN and Stacking, fully demonstrating the model's stability and accuracy consistency in control scenarios involving trace impurity elements. This advantage mainly stems from SGCT-Net's ability to accurately capture the cross-features of weak signals and strongly coupled structures: its graph structure perception module constructs a fully connected variable graph to explicitly model the implicit coupling relationships between low-concentration elements such as phosphorus and sulfur and upstream and downstream process variables, improving the structural integrity of feature representation; simultaneously, the feature soft selection mechanism combines sparse regularization and multi-scale gating to highlight phosphorus and sulfur-sensitive variables in high-dimensional inputs, avoiding redundant interference.

[0128] Error distribution curve as shown Figure 11 and Figure 12 As shown, the error density distribution of each model in the final task of predicting phosphorus and sulfur content is illustrated.

[0129] from Figure 11It can be seen that the prediction error of SGCT-Net is most concentrated around 0, with the highest curve concentration, and the vast majority of sample errors fall within the interval of [–0.01%, +0.01%], and the tail on both sides decays rapidly. This indicates that the model has strong error control capability in predicting the end-point phosphorus content. In contrast, the error distribution of PCA-BP and MLI-GA-BPNN is wider, and the density peak deviates from 0, with a systematic overestimation trend; although the concentration of Stacking is acceptable, there are still many samples above the error of ±0.01%, indicating that its boundary prediction is unstable. Figure 12 In the end-point sulfur element prediction, SGCT-Net also has the most compact error distribution, with the vast majority of samples distributed within the interval of [–0.005%, +0.005%], and the proportion of samples with errors close to 0 is significantly higher than other models. This distribution characteristic shows that SGCT-Net has higher robustness and consistency when dealing with low-concentration and high-sensitivity variables. In comparison, the curves of PCA-BP and MLI-GA-BPNN shift to the right, with a significant overestimation problem; although Stacking is slightly better in error shrinkage, its density in the central region is still lower than that of SGCT-Net. The performance of SGCT-Net in this task largely depends on its multi-granularity time modeling structure, which effectively improves the model's perception ability of complex dynamic processes by capturing small trend changes in the furnace evolution.

[0130] From Figure 13 It can be seen that SGCT-Net leads in the "phosphorus and sulfur double hit" ratio, with more than 80%, which is 10-25 percentage points higher than other models, indicating that it has the strongest prediction ability for trace impurity elements. At the same time, in the "all-miss" category, SGCT-Net has the lowest proportion, further verifying its stability and generalization ability in complex variable modeling. This performance advantage is mainly due to the synergistic effect of the graph structure attention mechanism and the multi-granularity time modeling module in SGCT-Net: the former can identify potential coupled variable combinations between phosphorus and sulfur, and the latter can capture the implicit evolution trend in the furnace sequence, making the model have stronger linkage perception and error control capability when predicting two low-concentration and high-sensitivity variables simultaneously.

[0131] The present application sets up the third group of experiments, and carries out evaluation for the prediction task of end-point manganese element content. As a typical alloying control element, the end-point content of manganese not only affects the subsequent alloy feeding amount, but also has a direct impact on the composition and cost control of steel. Compared with phosphorus and sulfur, the manganese content is higher but fluctuates sharply, and is influenced by multiple factors such as temperature, slag composition and reaction time, so the modeling process is still challenging. Therefore, the following three models are selected as comparison methods in this experiment:

[0132] IWOA-LSSVM: This method uses the improved whale optimization algorithm (IWOA) to optimize the LSSVM model parameters to improve the prediction accuracy of the end-point manganese content. By introducing cosine disturbance and nonlinear convergence mechanism, the fitting and generalization ability of the model under nonlinear conditions is enhanced.

[0133] NN-Static: This method is based on feedforward neural network to build a static prediction model, the input is multiple process parameters, the output is the end-point manganese content, through BP algorithm training, to realize the nonlinear modeling and prediction of manganese content.

[0134] SSA-LSTM: This method uses Spiral Search Algorithm (Spiral Search Algorithm, SSA) to globally optimize the hyperparameters of LSTM model to improve its fitting ability and prediction performance in converter smelting data.

[0135] The experimental results are shown in Table 5.

[0136] Table 5 Comparison of end-point manganese content prediction results

[0137]

[0138] From the results in Table 5, SGCT-Net also achieved overall optimal performance in the end-point manganese content prediction task. In and two error indicators, SGCT-Net is reduced by 18.9% and 17.9% respectively compared with SSA-LSTM, and compared with traditional static neural network model and IWOA-LSSVM based on optimized support vector machine is significantly superior. Within the error range of ±0.05%, the accuracy of SGCT-Net is as high as 84.41%, which is 8.46 percentage points higher than SSA-LSTM, fully demonstrating its significant advantage in high volatility alloy element modeling. Thanks to the multi-granularity time series modeling module for modeling the reaction rhythm and residual effect in the previous heats, the model shows stronger adaptability and anti-interference ability when facing manganese, a "high volatility, strong coupling, high sensitivity" alloy variable, providing reliable support for fine control in the actual steelmaking process.

[0139] The error distribution curve is shown in Figure 14 , which shows the error density distribution of each model in the end-point manganese content prediction task.

[0140] From Figure 14It can be seen that the prediction error of SGCT-Net is highly concentrated around 0, mainly distributed in the interval [–0.05%, +0.05%], with the highest density in the center, indicating that the model has stronger error compression and abnormal control capability in the prediction of end-point manganese content. In contrast, the error distribution of SSA-LSTM and IWOA-LSSVM is wider, especially SSA-LSTM still has a large density in the tail on both sides, showing higher sensitivity to complex working conditions; the error fluctuation of NN-Static is larger, and the density curve is not smooth, indicating that its prediction stability is weak. The reason why SGCT-Net performs better in the modeling of manganese, a high-volatility and strongly coupled variable, is mainly due to the multi-granularity time series modeling module in it, which can capture the residual information between heats.

[0141] To systematically evaluate the specific contribution of each core module in the proposed model to the end-point prediction performance, three groups of ablation experiments were designed to remove and replace each module one by one, targeting the feature selection mechanism, spatial structure modeling ability, and time series modeling ability, respectively. The effectiveness and necessity of each module were verified by changes in model performance. The specific comparison model configurations are as follows:

[0142] SGCT-Net (complete model): the end-point state prediction model proposed in this application, which includes core modules: target difference-oriented feature soft selection, multi-scale spatial structure perception, multi-granularity time series modeling, and spatio-temporal feature adaptive fusion. The model has comprehensive capabilities such as dynamic feature modeling, complex variable structure expression, and heat evolution trend perception.

[0143] SGCT-Net-noSelector (remove feature soft selection module): replace the feature soft selection module with a static feature selection method based on Pearson correlation coefficient, select the input features with the highest correlation for each prediction target, and verify the modeling advantages of the learnable feature selection mechanism in the high-dimensional coupled scenario.

[0144] SGCT-Net-noSpatial (remove spatial structure perception module): replace the multi-scale spatial structure perception branch with a single-scale shallow perceptron, weaken the model's ability to model complex variable combination patterns within the current heat, and evaluate the contribution of spatial multi-scale structure extraction to local pattern recognition and prediction performance.

[0145] SGCT-Net-noTemporal (remove time series modeling module): replace the multi-granularity time series modeling branch with a normal fully connected network, no longer use the historical heat sequence for context modeling, test the influence of cross-heat dynamic trend perception ability on prediction stability and robustness.

[0146] The ablation experiment results are shown in Table 6.

[0147] Table 6 Ablation experiment results table

[0148]

[0149] From the ablation experiment results in Table 6, it can be seen that the complete model SGCT-Net performs best in each prediction task, fully verifying the effectiveness and synergy of each core module in improving prediction accuracy and robustness. In contrast, the simplified model that removes any module has a significant performance degradation in different indicators, indicating that each sub-module plays an irreplaceable role in the overall architecture of the model.

[0150] Specifically, the SGCT-Net-noSelector model generally performs worse than the complete model in multiple prediction tasks, indicating that the feature soft selection module for target difference can effectively distinguish between task-related and irrelevant features in the high-coupling input scenario, improving the expression ability of key variables under different prediction targets. In contrast, the traditional Pearson correlation screening method lacks context modeling capability and is difficult to deal with nonlinear coupling and feature redundancy problems between variables. SGCT-Net-noSpatial has a significant increase in error in end temperature and manganese content, which are more sensitive to local process changes, reflecting the importance of the multi-scale spatial structure perception module in extracting complex variable interactions and short-term process disturbance patterns within the heat. It is irreplaceable, especially in modeling scenarios dominated by local information. The SGCT-Net-noTemporal model has the most obvious performance decline in phosphorus, sulfur, and manganese element prediction tasks with strong volatility and significant historical dependence, indicating that the multi-granularity temporal dependence modeling module plays a key role in describing cross-heat evolution trends and slow-changing laws, effectively improving the stability and trend response ability of the model.

[0151] In summary, the ablation experiment systematically verifies the necessity and synergistic gain of the above modules from the dimensions of feature selection, structure perception, and temporal modeling. Each module complements each other in modeling high-dimensional complex variables, capturing process disturbance features, and identifying temporal evolution laws, constituting the key modeling support of SGCT-Net, significantly improving the robustness and generalization ability of the model under complex working conditions, and providing a solid prediction foundation for subsequent "oxygen blowing - auxiliary material - alloy" three-stage collaborative optimization.

[0152] Therefore, in order to effectively deal with the problem that the traditional endpoint state prediction method is difficult to cope with the coupling relationship of high-dimensional variables and the evolution trend between heats, the application proposes an endpoint state prediction model SGCT-Net which fuses feature selection and space-time structure modeling. The model is composed of four key modules: through the feature soft selection module oriented to target difference, the key variable identification ability under multi-objective is improved; with the help of the multi-scale space structure perception module, the complex local variable combination pattern in the current heat is extracted; the multi-granularity time sequence dependence modeling module is used to describe the historical process evolution characteristics; and the unified expression and dynamic weighted fusion of cross-dimensional information are realized in the space-time feature adaptive fusion module. A number of experiments carried out on real production data of typical steel grades show that SGCT-Net realizes the optimal performance in the prediction tasks of temperature, carbon, phosphorus, sulfur and manganese five types of endpoint variables, the average error is reduced by 18.9% at most, the multi-objective hit rate is significantly improved, and excellent prediction accuracy, stability and engineering adaptability are shown. The model lays a solid prediction foundation and data support for the subsequent multi-objective collaborative optimization of the three-stage process of "oxygen blowing - auxiliary material - alloy".

[0153] In addition, referring to Figure 2 According to a second aspect of the present embodiment, a storage medium is provided. The storage medium includes a stored program, wherein the program is executed by a processor to perform the method of any one of the preceding aspects when the program is run.

[0154] It should be noted that, for the foregoing method embodiments, in order to simply describe, they are all expressed as a series of action combinations, but those skilled in the art should know that the present application is not limited by the action sequence described, because according to the present application, certain steps can be performed in other sequences or simultaneously. Secondly, those skilled in the art should know that the embodiments described in the specification all belong to preferred embodiments, and the actions and modules involved are not necessarily essential to the present application.

[0155] From the above description of the embodiments, those skilled in the art can clearly understand that the method according to the above embodiments can be realized by means of software and the necessary general hardware platform, of course, it can also be realized by hardware, but in many cases the former is a better embodiment. Based on such understanding, the technical solutions of the present application can be embodied in the form of a software product, which is stored in a storage medium (such as ROM / RAM, magnetic disk, optical disk) and includes a plurality of instructions for making a terminal device (which can be a mobile phone, computer, server, or network device, etc.) execute the method described in each embodiment of the present application.

[0156] Example 2

[0157] Figure 15A converter steelmaking endpoint prediction device according to the present embodiment is shown, which corresponds to the method according to embodiment 1. Referring to Figure 15 As shown, the device comprises: a target difference-oriented feature soft selection module 1510, configured to perform multi-scale importance modeling on the multi-dimensional process features of the current heat, to generate importance weights of each dimension of process features; model the structural coupling relationship between the multi-dimensional process features through a graph attention mechanism, to generate structure-aware weights of each dimension of process features; fuse the importance weights and the structure-aware weights to obtain feature weights of each dimension of process features; multiply each dimension of process features with the corresponding feature weights to generate a feature regulation vector of the current heat; a multi-scale spatial structure-aware module 1520, configured to perform multi-scale convolution operation on the feature regulation vector of the current heat, to extract local pattern features under different receptive fields, and output a spatial feature vector of the current heat after self-attention mechanism enhancement; a multi-granularity temporal dependency modeling module 1530, configured to use multiple groups of Transformer encoders to process heat feature sequences of different historical window lengths respectively, to extract cross-heat temporal feature vectors; wherein the heat feature sequence is composed of the feature regulation vectors corresponding to the current heat and the previous k heats; and a spatio-temporal feature adaptive fusion prediction module 1540, configured to perform dimension alignment on the spatial feature vector and the temporal feature vector, to perform weighted fusion of each dimension through learnable fusion weights, and to output a prediction value of the steelmaking endpoint state of the current heat through a regression prediction network.

[0158] According to the present embodiment, first, the multi-scale gating weight generation and graph structure attention coupling analysis are performed on the multi-dimensional steelmaking process characteristics of the current heat, the feature regulation vector is generated to realize high-dimensional variable decoupling and target difference perception; then the local process combination mode is extracted through multi-core parallel convolution, and the spatial feature vector is constructed after the self-attention enhancement of the channel independent; at the same time, the multi-granularity Transformer encoder is used to process the historical heat sequence composed of the feature regulation vectors of the historical heats, and the time sequence feature vector for representing the process evolution across heats is extracted. Finally, the spatio-temporal features are dynamically gated and weighted, and the endpoint temperature and multi-element content prediction values are simultaneously output through the regression network. Among them, the multi-scale gating network distinguishes the key variable dependence of different prediction targets, and the graph attention mechanism explicitly models the physical correlation between process variables, realizes the mining of key feature response structure for different output tasks, dynamically adjusts the importance distribution of each variable, improves the key variable identification ability under multi-objective, and solves the high-dimensional variable coupling and target-dependent heterogeneous problems; the convolutional neural network is used to extract the structural mode of the process variables of the current heat, representing the static dependence of the "spatial dimension", at the same time, the Transformer is introduced to model the historical evolution process across heats, capturing the dynamic association of the "time dimension", solving the cross-heat information forgetting problem; through the adaptive fusion mechanism, the spatio-temporal feature contribution weight is adjusted according to the dynamic working condition, realizing the deep fusion of multi-source information and the unified modeling of the endpoint state, supporting the multi-objective collaborative optimization. Thus, the dynamic decoupling of strongly coupled process variables, the deep mining of cross-heat evolution information, and the collaborative and accurate prediction of endpoint temperature and multi-element content are realized. Further, the technical problems in the prior art that the traditional converter steelmaking endpoint state prediction method cannot effectively cope with the high-dimensional variable coupling relationship, the heterogeneous dependence structure of different prediction targets, and the forgetting of cross-heat evolution information are solved.

[0159] Embodiment 3

[0160] Figure 16 A converter steelmaking endpoint prediction device according to the present embodiment is shown, which corresponds to the method according to embodiment 1. Referring to Figure 16As shown, the device comprises: a processor 1610; and a memory 1620 connected with the processor 1610, for providing the processor 1610 with instructions to process the following processing steps: multi-scale importance modeling is performed on multi-dimensional process characteristics of the current heat, to generate importance weights of each dimension of process characteristics; a structure coupling relationship between the multi-dimensional process characteristics is modeled through a graph attention mechanism, to generate structure perception weights of each dimension of process characteristics; the importance weights and the structure perception weights are fused to obtain feature weights of each dimension of process characteristics; each dimension of process characteristics is multiplied by the corresponding feature weight to generate a feature regulation vector of the current heat; a multi-scale convolution operation is performed on the feature regulation vector of the current heat to extract local mode features under different receptive fields, and after being enhanced by a self-attention mechanism, a spatial feature vector of the current heat is output; a plurality of groups of Transformer encoders are used to process heat feature sequences of different historical window lengths respectively to extract time sequence feature vectors across heats; wherein the heat feature sequence is composed of the feature regulation vectors corresponding to the current heat and the previous k heats; and the spatial feature vector and the time sequence feature vector are dimensionally aligned, weighted and fused dimension by dimension through learnable fusion weights, and a regression prediction network is used to output a prediction value of a steelmaking endpoint state of the current heat.

[0161] Thus, according to the present embodiment, firstly, multi-scale gating weight generation and graph structure attention coupling analysis are performed on the multi-dimensional steelmaking process characteristics of the current heat, a feature regulation vector is generated to realize high-dimensional variable decoupling and target difference perception; then, local process combination patterns are extracted through multi-core parallel convolution, and spatial feature vectors are constructed after independent self-attention enhancement of the channels; at the same time, a multi-granularity Transformer encoder is used to process the historical heat sequence composed of the feature regulation vectors of the historical heats, and a time sequence feature vector for representing process evolution across heats is extracted. Finally, the spatio-temporal features are dynamically gated and weighted, and the endpoint temperature and multi-element content prediction values are simultaneously output through the regression network. Among them, the multi-scale gating network distinguishes the key variable dependence of different prediction targets, combines the graph attention mechanism to explicitly model the physical correlation between process variables, realizes the mining of key feature response structures for different output tasks, dynamically adjusts the importance distribution of each variable, improves the key variable identification capability under multi-target, and solves the high-dimensional variable coupling and target-dependent heterogeneity problem; the convolutional neural network is used to extract the structural patterns of the process variables of the current heat to represent the static dependence of the "spatial dimension", at the same time, the Transformer is introduced to model the historical evolution process across heats to capture the dynamic correlation of the "time dimension", and the cross-heat information forgetting problem is solved; through the adaptive fusion mechanism, the spatio-temporal feature contribution weight is adjusted according to the dynamic working condition, realizing the deep fusion of multi-source information and the unified modeling of the endpoint state, supporting multi-target collaborative optimization. Thus, the dynamic decoupling of strongly coupled process variables, the deep mining of cross-heat evolution information, and the collaborative and accurate prediction of endpoint temperature and multi-element content are realized. Further, the technical problems in the prior art that the traditional converter steelmaking endpoint state prediction method cannot effectively cope with the high-dimensional variable coupling relationship, the heterogeneous dependence structure of different prediction targets, and the forgetting of cross-heat evolution information are solved.

[0162] The above-mentioned serial numbers of the embodiments of the present application are only for description, and do not represent the advantages and disadvantages of the embodiments.

[0163] In the above-mentioned embodiments of the present application, the description of each embodiment has its own emphasis, and the parts not described in detail in a certain embodiment can be referred to the related description of other embodiments.

[0164] In several embodiments provided in the present application, it should be understood that the disclosed technical content can be implemented by other ways. Among them, the above-mentioned device embodiments are only schematic, for example, the division of the units is only a logical function division, and actual implementation can have another division way, for example, multiple units or components can be combined or integrated into another system, or some features can be ignored or not executed. In addition, the coupling or direct coupling or communication connection between the displayed or discussed each other can be through some interface, indirect coupling or communication connection between units or modules, which can be electrical or other forms.

[0165] The units described as separate components may or may not be physically separate, and the components displayed as units may or may not be physical units, i.e., may be located in one place, or may be distributed to multiple network units. Part or all of the units can be selected according to actual needs to achieve the purpose of the embodiment scheme.

[0166] In addition, each functional unit in each embodiment of the present application can be integrated in one processing unit, or each unit can exist physically, or two or more units can be integrated in one unit. The integrated unit can be realized in the form of hardware or in the form of a software functional unit.

[0167] The integrated unit, if realized in the form of a software functional unit and sold or used as an independent product, can be stored in a computer readable storage medium. Based on this understanding, the technical solutions of the present application, essentially or the part that contributes to the prior art, or all or part of the technical solutions can be embodied in the form of a software product. The computer software product is stored in a storage medium, including a number of instructions to make a computer device (which can be a personal computer, a server or a network device, etc.) execute all or part of the steps of the method described in each embodiment of the present application. The foregoing storage medium includes: a U disk, a read-only memory (ROM, Read-Only Memory), a random access memory (RAM, Random Access Memory), a mobile hard disk, a magnetic disk or an optical disk, and various program code storage media.

[0168] The above is only the preferred embodiment of the present application, and it should be noted that for those skilled in the art, without departing from the principles of the present application, a number of improvements and refinements can be made, and these improvements and refinements should be considered as the protection scope of the present application.

Claims

1. A method of predicting the end point of a converter steelmaking process, characterised in that, The method comprises the following steps: multi-scale importance modeling of multi-dimensional process characteristics of the current heat, to generate importance weights of each dimension process characteristic; modeling the structural coupling relationship between multi-dimensional process characteristics by graph attention mechanism, to generate structure perception weights of each dimension process characteristic; fusion of importance weights and structure perception weights to obtain feature weights of each dimension process characteristic; multiplying each dimension process characteristic with the corresponding feature weight to generate a feature regulation vector of the current heat; performing multi-scale convolution operation on the feature regulation vector of the current heat to extract local pattern features under different receptive fields, and outputting a spatial feature vector of the current heat after enhancement by self-attention mechanism; using multiple groups of Transformer encoders to process heat feature sequences of different historical window lengths respectively to extract time sequence feature vectors across heats; wherein the heat feature sequence is composed of the feature regulation vectors corresponding to the current heat and the previous k heats; dimension alignment of the spatial feature vector and the time sequence feature vector, and weighted fusion of each dimension by a learnable fusion weight, and outputting a prediction value of the steelmaking endpoint state of the current heat by a regression prediction network; wherein the operation of multi-scale importance modeling of multi-dimensional process characteristics of the current heat to generate importance weights of each dimension process characteristic comprises: dividing the multi-dimensional process characteristics of the current heat into local scale features, global scale features and statistical scale features; wherein the local scale features are variables dynamically regulated during the blowing process, the global scale features are static variables reflecting the structure of the furnace charge, and the statistical scale features are statistical variables representing the evolution trend of the process; generating importance weights of the local scale features by a first type of gated subnetwork, generating importance weights of the global scale features by a second type of gated subnetwork, and generating importance weights of the statistical scale features by a third type of gated subnetwork; wherein the first type of gated subnetwork adopts a single-layer nonlinear transformation structure, the second type of gated subnetwork adopts a multi-layer perception structure, and the third type of gated subnetwork adopts a Softplus activation function.

2. The method of claim 1, wherein, The operation of modeling the structural coupling relationship between multi-dimensional process characteristics by graph attention mechanism to generate structure perception weights of each dimension process characteristic comprises: taking each dimension feature in the multi-dimensional process characteristics of the current heat as a graph node, establishing a connection edge between any two nodes to form a fully connected graph structure; inputting the fully connected graph structure into a graph attention network to generate structure perception weights of each dimension process characteristic.

3. The method of claim 1, wherein, The operation of fusion of importance weights and structure perception weights to obtain feature weights of each dimension process characteristic comprises: for each dimension process characteristic, dynamically fusing the importance weights and the structure perception weights by a dedicated learnable fusion weight to obtain the corresponding feature weight.

4. The method of claim 1, wherein, The operation of performing multi-scale convolution operation on the feature regulation vector of the current heat to extract local pattern features under different receptive fields, and outputting a spatial feature vector of the current heat after enhancement by self-attention mechanism comprises: extending the feature regulation vector of the current heat into a one-dimensional feature map; performing convolution operation on the one-dimensional feature map in parallel to obtain n groups of feature maps; stitching the n groups of feature maps along the channel dimension into a fusion feature map; An independent self-attention network for the fusion feature map input channel is used to flatten the attention-enhanced fusion feature map into a one-dimensional vector as the spatial feature vector of the current heat.

5. The method of claim 1, wherein, The operation of extracting the cross-heat time sequence feature vector by using multiple groups of Transformer encoders to process heat feature sequences of different historical window lengths, includes: Multiple different window lengths are set, and for each window length, a corresponding heat feature sequence is constructed; wherein the heat feature sequence includes the feature regulation vector corresponding to the current heat and its previous k heats, and k is different in different heat feature sequences; Each heat feature sequence is input into an independent Transformer encoder for processing, and a feature representation corresponding to the window length is output; The feature representations corresponding to the multiple different window lengths are spliced in the feature dimension to obtain the cross-heat time sequence feature vector.

6. The method of claim 1, wherein, The operation of dimension alignment of the spatial feature vector and the time sequence feature vector, and weighted fusion of each dimension through a learnable fusion weight, and output of the prediction value of the steelmaking end state of the current heat by the regression prediction network, includes: The spatial feature vector and the time sequence feature vector are projected linearly to compress or expand to the same dimension; The aligned spatial feature vector and the time sequence feature vector are spliced, and then input into a gating network to generate a learnable fusion weight vector; The spatial feature vector and the time sequence feature vector are weighted and fused in each dimension according to the generated learnable fusion weight vector to obtain a fusion feature vector; The fusion feature vector is input into a regression prediction network to output the prediction value of the steelmaking end state of the current heat.

7. A storage medium, characterized by The storage medium includes a stored program, wherein the program is executed by the processor to perform the method of any one of claims 1-6 when the program is running.

8. A converter steelmaking end point prediction device characterized by comprising: It includes: A feature soft selection module oriented to target differences is used to model the multi-scale importance of the multi-dimensional process features of the current heat to generate importance weights of each process feature; A graph attention mechanism is used to model the structural coupling relationship between the multi-dimensional process features to generate structural perception weights of each process feature; The importance weights and the structural perception weights are fused to obtain feature weights of each process feature; Each process feature is multiplied by the corresponding feature weight to generate a feature regulation vector of the current heat; A multi-scale spatial structure perception module is used to perform multi-scale convolution operations on the feature regulation vector of the current heat to extract local pattern features under different receptive fields, and output a spatial feature vector of the current heat after being enhanced by a self-attention mechanism; A multi-granularity time sequence dependence modeling module is used to extract cross-heat time sequence feature vectors by using multiple groups of Transformer encoders to process heat feature sequences of different historical window lengths; wherein the heat feature sequence is composed of the feature regulation vector corresponding to the current heat and its previous k heats; A spatio-temporal feature adaptive fusion prediction module is used to align the dimensions of the spatial feature vector and the time sequence feature vector, and to perform weighted fusion of each dimension through a learnable fusion weight, and to output the prediction value of the steelmaking end state of the current heat by a regression prediction network. The operation of performing multi-scale importance modeling on the multi-dimensional process characteristics of the current heat, to generate importance weights of each process characteristic, includes: The multi-dimensional process characteristics of the current heat are divided into local scale characteristics, global scale characteristics and statistical scale characteristics; wherein the local scale characteristics are variables dynamically regulated in the blowing process, the global scale characteristics are static variables reflecting the burden structure, and the statistical scale characteristics are statistical variables representing the process evolution trend; The importance weights of the local scale characteristics are generated through a first type of gated sub-network; the importance weights of the global scale characteristics are generated through a second type of gated sub-network; and the importance weights of the statistical scale characteristics are generated through a third type of gated sub-network; The first type of gated sub-network adopts a single-layer nonlinear transformation structure, the second type of gated sub-network adopts a multi-layer perception structure, and the third type of gated sub-network adopts a Softplus activation function.

9. A converter steelmaking end point prediction device characterized by comprising: It includes: a processor; a memory connected with the processor, for providing the processor with instructions to process the following processing steps: performing multi-scale importance modeling on the multi-dimensional process characteristics of the current heat, to generate importance weights of each process characteristic; modeling the structural coupling relationship between the multi-dimensional process characteristics through a graph attention mechanism, to generate structure perception weights of each process characteristic; fusing the importance weights and the structure perception weights to obtain feature weights of each process characteristic; multiplying each process characteristic with the corresponding feature weight to generate a feature regulation vector of the current heat; performing multi-scale convolution operations on the feature regulation vector of the current heat to extract local pattern features under different receptive fields, and outputting a spatial feature vector of the current heat after enhancement by a self-attention mechanism; using multiple groups of Transformer encoders to process heat feature sequences of different historical window lengths respectively, to extract time sequence feature vectors across heats; wherein the heat feature sequence is composed of the feature regulation vectors corresponding to the current heat and the previous k heats; aligning the dimensions of the spatial feature vector and the time sequence feature vector, performing weighted fusion dimension by dimension through a learnable fusion weight, and outputting a prediction value of the steelmaking endpoint state of the current heat through a regression prediction network. The operation of performing multi-scale importance modeling on the multi-dimensional process characteristics of the current heat, to generate importance weights of each process characteristic, includes: The multi-dimensional process characteristics of the current heat are divided into local scale characteristics, global scale characteristics and statistical scale characteristics; wherein the local scale characteristics are variables dynamically regulated in the blowing process, the global scale characteristics are static variables reflecting the burden structure, and the statistical scale characteristics are statistical variables representing the process evolution trend; The importance weights of the local scale characteristics are generated through a first type of gated sub-network; the importance weights of the global scale characteristics are generated through a second type of gated sub-network; and the importance weights of the statistical scale characteristics are generated through a third type of gated sub-network; The first type of gated sub-network adopts a single-layer nonlinear transformation structure, the second type of gated sub-network adopts a multi-layer perception structure, and the third type of gated sub-network adopts a Softplus activation function.

Images